From 65f06a63ee4d5f563e6446fc9f122d85db9f2a96 Mon Sep 17 00:00:00 2001
From: Hashim Sharif <hsharif3@tyler.cs.illinois.edu>
Date: Tue, 2 Jul 2019 16:27:36 -0500
Subject: [PATCH] Adding Keras sources and Frontend from Github to Gitlab
---
llvm/projects/keras/README.md | 17 +
llvm/projects/keras/frontend/__init__.py | 0
.../keras/frontend/approxhpvm_translator.py | 878 ++++++++++++++
.../keras/frontend/hpvm_dfg_translator.py | 568 +++++++++
.../keras/frontend/promise_translator.py | 1079 +++++++++++++++++
llvm/projects/keras/frontend/setup.py | 12 +
llvm/projects/keras/frontend/weight_utils.py | 125 ++
llvm/projects/keras/keras_environment.yml | 331 +++++
llvm/projects/keras/setup.py | 12 +
llvm/projects/keras/src/alexnet.py | 173 +++
llvm/projects/keras/src/alexnet2.py | 243 ++++
llvm/projects/keras/src/lenet.py | 92 ++
llvm/projects/keras/src/resnet.py | 571 +++++++++
llvm/projects/keras/src/vgg16_cifar10.py | 241 ++++
llvm/projects/keras/src/vgg16_cifar100.py | 243 ++++
15 files changed, 4585 insertions(+)
create mode 100644 llvm/projects/keras/README.md
create mode 100644 llvm/projects/keras/frontend/__init__.py
create mode 100644 llvm/projects/keras/frontend/approxhpvm_translator.py
create mode 100644 llvm/projects/keras/frontend/hpvm_dfg_translator.py
create mode 100644 llvm/projects/keras/frontend/promise_translator.py
create mode 100644 llvm/projects/keras/frontend/setup.py
create mode 100644 llvm/projects/keras/frontend/weight_utils.py
create mode 100644 llvm/projects/keras/keras_environment.yml
create mode 100644 llvm/projects/keras/setup.py
create mode 100644 llvm/projects/keras/src/alexnet.py
create mode 100644 llvm/projects/keras/src/alexnet2.py
create mode 100644 llvm/projects/keras/src/lenet.py
create mode 100644 llvm/projects/keras/src/resnet.py
create mode 100644 llvm/projects/keras/src/vgg16_cifar10.py
create mode 100644 llvm/projects/keras/src/vgg16_cifar100.py
diff --git a/llvm/projects/keras/README.md b/llvm/projects/keras/README.md
new file mode 100644
index 0000000000..1f1133adb0
--- /dev/null
+++ b/llvm/projects/keras/README.md
@@ -0,0 +1,17 @@
+
+
+## Exporting Conda Environment:
+
+conda activate ${ENV_TO_EXPORT}
+conda env export > environment.yml
+
+## Importing Conda Environment:
+
+conda create -f environment.yml
+
+## Building and Installing Frontend:
+
+python setup.py build
+python setup.py install
+
+
diff --git a/llvm/projects/keras/frontend/__init__.py b/llvm/projects/keras/frontend/__init__.py
new file mode 100644
index 0000000000..e69de29bb2
diff --git a/llvm/projects/keras/frontend/approxhpvm_translator.py b/llvm/projects/keras/frontend/approxhpvm_translator.py
new file mode 100644
index 0000000000..2cb8c5d741
--- /dev/null
+++ b/llvm/projects/keras/frontend/approxhpvm_translator.py
@@ -0,0 +1,878 @@
+
+import sys
+import numpy as np
+from frontend.promise_translator import PromiseRtTranslator
+from frontend.hpvm_dfg_translator import HPVMTranslator
+from frontend.weight_utils import dumpLabels, dumpData, dumpConvWeights, dumpFcWeights, dumpFcBias
+import keras
+import os
+
+
+
+class DFG:
+
+ root_set = False;
+
+ def __init__(self):
+ self.node_map = {}
+ self.root_node = None
+ self.last_node = None
+
+
+ def hasSingleInput(self, layer):
+ layer_name = layer.__class__.__name__
+
+ singleInLayers = {}
+ singleInLayers["Conv2D"] = True
+ singleInLayers["Dense"] = True
+ singleInLayers["MaxPooling2D"] = True
+ singleInLayers["Activation"] = True
+ singleInLayers["BatchNormalization"] = True
+ singleInLayers["Flatten"] = True
+
+ if layer_name in singleInLayers:
+ return True
+
+ return False
+
+
+ def hasMultipleInputs(self, layer):
+ layer_name = layer.__class__.__name__
+
+ multipleInLayers = {}
+ multipleInLayers["Add"] = True
+
+ if layer_name in multipleInLayers:
+ return True
+
+ return False
+
+
+
+ def add_dfg_edge(self, inbound_node_name, dfg_node):
+
+ inbound_node_name = inbound_node_name.split(":")[0]
+ inbound_node_name = inbound_node_name.split("/")[0]
+ if inbound_node_name in self.node_map:
+ inbound_node = self.node_map[inbound_node_name]
+ print (inbound_node_name, " found!")
+ inbound_node.add_output(dfg_node)
+ dfg_node.add_input(inbound_node)
+
+ else:
+ print ("--inbound node NOT FOUND!")
+
+
+
+
+ def add_to_graph(self, layer):
+ dfg_node = DFGNode(layer)
+ if not self.root_set:
+ self.root_node = dfg_node
+ self.root_set = True # DFG root node is now set
+
+ if self.hasMultipleInputs(layer):
+ for j in range(len(layer.input)):
+ print(type(layer.input[j]))
+ print(layer.input[j].op.name)
+ self.add_dfg_edge(layer.input[j].op.name, dfg_node)
+
+ else:
+ #if self.hasSingleInput(layer):
+ print (layer.input.name)
+ self.add_dfg_edge(layer.input.name, dfg_node)
+
+ # Adding DFG node to name mapping
+ self.node_map[layer.name] = dfg_node
+
+
+ # Check if all predecessor nodes have been visited thus far - reverse postorder traversal
+ def predVisited(self, cur_node, visited_nodes):
+ for input_node in cur_node.inputs:
+ if input_node.layer_name not in visited_nodes:
+ return False;
+
+ # All predecessors are visited
+ return True
+
+
+ def traverseNode(self, cur_node, visited_nodes):
+
+ # Skip visited nodes
+ if cur_node.layer_name in visited_nodes:
+ return
+
+ if self.predVisited(cur_node, visited_nodes):
+ print(cur_node.layer_type)
+ print(cur_node.layer_name)
+ visited_nodes[cur_node.layer_name] = True
+
+ # Invoking traversal on outbound nodes
+ for output_node in cur_node.outputs:
+ self.traverseNode(output_node, visited_nodes)
+
+ # NOTE: Assuming that no outbound edges implies the last node in the graph
+ if len(cur_node.outputs) == 0:
+ self.last_node = cur_node
+
+
+ # Print the DFG in reverse postorder
+ def printDFG(self):
+
+ print ("\n\n ****** Printing DFG ******* \n\n")
+ visited_nodes = {}
+ # Starting traversal at the DFG root node
+ self.traverseNode(self.root_node, visited_nodes)
+
+
+
+
+class DFGNode:
+
+ def add_output(self, output_node):
+ self.outputs.append(output_node)
+
+ def add_input(self, input_node):
+ self.inputs.append(input_node)
+
+
+ def __init__(self, layer):
+
+ self.inputs = []
+ self.outputs = []
+
+ layer_type = layer.__class__.__name__
+ self.layer_type = layer_type # layer type e.g., conv2d, add, dense
+ self.layer_name = layer.name # unique layer identifier
+ print (self.layer_name)
+
+ if layer_type == "Conv2D" or layer_type == "Dense":
+ self.weights = layer.get_weights()[0]
+ print("\t", self.weights.shape)
+ self.use_bias = layer.use_bias
+
+ if layer.use_bias:
+ self.use_bias = layer.use_bias
+ self.bias_weights = layer.get_weights()[1]
+ print("\t", self.bias_weights.shape)
+
+ if layer_type == "Conv2D":
+ self.padding = layer.padding
+ self.strides = layer.strides
+ print("\t", self.strides)
+ print("\tPadding = ", self.padding)
+
+ if layer_type == "MaxPooling2D" or layer_type == "AveragePooling2D":
+ self.pool_size = layer.pool_size
+ self.strides = layer.strides
+ print("\t pool_size = ", self.pool_size)
+ print("\t strides = ", self.strides)
+
+ if layer_type == "Conv2D" or layer_type == "Dense" or layer_type == "Activation":
+ self.activation_type = layer.activation.__name__
+ print ("\t Activation = ", self.activation_type)
+
+
+ if layer_type == "ZeroPadding2D":
+ print ("***ZeroPaddding \n");
+ self.padding = layer.padding
+ print ("padding = ", self.padding);
+
+
+
+class TensorRtTranslator:
+
+ def __init__(self, dfg):
+ self.dfg = dfg
+ self.output_map = {}
+ self.counter = 0
+ self.weight_str = ""
+ self.program_str = ""
+ self.input_str = ""
+ self.filter_names = {}
+
+
+ def getWeightStr(self):
+ return self.weight_str
+
+
+ def getInputStr(self):
+ return self.input_str
+
+
+ def getFilterNames(self):
+ return self.filter_names
+
+
+ def getWeightVarName(self, weights):
+
+ output_var_name = "weights_" + str(self.w_counter)
+ self.w_counter += 1
+ self.filter_names[weights] = output_var_name
+
+ return output_var_name
+
+
+ def getVariableName(self, cur_node):
+
+ output_var_name = "var_" + str(self.counter)
+ self.counter += 1
+ self.output_map[cur_node.layer_name] = output_var_name
+
+ return output_var_name
+
+
+ def isSkipLayer(self, layer_type):
+
+ skip_layers = {}
+ skip_layers["Flatten"] = 0
+ skip_layers["Dropout"] = 0
+ skip_layers["ZeroPadding2D"] = 0
+
+ if layer_type in skip_layers:
+ return True
+ else:
+ return False
+
+
+ # NOTE: returns the previous DFG node ignoring "Flatten", "Dropout" Layers
+ def getPrevActiveLayer(self, cur_node):
+
+ pred_layer_type = cur_node.inputs[0].layer_type
+ # FIXME: Assuming the 'inference' phase - hence skipping Dropout
+ #if pred_layer_type == "Flatten" or pred_layer_type == "Dropout":
+ if self.isSkipLayer(pred_layer_type):
+ cur_node = self.getPrevActiveLayer(cur_node.inputs[0])
+ return cur_node
+ else:
+ return cur_node
+
+
+
+ def getSingleInputName(self, cur_node):
+
+ print (cur_node.layer_name)
+ # Assumption: If no inputs, the previous layer must be input layer
+ if len(cur_node.inputs) == 0:
+ return "input"
+
+ print ("Input_type = ", cur_node.inputs[0].layer_type)
+
+ pred_layer_type = cur_node.inputs[0].layer_type
+ # FIXME: Assuming the 'inference' phase - hence skipping Dropout
+ #if pred_layer_type == "Flatten" or pred_layer_type == "Dropout":
+ if self.isSkipLayer(pred_layer_type):
+ cur_node = self.getPrevActiveLayer(cur_node)
+
+ if cur_node.inputs[0].layer_type == "InputLayer":
+ return "input"
+
+ # get input to the layer
+ input_node_name = cur_node.inputs[0].layer_name # get the input layer ID
+
+ input_var_name = ""
+ if input_node_name in self.output_map:
+ input_var_name = self.output_map[input_node_name]
+ else:
+ print ("Input Var not found - Aborting....")
+ sys.exit(0)
+
+ return input_var_name
+
+
+
+ def getPrevLayerPadding(self, cur_node):
+
+ print (cur_node.layer_name)
+ # Assumption: If no inputs, the previous layer must be input layer
+ if len(cur_node.inputs) == 0:
+ return None
+
+ print ("Input_type = ", cur_node.inputs[0].layer_type)
+ if cur_node.inputs[0].layer_type == "ZeroPadding2D":
+ pred_padding = cur_node.inputs[0].padding
+ return pred_padding
+
+ return None
+
+
+
+ def getMultipleInputNames(self, cur_node):
+
+ var_names = []
+ for i in range(len(cur_node.inputs)):
+ # get input to the layer
+ input_node_name = cur_node.inputs[i].layer_name # get the input layer ID
+
+ input_var_name = ""
+ if input_node_name in self.output_map:
+ input_var_name = self.output_map[input_node_name]
+ var_names.append(input_var_name)
+ else:
+ print ("Input Var not found - Aborting....")
+ sys.exit(0)
+
+ return var_names
+
+
+
+ def hasBiasAdd(self, cur_node):
+
+ if cur_node.layer_type == "Conv2D" or cur_node.layer_type == "Dense":
+ return cur_node.use_bias
+
+ return False
+
+
+ def hasActivation(self, cur_node):
+
+ if cur_node.layer_type == "Conv2D" or cur_node.layer_type == "Dense":
+ return cur_node.activation_type != "linear"
+
+ return False
+
+
+ def genActivationCall(self, input_var, output_var, activation_type):
+
+ func_name = ""
+ if activation_type == "tanh":
+ func_name = "Tanh"
+
+ if activation_type == "relu":
+ func_name = "Relu"
+
+ if activation_type == "softmax":
+ func_name = "Softmax"
+
+ inst_str = "void* " + output_var + " = "
+ inst_str += "tensor" + func_name + "(" + input_var + "); \n"
+
+ return inst_str
+
+
+
+ def genNodeCalls(self, cur_node):
+
+ out_var_name1 = self.getVariableName(cur_node)
+ layer_type = cur_node.layer_type
+
+ if layer_type == "Conv2D":
+ input_var_name = self.getSingleInputName(cur_node)
+ weights = cur_node.weights
+ strides = cur_node.strides
+
+ padding = 0
+ if cur_node.padding.strip() == "valid":
+ padding = 0
+ else:
+ padding = cur_node.padding
+ padding = int((weights.shape[0] - 1) / 2)
+
+ prev_padding = self.getPrevLayerPadding(cur_node)
+ if prev_padding != None:
+ # FIXME: currently only supporting symmetric padding
+ padding = prev_padding[0][0]
+
+ inst_str = "void* " + out_var_name1 + " = "
+ inst_str += "tensorConvolution(" + input_var_name + ", "
+ inst_str += cur_node.layer_name + "_w, "
+ inst_str += str(padding) + ", "
+ inst_str += str(padding) + ", "
+ inst_str += str(strides[0]) + ", "
+ inst_str += str(strides[1]) + ", "
+ inst_str += "1, 0); \n"
+
+ self.program_str += inst_str
+
+
+ if layer_type == "Dense":
+ input_var_name = self.getSingleInputName(cur_node)
+
+ weights = cur_node.weights
+ inst_str = "void* " + out_var_name1 + " = "
+ inst_str += "tensorGemmGPU(" + input_var_name + ", "
+ inst_str += cur_node.layer_name + "_w"
+ inst_str += "); \n"
+
+ self.program_str += inst_str
+
+
+ if self.hasBiasAdd(cur_node):
+ out_var_name2 = self.getVariableName(cur_node)
+
+ inst_str = "void* " + out_var_name2 + " = "
+ inst_str += "tensorAdd(" + out_var_name1 + ", "
+ inst_str += cur_node.layer_name + "_b"
+ inst_str += "); \n"
+
+ self.program_str += inst_str
+
+ # NOTE: Changing output variable
+ out_var_name1 = out_var_name2
+
+
+ if self.hasActivation(cur_node):
+ activation_type = cur_node.activation_type
+ out_var_name3 = self.getVariableName(cur_node)
+
+ inst_str = self.genActivationCall(out_var_name1, out_var_name3, activation_type)
+ self.program_str += inst_str
+
+
+ if layer_type == "Activation":
+ input_var_name = self.getSingleInputName(cur_node)
+
+ inst_str = self.genActivationCall(input_var_name, out_var_name1, cur_node.activation_type)
+ self.program_str += inst_str
+
+
+ if layer_type == "BatchNormalization":
+ input_var_name = self.getSingleInputName(cur_node)
+
+ inst_str = "void* " + out_var_name1 + " = "
+ inst_str += "tensorBatchNormalization(" + input_var_name + "); \n"
+ self.program_str += inst_str
+
+
+ if layer_type == "Add":
+ input_vars = self.getMultipleInputNames(cur_node)
+
+ inst_str = "void* " + out_var_name1 + " = "
+ inst_str += "tensorAdd(" + input_vars[0] + ", " + input_vars[1] + "); \n"
+ self.program_str += inst_str
+
+
+ if layer_type == "MaxPooling2D" or layer_type == "AveragePooling2D":
+ input_var_name = self.getSingleInputName(cur_node)
+
+ pool_size = cur_node.pool_size
+ strides = cur_node.strides
+ # FIXME: Non-same padding is *NOT* currently supported
+ padding = 0
+ pool_type = 0
+ if layer_type == "MaxPooling2D":
+ pool_type = "0"
+ if layer_type == "AveragePooling2D":
+ pool_type = "1"
+
+ # tensorPooling(input, pool_type, pool_h, pool_w, v_pad, h_pad, v_stride, h_stride)
+ inst_str = "void* " + out_var_name1 + " = "
+ inst_str += "tensorPooling(" + input_var_name + "," + pool_type + "," + str(pool_size[0]) + "," + str(pool_size[1])
+ inst_str += "," + str(padding) + "," + str(padding) + "," + str(strides[0]) + "," + str(strides[1])
+ inst_str += "); \n"
+ self.program_str += inst_str
+
+
+
+
+
+ def codegenNode(self, dfg, cur_node, visited_nodes):
+
+ # Skip visited nodes
+ if cur_node.layer_name in visited_nodes:
+ return
+
+ if dfg.predVisited(cur_node, visited_nodes):
+
+ visited_nodes[cur_node.layer_name] = True
+
+ self.genNodeCalls(cur_node)
+
+ # Invoking traversal on outbound nodes
+ for output_node in cur_node.outputs:
+ self.codegenNode(dfg, output_node, visited_nodes)
+
+
+ # Print the DFG in reverse postorder
+ def codegen(self, dfg):
+
+ print ("\n\n ****** Codegen for HPVM Tensor Rt ******* \n\n")
+ visited_nodes = {}
+ # Starting traversal at the DFG root node
+ self.codegenNode(dfg, dfg.root_node, visited_nodes)
+
+
+
+
+ def dump_weights(self, model, prefix):
+
+ layer_count = 0
+ for i in range(len(model.layers)):
+ layer = model.layers[i]
+ layer_type = layer.__class__.__name__
+ layer_name = layer.name
+
+ if layer_type == "Conv2D":
+ weights = layer.get_weights()[0]
+ w_name = layer_name + "_w"
+
+ self.filter_names[w_name] = 1
+ print (weights.shape, w_name)
+
+ N = weights.shape[3]
+ C = weights.shape[2]
+ H = weights.shape[1]
+ W = weights.shape[0]
+
+ #unique_file_name = "conv" + str(layer_count) + ".bin"
+ unique_file_name = w_name + ".bin"
+ dumpConvWeights(prefix + unique_file_name, weights, N, C, H, W)
+
+ file_path = w_name + "_path"
+ file_path_str = "std::string " + file_path + " = " + " dir_prefix + std::string(\""
+ file_path_str += unique_file_name + "\"); \n"
+ self.weight_str += file_path_str
+
+ # FIXME: Be flexible for datatypes (currently only FP32 weights)
+ # NOTE: '0' specified for floating point type
+ self.weight_str += "void* " + w_name + " = " + " readTrainedWeights("
+ self.weight_str += file_path + ".c_str(), 0," + str(N) + "," + str(C) + "," + str(H) + "," + str(W)
+ self.weight_str += "); \n"
+
+
+ if layer.use_bias:
+ bias_weights = layer.get_weights()[1]
+ b_name = layer_name + "_b"
+
+ self.filter_names[b_name] = 1
+ print (bias_weights.shape, b_name)
+
+ #unique_file_name = "conv_bias" + str(layer_count) + ".bin"
+ unique_file_name = b_name + ".bin"
+ dumpFcBias(prefix + unique_file_name, bias_weights, bias_weights.shape[0])
+
+ file_path = b_name + "_path"
+ file_path_str = "std::string " + file_path + " = " + " dir_prefix + std::string(\""
+ file_path_str += unique_file_name + "\"); \n"
+ self.weight_str += file_path_str
+
+ C = bias_weights.shape[0]
+
+ self.weight_str += "void* " + b_name + " = " + " readTrainedWeights("
+ self.weight_str += file_path + ".c_str(), 0,1," + str(C) + ",1,1); \n"
+
+
+ if layer_type == "Dense":
+ weights = layer.get_weights()[0]
+ w_name = layer_name + "_w"
+
+ self.filter_names[w_name] = 1
+ print (weights.shape, w_name)
+
+ H = weights.shape[0]
+ W = weights.shape[1]
+
+ #unique_file_name = "fc" + str(layer_count) + ".bin"
+ unique_file_name = w_name + ".bin"
+ dumpFcWeights(prefix + unique_file_name, weights, H, W)
+
+ file_path = w_name + "_path"
+ file_path_str = "std::string " + file_path + " = " + " dir_prefix + std::string(\""
+ file_path_str += unique_file_name + "\"); \n"
+ self.weight_str += file_path_str
+
+ self.weight_str += "void* " + w_name + " = " + " readTrainedWeights("
+ self.weight_str += file_path + ".c_str(), 0,1,1," + str(H) + "," + str(W) + "); \n"
+
+
+ if layer.use_bias:
+ bias_weights = layer.get_weights()[1]
+ b_name = layer_name + "_b"
+
+ self.filter_names[b_name] = 1
+ print (bias_weights.shape, b_name)
+
+ unique_file_name = b_name + ".bin"
+ dumpFcBias(prefix + unique_file_name, bias_weights, bias_weights.shape[0])
+
+ file_path = b_name + "_path"
+ file_path_str = "std::string " + file_path + " = " + " dir_prefix + std::string(\""
+ file_path_str += unique_file_name + "\"); \n"
+ self.weight_str += file_path_str
+
+ C = bias_weights.shape[0]
+
+ self.weight_str += "void* " + b_name + " = " + " readTrainedWeights("
+ self.weight_str += file_path + ".c_str(), 0,1," + str(C) + ",1,1); \n"
+
+ layer_count += 1
+
+
+
+ def add_header(self):
+
+ headers = "\n#include <stdio.h> \n"
+ headers += "#include <stdlib.h> \n"
+ headers += "#include <unistd.h> \n"
+ headers += "#include <fcntl.h> \n"
+ headers += "#include <sys/types.h> \n"
+ headers += "#include <sys/stat.h> \n"
+ headers += "#include <string.h> \n"
+
+ headers += "#include \"../../tensor_runtime/include/tensor_runtime.h\" \n"
+ headers += "#include \"../include/utils.h\" \n\n"
+
+ main_func = "int main(){ \n\n"
+
+ initialization = "llvm_hpvm_initTensorRt(0); \n\n"
+
+ self.program_str += headers
+ self.program_str += main_func
+ self.program_str += initialization
+
+
+
+ def add_footer(self, test_data):
+
+ if test_data is not None and self.dfg.last_node is not None:
+ last_node = self.dfg.last_node
+ output_var = self.output_map[last_node.layer_name]
+ #accuracy_call = "\ncomputeAccuracy2(labels," + str(len(test_data)) + "," + output_var + "); \n"
+ #self.program_str += accuracy_call
+
+
+ destructors = "\nllvm_hpvm_cleanupTensorRt(); \n"
+ self.program_str += destructors
+
+ end_main = "\nreturn 0; \n\n}\n"
+ self.program_str += end_main
+
+ return 0
+
+
+
+ def genInputCalls(self, test_data, test_labels, weights_dir):
+
+ dumpData(weights_dir + "input.bin", test_data)
+
+ N = test_data.shape[0]
+ C = test_data.shape[1]
+ H = test_data.shape[2]
+ W = test_data.shape[3]
+
+ file_path = "input_path"
+ file_path_str = "std::string " + file_path + " = " + " dir_prefix + std::string(\""
+ file_path_str += "input.bin\"); \n"
+ self.weight_str += file_path_str
+
+ self.input_str += "void* input = readTrainedWeights("
+ self.input_str += file_path + ".c_str(), 0," + str(N) + "," + str(C) + ","
+ self.input_str += str(H) + "," + str(W) + "); \n"
+
+ #self.weight_str += self.input_str
+
+ # Adding input to the filter map
+ self.filter_names["input"] = 1
+
+ dumpLabels(weights_dir + "labels.bin", test_labels)
+
+ file_path = "labels_path"
+ file_path_str = "std::string " + file_path + " = " + " dir_prefix + std::string(\""
+ file_path_str += "labels.bin\"); \n"
+ self.weight_str += file_path_str
+
+ self.input_str += "uint8_t* labels = readLabels("
+ self.input_str += file_path + ".c_str()," + str(test_labels.shape[0]) + "); \n"
+
+
+
+ def genBatchLoop(self, x_test):
+
+ N = x_test.shape[0]
+ C = x_test.shape[1]
+ H = x_test.shape[2]
+ W = x_test.shape[3]
+
+ loop_str = ""
+ loop_str += "\nstartMemTracking(); \n\n"
+
+ loop_str += "int test_input_size = " + str(N) + "; \n"
+ loop_str += "int batch_size = " + str(N) + "; \n"
+ loop_str += "int batch_count = test_input_size / batch_size; \n"
+ loop_str += "float final_accuracy = 0.0; \n\n"
+
+ loop_str += "for(int i = 0; i < batch_count; i++){ \n\n"
+ loop_str += "int start = i * batch_size; \n"
+ loop_str += "int end = (i + 1) * batch_size; \n"
+
+ loop_str += "\nvoid* input = readInputBatch(input_path.c_str(),0,start,end,"
+ loop_str += str(C) + "," + str(H) + "," + str(W) + "); \n\n"
+
+ self.program_str += loop_str
+
+
+
+ def endBatchLoop(self):
+
+ end_loop_str = ""
+ end_loop_str += "\nuint8_t* labels = readLabelsBatch(labels_path.c_str(),start,end); \n"
+
+ last_node = self.dfg.last_node
+ output_var = self.output_map[last_node.layer_name]
+ accuracy_call = "\nfloat accuracy = computeAccuracy2(labels, batch_size, " + output_var + "); \n"
+ end_loop_str += accuracy_call
+
+ #end_loop_str += "float accuracy = computeAccuracy2(labels, batch_size, var_60); "
+ end_loop_str += "final_accuracy += accuracy; \n"
+ end_loop_str += "freeBatchMemory(); \n "
+ end_loop_str += "\n}\n\n"
+
+ end_loop_str += "final_accuracy = final_accuracy / batch_count; \n"
+ end_loop_str += "dumpFinalAccuracy(final_accuracy); \n\n"
+
+ self.program_str += end_loop_str
+
+
+
+
+ def generateSourceProgram(self, dir_prefix):
+
+ f = open(dir_prefix + "/src.cc", "w+")
+ f.write(self.program_str)
+ f.close()
+
+
+
+ def translate(self, model, weights_dir, test_data, test_labels):
+
+ self.add_header()
+
+ dir_path = "std::string dir_prefix = std::string(\"" + weights_dir + "\"); \n"
+ self.weight_str += dir_path
+
+ if test_data is not None:
+ self.genInputCalls(test_data, test_labels, weights_dir)
+
+ self.dump_weights(model, weights_dir)
+ self.program_str += "\n" + self.weight_str + "\n\n"
+
+ self.genBatchLoop(test_data)
+
+ self.codegen(self.dfg)
+
+ self.endBatchLoop()
+
+ self.add_footer(test_data);
+
+ self.generateSourceProgram(weights_dir)
+
+
+
+
+
+def reloadModelParams(model, reload_dir, x_test, y_test):
+
+ print ("\n\n*****NOTE: Reloading pre-trained weights \n")
+
+ score = model.evaluate(x_test, y_test, verbose=0)
+ print('Test loss2:', score[0])
+ print('Test accuracy2:', score[1])
+
+ for i in range(len(model.layers)):
+ layer = model.layers[i]
+ layer_name = layer.name
+ print ("*layer_name = ", layer_name)
+
+ if "conv" not in layer_name and "dense" not in layer_name:
+ continue
+
+ w_path = reload_dir + layer_name + "_w.bin"
+ b_path = reload_dir + layer_name + "_b.bin"
+
+ print ("** w_path = ", w_path)
+ print ("** b_path = ", b_path)
+
+ w_arr = np.fromfile(w_path, dtype='float32')
+ b_arr = np.fromfile(b_path, dtype='float32')
+
+ w_shape = layer.get_weights()[0].shape
+ b_shape = layer.get_weights()[1].shape
+
+ if "conv" in layer_name:
+ w_nchw_shape = (w_shape[3], w_shape[2], w_shape[0], w_shape[1])
+ w_arr = np.reshape(w_arr, w_nchw_shape)
+ b_arr = np.reshape(b_arr, b_shape)
+
+ w_arr = np.transpose(w_arr, (2,3,1,0))
+ print ("old_shape = ", w_shape, " new_shape = ", w_arr.shape)
+
+ if "dense" in layer_name:
+ w_arr = np.reshape(w_arr, w_shape)
+ b_arr = np.reshape(b_arr, b_shape)
+
+ weights = []
+ weights.append(w_arr)
+ weights.append(b_arr)
+ # NOTE: overriding weights
+ layer.set_weights(weights)
+
+ score = model.evaluate(x_test, y_test, verbose=0)
+ print('Test loss2:', score[0])
+ print('Test accuracy2:', score[1])
+
+
+def getUniquePath(weights_dir):
+
+ # Do not overwrite existing directories - create new with unique ID
+ if os.path.exists(weights_dir):
+ char_count = len(weights_dir)
+ if weights_dir[char_count - 1] == "/":
+ weights_dir = weights_dir[:char_count-1]
+
+ tokens = weights_dir.split("_")
+ last_tok = tokens[len(tokens) - 1]
+ if last_tok.isdigit():
+ id = int(last_tok)
+ id += 1
+ weights_dir = "_".join(tokens[:-1]) + "_" + str(id) + "/"
+ else:
+ weights_dir = "_".join(tokens) + "_1/"
+
+ weights_dir = getUniquePath(weights_dir)
+
+ #print (weights_dir)
+
+ return weights_dir
+
+
+#***** Top level External Function *******
+def translate_to_approxhpvm(model, weights_dir, test_data=None, test_labels=None,
+ num_classes=10, reload_dir=None):
+
+
+ weights_dir = getUniquePath(weights_dir)
+ os.mkdir(weights_dir)
+
+
+ if reload_dir is not None:
+ y_test = keras.utils.to_categorical(test_labels, num_classes)
+ reloadModelParams(model, reload_dir, test_data, y_test)
+
+
+ dfg = DFG()
+ for i in range(len(model.layers)):
+ layer = model.layers[i]
+ # NOTE: Add DNN layer to graph
+ dfg.add_to_graph(layer)
+
+ # Print DFG in reverse postorder
+ dfg.printDFG()
+
+
+ print ("test_data.shape = ", test_data.shape, "\n")
+ print ("test_labels.shape = ", test_labels.shape, "\n")
+
+ tensorRtTranslator = TensorRtTranslator(dfg)
+ tensorRtTranslator.translate(model, weights_dir, test_data, test_labels)
+ weight_str = tensorRtTranslator.getWeightStr()
+ input_str = tensorRtTranslator.getInputStr()
+
+
+ promiseRtTranslator = PromiseRtTranslator(dfg, weight_str)
+ promiseRtTranslator.translate(model, weights_dir, test_data)
+
+ filter_names = tensorRtTranslator.getFilterNames()
+ hpvmTranslator = HPVMTranslator(dfg, weight_str, input_str, filter_names)
+ hpvmTranslator.translate(model, weights_dir, test_data)
+
+
diff --git a/llvm/projects/keras/frontend/hpvm_dfg_translator.py b/llvm/projects/keras/frontend/hpvm_dfg_translator.py
new file mode 100644
index 0000000000..8f33a4c5e0
--- /dev/null
+++ b/llvm/projects/keras/frontend/hpvm_dfg_translator.py
@@ -0,0 +1,568 @@
+
+import sys
+
+
+class HPVMTranslator:
+
+ def __init__(self, dfg, weight_str, input_str, filter_names):
+ self.dfg = dfg
+ self.output_map = {}
+ self.counter = 0
+ self.weight_str = weight_str
+ self.input_str = input_str
+ self.filter_names = filter_names
+ self.node_str = ""
+ self.root_str = ""
+ self.root_struct_str = ""
+ self.main_func_str = ""
+ self.file_header_str = ""
+ self.hpvm_node_names = {}
+
+
+
+
+
+ def getVariableName(self, cur_node):
+ output_var_name = "var_" + str(self.counter)
+ self.counter += 1
+ self.output_map[cur_node.layer_name] = output_var_name
+ self.hpvm_node_names[output_var_name] = 1
+
+ return output_var_name
+
+
+
+ def isSkipLayer(self, layer_type):
+ skip_layers = {}
+ skip_layers["Flatten"] = 0
+ skip_layers["Dropout"] = 0
+ skip_layers["ZeroPadding2D"] = 0
+
+ if layer_type in skip_layers:
+ return True
+ else:
+ return False
+
+
+
+ # NOTE: returns the previous DFG node ignoring "Flatten", "Dropout" Layers
+ def getPrevActiveLayer(self, cur_node):
+ pred_layer_type = cur_node.inputs[0].layer_type
+ # FIXME: Assuming the 'inference' phase - hence skipping Dropout
+ #if pred_layer_type == "Flatten" or pred_layer_type == "Dropout":
+ if self.isSkipLayer(pred_layer_type):
+ cur_node = self.getPrevActiveLayer(cur_node.inputs[0])
+ return cur_node
+ else:
+ return cur_node
+
+
+ def getSingleInputName(self, cur_node):
+ print (cur_node.layer_name)
+ # Assumption: If no inputs, the previous layer must be input layer
+ if len(cur_node.inputs) == 0:
+ return "input"
+
+ print ("Input_type = ", cur_node.inputs[0].layer_type)
+ pred_layer_type = cur_node.inputs[0].layer_type
+ # FIXME: Assuming the 'inference' phase - hence skipping Dropout
+ #if pred_layer_type == "Flatten" or pred_layer_type == "Dropout":
+ if self.isSkipLayer(pred_layer_type):
+ cur_node = self.getPrevActiveLayer(cur_node)
+
+ if cur_node.inputs[0].layer_type == "InputLayer":
+ return "input"
+
+ # get input to the layer
+ input_node_name = cur_node.inputs[0].layer_name # get the input layer ID
+
+ input_var_name = ""
+ if input_node_name in self.output_map:
+ input_var_name = self.output_map[input_node_name]
+ else:
+ print ("Input Var not found - Aborting....")
+ sys.exit(0)
+
+ return input_var_name
+
+
+
+ def getPrevLayerPadding(self, cur_node):
+ print (cur_node.layer_name)
+ # Assumption: If no inputs, the previous layer must be input layer
+ if len(cur_node.inputs) == 0:
+ return None
+
+ print ("Input_type = ", cur_node.inputs[0].layer_type)
+ if cur_node.inputs[0].layer_type == "ZeroPadding2D":
+ pred_padding = cur_node.inputs[0].padding
+ return pred_padding
+
+ return None
+
+
+
+ def getMultipleInputNames(self, cur_node):
+ var_names = []
+ for i in range(len(cur_node.inputs)):
+ # get input to the layer
+ input_node_name = cur_node.inputs[i].layer_name # get the input layer ID
+
+ input_var_name = ""
+ if input_node_name in self.output_map:
+ input_var_name = self.output_map[input_node_name]
+ var_names.append(input_var_name)
+ else:
+ print ("Input Var not found - Aborting....")
+ sys.exit(0)
+
+ return var_names
+
+
+
+ def hasBiasAdd(self, cur_node):
+ if cur_node.layer_type == "Conv2D" or cur_node.layer_type == "Dense":
+ return cur_node.use_bias
+
+ return False
+
+
+ def hasActivation(self, cur_node):
+ if cur_node.layer_type == "Conv2D" or cur_node.layer_type == "Dense":
+ return cur_node.activation_type != "linear"
+
+ return False
+
+
+ def genActivationCall(self, input_var, output_var, activation_type):
+ header_str = self.genNodeHeader(output_var, 1)
+ inst_str = header_str
+
+ func_name = "__visc__tensor_"
+ if activation_type == "tanh":
+ func_name += "tanh"
+
+ if activation_type == "relu":
+ func_name += "relu"
+
+ if activation_type == "softmax":
+ func_name += "softmax"
+
+ inst_str += " void* r = " + func_name + "(t1); \n"
+ footer_str = self.genNodeFooter(2)
+ inst_str += footer_str
+
+ return inst_str
+
+
+ def genNodeHeader(self, var_name, num_params):
+ node_header_str = "void " + var_name + "_node("
+ for i in range(num_params):
+ node_header_str += "void* t" + str(i + 1) + ", "
+ node_header_str += "size_t bytes_t" + str(i + 1)
+ if i < num_params - 1:
+ node_header_str += ", "
+
+ node_header_str += ") { \n"
+ node_header_str += " __visc__hint(visc::CUDNN_TARGET); \n"
+ node_header_str += " __visc__attributes(" + str(num_params) + ", "
+
+ for i in range(num_params):
+ node_header_str += "t" + str(i + 1)
+ if i < num_params - 1:
+ node_header_str += ", "
+
+ node_header_str += ", 0); \n\n"
+ return node_header_str
+
+
+ def genNodeFooter(self, num_params):
+ node_footer_str = " __visc__return("
+ node_footer_str += str(num_params) + ", "
+ node_footer_str += "r, "
+ node_footer_str += "(size_t) 0); \n"
+ node_footer_str += "}\n\n"
+
+ return node_footer_str
+
+
+
+ def genHpvmNodeEdges(self, out_var_name, input_var_name, input_var_name2):
+
+ print ("input_var_name2 = ", input_var_name2)
+ print ("input_var_name = ", input_var_name)
+
+ hpvm_edge_str = "\n void* " + out_var_name + " = "
+ hpvm_edge_str += "__visc__createNodeND(0, " + out_var_name + "_node); \n\n"
+
+ if input_var_name in self.filter_names:
+ input_index = self.filter_names[input_var_name]
+ index1 = input_index * 2
+ index2 = index1 + 1
+ hpvm_edge_str += " __visc__bindIn(" + out_var_name + ", " + str(index1) + ", 0, 0); \n"
+ hpvm_edge_str += " __visc__bindIn(" + out_var_name + ", " + str(index2) + ", 1, 0); \n"
+
+ elif input_var_name in self.hpvm_node_names:
+ #input_index = self.output_map[input_var_name]
+ hpvm_edge_str += " __visc__edge(" + input_var_name + ", " + out_var_name + ", 1, 0, 0, 0); \n"
+ hpvm_edge_str += " __visc__edge(" + input_var_name + ", " + out_var_name + ", 1, 1, 1, 0); \n"
+
+
+ if input_var_name2 in self.filter_names:
+ input_index = self.filter_names[input_var_name2]
+ index1 = input_index * 2
+ index2 = index1 + 1
+ hpvm_edge_str += " __visc__bindIn(" + out_var_name + ", " + str(index1) + ", 2, 0); \n"
+ hpvm_edge_str += " __visc__bindIn(" + out_var_name + ", " + str(index2) + ", 3, 0); \n"
+
+ elif input_var_name2 in self.hpvm_node_names:
+ #input_index = self.output_map[input_var_name2]
+ hpvm_edge_str += " __visc__edge(" + input_var_name2 + ", " + out_var_name + ", 1, 0, 2, 0); \n"
+ hpvm_edge_str += " __visc__edge(" + input_var_name2 + ", " + out_var_name + ", 1, 1, 3, 0); \n"
+
+
+ return hpvm_edge_str
+
+
+
+ def genDenseNode(self, cur_node):
+ out_var_name = self.getVariableName(cur_node)
+
+ header_str = self.genNodeHeader(out_var_name, 2)
+ inst_str = header_str
+ inst_str += " void *r = __visc__tensor_mul(t1, t2); \n"
+ footer_str = self.genNodeFooter(2)
+ inst_str += footer_str
+
+ #self.genHpvmNodeVar(out_var_name)
+ input_var_name = self.getSingleInputName(cur_node)
+ weight_name = cur_node.layer_name + "_w"
+
+ self.root_str += self.genHpvmNodeEdges(out_var_name, input_var_name, weight_name)
+
+ self.node_str += inst_str
+
+
+
+
+ def genConvNode(self, cur_node):
+ #input_var_name = self.getSingleInputName(cur_node)
+ out_var_name = self.getVariableName(cur_node)
+
+ header_str = self.genNodeHeader(out_var_name, 2)
+ inst_str = header_str
+
+ weights = cur_node.weights
+ strides = cur_node.strides
+
+ padding = 0
+ if cur_node.padding.strip() == "valid":
+ padding = 0
+ else:
+ padding = cur_node.padding
+ padding = int((weights.shape[0] - 1) / 2)
+
+ prev_padding = self.getPrevLayerPadding(cur_node)
+ if prev_padding != None:
+ # FIXME: currently only supporting symmetric padding
+ padding = prev_padding[0][0]
+
+ inst_str += " void *r = __visc__tensor_convolution(t1, t2, "
+ inst_str += str(padding) + ", "
+ inst_str += str(padding) + ", "
+ inst_str += str(strides[0]) + ", "
+ inst_str += str(strides[1])
+ inst_str += "); \n"
+
+ footer_str = self.genNodeFooter(2)
+ inst_str += footer_str
+
+ self.node_str += inst_str
+
+ input_var_name = self.getSingleInputName(cur_node)
+ weight_name = cur_node.layer_name + "_w"
+
+ self.root_str += self.genHpvmNodeEdges(out_var_name, input_var_name, weight_name)
+
+
+
+ def genBiasNode(self, cur_node):
+ input_var_name = self.output_map[cur_node.layer_name]
+ out_var_name = self.getVariableName(cur_node)
+
+ header_str = self.genNodeHeader(out_var_name, 2)
+ inst_str = header_str
+ inst_str += " void *r = __visc__tensor_add(t1, t2); \n"
+ footer_str = self.genNodeFooter(2)
+ inst_str += footer_str
+
+ self.node_str += inst_str
+
+ weight_name = cur_node.layer_name + "_b"
+ self.root_str += self.genHpvmNodeEdges(out_var_name, input_var_name, weight_name)
+
+
+
+ def genSubActivationNode(self, cur_node):
+ input_var_name = self.output_map[cur_node.layer_name]
+ out_var_name = self.getVariableName(cur_node)
+ activation_type = cur_node.activation_type
+
+ inst_str = self.genActivationCall(input_var_name, out_var_name, activation_type)
+
+ self.node_str += inst_str
+
+ self.root_str += self.genHpvmNodeEdges(out_var_name, input_var_name, "")
+
+
+
+ def genActivationNode(self, cur_node):
+ input_var_name = self.getSingleInputName(cur_node)
+ out_var_name = self.getVariableName(cur_node)
+ activation_type = cur_node.activation_type
+
+ inst_str = self.genActivationCall(input_var_name, out_var_name, activation_type)
+ self.node_str += inst_str
+
+ self.root_str += self.genHpvmNodeEdges(out_var_name, input_var_name, "")
+
+
+ def genAddNode(self, cur_node):
+ out_var_name = self.getVariableName(cur_node)
+
+ header_str = self.genNodeHeader(out_var_name, 2)
+ inst_str = header_str
+ inst_str += " void *r = __visc__tensor_add(t1, t2); \n"
+ footer_str = self.genNodeFooter(2)
+ inst_str += footer_str
+
+ self.node_str += inst_str
+
+ input_vars = self.getMultipleInputNames(cur_node)
+ self.root_str += self.genHpvmNodeEdges(out_var_name, input_vars[0], input_vars[1])
+
+
+
+ def genPoolNode(self, cur_node):
+ out_var_name = self.getVariableName(cur_node)
+
+ header_str = self.genNodeHeader(out_var_name, 1)
+ inst_str = header_str
+
+ pool_size = cur_node.pool_size
+ strides = cur_node.strides
+ # FIXME: Non-same padding is *NOT* currently supported
+ padding = 0
+ pool_type = 0
+ func_name = ""
+
+ layer_type = cur_node.layer_type
+ if layer_type == "MaxPooling2D":
+ func_name = "__visc__tensor_pool_max"
+ if layer_type == "AveragePooling2D":
+ func_name = "__visc__tensor_pool_avg"
+
+ inst_str += " void* r = " + func_name + "(t1, "
+ inst_str += str(pool_size[0]) + ", " + str(pool_size[1]) + ", "
+ inst_str += str(padding) + ", " + str(padding) + ", "
+ inst_str += str(strides[0]) + ", " + str(strides[1]) + "); \n"
+
+ footer_str = self.genNodeFooter(2)
+ inst_str += footer_str
+
+ self.node_str += inst_str
+
+ input_var_name = self.getSingleInputName(cur_node)
+ self.root_str += self.genHpvmNodeEdges(out_var_name, input_var_name, "")
+
+
+ # Checks for call-type and redirects to functions that generate HPVM nodes
+ def genHpvmNodes(self, cur_node):
+
+ layer_type = cur_node.layer_type
+
+ if layer_type == "Conv2D":
+ self.genConvNode(cur_node)
+
+ if layer_type == "Dense":
+ self.genDenseNode(cur_node)
+
+ if self.hasBiasAdd(cur_node):
+ self.genBiasNode(cur_node)
+
+ if self.hasActivation(cur_node):
+ self.genSubActivationNode(cur_node)
+
+ if layer_type == "Activation":
+ self.genActivationNode(cur_node)
+
+ if layer_type == "Add":
+ self.genAddNode(cur_node)
+
+ if layer_type == "MaxPooling2D" or layer_type == "AveragePooling2D":
+ self.genPoolNode(cur_node)
+
+
+
+
+ def codegenNode(self, dfg, cur_node, visited_nodes):
+
+ # Skip visited nodes
+ if cur_node.layer_name in visited_nodes:
+ return
+
+ if dfg.predVisited(cur_node, visited_nodes):
+ visited_nodes[cur_node.layer_name] = True
+
+ self.genHpvmNodes(cur_node)
+ # Invoking traversal on outbound nodes
+ for output_node in cur_node.outputs:
+ self.codegenNode(dfg, output_node, visited_nodes)
+
+
+
+ # Print the DFG in reverse postorder
+ def codegen(self, dfg):
+ print ("\n\n ****** Codegen for ApproxHPVM DFG Representation ******* \n\n")
+ visited_nodes = {}
+ # Starting traversal at the DFG root node
+ self.codegenNode(dfg, dfg.root_node, visited_nodes)
+
+
+
+
+ def genFileHeader(self):
+ headers = "\n#include <stdio.h> \n"
+ headers += "#include <stdlib.h> \n"
+ headers += "#include <unistd.h> \n"
+ headers += "#include <fcntl.h> \n"
+ headers += "#include <sys/stat.h> \n"
+ headers += "#include <cstring> \n"
+
+ headers += "#include <visc.h> \n"
+ headers += "#include <tensorTypes.h> \n"
+ headers += "#include <tensorUtils.h> \n\n"
+
+ self.file_header_str = headers
+
+
+
+ def genRootNodeHeader(self):
+ root_signature = "void root("
+ index = 0
+ for f_name in self.filter_names:
+ if index > 0:
+ root_signature += "\t "
+ self.filter_names[f_name] = index
+ root_signature += "void* " + f_name + ", "
+ root_signature += "size_t " + f_name + "_bytes"
+ if index < len(self.filter_names) - 1:
+ root_signature += ", \n"
+ index += 1
+
+ root_signature += "){ \n\n"
+ root_signature += "\n __visc__hint(visc::CPU_TARGET); \n"
+ root_signature += " __visc__attributes(" + str(len(self.filter_names)) + ", "
+
+ index = 0
+ for f_name in self.filter_names:
+ root_signature += f_name
+ if index < len(self.filter_names) - 1:
+ root_signature += ", "
+ index += 1
+
+ root_signature += ", 0); \n\n"
+
+ self.root_str += root_signature
+
+
+ def genRootNodeFooter(self):
+ last_node = self.dfg.last_node
+ output_var = self.output_map[last_node.layer_name]
+
+ # Binding output of last DFG node to the Root Node output
+ root_footer_str = "\n __visc__bindOut(" + output_var + ", 0, 0, 0); \n"
+ root_footer_str += " __visc__bindOut(" + output_var + ", 1, 1, 0); \n"
+ root_footer_str += "\n}\n\n"
+
+ self.root_str += root_footer_str
+
+
+
+ def genRootStructure(self):
+ root_struct = ""
+ root_struct += "struct ret_t {\n"
+ root_struct += " void* tensor; \n"
+ root_struct += " size_t bytes; \n"
+ root_struct += "}; \n\n"
+
+ root_struct += "typedef struct __attribute__((__packed__)) {\n"
+ for f_name in self.filter_names:
+ root_struct += " void* " + f_name + "; \n"
+ root_struct += " size_t " + f_name + "_bytes; \n"
+
+ root_struct += "\n struct ret_t r; \n"
+ root_struct += "}\nRootIn;\n\n"
+
+ self.root_struct_str += root_struct
+
+
+
+ def genMainFunction(self, test_data):
+ main_func_str = "int main(){ \n\n"
+ main_func_str += self.weight_str
+ main_func_str += self.input_str
+ main_func_str += "\n__visc__init(); \n"
+ main_func_str += "RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn))); \n\n"
+
+ for f_name in self.filter_names:
+ main_func_str += "args->" + f_name + " = " + f_name + "; \n"
+ main_func_str += "args->" + f_name + "_bytes = 0; \n"
+
+ main_func_str += "\nvoid* dfg = __visc__launch(0, root, (void*) args); \n\n"
+ main_func_str += "__visc__wait(dfg); \n\n"
+
+ main_func_str += "void *result = static_cast<RootIn*>(args)->input; \n"
+ main_func_str += "hpvm_request_tensor(result, 0); \n\n"
+ main_func_str += "__visc__cleanup(); \n "
+
+ main_func_str += "computeAccuracy2(labels, " + str(len(test_data)) + ", result); \n"
+ main_func_str += "return 0; \n\n"
+ main_func_str += "} \n"
+
+ self.main_func_str += main_func_str
+
+
+
+
+ def generateSourceProgram(self, dir_prefix):
+ #print (self.node_str)
+ #print (self.root_str)
+ #print (self.root_struct_str)
+ #print (self.main_func_str)
+
+ program_str = self.file_header_str + self.node_str + self.root_str
+ program_str += self.root_struct_str + self.main_func_str
+
+ print (program_str)
+
+ f = open(dir_prefix + "/approxhpvm_src.cc", "w+")
+ f.write(program_str)
+ f.close()
+
+
+
+ def translate(self, model, weights_dir, test_data):
+
+ self.genFileHeader()
+ self.genRootNodeHeader()
+ self.genRootStructure()
+ self.codegen(self.dfg)
+ self.genRootNodeFooter()
+ self.genMainFunction(test_data)
+
+ # dump generated program string to source file
+ self.generateSourceProgram(weights_dir)
+
+
+
diff --git a/llvm/projects/keras/frontend/promise_translator.py b/llvm/projects/keras/frontend/promise_translator.py
new file mode 100644
index 0000000000..a4843c5291
--- /dev/null
+++ b/llvm/projects/keras/frontend/promise_translator.py
@@ -0,0 +1,1079 @@
+
+
+import numpy as np
+import sys
+from keras import backend as K
+
+
+
+class State:
+
+ def __init__(self):
+ self.ops = []
+ self.op_string = ""
+ self.num_ops = 0
+
+ def clear(self):
+ self.ops = []
+ self.op_string = ""
+ self.num_ops = 0
+
+ def add(self, cur_node, layer_type):
+ self.ops.append(cur_node)
+ if self.op_string != "":
+ self.op_string += "_"
+
+ self.op_string += layer_type
+ self.num_ops += 1
+
+
+ def getFirstOp(self):
+ return self.ops[0]
+
+ def getLastOp(self):
+ return self.ops[self.num_ops - 1]
+
+ def getDenseOp(self):
+ for i in range(self.num_ops):
+ if self.ops[i].layer_type == "Dense":
+ return self.ops[i]
+ return None # Should be unreachable
+
+
+ def getConvOp(self):
+ for i in range(self.num_ops):
+ if self.ops[i].layer_type == "Conv2D":
+ return self.ops[i]
+ return None # Should be unreachable
+
+
+ def getPoolOp(self):
+ for i in range(self.num_ops):
+ layer_type = self.ops[i].layer_type
+ if layer_type == "MaxPooling2D" or layer_type == "AveragePooling2D":
+ return self.ops[i]
+ return None # Should be unreachable
+
+
+ def getPadOp(self):
+ for i in range(self.num_ops):
+ layer_type = self.ops[i].layer_type
+ if layer_type == "ZeroPadding2D":
+ return self.ops[i]
+ return None # Should be unreachable
+
+
+ def getActivationID(self):
+ activation_type = 'linear'
+ for i in range(self.num_ops):
+ cur_op = self.ops[i]
+ layer_type = cur_op.layer_type
+ if layer_type == "Dense" or layer_type == "Conv2D" or layer_type == "Activation":
+ activation_type = cur_op.activation_type
+
+ activation_id = -1;
+ if activation_type == "tanh":
+ activation_id = 0
+ if activation_type == "relu":
+ activation_id = 1
+
+ return activation_id
+
+
+ def isDense(self):
+ if "dense" in self.op_string:
+ return True
+ return False
+
+ def isConv(self):
+ if "conv" in self.op_string:
+ return True
+ return False
+
+ def isPool(self):
+ if "pool" in self.op_string and self.num_ops == 1:
+ return True
+ return False
+
+
+ def isActivation(self):
+ if "activation" in self.op_string and self.num_ops == 1:
+ return True
+ return False
+
+
+ def getPadding(self):
+ padding_op = self.getPadOp()
+ prev_padding = 0
+ if padding_op is not None:
+ prev_padding = padding_op.padding[0][0]
+
+ conv_op = self.getConvOp()
+ if conv_op is None:
+ print ("ERROR: Conv Op not found")
+ sys.exit(0)
+
+ K = conv_op.weights.shape[0]
+ padding = 0
+ padding_type = conv_op.padding.strip()
+ if padding_type == "valid":
+ padding = 0
+ else:
+ padding = int((K - 1) / 2)
+
+ padding = padding + prev_padding
+
+ return padding
+
+
+ def getPoolInfo(self):
+ pool_op = self.getPoolOp()
+ if pool_op is None:
+ return -1, [0,0]
+
+ pool_id = -1
+ layer_type = pool_op.layer_type
+ if layer_type == "MaxPooling2D":
+ pool_id = 0
+ if layer_type == "AveragePooling2D":
+ pool_id = 1
+
+ pool_size = pool_op.pool_size
+ return pool_id, pool_size
+
+
+ def getStrides(self):
+ conv_op = self.getConvOp()
+ strides = conv_op.strides
+ return strides
+
+
+
+
+class PromiseRtTranslator:
+
+ # NOTE: weight_str can be optinally passed
+ def __init__(self, dfg, weight_str = ""):
+ self.dfg = dfg
+ self.output_map = {}
+ self.visited_nodes = {}
+ self.counter = 0
+ self.weight_str = weight_str
+ self.program_str = ""
+ self.swing_value = 9 # FP32
+ self.quant_ranges = {}
+ # Used to generate PromiseSim Info
+ self.layer_str = ""
+ self.layer_size_str = ""
+ self.layer_input_sizes = {}
+ self.unique_op_types = {}
+ self.batch_size = 0
+
+
+ def getVariableName(self, cur_node):
+
+ output_var_name = "var_" + str(self.counter)
+ self.counter += 1
+ self.output_map[cur_node.layer_name] = output_var_name
+
+ return output_var_name
+
+
+ def isSkipLayer(self, layer_type):
+
+ skip_layers = {}
+ skip_layers["Flatten"] = 0
+ skip_layers["Dropout"] = 0
+ skip_layers["ZeroPadding2D"] = 0
+
+ if layer_type in skip_layers:
+ return True
+ else:
+ return False
+
+
+ def isForwardLayer(self, layer_type):
+
+ skip_layers = {}
+ skip_layers["Input"] = 0
+ skip_layers["InputLayer"] = 0
+ skip_layers["Flatten"] = 0
+ skip_layers["Dropout"] = 0
+ if layer_type in skip_layers:
+ return True
+ else:
+ return False
+
+
+
+ def appendLayerSizeStr(self, promise_layer_type, state):
+
+ central_op = None
+ if promise_layer_type == "Conv":
+ central_op = state.getConvOp()
+ if promise_layer_type == "FC":
+ central_op = state.getDenseOp()
+
+ first_op = state.getFirstOp()
+ layer_name = first_op.layer_name
+
+ print("** layer_name = ", layer_name)
+
+ unique_id = 0
+ if promise_layer_type not in self.unique_op_types:
+ self.unique_op_types[promise_layer_type] = 1
+ unique_id = 1
+ else:
+ unique_id = self.unique_op_types[promise_layer_type]
+ unique_id += 1
+ self.unique_op_types[promise_layer_type] = unique_id
+
+ unique_layer_name = promise_layer_type + str(unique_id)
+ if promise_layer_type == "Conv" or promise_layer_type == "FC":
+ self.layer_size_str += unique_layer_name + ","
+ else:
+ # Handling single tensor ops - NO Promise layer
+ self.layer_size_str += "#tensor" + unique_layer_name + "\n"
+ return
+
+
+ weights_shape = central_op.weights.shape
+ input_size = self.layer_input_sizes[layer_name]
+ #print ("layer_name = ", layer_name, " input_size = ", input_size)
+ N = self.batch_size
+ C = input_size[1]
+
+ if str(C) == "?":
+ C = weights_shape[0]
+
+ self.layer_size_str += str(N) + "," + str(C) + ","
+
+ if promise_layer_type == "Conv":
+ H = input_size[2]
+ W = input_size[3]
+ self.layer_size_str += str(H) + "," + str(W) + ","
+
+
+ H = weights_shape[0]
+ W = weights_shape[1]
+
+ if promise_layer_type == "Conv":
+ N = weights_shape[3]
+ C = weights_shape[2]
+ self.layer_size_str += str(N) + "," + str(C) + ","
+
+ self.layer_size_str += str(H) + "," + str(W)
+
+ self.layer_size_str += "\n"
+
+
+
+
+
+ def appendLayerString(self, promise_layer_type, state):
+
+ layer_str = ""
+ for op in state.ops:
+ op_type = op.layer_type
+ if op_type == "Conv2D":
+ layer_str += "conv "
+ if op.use_bias:
+ layer_str += "add "
+ if op.activation_type != "linear":
+ layer_str += "activation "
+
+ if op_type == "Dense":
+ layer_str += "dense "
+ if op.use_bias:
+ layer_str += "add "
+ if op.activation_type != "linear":
+ layer_str += "activation "
+
+ if "Pooling" in op_type:
+ layer_str += "pool "
+
+ if op_type == "Add":
+ layer_str += "add "
+
+ if op_type == "Activation":
+ layer_str += "activation "
+
+ layer_str += "\n"
+
+ self.layer_str += layer_str
+
+ self.appendLayerSizeStr(promise_layer_type, state)
+
+
+
+
+ # NOTE: returns the previous DFG node ignoring "Flatten", "Dropout" Layers
+ def getPrevActiveLayer(self, cur_node):
+
+ pred_layer_type = cur_node.inputs[0].layer_type
+ # FIXME: Assuming the 'inference' phase - hence skipping Dropout
+ #if pred_layer_type == "Flatten" or pred_layer_type == "Dropout":
+ if self.isSkipLayer(pred_layer_type):
+ cur_node = self.getPrevActiveLayer(cur_node.inputs[0])
+ return cur_node
+ else:
+ return cur_node
+
+
+ # Retrieve input name of the previous layer
+ def getInputLayerName(self, cur_node):
+
+ print (cur_node.layer_name)
+ # Assumption: If no inputs, the previous layer must be input layer
+ if len(cur_node.inputs) == 0:
+ return "input"
+
+ print ("Input_type = ", cur_node.inputs[0].layer_type)
+
+ pred_layer_type = cur_node.inputs[0].layer_type
+ # FIXME: Assuming the 'inference' phase - hence skipping Dropout
+ #if pred_layer_type == "Flatten" or pred_layer_type == "Dropout":
+ if self.isSkipLayer(pred_layer_type):
+ cur_node = self.getPrevActiveLayer(cur_node)
+
+ if cur_node.inputs[0].layer_type == "InputLayer":
+ return "input"
+
+ # get input to the layer
+ input_node_name = cur_node.inputs[0].layer_name # get the input layer ID
+
+ return input_node_name
+
+
+
+ # Retrieve input name of the previous layer
+ def getSingleInputName(self, cur_node):
+
+ print (cur_node.layer_name)
+ # Assumption: If no inputs, the previous layer must be input layer
+ if len(cur_node.inputs) == 0:
+ return "input"
+
+ print ("Input_type = ", cur_node.inputs[0].layer_type)
+
+ pred_layer_type = cur_node.inputs[0].layer_type
+ # FIXME: Assuming the 'inference' phase - hence skipping Dropout
+ #if pred_layer_type == "Flatten" or pred_layer_type == "Dropout":
+ if self.isSkipLayer(pred_layer_type):
+ cur_node = self.getPrevActiveLayer(cur_node)
+
+ if cur_node.inputs[0].layer_type == "InputLayer":
+ return "input"
+
+ # get input to the layer
+ input_node_name = cur_node.inputs[0].layer_name # get the input layer ID
+
+
+ input_var_name = ""
+ if input_node_name in self.output_map:
+ input_var_name = self.output_map[input_node_name]
+ else:
+ print ("Input Var not found - Aborting....")
+ sys.exit(0)
+
+ return input_var_name
+
+
+
+ # Used to retrieve inputs for "add" operation with multiple inputs
+ def getMultipleInputNames(self, cur_node):
+
+ var_names = []
+ for i in range(len(cur_node.inputs)):
+ # get input to the layer
+ input_node_name = cur_node.inputs[i].layer_name # get the input layer ID
+
+ input_var_name = ""
+ if input_node_name in self.output_map:
+ input_var_name = self.output_map[input_node_name]
+ var_names.append(input_var_name)
+ else:
+ print ("Input Var not found - Aborting....")
+ sys.exit(0)
+
+ return var_names
+
+
+
+ def getWeightNames(self, cur_node):
+
+ layer_name = cur_node.layer_name
+ w_name = layer_name + "_w"
+ b_name = layer_name + "_b"
+ # If Conv has no bias Add operation
+ if cur_node.use_bias == False:
+ b_name = "NULL"
+
+ return w_name, b_name
+
+
+ def getWeightRange(self, cur_node):
+
+ layer_type = cur_node.layer_type
+ if layer_type != "Dense" and layer_type != "Conv2D":
+ print ("ERROR: layer_type = ", layer_type , " does not have weights ")
+ sys.exit(0)
+
+ weights = cur_node.weights
+ min_val = np.amin(weights)
+ max_val = np.amax(weights)
+
+ return min_val, max_val
+
+
+ def getBiasRange(self, cur_node):
+
+ layer_type = cur_node.layer_type
+ if layer_type != "Dense" and layer_type != "Conv2D":
+ print ("ERROR: layer_type = ", layer_type , " does not have weights ")
+ sys.exit(0)
+
+ if cur_node.use_bias == False:
+ return 0, 0
+
+ bias_weights = cur_node.bias_weights
+ min_val = np.amin(bias_weights)
+ max_val = np.amax(bias_weights)
+
+ return min_val, max_val
+
+
+ # Returns the output value ranges for the input and output to a PROMISE layer
+ def getQuantRange(self, state):
+
+ first_op = state.getFirstOp()
+ last_op = state.getLastOp()
+
+ prev_layer_name = self.getInputLayerName(first_op)
+ cur_layer_name = last_op.layer_name
+
+ print ("prev_layer_name ", prev_layer_name , " cur_layer_name = ", cur_layer_name)
+
+ if prev_layer_name not in self.quant_ranges or cur_layer_name not in self.quant_ranges:
+ print ("ERROR: Layer_name = ", prev_layer_name ," or ", cur_layer_name, " not found in quant_range")
+ sys.exit(0)
+
+ input_quant_range = self.quant_ranges[prev_layer_name]
+ output_quant_range = self.quant_ranges[cur_layer_name]
+
+ print (input_quant_range)
+ print (output_quant_range)
+
+ return input_quant_range, output_quant_range
+
+
+
+ def genDenseLayer(self, state):
+
+ print ("\n\n Layer = ", state.op_string, "\n\n")
+
+ first_op = state.getFirstOp()
+ dense_op = state.getDenseOp()
+ last_op = state.getLastOp()
+
+ input_var = self.getSingleInputName(first_op)
+ output_var = self.getVariableName(last_op)
+
+ w_name, b_name = self.getWeightNames(dense_op)
+ w_min, w_max = self.getWeightRange(dense_op)
+ b_min, b_max = self.getBiasRange(dense_op)
+
+ activation_id = state.getActivationID()
+
+ # NOTE: retrieve the quantization ranges for inputs and ouputs
+ input_quant_range, output_quant_range = self.getQuantRange(state)
+
+ promise_layer_str = "void* " + output_var + " = FCLayer_PROMISE(" + input_var + ", "
+ promise_layer_str += str(input_quant_range[0]) + ", " + str(input_quant_range[1]) + ", "
+ promise_layer_str += w_name + ", " + str(w_min) + ", " + str(w_max) + ", "
+ promise_layer_str += b_name + ", " + str(b_min) + ", " + str(b_max) + ", "
+ promise_layer_str += str(activation_id) + ", "
+ promise_layer_str += str(output_quant_range[0]) + ", " + str(output_quant_range[1]) + ", "
+ promise_layer_str += str(self.swing_value)
+ promise_layer_str += "); \n"
+
+ print (promise_layer_str)
+
+ self.program_str += promise_layer_str
+
+
+ self.appendLayerString("FC", state)
+
+ state.clear()
+
+
+
+ def genConvLayer(self, state):
+
+ print ("\n\n Layer = ", state.op_string, "\n\n")
+
+ first_op = state.getFirstOp()
+ conv_op = state.getConvOp()
+ last_op = state.getLastOp()
+
+ input_var = self.getSingleInputName(first_op)
+ output_var = self.getVariableName(last_op)
+
+ w_name, b_name = self.getWeightNames(conv_op)
+ w_min, w_max = self.getWeightRange(conv_op)
+ b_min, b_max = self.getBiasRange(conv_op)
+
+ activation_id = state.getActivationID()
+ padding = state.getPadding()
+ pool_id, pool_size = state.getPoolInfo()
+ strides = state.getStrides()
+
+ # NOTE: retrieve the quantization ranges for inputs and ouputs
+ input_quant_range, output_quant_range = self.getQuantRange(state)
+
+ # NOTE: Assuming symmetric K*K pool size
+ promise_layer_str = "void* " + output_var + " = ConvLayer_PROMISE(" + input_var + ", "
+ promise_layer_str += str(input_quant_range[0]) + ", " + str(input_quant_range[1]) + ", "
+ promise_layer_str += w_name + ", " + str(w_min) + ", " + str(w_max) + ", "
+ promise_layer_str += b_name + ", " + str(b_min) + ", " + str(b_max) + ", "
+ promise_layer_str += str(padding) + ", " + str(padding) + ", "
+ promise_layer_str += str(strides[0]) + ", " + str(strides[1]) + ", "
+ promise_layer_str += str(pool_id) + ", " + str(pool_size[0]) + ", "
+ promise_layer_str += str(activation_id) + ", "
+ promise_layer_str += str(output_quant_range[0]) + ", " + str(output_quant_range[1]) + ", "
+ promise_layer_str += str(self.swing_value)
+ promise_layer_str += "); \n"
+
+ print (promise_layer_str)
+
+ self.program_str += promise_layer_str
+
+
+ self.appendLayerString("Conv", state)
+
+ state.clear()
+
+
+ def genSoftmaxLayer(self, state):
+ print ("\n\n Layer = ", state.op_string, "\n\n")
+
+ first_op = state.getFirstOp()
+ last_op = state.getLastOp()
+
+ input_var = self.getSingleInputName(first_op)
+ output_var = self.getVariableName(last_op)
+
+ promise_layer_str = "void* " + output_var + " = tensorSoftmax(" + input_var + "); \n"
+ print (promise_layer_str)
+
+ self.program_str += promise_layer_str
+
+ state.clear()
+
+
+ def genAddLayer(self, state):
+ print ("\n\n Layer = ", state.op_string, "\n\n")
+
+ first_op = state.getFirstOp()
+ last_op = state.getLastOp()
+
+ input_vars = self.getMultipleInputNames(first_op)
+ output_var = self.getVariableName(last_op)
+
+ promise_layer_str = "void* " + output_var + " = tensorAdd(" + input_vars[0]
+ promise_layer_str += ", " + input_vars[1] + "); \n"
+ print (promise_layer_str)
+
+ self.program_str += promise_layer_str
+
+
+ self.appendLayerString("Add", state)
+
+ state.clear()
+
+
+
+
+ def genActivationLayer(self, state):
+ print ("\n\n Layer = ", state.op_string, "\n\n")
+
+ first_op = state.getFirstOp()
+ input_var = self.getSingleInputName(first_op)
+ output_var = self.getVariableName(first_op)
+
+ activation_type = first_op.activation_type
+
+ func_name = ""
+ if activation_type == "tanh":
+ func_name = "Tanh"
+
+ if activation_type == "relu":
+ func_name = "Relu"
+
+ inst_str = "void* " + output_var + " = "
+ inst_str += "tensor" + func_name + "(" + input_var + "); \n"
+
+ self.program_str += inst_str
+
+
+ self.appendLayerString(func_name, state)
+
+ state.clear()
+
+
+ # FIXME: Only supporting single AveragePooling layers
+ def genPoolLayer(self, state):
+ print ("\n\n Layer = ", state.op_string, "\n\n")
+
+ # For single pool layer should be all same
+ pool_op = state.getPoolOp()
+
+ input_var = self.getSingleInputName(pool_op)
+ output_var = self.getVariableName(pool_op)
+
+ pool_size = pool_op.pool_size
+ strides = pool_op.strides
+ # FIXME: Same padding is *NOT* currently supported
+ padding = 0
+ pool_type = 0
+
+ layer_type = pool_op.layer_type
+ if layer_type == "MaxPooling2D":
+ pool_type = "0"
+ if layer_type == "AveragePooling2D":
+ pool_type = "1"
+
+ # tensorPooling(input, pool_type, pool_h, pool_w, v_pad, h_pad, v_stride, h_stride)
+ inst_str = "void* " + output_var + " = "
+ inst_str += "tensorPooling(" + input_var + "," + pool_type + "," + str(pool_size[0]) + "," + str(pool_size[1])
+ inst_str += "," + str(padding) + "," + str(padding) + "," + str(strides[0]) + "," + str(strides[1])
+ inst_str += "); \n"
+ self.program_str += inst_str
+
+
+ self.appendLayerString("Pooling", state)
+
+ state.clear()
+
+
+
+
+
+ def genPreviousLayer(self, state):
+
+ if state.isDense():
+ self.genDenseLayer(state)
+
+ elif state.isConv():
+ self.genConvLayer(state)
+
+ elif state.isPool():
+ self.genPoolLayer(state)
+
+ elif state.isActivation():
+ self.genActivationLayer(state)
+
+
+ def shouldVisit(self, cur_node):
+ layer_name = cur_node.layer_name
+ # NOTE: visit a node if not already visited and all predecessors are visited
+ if self.dfg.predVisited(cur_node, self.visited_nodes) and layer_name not in self.visited_nodes:
+ return True
+ else:
+ return False
+
+
+ def handle_padding(self, cur_node, state):
+ if not self.shouldVisit(cur_node):
+ return
+
+ layer_name = cur_node.layer_name
+ print (layer_name)
+ self.visited_nodes[layer_name] = True
+
+ self.genPreviousLayer(state)
+
+ # Appending conv to state
+ state.add(cur_node, "padding")
+
+ self.traverseSuccessors(cur_node, state)
+
+
+
+ def handle_conv(self, cur_node, state):
+ if not self.shouldVisit(cur_node):
+ return
+
+ layer_name = cur_node.layer_name
+ print ("handle_conv", layer_name)
+ self.visited_nodes[layer_name] = True
+
+ self.genPreviousLayer(state)
+
+ # Appending conv to state
+ state.add(cur_node, "conv")
+
+ self.traverseSuccessors(cur_node, state)
+
+
+
+ def handle_add(self, cur_node, state):
+ if not self.shouldVisit(cur_node):
+ return
+
+ layer_name = cur_node.layer_name
+ print (layer_name)
+ self.visited_nodes[layer_name] = True
+
+ self.genPreviousLayer(state)
+
+ # Appending conv to state
+ state.add(cur_node, "add")
+
+ self.genAddLayer(state)
+
+ self.traverseSuccessors(cur_node, state)
+
+
+
+ def handle_activation(self, cur_node, state):
+ if not self.shouldVisit(cur_node):
+ return
+
+ layer_name = cur_node.layer_name
+ print (layer_name)
+ self.visited_nodes[layer_name] = True
+
+ # NOTE: If end of DNN
+ if cur_node.activation_type == "softmax":
+ self.genPreviousLayer(state)
+ state.add(cur_node, "activation")
+ self.genSoftmaxLayer(state)
+ # NOTE: return when observed end of DNN (softmax)
+ return
+
+ # Appending activation to state
+ state.add(cur_node, "activation")
+
+ self.traverseSuccessors(cur_node, state)
+
+
+ def handle_dense(self, cur_node, state):
+ if not self.shouldVisit(cur_node):
+ return
+
+ layer_name = cur_node.layer_name
+ print (layer_name)
+ self.visited_nodes[layer_name] = True
+
+ self.genPreviousLayer(state)
+
+ state.add(cur_node, "dense")
+
+ self.traverseSuccessors(cur_node, state)
+
+
+ def handle_pooling(self, cur_node, state):
+ if not self.shouldVisit(cur_node):
+ return
+
+ layer_name = cur_node.layer_name
+ print (layer_name)
+ self.visited_nodes[layer_name] = True
+
+ layer_type = cur_node.layer_type
+ if layer_type == "AveragePooling2D":
+ self.genPreviousLayer(state)
+
+ state.add(cur_node, "pool")
+
+ # NOTE: Will only generate pool layer if it is a standalone Pool (with no convolution)
+ # self.genPreviousLayer(state)
+
+ self.traverseSuccessors(cur_node, state)
+
+
+
+
+
+ def handleLayers(self, output_node, state):
+
+ layer_type = output_node.layer_type
+ print ("layer_type", layer_type)
+
+ if layer_type == "ZeroPadding2D":
+ self.handle_padding(output_node, state)
+
+ if layer_type == "Conv2D":
+ self.handle_conv(output_node, state)
+
+ if layer_type == "Dense":
+ self.handle_dense(output_node, state)
+
+ if layer_type == "Activation":
+ self.handle_activation(output_node, state)
+
+ if layer_type == "MaxPooling2D" or layer_type == "AveragePooling2D":
+ self.handle_pooling(output_node, state)
+
+ if layer_type == "Add":
+ self.handle_add(output_node, state)
+
+ if(self.isForwardLayer(layer_type)):
+ layer_name = output_node.layer_name
+ print ("NOTE: Skippping = ", layer_name)
+ self.visited_nodes[layer_name] = True
+ self.traverseSuccessors(output_node, state)
+
+
+
+ def traverseSuccessors(self, cur_node, state):
+
+ for output_node in cur_node.outputs:
+ self.handleLayers(output_node, state)
+
+
+
+ def findQuantizeRanges(self, model, x_test):
+
+ inp = model.input # input placeholder
+ outputs = [layer.output for layer in model.layers] # all layer outputs
+ functor = K.function([inp, K.learning_phase()], outputs ) # evaluation function
+
+ # Testing
+ layer_outs = functor([x_test, 1.])
+
+ # NOTE: Saving quant ranges for input
+ min_val = np.amin(x_test)
+ max_val = np.amax(x_test)
+ self.quant_ranges["input"] = (min_val, max_val)
+
+ ind = 0
+ for layer_out in layer_outs:
+ layer_name = model.layers[ind].name
+ print ("layer_name = ", layer_name)
+ #min_val = np.amin(layer_out)
+ #max_val = np.amax(layer_out)
+ min_val = np.percentile(layer_out, 0.5)
+ max_val = np.percentile(layer_out, 99.5)
+ print ("min_val = ", min_val, " max_val = ", max_val)
+
+ self.quant_ranges[layer_name] = (min_val, max_val)
+ ind += 1
+
+
+
+ def findLayerInputSizes(self, model, x_test):
+
+ self.batch_size = len(x_test)
+ for layer in model.layers:
+ layer_type = layer.__class__.__name__
+ if layer_type == "InputLayer" or layer_type == "Add":
+ continue
+
+ layer_name = layer.name
+ print ("layer_name = ", layer_name)
+ print ("layer_shape = ", layer.input.shape)
+ self.layer_input_sizes[layer_name] = layer.input.shape
+
+ #inp = model.input # input placeholder
+ #inputs = []
+ # functor = K.function([inp, K.learning_phase()], outputs ) # evaluation function
+ # functor2 = K.function([inp, K.learning_phase()], inputs ) # evaluation function
+ # Testing
+ # layer_outs = functor([x_test, 1.])
+ # layer_inputs = functor2([x_test, 1.])
+
+ #index = 0
+ #for layer_in in layer_inputs:
+ # print ("layer_in = ", layer_in.shape)
+ # layer_name = model.layers[index].name
+ # self.layer_input_sizes[layer_name] = layer_in.shape
+ # index += 1
+
+
+
+ def genExecutionLoop(self):
+
+ exec_loop = ""
+ exec_loop += "int total_runs = 100; \n"
+ exec_loop += "for (int i = 0 ; i < total_runs; i++){ \n\n"
+
+ return exec_loop
+
+
+ def endExecutionLoop(self):
+
+ end_exec_loop = "\n}\n"
+ return end_exec_loop
+
+
+ def genBatchLoop(self, x_test):
+
+ N = x_test.shape[0]
+ C = x_test.shape[1]
+ H = x_test.shape[2]
+ W = x_test.shape[3]
+
+ loop_str = ""
+ loop_str += "\nstartMemTracking(); \n\n"
+
+ loop_str += "int test_input_size = " + str(N) + "; \n"
+ loop_str += "int batch_size = " + str(N) + "; \n"
+ loop_str += "int batch_count = test_input_size / batch_size; \n"
+ loop_str += "float final_accuracy = 0.0; \n\n"
+
+ loop_str += "for(int i = 0; i < batch_count; i++){ \n\n"
+ loop_str += "\n\n" + self.weight_str + "\n\n"
+ loop_str += "int start = i * batch_size; \n"
+ loop_str += "int end = (i + 1) * batch_size; \n"
+
+ loop_str += "\nvoid* input = readInputBatch(input_path.c_str(),0,start,end,"
+ loop_str += str(C) + "," + str(H) + "," + str(W) + "); \n\n"
+
+ return loop_str
+
+
+
+ def endBatchLoop(self):
+
+ end_loop_str = ""
+ end_loop_str += "\nuint8_t* labels = readLabelsBatch(labels_path.c_str(),start,end); \n"
+
+ last_node = self.dfg.last_node
+ output_var = self.output_map[last_node.layer_name]
+ accuracy_call = "\nfloat accuracy = computeAccuracy2(labels, batch_size, " + output_var + "); \n"
+ end_loop_str += accuracy_call
+
+ #end_loop_str += "float accuracy = computeAccuracy2(labels, batch_size, var_60); "
+ end_loop_str += "final_accuracy += accuracy; \n"
+ end_loop_str += "freeBatchMemory(); \n "
+ end_loop_str += "\n}\n\n"
+
+ end_loop_str += "final_accuracy = final_accuracy / batch_count; \n"
+ end_loop_str += "dumpFinalAccuracy(final_accuracy); \n\n"
+
+ return end_loop_str
+
+
+
+
+ def genHeader(self):
+
+ headers = "\n#include <stdio.h> \n"
+ headers += "#include <stdlib.h> \n"
+ headers += "#include <unistd.h> \n"
+ headers += "#include <fcntl.h> \n"
+ headers += "#include <sys/types.h> \n"
+ headers += "#include <sys/stat.h> \n"
+ headers += "#include <string.h> \n"
+
+ headers += "#include \"../../../tensor_runtime/include/tensor_runtime.h\" \n"
+ headers += "#include \"../../include/utils.h\" \n\n"
+
+ main_func = "int main(){ \n\n"
+
+ initialization = "llvm_hpvm_initTensorRt(0); \n\n"
+
+ # Merging into one header string
+ header_str = headers
+ header_str += main_func
+ header_str += initialization
+
+ return header_str
+
+
+
+
+ def genFooter(self, test_data):
+
+ footer_str = ""
+ if test_data is not None and self.dfg.last_node is not None:
+ last_node = self.dfg.last_node
+ output_var = self.output_map[last_node.layer_name]
+ #accuracy_call = "\ncomputeAccuracy2(labels," + str(len(test_data)) + "," + output_var + "); \n"
+ #footer_str += accuracy_call
+
+ accuracy_call = "\ndumpExecutionAccuracies(); \n"
+ footer_str += accuracy_call
+
+ destructors = "\nllvm_hpvm_cleanupTensorRt(); \n"
+ footer_str += destructors
+
+ end_main = "\nreturn 0; \n\n}\n"
+ footer_str += end_main
+
+ return footer_str
+
+
+
+ def dumpLayerStr(self, dir_prefix):
+
+ f = open(dir_prefix + "/layer_composition.txt", "w+")
+ f.write(self.layer_str)
+ f.close()
+
+ f = open(dir_prefix + "/layers.txt", "w+")
+ f.write(self.layer_size_str)
+ f.close()
+
+
+
+ def dumpProgramString(self, final_str, dir_prefix):
+
+ f = open(dir_prefix + "/promise_src.cc", "w+")
+ f.write(final_str)
+ f.close()
+
+
+ def generateSourceProgram(self, weights_dir, x_test):
+
+ print(self.program_str)
+
+ final_str = ""
+ header_str = self.genHeader()
+ final_str += header_str
+
+ exec_loop = self.genExecutionLoop()
+ final_str += exec_loop
+
+ loop_str = self.genBatchLoop(x_test)
+ final_str += loop_str
+
+ #final_str += "\n\n" + self.weight_str + "\n\n"
+ final_str += self.program_str
+
+ end_loop_str = self.endBatchLoop()
+ final_str += end_loop_str
+
+ end_exec_loop = self.endExecutionLoop()
+ final_str += end_exec_loop
+
+ footer_str = self.genFooter(x_test)
+ final_str += footer_str
+ print (final_str)
+
+ self.dumpProgramString(final_str, weights_dir)
+
+
+
+ def translate(self, model, weights_dir, x_test):
+
+ print ("\n\n\n **** PromiseRT Translator ****** \n\n\n")
+ root_node = self.dfg.root_node
+ state = State()
+
+ self.findLayerInputSizes(model, x_test)
+
+ self.findQuantizeRanges(model, x_test)
+
+ #self.traverseSuccessors(root_node, state)
+ self.handleLayers(root_node, state)
+
+ print ("\n *** Generated PROMISE Layers **** \n ")
+
+ self.generateSourceProgram(weights_dir, x_test)
+
+ self.dumpLayerStr(weights_dir)
+
+ #print (self.layer_size_str)
diff --git a/llvm/projects/keras/frontend/setup.py b/llvm/projects/keras/frontend/setup.py
new file mode 100644
index 0000000000..9da7193379
--- /dev/null
+++ b/llvm/projects/keras/frontend/setup.py
@@ -0,0 +1,12 @@
+
+from setuptools import setup
+
+setup(
+ name='frontend',
+ version='1.0',
+ description='ApproxHPVM frontend modules',
+ author='Hashim',
+ author_email='hsharif3@illinois.edu',
+ packages=['frontend'],
+ install_requires=[],
+)
diff --git a/llvm/projects/keras/frontend/weight_utils.py b/llvm/projects/keras/frontend/weight_utils.py
new file mode 100644
index 0000000000..a04827dd7b
--- /dev/null
+++ b/llvm/projects/keras/frontend/weight_utils.py
@@ -0,0 +1,125 @@
+
+import numpy as np
+import struct
+import random
+
+
+def dumpLabels(file_name, Y_test):
+
+ f = open(file_name, "wb")
+
+ labels_map = {}
+ for label in Y_test:
+ label_val = 0
+ if len(Y_test.shape) > 1:
+ label_val = np.int8(label[0])
+ else:
+ label_val = np.int8(label)
+
+ if label_val not in labels_map:
+ labels_map[label_val] = 0
+ labels_map[label_val] += 1
+
+ f.write(label_val)
+
+ f.close()
+
+
+
+def dumpData(file_name, X_test):
+
+ N = X_test.shape[0]
+ C = X_test.shape[1]
+ H = X_test.shape[2]
+ W = X_test.shape[3]
+
+ print ("*DumpData")
+ print("-min_val = ", np.amin(X_test))
+ print("-max_val = ", np.amax(X_test))
+
+ f = open(file_name, "wb")
+ for i in range(N):
+ for j in range(C):
+ for k in range(H):
+ for l in range(W):
+ val = struct.unpack("f", struct.pack("f", X_test[i][j][k][l]))
+ f.write(np.float32(val[0]))
+
+ f.close()
+
+
+def dumpConvWeights(file_name, weights, N, C, H, W):
+
+ print (weights.shape)
+ print ("*DumpConvWeights")
+ print("-min_val = ", np.amin(weights))
+ print("-max_val = ", np.amax(weights))
+
+
+ f = open(file_name, "wb")
+ for i in range(N):
+ for j in range(C):
+ for k in range(H):
+ for l in range(W):
+ f.write(weights[k][l][j][i])
+
+ f.close()
+
+
+
+def dumpFcWeights(file_name, weights, H, W):
+
+ print (weights.shape)
+ print ("*DumpFcWeights")
+ print("-min_val = ", np.amin(weights))
+ print("-max_val = ", np.amax(weights))
+
+
+ f = open(file_name, "wb")
+ for i in range(H):
+ for j in range(W):
+ f.write(weights[i][j])
+
+ f.close()
+
+
+
+def dumpFcBias(file_name, bias, W):
+
+ print (bias.shape)
+ print ("*DumpFcBias")
+ print("-min_val = ", np.amin(bias))
+ print("-max_val = ", np.amax(bias))
+
+
+ f = open(file_name, "wb")
+ for i in range(W):
+ f.write(bias[i])
+
+ f.close()
+
+
+
+def dumpCalibrationData(file_name, X_train, labels_fname, train_labels):
+
+ combined_list = []
+ for i in range(len(X_train)):
+ tup = (X_train[i], train_labels[i])
+ combined_list.append(tup)
+
+ np.random.shuffle(combined_list)
+ #X_calibration = X_train[0:5000]
+
+ data_list = []
+ labels_list = []
+ for i in range(5000):
+ tup = combined_list[i]
+ data_list.append(tup[0])
+ labels_list.append(tup[1])
+
+ data_list = np.array(data_list)
+ labels_list = np.array(labels_list)
+
+ dumpData(file_name, data_list)
+ dumpLabels(labels_fname, labels_list)
+
diff --git a/llvm/projects/keras/keras_environment.yml b/llvm/projects/keras/keras_environment.yml
new file mode 100644
index 0000000000..1c9942f13d
--- /dev/null
+++ b/llvm/projects/keras/keras_environment.yml
@@ -0,0 +1,331 @@
+name: prakalp_keras
+channels:
+ - pytorch
+ - conda-forge
+ - defaults
+dependencies:
+ - absl-py=0.6.1=py36_0
+ - anaconda-project=0.8.2=py36_0
+ - asn1crypto=0.24.0=py36_0
+ - automat=0.7.0=py36_0
+ - babel=2.6.0=py36_0
+ - backports=1.0=py36_1
+ - backports.os=0.1.1=py36_0
+ - beautifulsoup4=4.6.3=py36_0
+ - bkcharts=0.2=py36_0
+ - blaze=0.11.3=py36_0
+ - conda=4.5.11=py36_0
+ - conda-env=2.6.0=1
+ - contextlib2=0.5.5=py36_0
+ - cycler=0.10.0=py36_0
+ - dill=0.2.8.2=py36_0
+ - docutils=0.14=py36_0
+ - entrypoints=0.2.3=py36_2
+ - et_xmlfile=1.0.1=py36_0
+ - idna=2.7=py36_0
+ - imageio=2.4.1=py36_0
+ - importlib_metadata=0.6=py36_0
+ - ipython_genutils=0.2.0=py36_0
+ - isort=4.3.4=py36_0
+ - jdcal=1.4=py36_0
+ - jedi=0.13.1=py36_0
+ - jinja2=2.10=py36_0
+ - jmespath=0.9.3=py36_0
+ - jsonschema=2.6.0=py36_0
+ - jupyter_client=5.2.3=py36_0
+ - jupyterlab=0.35.3=py36_0
+ - keyring=16.0.0=py36_0
+ - libgcc=7.2.0=h69d50b8_2
+ - libgfortran=3.0.0=1
+ - locket=0.2.0=py36_1
+ - more-itertools=4.3.0=py36_0
+ - nbconvert=5.3.1=py36_0
+ - nbformat=4.4.0=py36_0
+ - nose=1.3.7=py36_2
+ - notebook=5.7.0=py36_0
+ - numpydoc=0.8.0=py36_0
+ - odo=0.5.1=py36_0
+ - pathlib2=2.3.2=py36_0
+ - pexpect=4.6.0=py36_0
+ - pickleshare=0.7.5=py36_0
+ - ply=3.11=py36_0
+ - ptyprocess=0.6.0=py36_0
+ - pycodestyle=2.4.0=py36_0
+ - pygments=2.2.0=py36_0
+ - pylint=2.1.1=py36_0
+ - pyopenssl=18.0.0=py36_0
+ - qtconsole=4.4.2=py36_0
+ - requests=2.19.1=py36_0
+ - s3transfer=0.1.13=py36_0
+ - secretstorage=3.1.0=py36_0
+ - setuptools=40.5.0=py36_0
+ - singledispatch=3.4.0.3=py36_0
+ - six=1.11.0=py36_1
+ - snowballstemmer=1.2.1=py36_0
+ - sortedcollections=1.0.1=py36_0
+ - sphinx=1.8.1=py36_0
+ - spyder=3.3.1=py36_1
+ - sympy=1.3=py36_0
+ - tblib=1.3.2=py36_0
+ - termcolor=1.1.0=py36_1
+ - terminado=0.8.1=py36_1
+ - testpath=0.4.2=py36_0
+ - torchvision=0.2.1=py36_0
+ - traitlets=4.3.2=py36_0
+ - typing=3.6.4=py36_0
+ - unicodecsv=0.14.1=py36_0
+ - urllib3=1.23=py36_0
+ - wcwidth=0.1.7=py36_0
+ - wheel=0.32.2=py36_0
+ - widgetsnbextension=3.4.2=py36_0
+ - xlwt=1.3.0=py36_0
+ - _license=1.1=py36_1
+ - _tflow_select=2.1.0=gpu
+ - alabaster=0.7.12=py36_0
+ - anaconda-client=1.7.2=py36_0
+ - anaconda=custom=py36hbbc8b67_0
+ - anaconda-navigator=1.9.2=py36_0
+ - appdirs=1.4.3=py36h28b3542_0
+ - astor=0.7.1=py36_0
+ - astroid=2.0.4=py36_0
+ - astropy=3.0.5=py36h7b6447c_0
+ - atomicwrites=1.2.1=py36_0
+ - attrs=18.2.0=py36h28b3542_0
+ - backcall=0.1.0=py36_0
+ - backports.shutil_get_terminal_size=1.0.0=py36_2
+ - bitarray=0.8.3=py36h14c3975_0
+ - blas=1.0=mkl
+ - bleach=3.0.2=py36_0
+ - blosc=1.14.4=hdbcaa40_0
+ - bokeh=1.0.1=py36_0
+ - boto=2.49.0=py36_0
+ - boto3=1.9.35=py36_0
+ - botocore=1.12.35=py36_0
+ - bottleneck=1.2.1=py36h035aef0_1
+ - bz2file=0.98=py36_1
+ - bzip2=1.0.6=h14c3975_5
+ - ca-certificates=2018.03.07=0
+ - cairo=1.14.12=h8948797_3
+ - certifi=2018.10.15=py36_0
+ - cffi=1.11.5=py36he75722e_1
+ - chardet=3.0.4=py36_1
+ - chest=0.2.3=py36_1
+ - click=7.0=py36_0
+ - cloudpickle=0.6.1=py36_0
+ - clyent=1.2.2=py36_1
+ - colorama=0.4.0=py36_0
+ - configobj=5.0.6=py36_1
+ - constantly=15.1.0=py36h28b3542_0
+ - cryptography=2.3.1=py36hc365091_0
+ - cudatoolkit=9.0=h13b8566_0
+ - cudnn=7.1.2=cuda9.0_0
+ - cupti=9.0.176=0
+ - curl=7.61.0=h84994c4_0
+ - cython=0.29=py36he6710b0_0
+ - cytoolz=0.9.0.1=py36h14c3975_1
+ - dask=0.20.0=py36_0
+ - dask-core=0.20.0=py36_0
+ - datashape=0.5.4=py36_1
+ - dbus=1.13.2=h714fa37_1
+ - decorator=4.3.0=py36_0
+ - defusedxml=0.5.0=py36_1
+ - distributed=1.24.0=py36_0
+ - expat=2.2.6=he6710b0_0
+ - fastcache=1.0.2=py36h14c3975_2
+ - filelock=3.0.10=py36_0
+ - flask=1.0.2=py36_1
+ - flask-cors=3.0.6=py36_0
+ - fontconfig=2.13.0=h9420a91_0
+ - freetype=2.9.1=h8a8886c_1
+ - fribidi=1.0.5=h7b6447c_0
+ - gast=0.2.0=py36_0
+ - gensim=3.4.0=py36h14c3975_0
+ - get_terminal_size=1.0.0=haa9412d_0
+ - gevent=1.3.7=py36h7b6447c_1
+ - glib=2.56.2=hd408876_0
+ - glob2=0.6=py36_1
+ - gmp=6.1.2=h6c8ec71_1
+ - gmpy2=2.0.8=py36h10f8cd9_2
+ - graphite2=1.3.12=h23475e2_2
+ - greenlet=0.4.15=py36h7b6447c_0
+ - grpcio=1.12.1=py36hdbcaa40_0
+ - gst-plugins-base=1.14.0=hbbd80ab_1
+ - gstreamer=1.14.0=hb453b48_1
+ - h5py=2.8.0=py36h989c5e5_3
+ - harfbuzz=1.8.8=hffaf4a1_0
+ - hdf5=1.10.2=hba1933b_1
+ - heapdict=1.0.0=py36_2
+ - html5lib=1.0.1=py36_0
+ - hyperlink=18.0.0=py36_0
+ - icu=58.2=h9c2bf20_1
+ - imagesize=1.1.0=py36_0
+ - incremental=17.5.0=py36_0
+ - intel-openmp=2019.0=118
+ - ipykernel=5.1.0=py36h39e3cac_0
+ - ipython=7.1.1=py36h39e3cac_0
+ - ipywidgets=7.4.2=py36_0
+ - itsdangerous=1.1.0=py36_0
+ - jbig=2.1=hdba287a_0
+ - jeepney=0.4=py36_0
+ - jpeg=9b=h024ee3a_2
+ - jupyter=1.0.0=py36_7
+ - jupyter_console=6.0.0=py36_0
+ - jupyter_core=4.4.0=py36_0
+ - jupyterlab_launcher=0.13.1=py36_0
+ - jupyterlab_server=0.2.0=py36_0
+ - keras=2.1.6=py36_0
+ - keras-applications=1.0.6=py36_0
+ - keras-preprocessing=1.0.5=py36_0
+ - kiwisolver=1.0.1=py36hf484d3e_0
+ - lazy-object-proxy=1.3.1=py36h14c3975_2
+ - libcurl=7.61.0=h1ad7b7a_0
+ - libedit=3.1.20170329=h6b74fdf_2
+ - libffi=3.2.1=hd88cf55_4
+ - libgcc-ng=8.2.0=hdf63c60_1
+ - libgfortran-ng=7.3.0=hdf63c60_0
+ - libiconv=1.15=h63c8f33_5
+ - libpng=1.6.35=hbc83047_0
+ - libprotobuf=3.6.1=hd408876_0
+ - libsodium=1.0.16=h1bed415_0
+ - libssh2=1.8.0=h9cfc8f7_4
+ - libstdcxx-ng=8.2.0=hdf63c60_1
+ - libtiff=4.0.9=he85c1e1_2
+ - libtool=2.4.6=h7b6447c_5
+ - libuuid=1.0.3=h1bed415_2
+ - libxcb=1.13=h1bed415_1
+ - libxml2=2.9.8=h26e45fe_1
+ - libxslt=1.1.32=h1312cb7_0
+ - llvmlite=0.25.0=py36hd408876_0
+ - lxml=4.2.5=py36hefd8a0e_0
+ - lzo=2.10=h49e0be7_2
+ - markdown=3.0.1=py36_0
+ - markupsafe=1.0=py36h14c3975_1
+ - matplotlib=3.0.1=py36h5429711_0
+ - mccabe=0.6.1=py36_1
+ - mistune=0.8.4=py36h7b6447c_0
+ - mkl=2018.0.3=1
+ - mkl-service=1.1.2=py36h90e4bf4_5
+ - mkl_fft=1.0.6=py36h7dd41cf_0
+ - mkl_random=1.0.1=py36h4414c95_1
+ - mpc=1.1.0=h10f8cd9_1
+ - mpfr=4.0.1=hdf1c602_3
+ - mpmath=1.0.0=py36_2
+ - msgpack-python=0.5.6=py36h6bb024c_1
+ - multipledispatch=0.6.0=py36_0
+ - navigator-updater=0.2.1=py36_0
+ - nccl=1.3.5=cuda9.0_0
+ - ncurses=6.1=hf484d3e_0
+ - networkx=2.2=py36_1
+ - ninja=1.8.2=py36h6bb024c_1
+ - nltk=3.3.0=py36_0
+ - numba=0.40.0=py36h962f231_0
+ - numexpr=2.6.8=py36hd89afb7_0
+ - numpy=1.15.3=py36h1d66e8a_0
+ - numpy-base=1.15.3=py36h81de0dd_0
+ - olefile=0.46=py36_0
+ - openpyxl=2.5.9=py36_0
+ - openssl=1.0.2p=h14c3975_0
+ - packaging=18.0=py36_0
+ - pandas=0.23.4=py36h04863e7_0
+ - pandoc=2.2.3.2=0
+ - pandocfilters=1.4.2=py36_1
+ - pango=1.42.4=h049681c_0
+ - parso=0.3.1=py36_0
+ - partd=0.3.9=py36_0
+ - patchelf=0.9=he6710b0_3
+ - path.py=11.5.0=py36_0
+ - patsy=0.5.1=py36_0
+ - pcre=8.42=h439df22_0
+ - pep8=1.7.1=py36_0
+ - pillow=5.3.0=py36h34e0f95_0
+ - pip=18.1=py36_0
+ - pixman=0.34.0=hceecf20_3
+ - pkginfo=1.4.2=py36_1
+ - pluggy=0.8.0=py36_0
+ - prometheus_client=0.4.2=py36_0
+ - prompt_toolkit=2.0.7=py36_0
+ - protobuf=3.6.1=py36he6710b0_0
+ - psutil=5.4.8=py36h7b6447c_0
+ - py=1.7.0=py36_0
+ - pyasn1=0.4.4=py36h28b3542_0
+ - pyasn1-modules=0.2.2=py36_0
+ - pycosat=0.6.3=py36h14c3975_0
+ - pycparser=2.19=py36_0
+ - pycrypto=2.6.1=py36h14c3975_9
+ - pycurl=7.43.0.2=py36hb7f436b_0
+ - pyflakes=2.0.0=py36_0
+ - pyhamcrest=1.9.0=py36_2
+ - pyodbc=4.0.24=py36he6710b0_0
+ - pyparsing=2.2.2=py36_0
+ - pyqt=5.9.2=py36h05f1152_2
+ - pysocks=1.6.8=py36_0
+ - pytables=3.4.4=py36ha205bf6_0
+ - pytest=3.9.3=py36_0
+ - pytest-arraydiff=0.2=py36h39e3cac_0
+ - pytest-astropy=0.4.0=py36_0
+ - pytest-doctestplus=0.1.3=py36_0
+ - pytest-openfiles=0.3.0=py36_0
+ - pytest-remotedata=0.3.1=py36_0
+ - python=3.6.6=h6e4f718_2
+ - python-dateutil=2.7.5=py36_0
+ - pytorch=0.4.1=py36ha74772b_0
+ - pytz=2018.7=py36_0
+ - pywavelets=1.0.1=py36hdd07704_0
+ - pyyaml=3.13=py36h14c3975_0
+ - pyzmq=17.1.2=py36h14c3975_0
+ - qt=5.9.6=h8703b6f_2
+ - qtawesome=0.5.2=py36_0
+ - qtpy=1.5.2=py36_0
+ - readline=7.0=h7b6447c_5
+ - redis=5.0.0=h7b6447c_0
+ - redis-py=2.10.6=py36_0
+ - rope=0.11.0=py36_0
+ - ruamel_yaml=0.15.46=py36h14c3975_0
+ - scikit-image=0.14.0=py36hf484d3e_1
+ - scikit-learn=0.20.0=py36h4989274_1
+ - scipy=1.1.0=py36hfa4b5c9_1
+ - seaborn=0.9.0=py36_0
+ - send2trash=1.5.0=py36_0
+ - service_identity=17.0.0=py36h28b3542_0
+ - simplegeneric=0.8.1=py36_2
+ - sip=4.19.8=py36hf484d3e_0
+ - smart_open=1.7.1=py36_0
+ - snappy=1.1.7=hbae5bb6_3
+ - sockjs-tornado=1.0.6=py36_0
+ - sortedcontainers=2.0.5=py36_0
+ - sphinxcontrib=1.0=py36_1
+ - sphinxcontrib-websupport=1.1.0=py36_1
+ - spyder-kernels=0.2.6=py36_0
+ - sqlalchemy=1.2.12=py36h7b6447c_0
+ - sqlite=3.25.2=h7b6447c_0
+ - statsmodels=0.9.0=py36h035aef0_0
+ - tensorboard=1.11.0=py36hf484d3e_0
+ - tensorflow=1.11.0=gpu_py36h4459f94_0
+ - tensorflow-base=1.11.0=gpu_py36h8e0ae2d_0
+ - tensorflow-gpu=1.11.0=h0d30ee6_0
+ - tk=8.6.8=hbc83047_0
+ - toolz=0.9.0=py36_0
+ - tornado=5.1.1=py36h7b6447c_0
+ - tqdm=4.28.1=py36h28b3542_0
+ - twisted=18.9.0=py36h7b6447c_0
+ - typed-ast=1.1.0=py36h14c3975_0
+ - unixodbc=2.3.7=h14c3975_0
+ - webencodings=0.5.1=py36_1
+ - werkzeug=0.14.1=py36_0
+ - wrapt=1.10.11=py36h14c3975_2
+ - xlrd=1.1.0=py36_1
+ - xlsxwriter=1.1.2=py36_0
+ - xz=5.2.4=h14c3975_4
+ - yaml=0.1.7=had09818_2
+ - zeromq=4.2.5=hf484d3e_1
+ - zict=0.1.3=py36_0
+ - zlib=1.2.11=ha838bed_2
+ - zope=1.0=py36_1
+ - zope.interface=4.6.0=py36h7b6447c_0
+ - cuda91=1.0=h4c16780_0
+ - pip:
+ - msgpack==0.5.6
+ - tables==3.4.4
+ - torch==0.4.1
+prefix: /home/hsharif3/anaconda2/envs/prakalp_keras
+
diff --git a/llvm/projects/keras/setup.py b/llvm/projects/keras/setup.py
new file mode 100644
index 0000000000..9da7193379
--- /dev/null
+++ b/llvm/projects/keras/setup.py
@@ -0,0 +1,12 @@
+
+from setuptools import setup
+
+setup(
+ name='frontend',
+ version='1.0',
+ description='ApproxHPVM frontend modules',
+ author='Hashim',
+ author_email='hsharif3@illinois.edu',
+ packages=['frontend'],
+ install_requires=[],
+)
diff --git a/llvm/projects/keras/src/alexnet.py b/llvm/projects/keras/src/alexnet.py
new file mode 100644
index 0000000000..b1f29bfc05
--- /dev/null
+++ b/llvm/projects/keras/src/alexnet.py
@@ -0,0 +1,173 @@
+
+import numpy as np
+
+from keras.datasets import cifar10
+from keras.models import Sequential
+from keras.layers.core import Dense, Dropout, Flatten, Activation
+from keras.layers.convolutional import Conv2D
+from keras.optimizers import Adam
+from keras.layers.pooling import MaxPooling2D
+from keras.utils.np_utils import to_categorical
+from keras.preprocessing.image import ImageDataGenerator
+from keras import backend as K
+from keras import regularizers
+from keras.callbacks import LearningRateScheduler
+import sys
+import struct
+import keras
+import numpy as np
+import os
+from frontend.approxhpvm_translator import translate_to_approxhpvm
+from frontend.weight_utils import dumpCalibrationData
+
+
+
+def lr_schedule(epoch):
+ lrate = 0.001
+ if epoch > 20:
+ lrate = 0.0005
+ if epoch > 40:
+ lrate = 0.0003
+ if epoch > 60:
+ lrate = 0.0001
+ if epoch > 80:
+ lrate = 0.00005
+
+ return lrate
+
+
+
+def buildModel2():
+
+ activation_type = "tanh"
+ weight_decay = 1e-4
+
+ model = Sequential()
+ model.add(Conv2D(64, kernel_size=(11, 11), activation=activation_type,
+ input_shape=(3, 32, 32), padding = 'same',
+ kernel_regularizer=regularizers.l2(weight_decay) ))
+ model.add(MaxPooling2D(pool_size=(2, 2), strides=(2,2) ))
+ model.add(Dropout(0.2))
+ model.add(Conv2D(192, kernel_size=(5, 5), activation=activation_type, padding = 'same',
+ kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(MaxPooling2D(pool_size=(2, 2), strides=(2,2) ))
+ model.add(Dropout(0.3))
+
+ model.add(Conv2D(384, kernel_size=(3, 3), activation=activation_type, padding = 'same',
+ kernel_regularizer=regularizers.l2(weight_decay) ))
+ model.add(Conv2D(256, kernel_size=(3, 3), activation=activation_type, padding = 'same',
+ kernel_regularizer=regularizers.l2(weight_decay) ))
+ model.add(Conv2D(256, kernel_size=(3, 3), activation=activation_type, padding = 'same',
+ kernel_regularizer=regularizers.l2(weight_decay) ))
+ model.add(MaxPooling2D(pool_size=(2, 2), strides=(2,2) ))
+ model.add(Dropout(0.4))
+
+ model.add(Flatten())
+ #model.add(Flatten())
+ #model.add(Dense(256))
+ model.add(Dense(10))
+ model.add(Activation('softmax'))
+
+ return model
+
+
+
+def buildModel():
+
+ model = Sequential()
+ model.add(Conv2D(128, kernel_size=(3, 3), activation='tanh', input_shape=(3, 32, 32), padding = 'same'))
+ model.add(Conv2D(256, kernel_size=(3, 3), activation='tanh', padding = 'same'))
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+ #model.add(Dropout(0.25))
+
+ model.add(Conv2D(256, kernel_size=(3, 3), activation='tanh', padding = 'same'))
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+
+ model.add(Conv2D(256, kernel_size=(3, 3), activation='tanh', padding = 'same'))
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+ model.add(Conv2D(256, kernel_size=(3, 3), activation='tanh', padding = 'same'))
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+ #model.add(Dropout(0.25))
+
+ model.add(Flatten())
+ #model.add(Flatten())
+ model.add(Dense(4096, activation='tanh'))
+ #model.add(Dropout(0.5))
+ model.add(Dense(2048, activation='tanh'))
+ model.add(Dense(10, activation='tanh'))
+ model.add(Activation('softmax'))
+
+ return model
+
+
+
+
+def trainModel(model):
+
+ (X_train, Y_train), (X_test, Y_test) = cifar10.load_data()
+ test_labels = Y_test
+ train_labels = Y_train
+
+ #X_train = X_train.astype('float32')
+ #X_test = X_test.astype('float32')
+ X_train = X_train / 255.0
+ X_test = X_test / 255.0
+
+ mean = np.mean(X_train,axis=(0,1,2,3))
+ std = np.std(X_train,axis=(0,1,2,3))
+ X_train = (X_train-mean)/(std+1e-7)
+ X_test = (X_test-mean)/(std+1e-7)
+
+ dir_prefix = "/home/hsharif3/Gitlab/hpvm/llvm/projects/hpvm-tensor-rt/model_params/alexnet_cifar10/"
+
+ #opt_rms = keras.optimizers.rmsprop(lr=0.001,decay=1e-6)
+ # Compile the model
+ model.compile(loss='categorical_crossentropy',
+ optimizer=Adam(lr=0.0001, decay=1e-6),
+ #optimizer = opt_rms,
+ metrics=['accuracy'])
+
+ #print to_categorical(Y_train, 10)
+ print (to_categorical(Y_train))
+
+
+ datagen = ImageDataGenerator(
+ rotation_range=15,
+ width_shift_range=0.1,
+ height_shift_range=0.1,
+ horizontal_flip=True,
+ )
+ datagen.fit(X_train)
+
+
+ model.fit(X_train, to_categorical(Y_train, 10),
+ batch_size=128,
+ shuffle=True,
+ epochs = 1,
+ #epochs=100,
+ validation_data=(X_test, to_categorical(Y_test, 10)), callbacks=[LearningRateScheduler(lr_schedule)])
+
+ # Evaluate the model
+ scores = model.evaluate(X_test, to_categorical(Y_test, 10))
+
+ print('Loss: %.3f' % scores[0])
+ print('Accuracy: %.3f' % scores[1])
+
+ print ("*** TRAINED MODEL ****\n")
+
+
+ #dumpCalibrationData("calibration_data/alexnet_calib.bin", X_train,
+ # "calibration_data/alexnet_train_labels.bin", train_labels)
+
+ translate_to_approxhpvm(model, "data/alexnet_cifar10/", X_test, test_labels, 10)
+
+
+
+if __name__ == "__main__":
+
+ os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+ # Changing to NCHW format
+ K.set_image_data_format('channels_first')
+
+ model = buildModel2()
+ trainModel(model)
diff --git a/llvm/projects/keras/src/alexnet2.py b/llvm/projects/keras/src/alexnet2.py
new file mode 100644
index 0000000000..812b212165
--- /dev/null
+++ b/llvm/projects/keras/src/alexnet2.py
@@ -0,0 +1,243 @@
+
+import keras
+from keras.models import Sequential
+from keras.utils import np_utils
+from keras.preprocessing.image import ImageDataGenerator
+from keras.layers import Dense, Activation, Flatten, Dropout, BatchNormalization
+from keras.layers import Conv2D, MaxPooling2D
+from keras.datasets import cifar10
+from keras import regularizers
+from keras.callbacks import LearningRateScheduler
+import numpy as np
+import os
+import struct
+from keras import backend as K
+from approxhpvm_translator import translate_to_approxhpvm
+
+
+
+def dumpWeights(file_name, weights, N, H, W, C):
+ # NOTE: Writing the NHWC weights array as NCHW
+ f = open(file_name, "wb")
+ for i in range(N):
+ for j in range(C):
+ for k in range(H):
+ for l in range(W):
+ f.write(weights[i][k][l][j])
+
+ f.close()
+
+
+def dumpConvWeights(file_name, weights, N, C, H, W):
+
+ print (weights.shape)
+
+ f = open(file_name, "wb")
+ for i in range(N):
+ for j in range(C):
+ for k in range(H):
+ for l in range(W):
+ f.write(weights[k][l][j][i])
+ f.close()
+
+
+
+def dumpFcWeights(file_name, weights, H, W):
+
+ print (weights.shape)
+
+ f = open(file_name, "wb")
+ for i in range(H):
+ for j in range(W):
+ f.write(weights[i][j])
+ f.close()
+
+
+def dumpFcBias(file_name, bias, W):
+
+ print (bias.shape)
+
+ f = open(file_name, "wb")
+ for i in range(W):
+ f.write(bias[i])
+ f.close()
+
+
+def dumpLabels(file_name, Y_test):
+
+ f = open(file_name, "wb")
+
+ labels_map = {}
+ for label in Y_test:
+ label_val = np.int8(label[0])
+ if label_val not in labels_map:
+ labels_map[label_val] = 0
+ labels_map[label_val] += 1
+
+ f.write(label_val)
+
+ f.close()
+
+
+
+def dumpData(X_test, file_name, N, C, H, W):
+
+ print (X_test.shape)
+
+ f = open(file_name, "wb")
+ for i in range(N):
+ for j in range(C):
+ for k in range(H):
+ for l in range(W):
+ val = struct.unpack("f", struct.pack("f", X_test[i][j][k][l]))
+ f.write(np.float32(val[0]))
+
+ f.close()
+
+
+
+
+
+def lr_schedule(epoch):
+ lrate = 0.001
+ if epoch > 75:
+ lrate = 0.0005
+ if epoch > 100:
+ lrate = 0.0003
+ return lrate
+
+
+def lr_schedule2(epoch):
+ lrate = 0.0005
+ if epoch > 100:
+ lrate = 0.0003
+ if epoch > 200:
+ lrate = 0.0002
+ if epoch > 250:
+ lrate = 0.0001
+ if epoch > 300:
+ lrate = 0.00003
+
+ return lrate
+
+
+K.set_image_data_format('channels_first')
+
+(x_train, y_train), (x_test, y_test) = cifar10.load_data()
+test_labels = y_test
+x_train = x_train.astype('float32')
+x_test = x_test.astype('float32')
+
+#z-score
+mean = np.mean(x_train,axis=(0,1,2,3))
+std = np.std(x_train,axis=(0,1,2,3))
+x_train = (x_train-mean)/(std+1e-7)
+x_test = (x_test-mean)/(std+1e-7)
+
+
+# Dumping test data and test labels
+dir_prefix = "/home/hsharif3/Gitlab/hpvm/llvm/projects/hpvm-tensor-rt/model_params/alexnet2_cifar10/"
+
+dumpLabels(dir_prefix + "test_labels.bin", y_test)
+dumpData(x_test, dir_prefix + "norm_cifar_input.bin", 10000, 3, 32, 32)
+
+
+
+num_classes = 10
+y_train = np_utils.to_categorical(y_train,num_classes)
+y_test = np_utils.to_categorical(y_test,num_classes)
+
+weight_decay = 1e-4
+activation_type = 'tanh'
+
+
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+
+
+model = Sequential()
+model.add(Conv2D(32, (3,3), padding='same', kernel_regularizer=regularizers.l2(weight_decay), input_shape=x_train.shape[1:]))
+model.add(Activation(activation_type))
+#model.add(BatchNormalization())
+model.add(Conv2D(32, (3,3), padding='same', kernel_regularizer=regularizers.l2(weight_decay)))
+model.add(Activation(activation_type))
+#model.add(BatchNormalization())
+model.add(MaxPooling2D(pool_size=(2,2)))
+model.add(Dropout(0.2))
+
+model.add(Conv2D(64, (3,3), padding='same', kernel_regularizer=regularizers.l2(weight_decay)))
+model.add(Activation(activation_type))
+#model.add(BatchNormalization())
+model.add(Conv2D(64, (3,3), padding='same', kernel_regularizer=regularizers.l2(weight_decay)))
+model.add(Activation(activation_type))
+#model.add(BatchNormalization())
+model.add(MaxPooling2D(pool_size=(2,2)))
+model.add(Dropout(0.3))
+
+model.add(Conv2D(128, (3,3), padding='same', kernel_regularizer=regularizers.l2(weight_decay)))
+model.add(Activation(activation_type))
+#model.add(BatchNormalization())
+model.add(Conv2D(128, (3,3), padding='same', kernel_regularizer=regularizers.l2(weight_decay)))
+model.add(Activation(activation_type))
+#model.add(BatchNormalization())
+model.add(MaxPooling2D(pool_size=(2,2)))
+model.add(Dropout(0.4))
+
+model.add(Flatten())
+model.add(Dense(num_classes))
+model.add(Activation('softmax'))
+model.summary()
+
+#data augmentation
+datagen = ImageDataGenerator(
+ rotation_range=15,
+ width_shift_range=0.1,
+ height_shift_range=0.1,
+ horizontal_flip=True,
+ )
+
+datagen.fit(x_train)
+
+
+#training
+batch_size = 64
+
+opt_rms = keras.optimizers.rmsprop(lr=0.001,decay=1e-6)
+model.compile(loss='categorical_crossentropy', optimizer=opt_rms, metrics=['accuracy'])
+model.fit_generator(datagen.flow(x_train, y_train, batch_size=batch_size),\
+ steps_per_epoch=x_train.shape[0] // batch_size, #epochs=350,\
+ epochs=1,
+ verbose=1,validation_data=(x_test,y_test),callbacks=[LearningRateScheduler(lr_schedule2)])
+#save to disk
+model_json = model.to_json()
+with open('model.json', 'w') as json_file:
+ json_file.write(model_json)
+ model.save_weights('model.h5')
+
+#testing
+scores = model.evaluate(x_test, y_test, batch_size=128, verbose=1)
+print('\nTest result: %.3f loss: %.3f' % (scores[1]*100,scores[0]))
+
+
+translate_to_approxhpvm(model, "alexnet2_cifar10_test/", x_test, test_labels, "alexnet2_cifar10/", y_test)
+sys.exit(0)
+
+
+
+params = model.get_weights()
+dumpConvWeights(dir_prefix + "conv1.bin", params[0], 32, 3, 3, 3)
+dumpFcBias(dir_prefix + "conv1_bias.bin", params[1], 32)
+dumpConvWeights(dir_prefix + "conv2.bin", params[2], 32, 32, 3, 3)
+dumpFcBias(dir_prefix + "conv2_bias.bin", params[3], 32)
+dumpConvWeights(dir_prefix + "conv3.bin", params[4], 64, 32, 3, 3)
+dumpFcBias(dir_prefix + "conv3_bias.bin", params[5], 64)
+dumpConvWeights(dir_prefix + "conv4.bin", params[6], 64, 64, 3, 3)
+dumpFcBias(dir_prefix + "conv4_bias.bin", params[7], 64)
+dumpConvWeights(dir_prefix + "conv5.bin", params[8], 128, 64, 3, 3)
+dumpFcBias(dir_prefix + "conv5_bias.bin", params[9], 128)
+dumpConvWeights(dir_prefix + "conv6.bin", params[10], 128, 128, 3, 3)
+dumpFcBias(dir_prefix + "conv6_bias.bin", params[11], 128)
+
+dumpFcWeights(dir_prefix + "fc1.bin", params[12], 2048, 10)
+dumpFcBias(dir_prefix + "fc1_bias.bin", params[13], 10)
+
+
diff --git a/llvm/projects/keras/src/lenet.py b/llvm/projects/keras/src/lenet.py
new file mode 100644
index 0000000000..3fa8123bd1
--- /dev/null
+++ b/llvm/projects/keras/src/lenet.py
@@ -0,0 +1,92 @@
+
+import sys
+import keras
+from keras.datasets import mnist
+from keras.models import Sequential
+from keras.layers import Dense, Dropout, Flatten, Activation
+from keras.layers import Conv2D, MaxPooling2D, ZeroPadding2D
+from keras import backend as K
+from frontend.approxhpvm_translator import translate_to_approxhpvm
+
+
+batch_size = 128
+num_classes = 10
+epochs = 5
+
+# input image dimensions
+img_rows, img_cols = 28, 28
+
+
+
+if __name__ == "__main__":
+
+ K.set_image_data_format('channels_first')
+
+ # the data, split between train and test sets
+ (x_train, y_train), (x_test, y_test) = mnist.load_data()
+ test_labels = y_test
+
+
+ if K.image_data_format() == 'channels_first':
+ x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
+ x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
+ input_shape = (1, img_rows, img_cols)
+ else:
+ x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
+ x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
+ input_shape = (img_rows, img_cols, 1)
+
+
+ print(K.image_data_format())
+
+ x_train = x_train.astype('float32')
+ x_test = x_test.astype('float32')
+ x_train /= 255
+ x_test /= 255
+ print('x_train shape:', x_train.shape)
+ print(x_train.shape[0], 'train samples')
+ print(x_test.shape[0], 'test samples')
+
+ # convert class vectors to binary class matrices
+ y_train = keras.utils.to_categorical(y_train, num_classes)
+ y_test = keras.utils.to_categorical(y_test, num_classes)
+
+
+ activation_type = 'relu'
+
+ model = Sequential()
+ model.add(Conv2D(32, kernel_size=(5, 5),
+ activation=activation_type,
+ padding = 'same',
+ input_shape=input_shape))
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+ model.add(Conv2D(64, (5, 5), activation=activation_type, padding = 'same'))
+ model.add(ZeroPadding2D(padding = (1,1)))
+ model.add(Conv2D(64, (3, 3), strides = (2,2), activation=activation_type) )
+ model.add(Flatten())
+ model.add(Dense(1024, activation=activation_type))
+ model.add(Dense(num_classes, activation=activation_type))
+ model.add(Activation('softmax'))
+
+
+
+ model.compile(loss=keras.losses.categorical_crossentropy,
+ optimizer=keras.optimizers.Adadelta(),
+ metrics=['accuracy'])
+
+ model.fit(x_train, y_train,
+ batch_size=batch_size,
+ epochs=1,
+ verbose=1,
+ validation_data=(x_test, y_test))
+
+
+ score = model.evaluate(x_test, y_test, verbose=0)
+ print('Test loss:', score[0])
+ print('Test accuracy:', score[1])
+
+
+ model.summary()
+
+ translate_to_approxhpvm(model, "data/lenet_hpvm_batch/", x_test, test_labels, 10)
+
diff --git a/llvm/projects/keras/src/resnet.py b/llvm/projects/keras/src/resnet.py
new file mode 100644
index 0000000000..6afa1c50fa
--- /dev/null
+++ b/llvm/projects/keras/src/resnet.py
@@ -0,0 +1,571 @@
+"""
+#Trains a ResNet on the CIFAR10 dataset.
+
+ResNet v1:
+[Deep Residual Learning for Image Recognition
+](https://arxiv.org/pdf/1512.03385.pdf)
+
+ResNet v2:
+[Identity Mappings in Deep Residual Networks
+](https://arxiv.org/pdf/1603.05027.pdf)
+
+
+Model|n|200-epoch accuracy|Original paper accuracy |sec/epoch GTX1080Ti
+:------------|--:|-------:|-----------------------:|---:
+ResNet20 v1| 3| 92.16 %| 91.25 %|35
+ResNet32 v1| 5| 92.46 %| 92.49 %|50
+ResNet44 v1| 7| 92.50 %| 92.83 %|70
+ResNet56 v1| 9| 92.71 %| 93.03 %|90
+ResNet110 v1| 18| 92.65 %| 93.39+-.16 %|165
+ResNet164 v1| 27| - %| 94.07 %| -
+ResNet1001 v1|N/A| - %| 92.39 %| -
+
+
+
+Model|n|200-epoch accuracy|Original paper accuracy |sec/epoch GTX1080Ti
+:------------|--:|-------:|-----------------------:|---:
+ResNet20 v2| 2| - %| - %|---
+ResNet32 v2|N/A| NA %| NA %| NA
+ResNet44 v2|N/A| NA %| NA %| NA
+ResNet56 v2| 6| 93.01 %| NA %|100
+ResNet110 v2| 12| 93.15 %| 93.63 %|180
+ResNet164 v2| 18| - %| 94.54 %| -
+ResNet1001 v2|111| - %| 95.08+-.14 %| -
+"""
+
+from __future__ import print_function
+import keras
+from keras.layers import Dense, Conv2D, BatchNormalization, Activation
+from keras.layers import AveragePooling2D, Input, Flatten, ZeroPadding2D
+from keras.optimizers import Adam
+from keras.callbacks import ModelCheckpoint, LearningRateScheduler
+from keras.callbacks import ReduceLROnPlateau
+from keras.preprocessing.image import ImageDataGenerator
+from keras.regularizers import l2
+from keras import backend as K
+from keras.models import Model
+from keras.datasets import cifar10
+from keras import backend as K
+import numpy as np
+import os
+import sys
+from approxhpvm_translator import translate_to_approxhpvm
+from weight_utils import dumpCalibrationData
+
+
+
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+
+
+K.set_image_data_format('channels_first')
+
+
+# Training parameters
+batch_size = 32 # orig paper trained all networks with batch_size=128
+#---- epochs = 200
+epochs = 2
+data_augmentation = True
+num_classes = 10
+
+# Subtracting pixel mean improves accuracy
+subtract_pixel_mean = True
+
+# Model parameter
+# ----------------------------------------------------------------------------
+# | | 200-epoch | Orig Paper| 200-epoch | Orig Paper| sec/epoch
+# Model | n | ResNet v1 | ResNet v1 | ResNet v2 | ResNet v2 | GTX1080Ti
+# |v1(v2)| %Accuracy | %Accuracy | %Accuracy | %Accuracy | v1 (v2)
+# ----------------------------------------------------------------------------
+# ResNet20 | 3 (2)| 92.16 | 91.25 | ----- | ----- | 35 (---)
+# ResNet32 | 5(NA)| 92.46 | 92.49 | NA | NA | 50 ( NA)
+# ResNet44 | 7(NA)| 92.50 | 92.83 | NA | NA | 70 ( NA)
+# ResNet56 | 9 (6)| 92.71 | 93.03 | 93.01 | NA | 90 (100)
+# ResNet110 |18(12)| 92.65 | 93.39+-.16| 93.15 | 93.63 | 165(180)
+# ResNet164 |27(18)| ----- | 94.07 | ----- | 94.54 | ---(---)
+# ResNet1001| (111)| ----- | 92.39 | ----- | 95.08+-.14| ---(---)
+# ---------------------------------------------------------------------------
+n = 3
+
+# Model version
+# Orig paper: version = 1 (ResNet v1), Improved ResNet: version = 2 (ResNet v2)
+version = 1
+
+# Computed depth from supplied model parameter n
+if version == 1:
+ depth = n * 6 + 2
+elif version == 2:
+ depth = n * 9 + 2
+
+# Model name, depth and version
+model_type = 'ResNet%dv%d' % (depth, version)
+
+# Load the CIFAR10 data.
+(x_train, y_train), (x_test, y_test) = cifar10.load_data()
+test_labels = y_test
+train_labels = y_train
+
+# Input image dimensions.
+input_shape = x_train.shape[1:]
+
+# Normalize data.
+x_train = x_train.astype('float32') / 255
+x_test = x_test.astype('float32') / 255
+
+# If subtract pixel mean is enabled
+if subtract_pixel_mean:
+ x_train_mean = np.mean(x_train, axis=0)
+ x_train -= x_train_mean
+ x_test -= x_train_mean
+
+print('x_train shape:', x_train.shape)
+print(x_train.shape[0], 'train samples')
+print(x_test.shape[0], 'test samples')
+print('y_train shape:', y_train.shape)
+
+# Convert class vectors to binary class matrices.
+y_train = keras.utils.to_categorical(y_train, num_classes)
+y_test = keras.utils.to_categorical(y_test, num_classes)
+
+
+
+
+
+
+def lr_schedule(epoch):
+ """Learning Rate Schedule
+
+ Learning rate is scheduled to be reduced after 80, 120, 160, 180 epochs.
+ Called automatically every epoch as part of callbacks during training.
+
+ # Arguments
+ epoch (int): The number of epochs
+
+ # Returns
+ lr (float32): learning rate
+ """
+ lr = 1e-3
+ if epoch > 180:
+ lr *= 0.5e-3
+ elif epoch > 160:
+ lr *= 1e-3
+ elif epoch > 120:
+ lr *= 1e-2
+ elif epoch > 80:
+ lr *= 1e-1
+ print('Learning rate: ', lr)
+ return lr
+
+
+def resnet_layer(inputs,
+ num_filters=16,
+ kernel_size=3,
+ strides=1,
+ activation='relu',
+ batch_normalization=True,
+ conv_first=True):
+ """2D Convolution-Batch Normalization-Activation stack builder
+
+ # Arguments
+ inputs (tensor): input tensor from input image or previous layer
+ num_filters (int): Conv2D number of filters
+ kernel_size (int): Conv2D square kernel dimensions
+ strides (int): Conv2D square stride dimensions
+ activation (string): activation name
+ batch_normalization (bool): whether to include batch normalization
+ conv_first (bool): conv-bn-activation (True) or
+ bn-activation-conv (False)
+
+ # Returns
+ x (tensor): tensor as input to the next layer
+ """
+ conv = Conv2D(num_filters,
+ kernel_size=kernel_size,
+ strides=strides,
+ padding='valid', # NOTE: using valid convs with explicit pad operation
+ kernel_initializer='he_normal',
+ kernel_regularizer=l2(1e-4))
+
+ padding_value = int((kernel_size - 1) / 2)
+ zero_padding = ZeroPadding2D(padding = (padding_value, padding_value))
+
+ # FIXME: Temporarily disabled batch normalization
+ batch_normalization = False
+
+ x = inputs
+ x = zero_padding(x)
+ if conv_first:
+ x = conv(x)
+ if batch_normalization:
+ x = BatchNormalization()(x)
+ if activation is not None:
+ x = Activation(activation)(x)
+ else:
+ if batch_normalization:
+ x = BatchNormalization()(x)
+ if activation is not None:
+ x = Activation(activation)(x)
+ x = conv(x)
+ return x
+
+
+
+def resnet_v0(input_shape, depth, num_classes=10):
+ """ResNet Version 1 Model builder [a]
+
+ Stacks of 2 x (3 x 3) Conv2D-BN-ReLU
+ Last ReLU is after the shortcut connection.
+ At the beginning of each stage, the feature map size is halved (downsampled)
+ by a convolutional layer with strides=2, while the number of filters is
+ doubled. Within each stage, the layers have the same number filters and the
+ same number of filters.
+ Features maps sizes:
+ stage 0: 32x32, 16
+ stage 1: 16x16, 32
+ stage 2: 8x8, 64
+ The Number of parameters is approx the same as Table 6 of [a]:
+ ResNet20 0.27M
+ ResNet32 0.46M
+ ResNet44 0.66M
+ ResNet56 0.85M
+ ResNet110 1.7M
+
+ # Arguments
+ input_shape (tensor): shape of input image tensor
+ depth (int): number of core convolutional layers
+ num_classes (int): number of classes (CIFAR10 has 10)
+
+ # Returns
+ model (Model): Keras model instance
+ """
+ if (depth - 2) % 6 != 0:
+ raise ValueError('depth should be 6n+2 (eg 20, 32, 44 in [a])')
+ # Start model definition.
+ num_filters = 16
+ num_res_blocks = int((depth - 2) / 6)
+
+ inputs = Input(shape=input_shape)
+ x = resnet_layer(inputs=inputs)
+ # Instantiate the stack of residual units
+ for stack in range(3):
+ for res_block in range(num_res_blocks):
+ strides = 1
+ if stack > 0 and res_block == 0: # first layer but not first stack
+ strides = 2 # downsample
+ y = resnet_layer(inputs=x,
+ num_filters=num_filters,
+ strides=strides)
+ y = resnet_layer(inputs=y,
+ num_filters=num_filters,
+ activation=None)
+ if stack > 0 and res_block == 0: # first layer but not first stack
+ # linear projection residual shortcut connection to match
+ # changed dims
+ x = resnet_layer(inputs=x,
+ num_filters=num_filters,
+ kernel_size=1,
+ strides=strides,
+ activation=None,
+ batch_normalization=False)
+ x = keras.layers.add([x, y])
+ x = Activation('relu')(x)
+ num_filters *= 1
+
+ # Add classifier on top.
+ # v1 does not use BN after last shortcut connection-ReLU
+ #-- x = AveragePooling2D(pool_size=8)(x)
+ y = Flatten()(x)
+ x = Dense(64)(y)
+ outputs = Dense(num_classes,
+ activation='softmax',
+ kernel_initializer='he_normal')(x)
+
+ # Instantiate model.
+ model = Model(inputs=inputs, outputs=outputs)
+ return model
+
+
+def resnet_v1_1(input_shape, depth, num_classes=10):
+ """ResNet Version 1 Model builder [a]
+
+ Stacks of 2 x (3 x 3) Conv2D-BN-ReLU
+ Last ReLU is after the shortcut connection.
+ At the beginning of each stage, the feature map size is halved (downsampled)
+ by a convolutional layer with strides=2, while the number of filters is
+ doubled. Within each stage, the layers have the same number filters and the
+ same number of filters.
+ Features maps sizes:
+ stage 0: 32x32, 16
+ stage 1: 16x16, 32
+ stage 2: 8x8, 64
+ The Number of parameters is approx the same as Table 6 of [a]:
+ ResNet20 0.27M
+ ResNet32 0.46M
+ ResNet44 0.66M
+ ResNet56 0.85M
+ ResNet110 1.7M
+
+ # Arguments
+ input_shape (tensor): shape of input image tensor
+ depth (int): number of core convolutional layers
+ num_classes (int): number of classes (CIFAR10 has 10)
+
+ # Returns
+ model (Model): Keras model instance
+ """
+ if (depth - 2) % 6 != 0:
+ raise ValueError('depth should be 6n+2 (eg 20, 32, 44 in [a])')
+ # Start model definition.
+ num_filters = 16
+ num_res_blocks = int((depth - 2) / 6)
+
+ inputs = Input(shape=input_shape)
+ x = resnet_layer(inputs=inputs)
+ # Instantiate the stack of residual units
+ for stack in range(3):
+ for res_block in range(num_res_blocks):
+ strides = 1
+ if stack > 0 and res_block == 0: # first layer but not first stack
+ strides = 2 # downsample
+ y = resnet_layer(inputs=x,
+ num_filters=num_filters,
+ strides=strides)
+ y = resnet_layer(inputs=y,
+ num_filters=num_filters,
+ activation=None)
+ if stack > 0 and res_block == 0: # first layer but not first stack
+ # linear projection residual shortcut connection to match
+ # changed dims
+ x = resnet_layer(inputs=x,
+ num_filters=num_filters,
+ kernel_size=1,
+ strides=strides,
+ activation=None,
+ batch_normalization=False)
+ x = keras.layers.add([x, y])
+ x = Activation('relu')(x)
+ num_filters *= 2
+
+
+ x = AveragePooling2D(pool_size=8)(x)
+ y = Flatten()(x)
+ outputs = Dense(num_classes,
+ #activation='softmax',
+ kernel_initializer='he_normal')(y)
+
+ outputs = Activation('softmax')(outputs)
+
+
+ # Instantiate model.
+ model = Model(inputs=inputs, outputs=outputs)
+ return model
+
+
+
+def resnet_v2(input_shape, depth, num_classes=10):
+ """ResNet Version 2 Model builder [b]
+
+ Stacks of (1 x 1)-(3 x 3)-(1 x 1) BN-ReLU-Conv2D or also known as
+ bottleneck layer
+ First shortcut connection per layer is 1 x 1 Conv2D.
+ Second and onwards shortcut connection is identity.
+ At the beginning of each stage, the feature map size is halved (downsampled)
+ by a convolutional layer with strides=2, while the number of filter maps is
+ doubled. Within each stage, the layers have the same number filters and the
+ same filter map sizes.
+ Features maps sizes:
+ conv1 : 32x32, 16
+ stage 0: 32x32, 64
+ stage 1: 16x16, 128
+ stage 2: 8x8, 256
+
+ # Arguments
+ input_shape (tensor): shape of input image tensor
+ depth (int): number of core convolutional layers
+ num_classes (int): number of classes (CIFAR10 has 10)
+
+ # Returns
+ model (Model): Keras model instance
+ """
+ if (depth - 2) % 9 != 0:
+ raise ValueError('depth should be 9n+2 (eg 56 or 110 in [b])')
+ # Start model definition.
+ num_filters_in = 16
+ num_res_blocks = int((depth - 2) / 9)
+
+ inputs = Input(shape=input_shape)
+ # v2 performs Conv2D with BN-ReLU on input before splitting into 2 paths
+ x = resnet_layer(inputs=inputs,
+ num_filters=num_filters_in,
+ conv_first=True)
+
+ # Instantiate the stack of residual units
+ for stage in range(3):
+ for res_block in range(num_res_blocks):
+ activation = 'relu'
+ batch_normalization = True
+ strides = 1
+ if stage == 0:
+ num_filters_out = num_filters_in * 4
+ if res_block == 0: # first layer and first stage
+ activation = None
+ batch_normalization = False
+ else:
+ num_filters_out = num_filters_in * 2
+ if res_block == 0: # first layer but not first stage
+ strides = 2 # downsample
+
+ # bottleneck residual unit
+ y = resnet_layer(inputs=x,
+ num_filters=num_filters_in,
+ kernel_size=1,
+ strides=strides,
+ activation=activation,
+ batch_normalization=batch_normalization,
+ conv_first=False)
+ y = resnet_layer(inputs=y,
+ num_filters=num_filters_in,
+ conv_first=False)
+ y = resnet_layer(inputs=y,
+ num_filters=num_filters_out,
+ kernel_size=1,
+ conv_first=False)
+ if res_block == 0:
+ # linear projection residual shortcut connection to match
+ # changed dims
+ x = resnet_layer(inputs=x,
+ num_filters=num_filters_out,
+ kernel_size=1,
+ strides=strides,
+ activation=None,
+ batch_normalization=False)
+ x = keras.layers.add([x, y])
+
+ num_filters_in = num_filters_out
+
+ # Add classifier on top.
+ # v2 has BN-ReLU before Pooling
+ x = BatchNormalization()(x)
+ x = Activation('relu')(x)
+ x = AveragePooling2D(pool_size=8)(x)
+ y = Flatten()(x)
+ outputs = Dense(num_classes,
+ activation='softmax',
+ kernel_initializer='he_normal')(y)
+
+ # Instantiate model.
+ model = Model(inputs=inputs, outputs=outputs)
+ return model
+
+depth = 20
+
+if version == 2:
+ model = resnet_v2(input_shape=input_shape, depth=depth)
+else:
+ model = resnet_v1_1(input_shape=input_shape, depth=depth)
+
+
+model.compile(loss='categorical_crossentropy',
+ optimizer=Adam(lr=lr_schedule(0)),
+ metrics=['accuracy'])
+model.summary()
+print(model_type)
+
+# Prepare model model saving directory.
+save_dir = os.path.join(os.getcwd(), 'saved_models')
+model_name = 'cifar10_%s_model.{epoch:03d}.h5' % model_type
+if not os.path.isdir(save_dir):
+ os.makedirs(save_dir)
+filepath = os.path.join(save_dir, model_name)
+
+# Prepare callbacks for model saving and for learning rate adjustment.
+checkpoint = ModelCheckpoint(filepath=filepath,
+ monitor='val_acc',
+ verbose=1,
+ save_best_only=True)
+
+lr_scheduler = LearningRateScheduler(lr_schedule)
+
+lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1),
+ cooldown=0,
+ patience=5,
+ min_lr=0.5e-6)
+
+callbacks = [checkpoint, lr_reducer, lr_scheduler]
+
+# Run training, with or without data augmentation.
+if not data_augmentation:
+ print('Not using data augmentation.')
+ model.fit(x_train, y_train,
+ batch_size=batch_size,
+ epochs=epochs,
+ validation_data=(x_test, y_test),
+ shuffle=True,
+ callbacks=callbacks)
+else:
+ print('Using real-time data augmentation.')
+ # This will do preprocessing and realtime data augmentation:
+ datagen = ImageDataGenerator(
+ # set input mean to 0 over the dataset
+ featurewise_center=False,
+ # set each sample mean to 0
+ samplewise_center=False,
+ # divide inputs by std of dataset
+ featurewise_std_normalization=False,
+ # divide each input by its std
+ samplewise_std_normalization=False,
+ # apply ZCA whitening
+ zca_whitening=False,
+ # epsilon for ZCA whitening
+ zca_epsilon=1e-06,
+ # randomly rotate images in the range (deg 0 to 180)
+ rotation_range=0,
+ # randomly shift images horizontally
+ width_shift_range=0.1,
+ # randomly shift images vertically
+ height_shift_range=0.1,
+ # set range for random shear
+ shear_range=0.,
+ # set range for random zoom
+ zoom_range=0.,
+ # set range for random channel shifts
+ channel_shift_range=0.,
+ # set mode for filling points outside the input boundaries
+ fill_mode='nearest',
+ # value used for fill_mode = "constant"
+ cval=0.,
+ # randomly flip images
+ horizontal_flip=True,
+ # randomly flip images
+ vertical_flip=False,
+ # set rescaling factor (applied before any other transformation)
+ rescale=None,
+ # set function that will be applied on each input
+ preprocessing_function=None,
+ # image data format, either "channels_first" or "channels_last"
+ data_format="channels_first",
+ # fraction of images reserved for validation (strictly between 0 and 1)
+ validation_split=0.0)
+
+ # Compute quantities required for featurewise normalization
+ # (std, mean, and principal components if ZCA whitening is applied).
+ datagen.fit(x_train)
+
+ # Fit the model on the batches generated by datagen.flow().
+ model.fit_generator(datagen.flow(x_train, y_train, batch_size=batch_size),
+ validation_data=(x_test, y_test),
+ epochs=epochs, verbose=1, workers=4,
+ callbacks=callbacks)
+
+# Score trained model.
+scores = model.evaluate(x_test, y_test, verbose=1)
+print('Test loss:', scores[0])
+print('Test accuracy:', scores[1])
+
+
+dumpCalibrationData("calibration_data/resnet18_calib.bin", x_train,
+ "calibration_data/resnet18_train_labels.bin", train_labels)
+sys.exit(0)
+
+#translate_to_approxhpvm(model, "resnet18_cifar10_hpvm/", x_test, test_labels)
+
+translate_to_approxhpvm(model, "resnet_test/", x_test, test_labels, 'resnet18_cifar10_promise/', y_test)
diff --git a/llvm/projects/keras/src/vgg16_cifar10.py b/llvm/projects/keras/src/vgg16_cifar10.py
new file mode 100644
index 0000000000..64e5d36e78
--- /dev/null
+++ b/llvm/projects/keras/src/vgg16_cifar10.py
@@ -0,0 +1,241 @@
+
+
+from __future__ import print_function
+import keras
+from keras.datasets import cifar10
+from keras.preprocessing.image import ImageDataGenerator
+from keras.models import Sequential
+from keras.layers import Dense, Dropout, Activation, Flatten
+from keras.layers import Conv2D, MaxPooling2D, BatchNormalization
+from keras import optimizers
+import numpy as np
+from keras.layers.core import Lambda
+from keras import backend as K
+from keras import regularizers
+import os
+import sys
+from frontend.approxhpvm_translator import translate_to_approxhpvm
+from frontend.weight_utils import dumpCalibrationData
+
+
+
+class cifar10vgg:
+ def __init__(self,train=True):
+ self.num_classes = 10
+ self.weight_decay = 0.0005
+ self.x_shape = [3,32,32]
+
+ self.model = self.build_model()
+ if train:
+ self.model = self.train(self.model)
+ else:
+ self.model.load_weights('cifar10vgg.h5')
+
+
+ def build_model(self):
+ # Build the network of vgg for 10 classes with massive dropout and weight decay as described in the paper.
+
+ model = Sequential()
+ weight_decay = self.weight_decay
+
+ model.add(Conv2D(64, (3, 3), padding='same',
+ input_shape=self.x_shape,kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.3))
+
+ model.add(Conv2D(64, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ # model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+
+ model.add(Conv2D(128, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(128, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+
+ model.add(Conv2D(256, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(256, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ # model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(256, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ # model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ # model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ # model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ # model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ # model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ # model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ # model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+ model.add(Dropout(0.5))
+
+ model.add(Flatten())
+ model.add(Dense(512,kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ # model.add(BatchNormalization())
+
+ model.add(Dropout(0.5))
+ model.add(Dense(self.num_classes))
+ model.add(Activation('softmax'))
+ return model
+
+
+ def normalize(self,X_train,X_test):
+ #this function normalize inputs for zero mean and unit variance
+ # it is used when training a model.
+ # Input: training set and test set
+ # Output: normalized training set and test set according to the trianing set statistics.
+ mean = np.mean(X_train,axis=(0,1,2,3))
+ std = np.std(X_train, axis=(0, 1, 2, 3))
+ X_train = (X_train-mean)/(std+1e-7)
+ X_test = (X_test-mean)/(std+1e-7)
+ return X_train, X_test
+
+
+ def normalize_production(self,x):
+ #this function is used to normalize instances in production according to saved training set statistics
+ # Input: X - a training set
+ # Output X - a normalized training set according to normalization constants.
+
+ #these values produced during first training and are general for the standard cifar10 training set normalization
+ mean = 120.707
+ std = 64.15
+ return (x-mean)/(std+1e-7)
+
+
+ def predict(self,x,normalize=True,batch_size=50):
+ if normalize:
+ x = self.normalize_production(x)
+ return self.model.predict(x,batch_size)
+
+
+ def train(self,model):
+
+ #training parameters
+ batch_size = 128
+ #maxepoches = 250
+ #maxepoches = 250
+ maxepoches = 30
+ learning_rate = 0.01
+ lr_decay = 1e-6
+ lr_drop = 20
+ # The data, shuffled and split between train and test sets:
+ (x_train, y_train), (x_test, y_test) = cifar10.load_data()
+ x_train = x_train.astype('float32')
+ x_test = x_test.astype('float32')
+ x_train, x_test = self.normalize(x_train, x_test)
+
+ y_train = keras.utils.to_categorical(y_train, self.num_classes)
+ y_test = keras.utils.to_categorical(y_test, self.num_classes)
+
+ def lr_scheduler(epoch):
+ return learning_rate * (0.5 ** (epoch // lr_drop))
+ reduce_lr = keras.callbacks.LearningRateScheduler(lr_scheduler)
+
+ #data augmentation
+ datagen = ImageDataGenerator(
+ featurewise_center=False, # set input mean to 0 over the dataset
+ samplewise_center=False, # set each sample mean to 0
+ featurewise_std_normalization=False, # divide inputs by std of the dataset
+ samplewise_std_normalization=False, # divide each input by its std
+ zca_whitening=False, # apply ZCA whitening
+ rotation_range=15, # randomly rotate images in the range (degrees, 0 to 180)
+ width_shift_range=0.1, # randomly shift images horizontally (fraction of total width)
+ height_shift_range=0.1, # randomly shift images vertically (fraction of total height)
+ horizontal_flip=True, # randomly flip images
+ vertical_flip=False) # randomly flip images
+ # (std, mean, and principal components if ZCA whitening is applied).
+ datagen.fit(x_train)
+
+
+
+ #optimization details
+ sgd = optimizers.SGD(lr=learning_rate, decay=lr_decay, momentum=0.9, nesterov=True)
+ model.compile(loss='categorical_crossentropy', optimizer=sgd,metrics=['accuracy'])
+
+
+ # training process in a for loop with learning rate drop every 25 epoches.
+
+ historytemp = model.fit_generator(datagen.flow(x_train, y_train,
+ batch_size=batch_size),
+ steps_per_epoch=x_train.shape[0] // batch_size,
+ epochs=maxepoches,
+ validation_data=(x_test, y_test),callbacks=[reduce_lr],verbose=2)
+
+ model.save_weights('cifar10vgg.h5')
+ return model
+
+
+
+if __name__ == '__main__':
+
+ K.set_image_data_format('channels_first')
+
+ os.environ["CUDA_VISIBLE_DEVICES"] = "1"
+
+ (x_train, y_train), (x_test, y_test) = cifar10.load_data()
+ test_labels = y_test
+ train_labels = y_train
+ x_train = x_train.astype('float32')
+ x_test = x_test.astype('float32')
+
+ y_train = keras.utils.to_categorical(y_train, 10)
+ y_test = keras.utils.to_categorical(y_test, 10)
+
+ model = cifar10vgg()
+
+ predicted_x = model.predict(x_test)
+
+ norm_test = model.normalize_production(x_test)
+
+ # Normalizing train data before dumping
+ #x_train, x_test = model.normalize(x_train, x_test)
+ x_train = model.normalize_production(x_train)
+
+ # dumpCalibrationData("vgg16_cifar_calib.bin", x_train, "vgg16_train_labels.bin", train_labels)
+
+ translate_to_approxhpvm(model.model, "data/vgg16_cifar10/", norm_test, test_labels, 10)
+
+ residuals = np.argmax(predicted_x,1)!=np.argmax(y_test,1)
+
+ loss = sum(residuals)/len(residuals)
+ print("the validation 0/1 loss is: ",loss)
+
+
+
diff --git a/llvm/projects/keras/src/vgg16_cifar100.py b/llvm/projects/keras/src/vgg16_cifar100.py
new file mode 100644
index 0000000000..66fe6be669
--- /dev/null
+++ b/llvm/projects/keras/src/vgg16_cifar100.py
@@ -0,0 +1,243 @@
+
+from __future__ import print_function
+import os
+import keras
+from keras.datasets import cifar100
+from keras.preprocessing.image import ImageDataGenerator
+from keras.models import Sequential
+from keras.layers import Dense, Dropout, Activation, Flatten
+from keras.layers import Conv2D, MaxPooling2D
+from keras import optimizers
+import numpy as np
+from keras.layers.core import Lambda
+from keras import backend as K
+from keras import regularizers
+from approxhpvm_translator import translate_to_approxhpvm
+import sys
+from weight_utils import dumpCalibrationData
+
+
+
+class cifar100vgg:
+ def __init__(self,train=True):
+ self.num_classes = 100
+ self.weight_decay = 0.0005
+ self.x_shape = [3,32,32]
+
+ self.model = self.build_model()
+ if train:
+ self.model = self.train(self.model)
+ else:
+ self.model.load_weights('cifar100vgg.h5')
+
+
+ def build_model(self):
+ # Build the network of vgg for 10 classes with massive dropout and weight decay as described in the paper.
+
+ model = Sequential()
+ weight_decay = self.weight_decay
+
+ model.add(Conv2D(64, (3, 3), padding='same',
+ input_shape=self.x_shape,kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.3))
+
+ model.add(Conv2D(64, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+
+ model.add(Conv2D(128, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(128, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+
+ model.add(Conv2D(256, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(256, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(256, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+ model.add(Dropout(0.4))
+
+ model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+
+ model.add(MaxPooling2D(pool_size=(2, 2)))
+ model.add(Dropout(0.5))
+
+ model.add(Flatten())
+ model.add(Dense(512,kernel_regularizer=regularizers.l2(weight_decay)))
+ model.add(Activation('relu'))
+ #model.add(BatchNormalization())
+
+ model.add(Dropout(0.5))
+ model.add(Dense(self.num_classes))
+ model.add(Activation('softmax'))
+ return model
+
+
+ def normalize(self,X_train,X_test):
+ #this function normalize inputs for zero mean and unit variance
+ # it is used when training a model.
+ # Input: training set and test set
+ # Output: normalized training set and test set according to the trianing set statistics.
+ mean = np.mean(X_train,axis=(0,1,2,3))
+ std = np.std(X_train, axis=(0, 1, 2, 3))
+ print(mean)
+ print(std)
+ X_train = (X_train-mean)/(std+1e-7)
+ X_test = (X_test-mean)/(std+1e-7)
+ return X_train, X_test
+
+ def normalize_production(self,x):
+ #this function is used to normalize instances in production according to saved training set statistics
+ # Input: X - a training set
+ # Output X - a normalized training set according to normalization constants.
+
+ #these values produced during first training and are general for the standard cifar10 training set normalization
+ mean = 121.936
+ std = 68.389
+ return (x-mean)/(std+1e-7)
+
+ def predict(self,x,normalize=True,batch_size=50):
+ if normalize:
+ x = self.normalize_production(x)
+ return self.model.predict(x,batch_size)
+
+ def train(self,model):
+
+ #training parameters
+ batch_size = 128
+ #maxepoches = 250
+ #maxepoches = 400
+ maxepoches = 4
+ learning_rate = 0.05
+ lr_decay = 1e-6
+ lr_drop = 20
+
+ # The data, shuffled and split between train and test sets:
+ (x_train, y_train), (x_test, y_test) = cifar100.load_data()
+ x_train = x_train.astype('float32')
+ x_test = x_test.astype('float32')
+ x_train, x_test = self.normalize(x_train, x_test)
+
+ y_train = keras.utils.to_categorical(y_train, self.num_classes)
+ y_test = keras.utils.to_categorical(y_test, self.num_classes)
+
+
+ def lr_scheduler(epoch):
+ return learning_rate * (0.5 ** (epoch // lr_drop))
+ reduce_lr = keras.callbacks.LearningRateScheduler(lr_scheduler)
+
+
+ #data augmentation
+ datagen = ImageDataGenerator(
+ featurewise_center=False, # set input mean to 0 over the dataset
+ samplewise_center=False, # set each sample mean to 0
+ featurewise_std_normalization=False, # divide inputs by std of the dataset
+ samplewise_std_normalization=False, # divide each input by its std
+ zca_whitening=False, # apply ZCA whitening
+ rotation_range=15, # randomly rotate images in the range (degrees, 0 to 180)
+ width_shift_range=0.1, # randomly shift images horizontally (fraction of total width)
+ height_shift_range=0.1, # randomly shift images vertically (fraction of total height)
+ horizontal_flip=True, # randomly flip images
+ vertical_flip=False) # randomly flip images
+ # (std, mean, and principal components if ZCA whitening is applied).
+ datagen.fit(x_train)
+
+
+
+ #optimization details
+ sgd = optimizers.SGD(lr=learning_rate, decay=lr_decay, momentum=0.9, nesterov=True)
+ model.compile(loss='categorical_crossentropy', optimizer=sgd,metrics=['accuracy'])
+
+
+ # training process in a for loop with learning rate drop every 25 epoches.
+
+ historytemp = model.fit_generator(datagen.flow(x_train, y_train,
+ batch_size=batch_size),
+ steps_per_epoch=x_train.shape[0] // batch_size,
+ epochs=maxepoches,
+ validation_data=(x_test, y_test),callbacks=[reduce_lr],verbose=2)
+ model.save_weights('cifar100vgg.h5')
+ return model
+
+if __name__ == '__main__':
+
+ K.set_image_data_format('channels_first')
+ os.environ["CUDA_VISIBLE_DEVICES"] = "1"
+
+ (x_train, y_train), (x_test, y_test) = cifar100.load_data()
+ test_labels = y_test
+ train_labels = y_train
+
+ x_train = x_train.astype('float32')
+ x_test = x_test.astype('float32')
+
+ y_train = keras.utils.to_categorical(y_train, 100)
+ y_test = keras.utils.to_categorical(y_test, 100)
+
+ model = cifar100vgg()
+
+ predicted_x = model.predict(x_test)
+
+ norm_test = model.normalize_production(x_test)
+
+ x_train = model.normalize_production(x_train)
+
+ dumpCalibrationData("calibration_data/vgg16_cifar100_calib.bin", x_train,
+ "calibration_data/vgg16_cifar100_train_labels.bin", train_labels)
+ sys.exit(0)
+
+
+ translate_to_approxhpvm(model.model, "vgg16_cifar100_test/", norm_test, test_labels,
+ "vgg16_cifar100_front", y_test)
+
+
+ residuals = (np.argmax(predicted_x,1)!=np.argmax(y_test,1))
+ loss = sum(residuals)/len(residuals)
+ print("the validation 0/1 loss is: ",loss)
+
+
--
GitLab