Skip to content
Snippets Groups Projects
Commit 10a57c79 authored by shingjan's avatar shingjan
Browse files

change code gen and graph ir for tensor runtime

parent 6491cfe3
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
Showing
with 368 additions and 1632 deletions
hpvm/projects/onnx/frontend/approx_codegen.py
+ 18
− 17
View file @ 10a57c79
...@@ -5,7 +5,7 @@ class GraphCodeGen: ...@@ -5,7 +5,7 @@ class GraphCodeGen:
self._root = "" self._root = ""
self._root_struct = "" self._root_struct = ""
self._main_func = "" self._main_func = ""
def emitHeaders(self): def emitHeaders(self):
headers = "\n#include <stdio.h> \n" headers = "\n#include <stdio.h> \n"
headers += "#include <stdlib.h> \n" headers += "#include <stdlib.h> \n"
...@@ -17,7 +17,7 @@ class GraphCodeGen: ...@@ -17,7 +17,7 @@ class GraphCodeGen:
headers += "#include <tensorTypes.h> \n" headers += "#include <tensorTypes.h> \n"
headers += "#include <tensorUtils.h> \n\n" headers += "#include <tensorUtils.h> \n\n"
self._headers = headers self._headers = headers
def emitRoot(self): def emitRoot(self):
def emitRootNodeHeader(): def emitRootNodeHeader():
root_signature = "void root(" root_signature = "void root("
...@@ -27,29 +27,31 @@ class GraphCodeGen: ...@@ -27,29 +27,31 @@ class GraphCodeGen:
root_signature += "\t " root_signature += "\t "
self.filter_names[f_name] = index self.filter_names[f_name] = index
root_signature += "void* " + f_name + ", " root_signature += "void* " + f_name + ", "
root_signature += "size_t " + f_name + "_bytes" root_signature += "size_t " + f_name + "_bytes"
if index < len(self.filter_names) - 1: if index < len(self.filter_names) - 1:
root_signature += ", \n" root_signature += ", \n"
index += 1 index += 1
root_signature += "){ \n\n" root_signature += "){ \n\n"
root_signature += "\n __visc__hint(visc::CPU_TARGET); \n" root_signature += "\n __visc__hint(visc::CPU_TARGET); \n"
root_signature += " __visc__attributes(" + str(len(self.filter_names)) + ", " root_signature += " __visc__attributes(" + \
str(len(self.filter_names)) + ", "
index = 0 index = 0
for f_name in self.filter_names: for f_name in self.filter_names:
root_signature += f_name root_signature += f_name
if index < len(self.filter_names) - 1: if index < len(self.filter_names) - 1:
root_signature += ", " root_signature += ", "
index += 1 index += 1
root_signature += ", 0); \n\n" root_signature += ", 0); \n\n"
return root_signature return root_signature
def emitRootNodeFooter(self):
def emitRootNodeFooter(self):
last_node = self.dfg.last_node last_node = self.dfg.last_node
output_var = self.output_map[last_node.layer_name] output_var = self.output_map[last_node.layer_name]
# Binding output of last DFG node to the Root Node output # 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 = "\n __visc__bindOut(" + \
root_footer_str += " __visc__bindOut(" + output_var + ", 1, 1, 0); \n" output_var + ", 0, 0, 0); \n"
root_footer_str += " __visc__bindOut(" + \
output_var + ", 1, 1, 0); \n"
root_footer_str += "\n}\n\n" root_footer_str += "\n}\n\n"
return root_footer_str return root_footer_str
...@@ -71,14 +73,14 @@ class GraphCodeGen: ...@@ -71,14 +73,14 @@ class GraphCodeGen:
self._root_struct += emitRootStructure() self._root_struct += emitRootStructure()
self.codegen(self.dfg) self.codegen(self.dfg)
self._root += emitRootNodeFooter() self._root += emitRootNodeFooter()
def emitMainFunc(self, test_data): def emitMainFunc(self, test_data):
main_func_str = "int main(){ \n\n" main_func_str = "int main(){ \n\n"
main_func_str += self.weight_str main_func_str += self.weight_str
main_func_str += self.input_str main_func_str += self.input_str
main_func_str += "\n__visc__init(); \n" main_func_str += "\n__visc__init(); \n"
main_func_str += "RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn))); \n\n" main_func_str += "RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn))); \n\n"
for f_name in self.filter_names: for f_name in self.filter_names:
main_func_str += "args->" + f_name + " = " + f_name + "; \n" main_func_str += "args->" + f_name + " = " + f_name + "; \n"
main_func_str += "args->" + f_name + "_bytes = 0; \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 += "\nvoid* dfg = __visc__launch(0, root, (void*) args); \n\n"
...@@ -86,23 +88,22 @@ class GraphCodeGen: ...@@ -86,23 +88,22 @@ class GraphCodeGen:
main_func_str += "void *result = static_cast<RootIn*>(args)->input; \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 += "hpvm_request_tensor(result, 0); \n\n"
main_func_str += "__visc__cleanup(); \n " main_func_str += "__visc__cleanup(); \n "
main_func_str += "computeAccuracy3(labels, result); \n" main_func_str += "computeAccuracy3(labels, result); \n"
main_func_str += "return 0; \n\n" main_func_str += "return 0; \n\n"
main_func_str += "} \n" main_func_str += "} \n"
self._main_func += main_func_str self._main_func += main_func_str
def emitSource(self, dir_prefix): def emitSource(self, dir_prefix):
source = self._headers + self._nodes + self._root source = self._headers + self._nodes + self._root
source += self._root_struct + self._main_func source += self._root_struct + self._main_func
print (source) print(source)
f = open(dir_prefix + "/approxhpvm_src.cc", "w+") f = open(dir_prefix + "/approxhpvm_src.cc", "w+")
f.write(source) f.write(source)
f.close() f.close()
def codegen(self, model, weights_dir, test_data): def codegen(self, model, weights_dir, test_data):
self.emitHeaders() self.emitHeaders()
self.emitRoot() self.emitRoot()
self.emitMainFunc(test_data) self.emitMainFunc(test_data)
# dump generated program string to source file # dump generated program string to source file
self.emitSource(weights_dir) self.emitSource(weights_dir)
\ No newline at end of file
hpvm/projects/onnx/frontend/common.py
+ 20
− 0
View file @ 10a57c79
import sys import sys
import os import os
from onnx import numpy_helper
class Tensor(object):
def __init__(self, proto):
if not proto.name.strip():
raise ValueError("Tensor's name is required.")
self.name = proto.name
def __str__(self):
return "Tensor: {}\n".format(self.name)
__repr__ = __str__
# Can be either input or weight tensor
class WeightTensor(Tensor):
def __init__(self, input_proto):
Tensor.__init__(self, input_proto)
self.input_data = numpy_helper.to_array(input_proto)#.reshape(tuple(input_proto.dims))
class InputTensor(Tensor):
def __init__(self, inter_name):
self.name = inter_name
hpvm/projects/onnx/frontend/graph_builder.py
+ 97
− 80
View file @ 10a57c79
import sys
from onnx import numpy_helper
from graph_ir import Node
from common import InputTensor, WeightTensor
support_onnx_ops = {"DepthwiseConv2D" : None,
"Conv" : [2], # only 2d supported here
"MatMul" : None,
"MaxPool": [2], # only 2d supported here
"Activation" : None,
"BatchNormalization" : None,
"Flatten" : None,
"Add" : None,
"Relu" : None,
"Softmax" : None,
"Identity": None}
class DFG(object): class DFG(object):
root_set = False; root_set = False
def __init__(self): def __init__(self):
self.node_map = {} self.node_map = {}
...@@ -15,12 +33,10 @@ class DFG(object): ...@@ -15,12 +33,10 @@ class DFG(object):
"Flatten"} "Flatten"}
self.mutliInputLayers = {"Add"} self.mutliInputLayers = {"Add"}
def hasSingleInput(self, layer): def hasSingleInput(self, layer):
layer_name = layer.__class__.__name__ layer_name = layer.__class__.__name__
return layer_name in self.singleInputLayers return layer_name in self.singleInputLayers
def hasMultipleInputs(self, layer): def hasMultipleInputs(self, layer):
layer_name = layer.__class__.__name__ layer_name = layer.__class__.__name__
return layer_name in self.multiInputLayers return layer_name in self.multiInputLayers
...@@ -30,36 +46,37 @@ class DFG(object): ...@@ -30,36 +46,37 @@ class DFG(object):
inbound_node_name = inbound_node_name.split("/")[0] inbound_node_name = inbound_node_name.split("/")[0]
if inbound_node_name in self.node_map: if inbound_node_name in self.node_map:
inbound_node = self.node_map[inbound_node_name] inbound_node = self.node_map[inbound_node_name]
print (inbound_node_name, " found!") print(inbound_node_name, " found!")
inbound_node.add_output(dfg_node) inbound_node.add_output(dfg_node)
dfg_node.add_input(inbound_node) dfg_node.add_input(inbound_node)
else: else:
print ("--inbound node NOT FOUND!") print("--inbound node NOT FOUND!")
def add_to_graph(self, layer): def add_to_graph(self, layer):
dfg_node = DFGNode(layer) dfg_node = DFGNode(layer)
if not self.root_set: if not self.root_set:
self.root_node = dfg_node self.root_node = dfg_node
self.root_set = True # DFG root node is now set self.root_set = True # DFG root node is now set
if self.hasMultipleInputs(layer): if self.hasMultipleInputs(layer):
for j in range(len(layer.input)): for j in range(len(layer.input)):
print(type(layer.input[j])) print(type(layer.input[j]))
print(layer.input[j].op.name) print(layer.input[j].op.name)
self.add_dfg_edge(layer.input[j].op.name, dfg_node) self.add_dfg_edge(layer.input[j].op.name, dfg_node)
else: else:
print (layer.input.name) print(layer.input.name)
self.add_dfg_edge(layer.input.name, dfg_node) self.add_dfg_edge(layer.input.name, dfg_node)
# Adding DFG node to name mapping # Adding DFG node to name mapping
self.node_map[layer.name] = dfg_node self.node_map[layer.name] = dfg_node
# Check if all predecessor nodes have been visited thus far - reverse
# postorder traversal
# Check if all predecessor nodes have been visited thus far - reverse postorder traversal
def predVisited(self, cur_node, visited_nodes): def predVisited(self, cur_node, visited_nodes):
for input_node in cur_node.inputs: for input_node in cur_node.inputs:
if input_node.layer_name not in visited_nodes: if input_node.layer_name not in visited_nodes:
return False; return False
# All predecessors are visited # All predecessors are visited
return True return True
def traverseNode(self, cur_node, visited_nodes): def traverseNode(self, cur_node, visited_nodes):
...@@ -76,19 +93,20 @@ class DFG(object): ...@@ -76,19 +93,20 @@ class DFG(object):
for output_node in cur_node.outputs: for output_node in cur_node.outputs:
self.traverseNode(output_node, visited_nodes) self.traverseNode(output_node, visited_nodes)
# NOTE: Assuming that no outbound edges implies the last node in the graph # NOTE: Assuming that no outbound edges implies the last node in
# the graph
if len(cur_node.outputs) == 0: if len(cur_node.outputs) == 0:
self.last_node = cur_node self.last_node = cur_node
# Build and Print the DFG in reverse postorder
#Build and Print the DFG in reverse postorder
def buildDFG(self): def buildDFG(self):
print ("\n\n ****** Traversing and Printing DFG ******* \n\n") print("\n\n ****** Traversing and Printing DFG ******* \n\n")
visited_nodes = {} visited_nodes = {}
# Starting traversal at the DFG root node # Starting traversal at the DFG root node
self.traverseNode(self.root_node, visited_nodes) self.traverseNode(self.root_node, visited_nodes)
## This should be the place where partial evaluation happens # This should be the place where partial evaluation happens
def emitNode(self, layer): def emitNode(self, layer):
if layer.op_type == "Conv": if layer.op_type == "Conv":
return Conv2DNode() return Conv2DNode()
...@@ -108,22 +126,21 @@ class DFG(object): ...@@ -108,22 +126,21 @@ class DFG(object):
pass pass
else: else:
raise ValueError("Unsupported operator type!") raise ValueError("Unsupported operator type!")
sys.exit("Unsupported operator type!")
class GraphBuilder(object): class GraphBuilder(object):
def __init__(self, model, shape, dtype, opset): def __init__(self, model, shape, dtype, opset, weight_dir):
self._check_model(model) self._check_model(model)
self._nodes = {} self.model = model
self._params = {} self.dtype = dtype
self._renames = {} self.graph = model.graph
self._num_input = 0 self.opset = opset
self._num_param = 0 self.weight_dir = weight_dir
self._dtype = dtype self.shape = shape if shape else self._build_shape()
self._graph = model.graph
self._opset = opset
self._shape = shape if shape else self._build_shape()
self.program_str = ""
self.dfg = DFG() self.dfg = DFG()
self.tensors = dict()
################################################ ################################################
# Aux functions for graph building # Aux functions for graph building
################################################ ################################################
...@@ -140,18 +157,19 @@ class GraphBuilder(object): ...@@ -140,18 +157,19 @@ class GraphBuilder(object):
import warnings import warnings
warnings.warn(str(e)) warnings.warn(str(e))
except ImportError as e: except ImportError as e:
raise ImportError("Unable to import onnx.checker which is required {}".format(e)) raise ImportError(
"Unable to import onnx.checker which is required {}".format(e))
def _build_shape(self): def _build_shape(self):
shape = {} shape = {}
for input in self._graph.input: for input in self.graph.input:
# get type of input tensor # get type of input tensor
tensor_type = input.type.tensor_type tensor_type = input.type.tensor_type
# check if it has a shape: # check if it has a shape:
if (tensor_type.HasField("shape")): if (tensor_type.HasField("shape")):
shape[input.name] = tensor_type.shape shape[input.name] = tensor_type.shape
return shape return shape
def _parse_array(self, tensor_proto): def _parse_array(self, tensor_proto):
try: try:
from onnx.numpy_helper import to_array from onnx.numpy_helper import to_array
...@@ -160,15 +178,15 @@ class GraphBuilder(object): ...@@ -160,15 +178,15 @@ class GraphBuilder(object):
"Unable to import onnx which is required {}".format(e)) "Unable to import onnx which is required {}".format(e))
np_array = to_array(tensor_proto).reshape(tuple(tensor_proto.dims)) np_array = to_array(tensor_proto).reshape(tuple(tensor_proto.dims))
return np_array return np_array
def _parse_value_proto(self, value_proto): def _parse_value_proto(self, value_proto):
"""Parse ValueProto or raw str.""" """Parse ValueProto or raw str."""
try: try:
name = value_proto.name name = value_proto.name
except AttributeError: except AttributeError:
name = value_proto name = value_proto
return name return name
def _parse_dtype(self, value_proto, dtype): def _parse_dtype(self, value_proto, dtype):
"""Parse dtype.""" """Parse dtype."""
try: try:
...@@ -177,55 +195,54 @@ class GraphBuilder(object): ...@@ -177,55 +195,54 @@ class GraphBuilder(object):
except AttributeError: except AttributeError:
return dtype return dtype
def _support_check(self, node):
op_name = node.op_type
if op_name not in support_onnx_ops:
return False
else:
if support_onnx_ops[op_name] == None:
return True
else:
#print(type(node.attribute))
for attr in node.attribute:
# partially evaluate the kernel shape
if attr.name == 'kernel_shape':
return len(attr.ints) in support_onnx_ops[op_name]
return False
def _dump_weight(self, weight_tensor):
print("weights dump here")
################################################ ################################################
# Top level Graph Building functions # Top level Graph Building functions
# return the compile-ready graph # return the compilation-ready graph
################################################ ################################################
def build_graph(self): def build_graph(self):
# parse init tensors for weight_tensor in self.graph.initializer:
for init_tensor in self._graph.initializer: self.tensors[weight_tensor.name] = WeightTensor(weight_tensor)
if not init_tensor.name.strip(): for i in self.graph.input:
raise ValueError("Tensor's name is required.") if i.name not in self.tensors:
self._params[init_tensor.name] = self._parse_array(init_tensor) self.tensors[i.name] = InputTensor(i.name)
self._nodes[init_tensor.name] = Node(init_tensor.name, for node in self.graph.node:
self._params[init_tensor.name].shape,
self._params[init_tensor.name].dtype)
# parse inputs
# from onnx v0.2, GraphProto.input has type ValueInfoProto,
# and the name is 'i.name'
for i in self._graph.input:
i_name = self._parse_value_proto(i)
d_type = self._parse_dtype(i, 'float32')
if i_name in self._params:
# i is a param instead of input
self._num_param += 1
self._params[i_name] = self._params.pop(i_name)
self._nodes[i_name] = Node(i_name,
self._params[i_name].shape,
self._params[i_name].dtype)
else:
self._num_input += 1
if i_name in self._shape:
tshape = self._shape[i_name]
else:
raise ValueError("Must provide an input shape for `{0}`.".format(i_name))
if isinstance(self._dtype, dict):
dtype = self._dtype[i_name] if i_name in self._dtype else d_type
else:
dtype = d_type
self._nodes[i_name] = Node(i_name,
tshape,
dtype)
for node in self._graph.node:
op_name = node.op_type op_name = node.op_type
print("###############################") #print("###############################")
print(op_name) if not self._support_check(node):
raise ValueError(
"Operator not currently supported: `{0}`!".format(op_name))
#print("attribute: " + str(node.attribute)) #print("attribute: " + str(node.attribute))
#print("input: " + str(node.input)) #print("input: " + str(node.input))
#print("output: " + str(node.output)) #print("output: " + str(node.output))
#print(self._nodes) #print("###############################")
for i in node.input:
if i not in self.tensors:
raise ValueError(
"Compilation Interrupted for missing input!`{0}`.".format(i))
for i in node.output:
if i not in self.tensors:
self.tensors[i] = InputTensor(i)
for tensor in self.tensors.values():
if isinstance(tensor, WeightTensor):
print(tensor.name)
return self.dfg return self.dfg
\ No newline at end of file
hpvm/projects/onnx/frontend/graph_codegen.py
+ 14
− 12
View file @ 10a57c79
...@@ -4,7 +4,8 @@ import os ...@@ -4,7 +4,8 @@ import os
from graph_builder import * from graph_builder import *
from graph_ir import * from graph_ir import *
class GraphCodeGen(object): class GraphCodeGen(object):
def __init__(self, DFG): def __init__(self, DFG):
self.program_str = "" self.program_str = ""
...@@ -16,7 +17,7 @@ class GraphCodeGen(object): ...@@ -16,7 +17,7 @@ class GraphCodeGen(object):
def dump_weights(self): def dump_weights(self):
for init_tensor in self._graph.initializer: for init_tensor in self._graph.initializer:
#print(init_tensor) # print(init_tensor)
pass pass
def traverse_graph(self, cur_node, visited): def traverse_graph(self, cur_node, visited):
...@@ -33,7 +34,7 @@ class GraphCodeGen(object): ...@@ -33,7 +34,7 @@ class GraphCodeGen(object):
self.build_graph() self.build_graph()
visited_nodes = {} visited_nodes = {}
self.traverse_graph(self.dfg.root, visited) self.traverse_graph(self.dfg.root, visited)
def emit_header(self): def emit_header(self):
headers = "\n#include <stdio.h> \n" headers = "\n#include <stdio.h> \n"
headers += "#include <stdlib.h> \n" headers += "#include <stdlib.h> \n"
...@@ -53,9 +54,9 @@ class GraphCodeGen(object): ...@@ -53,9 +54,9 @@ class GraphCodeGen(object):
def emit_footer(self, test_data): def emit_footer(self, test_data):
if test_data is not None and self.dfg.last_node is not None: if test_data is not None and self.dfg.last_node is not None:
last_node = self.dfg.last_node last_node = self.dfg.last_node
output_var = self.output_map[last_node.layer_name] output_var = self.output_map[last_node.layer_name]
destructors = "\nllvm_hpvm_cleanupTensorRt(); \n" destructors = "\nllvm_hpvm_cleanupTensorRt(); \n"
end_main = "\nreturn 0; \n\n}\n" end_main = "\nreturn 0; \n\n}\n"
self.program_str += destructors self.program_str += destructors
...@@ -66,10 +67,10 @@ class GraphCodeGen(object): ...@@ -66,10 +67,10 @@ class GraphCodeGen(object):
C = x_test.shape[1] C = x_test.shape[1]
H = x_test.shape[2] H = x_test.shape[2]
W = x_test.shape[3] W = x_test.shape[3]
loop_str = "" loop_str = ""
loop_str += "\nstartMemTracking(); \n\n" loop_str += "\nstartMemTracking(); \n\n"
loop_str += "int test_input_size = " + str(N) + "; \n" loop_str += "int test_input_size = " + str(N) + "; \n"
loop_str += "int batch_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 += "int batch_count = test_input_size / batch_size; \n"
...@@ -79,7 +80,7 @@ class GraphCodeGen(object): ...@@ -79,7 +80,7 @@ class GraphCodeGen(object):
loop_str += "int start = i * batch_size; \n" loop_str += "int start = i * batch_size; \n"
loop_str += "int end = (i + 1) * 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 += "\nvoid* input = readInputBatch(input_path.c_str(),0,start,end,"
loop_str += str(C) + "," + str(H) + "," + str(W) + "); \n\n" loop_str += str(C) + "," + str(H) + "," + str(W) + "); \n\n"
self.program_str += loop_str self.program_str += loop_str
...@@ -89,7 +90,8 @@ class GraphCodeGen(object): ...@@ -89,7 +90,8 @@ class GraphCodeGen(object):
end_loop_str += "\nuint32_t* labels = readLabelsBatch3(labels_path.c_str(),start,end); \n" end_loop_str += "\nuint32_t* labels = readLabelsBatch3(labels_path.c_str(),start,end); \n"
last_node = self.dfg.last_node last_node = self.dfg.last_node
output_var = self.output_map[last_node.layer_name] output_var = self.output_map[last_node.layer_name]
accuracy_call = "\nfloat accuracy = computeAccuracy3(labels, " + output_var + "); \n" accuracy_call = "\nfloat accuracy = computeAccuracy3(labels, " + \
output_var + "); \n"
end_loop_str += accuracy_call end_loop_str += accuracy_call
#end_loop_str += "float accuracy = computeAccuracy2(labels, batch_size, var_60); " #end_loop_str += "float accuracy = computeAccuracy2(labels, batch_size, var_60); "
end_loop_str += "final_accuracy += accuracy; \n" end_loop_str += "final_accuracy += accuracy; \n"
...@@ -114,8 +116,8 @@ class GraphCodeGen(object): ...@@ -114,8 +116,8 @@ class GraphCodeGen(object):
def codegen(self, weights_dir, test_data, test_labels): def codegen(self, weights_dir, test_data, test_labels):
if os.path.exists(weights_dir): if os.path.exists(weights_dir):
raise ValueError("Weight dir existed. Compilation interrupted!") raise ValueError("Weight dir existed. Compilation interrupted!")
os.mkdir(weights_dir) os.mkdir(weights_dir)
self.emit_header() self.emit_header()
self.emit_weights(weights_dir) self.emit_weights(weights_dir)
self.emit_batch_loop(test_data) self.emit_batch_loop(test_data)
self.emit_graph() self.emit_graph()
......
hpvm/projects/onnx/frontend/graph_ir.py
+ 61
− 20
View file @ 10a57c79
################################################ ################################################
# Top Level DFGNode interface # Top Level DFGNode interface
################################################ ################################################
class DFGNode(object): class DFGNode(object):
def add_output(self, output_node): def add_output(self, output_node):
self.outputs.append(output_node) self.outputs.append(output_node)
def add_input(self, input_node): def add_input(self, input_node):
self.inputs.append(input_node) self.inputs.append(input_node)
def __init__(self, layer): def __init__(self, layer):
self.inputs = [] self.inputs = []
self.outputs = [] self.outputs = []
self.name = layer.name self.name = layer.name
self.op_type = layer.op_type self.op_type = layer.op_type
''' '''
Element wise operatos that is for activation function Element wise operatos that is for activation function
e.g. HardSigmoid, LeakyRelu, PRelu, Pow, Reciprocal, e.g. HardSigmoid, LeakyRelu, PRelu, Pow, Reciprocal,
Relu, Selu, Sigmoid, Softplus, Sqrt, ThresholdedRelu, Relu, Selu, Sigmoid, Softplus, Sqrt, ThresholdedRelu,
Abs, Ceil, Elu, Floor, Neg Abs, Ceil, Elu, Floor, Neg
''' '''
class ActivationNode(DFGNode): class ActivationNode(DFGNode):
pass pass
''' '''
ELement wise operators that is not for activation function. ELement wise operators that is not for activation function.
In other words, they are logical comparison operators In other words, they are logical comparison operators
e.g. And, Equal, Greater, GreaterOrEqual, Less, LessOrEqual, e.g. And, Equal, Greater, GreaterOrEqual, Less, LessOrEqual,
Or, Xor Or, Xor
''' '''
class LogicalOpNode(DFGNode): class LogicalOpNode(DFGNode):
pass pass
class Node(object): class Node(object):
def __init__(self, name, shape, dtype): def __init__(self, name, shape, dtype):
self._name = name self._name = name
self._shape = shape if shape else {} self._shape = shape if shape else {}
self._dtype = dtype self._dtype = dtype
def __str__(self): def __str__(self):
return "Node: " + self._name + " with shape: " + str(self._shape) + " and data type " + str(self._dtype) return "Node: " + self._name + " with shape: " + \
str(self._shape) + " and data type " + str(self._dtype)
__repr__ = __str__ __repr__ = __str__
################################################ ################################################
# Actually Implementation of Operators # Actually Implementation of Operators
################################################ ################################################
class AddNode(DFGNode): class AddNode(DFGNode):
def __init__(self, layer): def __init__(self, layer):
DFGNode.__init__(self, layer) DFGNode.__init__(self, layer)
def codegen(self): def codegen(self):
input_vars = self.getMultipleInputNames(cur_node) input_vars = self.getMultipleInputNames(cur_node)
inst_str = "void* " + out_var_name1 + " = " inst_str = "void* " + out_var_name1 + " = "
inst_str += "tensorAdd(" + input_vars[0] + ", " + input_vars[1] + "); \n" inst_str += "tensorAdd(" + \
input_vars[0] + ", " + input_vars[1] + "); \n"
return inst_str return inst_str
class MatMulNode(DFGNode): class MatMulNode(DFGNode):
def __init__(self, layer): def __init__(self, layer):
DFGNode.__init__(self, layer) DFGNode.__init__(self, layer)
def codegen(self): def codegen(self):
pass pass
class SoftMaxNode(DFGNode): class SoftMaxNode(DFGNode):
def __init__(self, layer): def __init__(self, layer):
DFGNode.__init__(self, layer) DFGNode.__init__(self, layer)
def codegen(self): def codegen(self):
pass pass
class Conv2DNode(DFGNode): class Conv2DNode(DFGNode):
def __init__(self, layer): def __init__(self, layer):
DFGNode.__init__(self, layer) DFGNode.__init__(self, layer)
...@@ -73,6 +95,7 @@ class Conv2DNode(DFGNode): ...@@ -73,6 +95,7 @@ class Conv2DNode(DFGNode):
self.strides = layer.strides self.strides = layer.strides
print("\t", self.strides) print("\t", self.strides)
print("\tPadding = ", self.padding) print("\tPadding = ", self.padding)
def codegen(self): def codegen(self):
input_var_name = self.getSingleInputName(cur_node) input_var_name = self.getSingleInputName(cur_node)
weights = cur_node.weights weights = cur_node.weights
...@@ -81,12 +104,12 @@ class Conv2DNode(DFGNode): ...@@ -81,12 +104,12 @@ class Conv2DNode(DFGNode):
if cur_node.padding.strip() == "valid": if cur_node.padding.strip() == "valid":
padding = 0 padding = 0
else: else:
padding = cur_node.padding padding = cur_node.padding
padding = int((weights.shape[0] - 1) / 2) padding = int((weights.shape[0] - 1) / 2)
prev_padding = self.getPrevLayerPadding(cur_node) prev_padding = self.getPrevLayerPadding(cur_node)
if prev_padding != None: if prev_padding is not None:
# FIXME: currently only supporting symmetric padding # FIXME: currently only supporting symmetric padding
padding = prev_padding[0][0] padding = prev_padding[0][0]
inst_str = "void* " + out_var_name1 + " = " inst_str = "void* " + out_var_name1 + " = "
inst_str += "tensorConvolution(" + input_var_name + ", " inst_str += "tensorConvolution(" + input_var_name + ", "
...@@ -102,11 +125,13 @@ class Conv2DNode(DFGNode): ...@@ -102,11 +125,13 @@ class Conv2DNode(DFGNode):
else: else:
inst_str += "1); \n" inst_str += "1); \n"
if strides[0] > 1 and cur_node.padding.strip() == "same": if strides[0] > 1 and cur_node.padding.strip() == "same":
print ("!ERROR: Same Padding not supported for Conv with Stride > 1") print("!ERROR: Same Padding not supported for Conv with Stride > 1")
print ("Use: ZeroPadding2D(padding=(" + str(padding) + "," + str(padding) + "));\n"); print("Use: ZeroPadding2D(padding=(" +
str(padding) + "," + str(padding) + "));\n")
sys.exit(0) sys.exit(0)
return inst_str return inst_str
class DepthwiseConv2DNode(DFGNode): class DepthwiseConv2DNode(DFGNode):
def __init__(self, layer): def __init__(self, layer):
DFGNode.__init__(self, layer) DFGNode.__init__(self, layer)
...@@ -117,6 +142,7 @@ class DepthwiseConv2DNode(DFGNode): ...@@ -117,6 +142,7 @@ class DepthwiseConv2DNode(DFGNode):
self.strides = layer.strides self.strides = layer.strides
print("\t", self.strides) print("\t", self.strides)
print("\tPadding = ", self.padding) print("\tPadding = ", self.padding)
def codegen(self): def codegen(self):
input_var_name = self.getSingleInputName(cur_node) input_var_name = self.getSingleInputName(cur_node)
weights = cur_node.weights weights = cur_node.weights
...@@ -125,12 +151,12 @@ class DepthwiseConv2DNode(DFGNode): ...@@ -125,12 +151,12 @@ class DepthwiseConv2DNode(DFGNode):
if cur_node.padding.strip() == "valid": if cur_node.padding.strip() == "valid":
padding = 0 padding = 0
else: else:
padding = cur_node.padding padding = cur_node.padding
padding = int((weights.shape[0] - 1) / 2) padding = int((weights.shape[0] - 1) / 2)
prev_padding = self.getPrevLayerPadding(cur_node) prev_padding = self.getPrevLayerPadding(cur_node)
if prev_padding != None: if prev_padding is not None:
# FIXME: currently only supporting symmetric padding # FIXME: currently only supporting symmetric padding
padding = prev_padding[0][0] padding = prev_padding[0][0]
inst_str = "void* " + out_var_name1 + " = " inst_str = "void* " + out_var_name1 + " = "
inst_str += "tensorConvolution(" + input_var_name + ", " inst_str += "tensorConvolution(" + input_var_name + ", "
...@@ -146,17 +172,20 @@ class DepthwiseConv2DNode(DFGNode): ...@@ -146,17 +172,20 @@ class DepthwiseConv2DNode(DFGNode):
else: else:
inst_str += "1); \n" inst_str += "1); \n"
if strides[0] > 1 and cur_node.padding.strip() == "same": if strides[0] > 1 and cur_node.padding.strip() == "same":
print ("!ERROR: Same Padding not supported for Conv with Stride > 1") print("!ERROR: Same Padding not supported for Conv with Stride > 1")
print ("Use: ZeroPadding2D(padding=(" + str(padding) + "," + str(padding) + "));\n"); print("Use: ZeroPadding2D(padding=(" +
str(padding) + "," + str(padding) + "));\n")
sys.exit(0) sys.exit(0)
return inst_str return inst_str
class DenseNode(DFGNode): class DenseNode(DFGNode):
def __init__(self, layer): def __init__(self, layer):
DFGNode.__init__(self, layer) DFGNode.__init__(self, layer)
self.weights = layer.get_weights()[0] self.weights = layer.get_weights()[0]
print("\t", self.weights.shape) print("\t", self.weights.shape)
self.use_bias = layer.use_bias self.use_bias = layer.use_bias
def codegen(self): def codegen(self):
input_var_name = self.getSingleInputName(cur_node) input_var_name = self.getSingleInputName(cur_node)
weights = cur_node.weights weights = cur_node.weights
...@@ -166,6 +195,7 @@ class DenseNode(DFGNode): ...@@ -166,6 +195,7 @@ class DenseNode(DFGNode):
inst_str += "); \n" inst_str += "); \n"
return inst_str return inst_str
class MaxPool2DNode(DFGNode): class MaxPool2DNode(DFGNode):
def __init__(self, layer): def __init__(self, layer):
DFGNode.__init__(self, layer) DFGNode.__init__(self, layer)
...@@ -173,12 +203,16 @@ class MaxPool2DNode(DFGNode): ...@@ -173,12 +203,16 @@ class MaxPool2DNode(DFGNode):
self.strides = layer.strides self.strides = layer.strides
print("\t pool_size = ", self.pool_size) print("\t pool_size = ", self.pool_size)
print("\t strides = ", self.strides) print("\t strides = ", self.strides)
def codegen(self): def codegen(self):
inst_str = "void* " + out_var_name1 + " = " inst_str = "void* " + out_var_name1 + " = "
inst_str += "tensorPooling(" + input_var_name + "," + pool_type + "," + str(pool_size[0]) + "," + str(pool_size[1]) inst_str += "tensorPooling(" + input_var_name + "," + \
inst_str += "," + str(padding) + "," + str(padding) + "," + str(strides[0]) + "," + str(strides[1]) 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" inst_str += "); \n"
class AveragePooling2DNode(DFGNode): class AveragePooling2DNode(DFGNode):
def __init__(self, layer): def __init__(self, layer):
DFGNode.__init__(self, layer) DFGNode.__init__(self, layer)
...@@ -186,22 +220,28 @@ class AveragePooling2DNode(DFGNode): ...@@ -186,22 +220,28 @@ class AveragePooling2DNode(DFGNode):
self.strides = layer.strides self.strides = layer.strides
print("\t pool_size = ", self.pool_size) print("\t pool_size = ", self.pool_size)
print("\t strides = ", self.strides) print("\t strides = ", self.strides)
def codegen(self): def codegen(self):
inst_str = "void* " + out_var_name1 + " = " inst_str = "void* " + out_var_name1 + " = "
inst_str += "tensorPooling(" + input_var_name + "," + pool_type + "," + str(pool_size[0]) + "," + str(pool_size[1]) inst_str += "tensorPooling(" + input_var_name + "," + \
inst_str += "," + str(padding) + "," + str(padding) + "," + str(strides[0]) + "," + str(strides[1]) 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" inst_str += "); \n"
return insr_str return insr_str
class ZeroPadding2DNode(DFGNode): class ZeroPadding2DNode(DFGNode):
def __init__(self, layer): def __init__(self, layer):
DFGNode.__init__(self, layer) DFGNode.__init__(self, layer)
print ("***ZeroPaddding \n"); print("***ZeroPaddding \n")
self.padding = layer.padding self.padding = layer.padding
print ("padding = ", self.padding) print("padding = ", self.padding)
def codegen(self): def codegen(self):
pass pass
class BatchNormalizationNode(DFGNode): class BatchNormalizationNode(DFGNode):
def __init__(self, layer): def __init__(self, layer):
DFGNode.__init__(self, layer) DFGNode.__init__(self, layer)
...@@ -210,6 +250,7 @@ class BatchNormalizationNode(DFGNode): ...@@ -210,6 +250,7 @@ class BatchNormalizationNode(DFGNode):
self.gamma = layer.gamma self.gamma = layer.gamma
self.moving_mean = layer.moving_mean self.moving_mean = layer.moving_mean
self.moving_variance = layer.moving_variance self.moving_variance = layer.moving_variance
def codegen(self): def codegen(self):
input_var_name = self.getSingleInputName(cur_node) input_var_name = self.getSingleInputName(cur_node)
inst_str = "void* " + out_var_name1 + " = " inst_str = "void* " + out_var_name1 + " = "
......
hpvm/projects/onnx/frontend/main.py
+ 11
− 11
View file @ 10a57c79
...@@ -4,42 +4,42 @@ import numpy as np ...@@ -4,42 +4,42 @@ import numpy as np
import onnx import onnx
import glob import glob
from onnx import numpy_helper, version_converter from onnx import version_converter
from onnxruntime.backend.backend import OnnxRuntimeBackend as backend #from onnxruntime.backend.backend import OnnxRuntimeBackend as backend
# onnx2hpvm modules # onnx2hpvm modules
from graph_builder import GraphBuilder from graph_builder import GraphBuilder
from graph_codegen import GraphCodeGen from graph_codegen import GraphCodeGen
# from approx_codegen import GraphCodeGen # from approx_codegen import GraphCodeGen
def convert_version(model):
def convert_version(model, new_version):
print('The model before conversion:\n{}'.format(model)) print('The model before conversion:\n{}'.format(model))
# A full list of supported adapters can be found here: # A full list of supported adapters can be found here:
# https://github.com/onnx/onnx/blob/master/onnx/version_converter.py#L21 # https://github.com/onnx/onnx/blob/master/onnx/version_converter.py#L21
# Apply the version conversion on the original model # Apply the version conversion on the original model
converted_model = version_converter.convert_version(model, 12) converted_model = version_converter.convert_version(model, new_version)
print('The model after conversion:\n{}'.format(converted_model)) print('The model after conversion:\n{}'.format(converted_model))
return converted_model return converted_model
def main(): def main():
model = onnx.load('../models/keras/alexnet.onnx') model = onnx.load('../models/keras/lenet.onnx')
weights_dir = './test_src' weights_dir = './test_src'
test_data_dir = '../models/mnist/test_data_set_0' # test_data_dir = '../models/mnist/test_data_set_0'
# converted_model = convert_version(model) # converted_model = convert_version(model)
graph = model.graph
try: try:
opset = model.opset_import[0].version if model.opset_import else 1 opset = model.opset_import[0].version if model.opset_import else 1
except AttributeError: except AttributeError:
opset = 1 # default opset version set to 1 if not specified opset = 1 # default opset version set to 1 if not specified
print("opset version: ", opset) print("opset version: ", opset)
gBuilder = GraphBuilder(model, None, "float32", opset) gBuilder = GraphBuilder(model, None, "float32", opset, weights_dir)
gCodegen = GraphCodeGen(gBuilder.build_graph()) gBuilder.build_graph()
gCodegen.codegen(weights_dir, test_data)#, test_labels) #gCodegen = GraphCodeGen(gBuilder.build_graph())
# gCodegen.codegen(weights_dir, test_data)#, test_labels)
if __name__ == "__main__": if __name__ == "__main__":
......
hpvm/projects/onnx/frontend/weight_utils.py 0 → 100644
+ 147
− 0
View file @ 10a57c79
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])
label_val = np.int32(label[0])
else:
#label_val = np.int8(label)
label_val = np.int32(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 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")
X_test.tofile(f)
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)
hpvm/projects/onnx/keras_ref/generated/vgg16/approxhpvm_src.cc deleted 100644 → 0
+ 0
− 982
View file @ 6491cfe3
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/stat.h>
#include <cstring>
#include <visc.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
void var_0_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_1_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_2_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_3_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_4_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_5_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_6_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
__visc__return(2, r, (size_t) 0);
}
void var_7_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_8_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_9_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_10_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_11_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_12_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_13_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
__visc__return(2, r, (size_t) 0);
}
void var_14_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_15_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_16_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_17_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_18_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_19_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_20_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_21_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_22_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_23_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
__visc__return(2, r, (size_t) 0);
}
void var_24_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_25_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_26_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_27_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_28_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_29_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_30_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_31_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_32_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_33_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
__visc__return(2, r, (size_t) 0);
}
void var_34_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_35_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_36_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_37_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_38_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_39_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_40_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
__visc__return(2, r, (size_t) 0);
}
void var_41_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_42_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_43_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
__visc__return(2, r, (size_t) 0);
}
void var_44_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_mul(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_45_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_46_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_relu(t1);
__visc__return(2, r, (size_t) 0);
}
void var_47_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_mul(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_48_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
__visc__return(2, r, (size_t) 0);
}
void var_49_node(void* t1, size_t bytes_t1) {
__visc__hint(visc::CUDNN_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_softmax(t1);
__visc__return(2, r, (size_t) 0);
}
void root(void* input, size_t input_bytes,
void* conv2d_1_w, size_t conv2d_1_w_bytes,
void* conv2d_1_b, size_t conv2d_1_b_bytes,
void* conv2d_2_w, size_t conv2d_2_w_bytes,
void* conv2d_2_b, size_t conv2d_2_b_bytes,
void* conv2d_3_w, size_t conv2d_3_w_bytes,
void* conv2d_3_b, size_t conv2d_3_b_bytes,
void* conv2d_4_w, size_t conv2d_4_w_bytes,
void* conv2d_4_b, size_t conv2d_4_b_bytes,
void* conv2d_5_w, size_t conv2d_5_w_bytes,
void* conv2d_5_b, size_t conv2d_5_b_bytes,
void* conv2d_6_w, size_t conv2d_6_w_bytes,
void* conv2d_6_b, size_t conv2d_6_b_bytes,
void* conv2d_7_w, size_t conv2d_7_w_bytes,
void* conv2d_7_b, size_t conv2d_7_b_bytes,
void* conv2d_8_w, size_t conv2d_8_w_bytes,
void* conv2d_8_b, size_t conv2d_8_b_bytes,
void* conv2d_9_w, size_t conv2d_9_w_bytes,
void* conv2d_9_b, size_t conv2d_9_b_bytes,
void* conv2d_10_w, size_t conv2d_10_w_bytes,
void* conv2d_10_b, size_t conv2d_10_b_bytes,
void* conv2d_11_w, size_t conv2d_11_w_bytes,
void* conv2d_11_b, size_t conv2d_11_b_bytes,
void* conv2d_12_w, size_t conv2d_12_w_bytes,
void* conv2d_12_b, size_t conv2d_12_b_bytes,
void* conv2d_13_w, size_t conv2d_13_w_bytes,
void* conv2d_13_b, size_t conv2d_13_b_bytes,
void* dense_1_w, size_t dense_1_w_bytes,
void* dense_1_b, size_t dense_1_b_bytes,
void* dense_2_w, size_t dense_2_w_bytes,
void* dense_2_b, size_t dense_2_b_bytes){
__visc__hint(visc::CPU_TARGET);
__visc__attributes(31, input, conv2d_1_w, conv2d_1_b, conv2d_2_w, conv2d_2_b, conv2d_3_w, conv2d_3_b, conv2d_4_w, conv2d_4_b, conv2d_5_w, conv2d_5_b, conv2d_6_w, conv2d_6_b, conv2d_7_w, conv2d_7_b, conv2d_8_w, conv2d_8_b, conv2d_9_w, conv2d_9_b, conv2d_10_w, conv2d_10_b, conv2d_11_w, conv2d_11_b, conv2d_12_w, conv2d_12_b, conv2d_13_w, conv2d_13_b, dense_1_w, dense_1_b, dense_2_w, dense_2_b, 0);
void* var_0 = __visc__createNodeND(0, var_0_node);
__visc__bindIn(var_0, 0, 0, 0);
__visc__bindIn(var_0, 1, 1, 0);
__visc__bindIn(var_0, 2, 2, 0);
__visc__bindIn(var_0, 3, 3, 0);
void* var_1 = __visc__createNodeND(0, var_1_node);
__visc__edge(var_0, var_1, 1, 0, 0, 0);
__visc__edge(var_0, var_1, 1, 1, 1, 0);
__visc__bindIn(var_1, 4, 2, 0);
__visc__bindIn(var_1, 5, 3, 0);
void* var_2 = __visc__createNodeND(0, var_2_node);
__visc__edge(var_1, var_2, 1, 0, 0, 0);
__visc__edge(var_1, var_2, 1, 1, 1, 0);
void* var_3 = __visc__createNodeND(0, var_3_node);
__visc__edge(var_2, var_3, 1, 0, 0, 0);
__visc__edge(var_2, var_3, 1, 1, 1, 0);
__visc__bindIn(var_3, 6, 2, 0);
__visc__bindIn(var_3, 7, 3, 0);
void* var_4 = __visc__createNodeND(0, var_4_node);
__visc__edge(var_3, var_4, 1, 0, 0, 0);
__visc__edge(var_3, var_4, 1, 1, 1, 0);
__visc__bindIn(var_4, 8, 2, 0);
__visc__bindIn(var_4, 9, 3, 0);
void* var_5 = __visc__createNodeND(0, var_5_node);
__visc__edge(var_4, var_5, 1, 0, 0, 0);
__visc__edge(var_4, var_5, 1, 1, 1, 0);
void* var_6 = __visc__createNodeND(0, var_6_node);
__visc__edge(var_5, var_6, 1, 0, 0, 0);
__visc__edge(var_5, var_6, 1, 1, 1, 0);
void* var_7 = __visc__createNodeND(0, var_7_node);
__visc__edge(var_6, var_7, 1, 0, 0, 0);
__visc__edge(var_6, var_7, 1, 1, 1, 0);
__visc__bindIn(var_7, 10, 2, 0);
__visc__bindIn(var_7, 11, 3, 0);
void* var_8 = __visc__createNodeND(0, var_8_node);
__visc__edge(var_7, var_8, 1, 0, 0, 0);
__visc__edge(var_7, var_8, 1, 1, 1, 0);
__visc__bindIn(var_8, 12, 2, 0);
__visc__bindIn(var_8, 13, 3, 0);
void* var_9 = __visc__createNodeND(0, var_9_node);
__visc__edge(var_8, var_9, 1, 0, 0, 0);
__visc__edge(var_8, var_9, 1, 1, 1, 0);
void* var_10 = __visc__createNodeND(0, var_10_node);
__visc__edge(var_9, var_10, 1, 0, 0, 0);
__visc__edge(var_9, var_10, 1, 1, 1, 0);
__visc__bindIn(var_10, 14, 2, 0);
__visc__bindIn(var_10, 15, 3, 0);
void* var_11 = __visc__createNodeND(0, var_11_node);
__visc__edge(var_10, var_11, 1, 0, 0, 0);
__visc__edge(var_10, var_11, 1, 1, 1, 0);
__visc__bindIn(var_11, 16, 2, 0);
__visc__bindIn(var_11, 17, 3, 0);
void* var_12 = __visc__createNodeND(0, var_12_node);
__visc__edge(var_11, var_12, 1, 0, 0, 0);
__visc__edge(var_11, var_12, 1, 1, 1, 0);
void* var_13 = __visc__createNodeND(0, var_13_node);
__visc__edge(var_12, var_13, 1, 0, 0, 0);
__visc__edge(var_12, var_13, 1, 1, 1, 0);
void* var_14 = __visc__createNodeND(0, var_14_node);
__visc__edge(var_13, var_14, 1, 0, 0, 0);
__visc__edge(var_13, var_14, 1, 1, 1, 0);
__visc__bindIn(var_14, 18, 2, 0);
__visc__bindIn(var_14, 19, 3, 0);
void* var_15 = __visc__createNodeND(0, var_15_node);
__visc__edge(var_14, var_15, 1, 0, 0, 0);
__visc__edge(var_14, var_15, 1, 1, 1, 0);
__visc__bindIn(var_15, 20, 2, 0);
__visc__bindIn(var_15, 21, 3, 0);
void* var_16 = __visc__createNodeND(0, var_16_node);
__visc__edge(var_15, var_16, 1, 0, 0, 0);
__visc__edge(var_15, var_16, 1, 1, 1, 0);
void* var_17 = __visc__createNodeND(0, var_17_node);
__visc__edge(var_16, var_17, 1, 0, 0, 0);
__visc__edge(var_16, var_17, 1, 1, 1, 0);
__visc__bindIn(var_17, 22, 2, 0);
__visc__bindIn(var_17, 23, 3, 0);
void* var_18 = __visc__createNodeND(0, var_18_node);
__visc__edge(var_17, var_18, 1, 0, 0, 0);
__visc__edge(var_17, var_18, 1, 1, 1, 0);
__visc__bindIn(var_18, 24, 2, 0);
__visc__bindIn(var_18, 25, 3, 0);
void* var_19 = __visc__createNodeND(0, var_19_node);
__visc__edge(var_18, var_19, 1, 0, 0, 0);
__visc__edge(var_18, var_19, 1, 1, 1, 0);
void* var_20 = __visc__createNodeND(0, var_20_node);
__visc__edge(var_19, var_20, 1, 0, 0, 0);
__visc__edge(var_19, var_20, 1, 1, 1, 0);
__visc__bindIn(var_20, 26, 2, 0);
__visc__bindIn(var_20, 27, 3, 0);
void* var_21 = __visc__createNodeND(0, var_21_node);
__visc__edge(var_20, var_21, 1, 0, 0, 0);
__visc__edge(var_20, var_21, 1, 1, 1, 0);
__visc__bindIn(var_21, 28, 2, 0);
__visc__bindIn(var_21, 29, 3, 0);
void* var_22 = __visc__createNodeND(0, var_22_node);
__visc__edge(var_21, var_22, 1, 0, 0, 0);
__visc__edge(var_21, var_22, 1, 1, 1, 0);
void* var_23 = __visc__createNodeND(0, var_23_node);
__visc__edge(var_22, var_23, 1, 0, 0, 0);
__visc__edge(var_22, var_23, 1, 1, 1, 0);
void* var_24 = __visc__createNodeND(0, var_24_node);
__visc__edge(var_23, var_24, 1, 0, 0, 0);
__visc__edge(var_23, var_24, 1, 1, 1, 0);
__visc__bindIn(var_24, 30, 2, 0);
__visc__bindIn(var_24, 31, 3, 0);
void* var_25 = __visc__createNodeND(0, var_25_node);
__visc__edge(var_24, var_25, 1, 0, 0, 0);
__visc__edge(var_24, var_25, 1, 1, 1, 0);
__visc__bindIn(var_25, 32, 2, 0);
__visc__bindIn(var_25, 33, 3, 0);
void* var_26 = __visc__createNodeND(0, var_26_node);
__visc__edge(var_25, var_26, 1, 0, 0, 0);
__visc__edge(var_25, var_26, 1, 1, 1, 0);
void* var_27 = __visc__createNodeND(0, var_27_node);
__visc__edge(var_26, var_27, 1, 0, 0, 0);
__visc__edge(var_26, var_27, 1, 1, 1, 0);
__visc__bindIn(var_27, 34, 2, 0);
__visc__bindIn(var_27, 35, 3, 0);
void* var_28 = __visc__createNodeND(0, var_28_node);
__visc__edge(var_27, var_28, 1, 0, 0, 0);
__visc__edge(var_27, var_28, 1, 1, 1, 0);
__visc__bindIn(var_28, 36, 2, 0);
__visc__bindIn(var_28, 37, 3, 0);
void* var_29 = __visc__createNodeND(0, var_29_node);
__visc__edge(var_28, var_29, 1, 0, 0, 0);
__visc__edge(var_28, var_29, 1, 1, 1, 0);
void* var_30 = __visc__createNodeND(0, var_30_node);
__visc__edge(var_29, var_30, 1, 0, 0, 0);
__visc__edge(var_29, var_30, 1, 1, 1, 0);
__visc__bindIn(var_30, 38, 2, 0);
__visc__bindIn(var_30, 39, 3, 0);
void* var_31 = __visc__createNodeND(0, var_31_node);
__visc__edge(var_30, var_31, 1, 0, 0, 0);
__visc__edge(var_30, var_31, 1, 1, 1, 0);
__visc__bindIn(var_31, 40, 2, 0);
__visc__bindIn(var_31, 41, 3, 0);
void* var_32 = __visc__createNodeND(0, var_32_node);
__visc__edge(var_31, var_32, 1, 0, 0, 0);
__visc__edge(var_31, var_32, 1, 1, 1, 0);
void* var_33 = __visc__createNodeND(0, var_33_node);
__visc__edge(var_32, var_33, 1, 0, 0, 0);
__visc__edge(var_32, var_33, 1, 1, 1, 0);
void* var_34 = __visc__createNodeND(0, var_34_node);
__visc__edge(var_33, var_34, 1, 0, 0, 0);
__visc__edge(var_33, var_34, 1, 1, 1, 0);
__visc__bindIn(var_34, 42, 2, 0);
__visc__bindIn(var_34, 43, 3, 0);
void* var_35 = __visc__createNodeND(0, var_35_node);
__visc__edge(var_34, var_35, 1, 0, 0, 0);
__visc__edge(var_34, var_35, 1, 1, 1, 0);
__visc__bindIn(var_35, 44, 2, 0);
__visc__bindIn(var_35, 45, 3, 0);
void* var_36 = __visc__createNodeND(0, var_36_node);
__visc__edge(var_35, var_36, 1, 0, 0, 0);
__visc__edge(var_35, var_36, 1, 1, 1, 0);
void* var_37 = __visc__createNodeND(0, var_37_node);
__visc__edge(var_36, var_37, 1, 0, 0, 0);
__visc__edge(var_36, var_37, 1, 1, 1, 0);
__visc__bindIn(var_37, 46, 2, 0);
__visc__bindIn(var_37, 47, 3, 0);
void* var_38 = __visc__createNodeND(0, var_38_node);
__visc__edge(var_37, var_38, 1, 0, 0, 0);
__visc__edge(var_37, var_38, 1, 1, 1, 0);
__visc__bindIn(var_38, 48, 2, 0);
__visc__bindIn(var_38, 49, 3, 0);
void* var_39 = __visc__createNodeND(0, var_39_node);
__visc__edge(var_38, var_39, 1, 0, 0, 0);
__visc__edge(var_38, var_39, 1, 1, 1, 0);
void* var_40 = __visc__createNodeND(0, var_40_node);
__visc__edge(var_39, var_40, 1, 0, 0, 0);
__visc__edge(var_39, var_40, 1, 1, 1, 0);
__visc__bindIn(var_40, 50, 2, 0);
__visc__bindIn(var_40, 51, 3, 0);
void* var_41 = __visc__createNodeND(0, var_41_node);
__visc__edge(var_40, var_41, 1, 0, 0, 0);
__visc__edge(var_40, var_41, 1, 1, 1, 0);
__visc__bindIn(var_41, 52, 2, 0);
__visc__bindIn(var_41, 53, 3, 0);
void* var_42 = __visc__createNodeND(0, var_42_node);
__visc__edge(var_41, var_42, 1, 0, 0, 0);
__visc__edge(var_41, var_42, 1, 1, 1, 0);
void* var_43 = __visc__createNodeND(0, var_43_node);
__visc__edge(var_42, var_43, 1, 0, 0, 0);
__visc__edge(var_42, var_43, 1, 1, 1, 0);
void* var_44 = __visc__createNodeND(0, var_44_node);
__visc__edge(var_43, var_44, 1, 0, 0, 0);
__visc__edge(var_43, var_44, 1, 1, 1, 0);
__visc__bindIn(var_44, 54, 2, 0);
__visc__bindIn(var_44, 55, 3, 0);
void* var_45 = __visc__createNodeND(0, var_45_node);
__visc__edge(var_44, var_45, 1, 0, 0, 0);
__visc__edge(var_44, var_45, 1, 1, 1, 0);
__visc__bindIn(var_45, 56, 2, 0);
__visc__bindIn(var_45, 57, 3, 0);
void* var_46 = __visc__createNodeND(0, var_46_node);
__visc__edge(var_45, var_46, 1, 0, 0, 0);
__visc__edge(var_45, var_46, 1, 1, 1, 0);
void* var_47 = __visc__createNodeND(0, var_47_node);
__visc__edge(var_46, var_47, 1, 0, 0, 0);
__visc__edge(var_46, var_47, 1, 1, 1, 0);
__visc__bindIn(var_47, 58, 2, 0);
__visc__bindIn(var_47, 59, 3, 0);
void* var_48 = __visc__createNodeND(0, var_48_node);
__visc__edge(var_47, var_48, 1, 0, 0, 0);
__visc__edge(var_47, var_48, 1, 1, 1, 0);
__visc__bindIn(var_48, 60, 2, 0);
__visc__bindIn(var_48, 61, 3, 0);
void* var_49 = __visc__createNodeND(0, var_49_node);
__visc__edge(var_48, var_49, 1, 0, 0, 0);
__visc__edge(var_48, var_49, 1, 1, 1, 0);
__visc__bindOut(var_49, 0, 0, 0);
__visc__bindOut(var_49, 1, 1, 0);
}
struct ret_t {
void* tensor;
size_t bytes;
};
typedef struct __attribute__((__packed__)) {
void* input;
size_t input_bytes;
void* conv2d_1_w;
size_t conv2d_1_w_bytes;
void* conv2d_1_b;
size_t conv2d_1_b_bytes;
void* conv2d_2_w;
size_t conv2d_2_w_bytes;
void* conv2d_2_b;
size_t conv2d_2_b_bytes;
void* conv2d_3_w;
size_t conv2d_3_w_bytes;
void* conv2d_3_b;
size_t conv2d_3_b_bytes;
void* conv2d_4_w;
size_t conv2d_4_w_bytes;
void* conv2d_4_b;
size_t conv2d_4_b_bytes;
void* conv2d_5_w;
size_t conv2d_5_w_bytes;
void* conv2d_5_b;
size_t conv2d_5_b_bytes;
void* conv2d_6_w;
size_t conv2d_6_w_bytes;
void* conv2d_6_b;
size_t conv2d_6_b_bytes;
void* conv2d_7_w;
size_t conv2d_7_w_bytes;
void* conv2d_7_b;
size_t conv2d_7_b_bytes;
void* conv2d_8_w;
size_t conv2d_8_w_bytes;
void* conv2d_8_b;
size_t conv2d_8_b_bytes;
void* conv2d_9_w;
size_t conv2d_9_w_bytes;
void* conv2d_9_b;
size_t conv2d_9_b_bytes;
void* conv2d_10_w;
size_t conv2d_10_w_bytes;
void* conv2d_10_b;
size_t conv2d_10_b_bytes;
void* conv2d_11_w;
size_t conv2d_11_w_bytes;
void* conv2d_11_b;
size_t conv2d_11_b_bytes;
void* conv2d_12_w;
size_t conv2d_12_w_bytes;
void* conv2d_12_b;
size_t conv2d_12_b_bytes;
void* conv2d_13_w;
size_t conv2d_13_w_bytes;
void* conv2d_13_b;
size_t conv2d_13_b_bytes;
void* dense_1_w;
size_t dense_1_w_bytes;
void* dense_1_b;
size_t dense_1_b_bytes;
void* dense_2_w;
size_t dense_2_w_bytes;
void* dense_2_b;
size_t dense_2_b_bytes;
struct ret_t r;
}
RootIn;
int main(){
std::string dir_prefix = std::string("data/vgg16_cifar10/");
std::string input_path = dir_prefix + std::string("input.bin");
std::string labels_path = dir_prefix + std::string("labels.bin");
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
void* conv2d_1_w = readTrainedWeights(conv2d_1_w_path.c_str(), 0,64,3,3,3);
std::string conv2d_1_b_path = dir_prefix + std::string("conv2d_1_b.bin");
void* conv2d_1_b = readTrainedWeights(conv2d_1_b_path.c_str(), 0,1,64,1,1);
std::string conv2d_2_w_path = dir_prefix + std::string("conv2d_2_w.bin");
void* conv2d_2_w = readTrainedWeights(conv2d_2_w_path.c_str(), 0,64,64,3,3);
std::string conv2d_2_b_path = dir_prefix + std::string("conv2d_2_b.bin");
void* conv2d_2_b = readTrainedWeights(conv2d_2_b_path.c_str(), 0,1,64,1,1);
std::string conv2d_3_w_path = dir_prefix + std::string("conv2d_3_w.bin");
void* conv2d_3_w = readTrainedWeights(conv2d_3_w_path.c_str(), 0,128,64,3,3);
std::string conv2d_3_b_path = dir_prefix + std::string("conv2d_3_b.bin");
void* conv2d_3_b = readTrainedWeights(conv2d_3_b_path.c_str(), 0,1,128,1,1);
std::string conv2d_4_w_path = dir_prefix + std::string("conv2d_4_w.bin");
void* conv2d_4_w = readTrainedWeights(conv2d_4_w_path.c_str(), 0,128,128,3,3);
std::string conv2d_4_b_path = dir_prefix + std::string("conv2d_4_b.bin");
void* conv2d_4_b = readTrainedWeights(conv2d_4_b_path.c_str(), 0,1,128,1,1);
std::string conv2d_5_w_path = dir_prefix + std::string("conv2d_5_w.bin");
void* conv2d_5_w = readTrainedWeights(conv2d_5_w_path.c_str(), 0,256,128,3,3);
std::string conv2d_5_b_path = dir_prefix + std::string("conv2d_5_b.bin");
void* conv2d_5_b = readTrainedWeights(conv2d_5_b_path.c_str(), 0,1,256,1,1);
std::string conv2d_6_w_path = dir_prefix + std::string("conv2d_6_w.bin");
void* conv2d_6_w = readTrainedWeights(conv2d_6_w_path.c_str(), 0,256,256,3,3);
std::string conv2d_6_b_path = dir_prefix + std::string("conv2d_6_b.bin");
void* conv2d_6_b = readTrainedWeights(conv2d_6_b_path.c_str(), 0,1,256,1,1);
std::string conv2d_7_w_path = dir_prefix + std::string("conv2d_7_w.bin");
void* conv2d_7_w = readTrainedWeights(conv2d_7_w_path.c_str(), 0,256,256,3,3);
std::string conv2d_7_b_path = dir_prefix + std::string("conv2d_7_b.bin");
void* conv2d_7_b = readTrainedWeights(conv2d_7_b_path.c_str(), 0,1,256,1,1);
std::string conv2d_8_w_path = dir_prefix + std::string("conv2d_8_w.bin");
void* conv2d_8_w = readTrainedWeights(conv2d_8_w_path.c_str(), 0,512,256,3,3);
std::string conv2d_8_b_path = dir_prefix + std::string("conv2d_8_b.bin");
void* conv2d_8_b = readTrainedWeights(conv2d_8_b_path.c_str(), 0,1,512,1,1);
std::string conv2d_9_w_path = dir_prefix + std::string("conv2d_9_w.bin");
void* conv2d_9_w = readTrainedWeights(conv2d_9_w_path.c_str(), 0,512,512,3,3);
std::string conv2d_9_b_path = dir_prefix + std::string("conv2d_9_b.bin");
void* conv2d_9_b = readTrainedWeights(conv2d_9_b_path.c_str(), 0,1,512,1,1);
std::string conv2d_10_w_path = dir_prefix + std::string("conv2d_10_w.bin");
void* conv2d_10_w = readTrainedWeights(conv2d_10_w_path.c_str(), 0,512,512,3,3);
std::string conv2d_10_b_path = dir_prefix + std::string("conv2d_10_b.bin");
void* conv2d_10_b = readTrainedWeights(conv2d_10_b_path.c_str(), 0,1,512,1,1);
std::string conv2d_11_w_path = dir_prefix + std::string("conv2d_11_w.bin");
void* conv2d_11_w = readTrainedWeights(conv2d_11_w_path.c_str(), 0,512,512,3,3);
std::string conv2d_11_b_path = dir_prefix + std::string("conv2d_11_b.bin");
void* conv2d_11_b = readTrainedWeights(conv2d_11_b_path.c_str(), 0,1,512,1,1);
std::string conv2d_12_w_path = dir_prefix + std::string("conv2d_12_w.bin");
void* conv2d_12_w = readTrainedWeights(conv2d_12_w_path.c_str(), 0,512,512,3,3);
std::string conv2d_12_b_path = dir_prefix + std::string("conv2d_12_b.bin");
void* conv2d_12_b = readTrainedWeights(conv2d_12_b_path.c_str(), 0,1,512,1,1);
std::string conv2d_13_w_path = dir_prefix + std::string("conv2d_13_w.bin");
void* conv2d_13_w = readTrainedWeights(conv2d_13_w_path.c_str(), 0,512,512,3,3);
std::string conv2d_13_b_path = dir_prefix + std::string("conv2d_13_b.bin");
void* conv2d_13_b = readTrainedWeights(conv2d_13_b_path.c_str(), 0,1,512,1,1);
std::string dense_1_w_path = dir_prefix + std::string("dense_1_w.bin");
void* dense_1_w = readTrainedWeights(dense_1_w_path.c_str(), 0,1,1,512,512);
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void* dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0,1,512,1,1);
std::string dense_2_w_path = dir_prefix + std::string("dense_2_w.bin");
void* dense_2_w = readTrainedWeights(dense_2_w_path.c_str(), 0,1,1,512,10);
std::string dense_2_b_path = dir_prefix + std::string("dense_2_b.bin");
void* dense_2_b = readTrainedWeights(dense_2_b_path.c_str(), 0,1,10,1,1);
void* input = readTrainedWeights(input_path.c_str(), 0,10000,3,32,32);
uint8_t* labels = readLabels(labels_path.c_str(),10000);
__visc__init();
RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn)));
args->input = input;
args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
args->conv2d_1_b_bytes = 0;
args->conv2d_2_w = conv2d_2_w;
args->conv2d_2_w_bytes = 0;
args->conv2d_2_b = conv2d_2_b;
args->conv2d_2_b_bytes = 0;
args->conv2d_3_w = conv2d_3_w;
args->conv2d_3_w_bytes = 0;
args->conv2d_3_b = conv2d_3_b;
args->conv2d_3_b_bytes = 0;
args->conv2d_4_w = conv2d_4_w;
args->conv2d_4_w_bytes = 0;
args->conv2d_4_b = conv2d_4_b;
args->conv2d_4_b_bytes = 0;
args->conv2d_5_w = conv2d_5_w;
args->conv2d_5_w_bytes = 0;
args->conv2d_5_b = conv2d_5_b;
args->conv2d_5_b_bytes = 0;
args->conv2d_6_w = conv2d_6_w;
args->conv2d_6_w_bytes = 0;
args->conv2d_6_b = conv2d_6_b;
args->conv2d_6_b_bytes = 0;
args->conv2d_7_w = conv2d_7_w;
args->conv2d_7_w_bytes = 0;
args->conv2d_7_b = conv2d_7_b;
args->conv2d_7_b_bytes = 0;
args->conv2d_8_w = conv2d_8_w;
args->conv2d_8_w_bytes = 0;
args->conv2d_8_b = conv2d_8_b;
args->conv2d_8_b_bytes = 0;
args->conv2d_9_w = conv2d_9_w;
args->conv2d_9_w_bytes = 0;
args->conv2d_9_b = conv2d_9_b;
args->conv2d_9_b_bytes = 0;
args->conv2d_10_w = conv2d_10_w;
args->conv2d_10_w_bytes = 0;
args->conv2d_10_b = conv2d_10_b;
args->conv2d_10_b_bytes = 0;
args->conv2d_11_w = conv2d_11_w;
args->conv2d_11_w_bytes = 0;
args->conv2d_11_b = conv2d_11_b;
args->conv2d_11_b_bytes = 0;
args->conv2d_12_w = conv2d_12_w;
args->conv2d_12_w_bytes = 0;
args->conv2d_12_b = conv2d_12_b;
args->conv2d_12_b_bytes = 0;
args->conv2d_13_w = conv2d_13_w;
args->conv2d_13_w_bytes = 0;
args->conv2d_13_b = conv2d_13_b;
args->conv2d_13_b_bytes = 0;
args->dense_1_w = dense_1_w;
args->dense_1_w_bytes = 0;
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
args->dense_2_w = dense_2_w;
args->dense_2_w_bytes = 0;
args->dense_2_b = dense_2_b;
args->dense_2_b_bytes = 0;
void* dfg = __visc__launch(0, root, (void*) args);
__visc__wait(dfg);
void *result = static_cast<RootIn*>(args)->input;
hpvm_request_tensor(result, 0);
__visc__cleanup();
computeAccuracy2(labels, 10000, result);
return 0;
}
hpvm/projects/onnx/keras_ref/generated/vgg16/src.cc deleted 100644 → 0
+ 0
− 164
View file @ 6491cfe3
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <string.h>
#include "../../tensor_runtime/include/tensor_runtime.h"
#include "../include/utils.h"
int main(){
llvm_hpvm_initTensorRt(0);
std::string dir_prefix = std::string("data/vgg16_cifar10/");
std::string input_path = dir_prefix + std::string("input.bin");
std::string labels_path = dir_prefix + std::string("labels.bin");
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
void* conv2d_1_w = readTrainedWeights(conv2d_1_w_path.c_str(), 0,64,3,3,3);
std::string conv2d_1_b_path = dir_prefix + std::string("conv2d_1_b.bin");
void* conv2d_1_b = readTrainedWeights(conv2d_1_b_path.c_str(), 0,1,64,1,1);
std::string conv2d_2_w_path = dir_prefix + std::string("conv2d_2_w.bin");
void* conv2d_2_w = readTrainedWeights(conv2d_2_w_path.c_str(), 0,64,64,3,3);
std::string conv2d_2_b_path = dir_prefix + std::string("conv2d_2_b.bin");
void* conv2d_2_b = readTrainedWeights(conv2d_2_b_path.c_str(), 0,1,64,1,1);
std::string conv2d_3_w_path = dir_prefix + std::string("conv2d_3_w.bin");
void* conv2d_3_w = readTrainedWeights(conv2d_3_w_path.c_str(), 0,128,64,3,3);
std::string conv2d_3_b_path = dir_prefix + std::string("conv2d_3_b.bin");
void* conv2d_3_b = readTrainedWeights(conv2d_3_b_path.c_str(), 0,1,128,1,1);
std::string conv2d_4_w_path = dir_prefix + std::string("conv2d_4_w.bin");
void* conv2d_4_w = readTrainedWeights(conv2d_4_w_path.c_str(), 0,128,128,3,3);
std::string conv2d_4_b_path = dir_prefix + std::string("conv2d_4_b.bin");
void* conv2d_4_b = readTrainedWeights(conv2d_4_b_path.c_str(), 0,1,128,1,1);
std::string conv2d_5_w_path = dir_prefix + std::string("conv2d_5_w.bin");
void* conv2d_5_w = readTrainedWeights(conv2d_5_w_path.c_str(), 0,256,128,3,3);
std::string conv2d_5_b_path = dir_prefix + std::string("conv2d_5_b.bin");
void* conv2d_5_b = readTrainedWeights(conv2d_5_b_path.c_str(), 0,1,256,1,1);
std::string conv2d_6_w_path = dir_prefix + std::string("conv2d_6_w.bin");
void* conv2d_6_w = readTrainedWeights(conv2d_6_w_path.c_str(), 0,256,256,3,3);
std::string conv2d_6_b_path = dir_prefix + std::string("conv2d_6_b.bin");
void* conv2d_6_b = readTrainedWeights(conv2d_6_b_path.c_str(), 0,1,256,1,1);
std::string conv2d_7_w_path = dir_prefix + std::string("conv2d_7_w.bin");
void* conv2d_7_w = readTrainedWeights(conv2d_7_w_path.c_str(), 0,256,256,3,3);
std::string conv2d_7_b_path = dir_prefix + std::string("conv2d_7_b.bin");
void* conv2d_7_b = readTrainedWeights(conv2d_7_b_path.c_str(), 0,1,256,1,1);
std::string conv2d_8_w_path = dir_prefix + std::string("conv2d_8_w.bin");
void* conv2d_8_w = readTrainedWeights(conv2d_8_w_path.c_str(), 0,512,256,3,3);
std::string conv2d_8_b_path = dir_prefix + std::string("conv2d_8_b.bin");
void* conv2d_8_b = readTrainedWeights(conv2d_8_b_path.c_str(), 0,1,512,1,1);
std::string conv2d_9_w_path = dir_prefix + std::string("conv2d_9_w.bin");
void* conv2d_9_w = readTrainedWeights(conv2d_9_w_path.c_str(), 0,512,512,3,3);
std::string conv2d_9_b_path = dir_prefix + std::string("conv2d_9_b.bin");
void* conv2d_9_b = readTrainedWeights(conv2d_9_b_path.c_str(), 0,1,512,1,1);
std::string conv2d_10_w_path = dir_prefix + std::string("conv2d_10_w.bin");
void* conv2d_10_w = readTrainedWeights(conv2d_10_w_path.c_str(), 0,512,512,3,3);
std::string conv2d_10_b_path = dir_prefix + std::string("conv2d_10_b.bin");
void* conv2d_10_b = readTrainedWeights(conv2d_10_b_path.c_str(), 0,1,512,1,1);
std::string conv2d_11_w_path = dir_prefix + std::string("conv2d_11_w.bin");
void* conv2d_11_w = readTrainedWeights(conv2d_11_w_path.c_str(), 0,512,512,3,3);
std::string conv2d_11_b_path = dir_prefix + std::string("conv2d_11_b.bin");
void* conv2d_11_b = readTrainedWeights(conv2d_11_b_path.c_str(), 0,1,512,1,1);
std::string conv2d_12_w_path = dir_prefix + std::string("conv2d_12_w.bin");
void* conv2d_12_w = readTrainedWeights(conv2d_12_w_path.c_str(), 0,512,512,3,3);
std::string conv2d_12_b_path = dir_prefix + std::string("conv2d_12_b.bin");
void* conv2d_12_b = readTrainedWeights(conv2d_12_b_path.c_str(), 0,1,512,1,1);
std::string conv2d_13_w_path = dir_prefix + std::string("conv2d_13_w.bin");
void* conv2d_13_w = readTrainedWeights(conv2d_13_w_path.c_str(), 0,512,512,3,3);
std::string conv2d_13_b_path = dir_prefix + std::string("conv2d_13_b.bin");
void* conv2d_13_b = readTrainedWeights(conv2d_13_b_path.c_str(), 0,1,512,1,1);
std::string dense_1_w_path = dir_prefix + std::string("dense_1_w.bin");
void* dense_1_w = readTrainedWeights(dense_1_w_path.c_str(), 0,1,1,512,512);
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void* dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0,1,512,1,1);
std::string dense_2_w_path = dir_prefix + std::string("dense_2_w.bin");
void* dense_2_w = readTrainedWeights(dense_2_w_path.c_str(), 0,1,1,512,10);
std::string dense_2_b_path = dir_prefix + std::string("dense_2_b.bin");
void* dense_2_b = readTrainedWeights(dense_2_b_path.c_str(), 0,1,10,1,1);
startMemTracking();
int test_input_size = 10000;
int batch_size = 10000;
int batch_count = test_input_size / batch_size;
float final_accuracy = 0.0;
for(int i = 0; i < batch_count; i++){
int start = i * batch_size;
int end = (i + 1) * batch_size;
void* input = readInputBatch(input_path.c_str(),0,start,end,3,32,32);
void* var_0 = tensorConvolution(input, conv2d_1_w, 1, 1, 1, 1, 1, 0);
void* var_1 = tensorAdd(var_0, conv2d_1_b);
void* var_2 = tensorRelu(var_1);
void* var_4 = tensorConvolution(var_2, conv2d_2_w, 1, 1, 1, 1, 1, 0);
void* var_5 = tensorAdd(var_4, conv2d_2_b);
void* var_6 = tensorRelu(var_5);
void* var_7 = tensorPooling(var_6,0,2,2,0,0,2,2);
void* var_8 = tensorConvolution(var_7, conv2d_3_w, 1, 1, 1, 1, 1, 0);
void* var_9 = tensorAdd(var_8, conv2d_3_b);
void* var_10 = tensorRelu(var_9);
void* var_12 = tensorConvolution(var_10, conv2d_4_w, 1, 1, 1, 1, 1, 0);
void* var_13 = tensorAdd(var_12, conv2d_4_b);
void* var_14 = tensorRelu(var_13);
void* var_15 = tensorPooling(var_14,0,2,2,0,0,2,2);
void* var_16 = tensorConvolution(var_15, conv2d_5_w, 1, 1, 1, 1, 1, 0);
void* var_17 = tensorAdd(var_16, conv2d_5_b);
void* var_18 = tensorRelu(var_17);
void* var_20 = tensorConvolution(var_18, conv2d_6_w, 1, 1, 1, 1, 1, 0);
void* var_21 = tensorAdd(var_20, conv2d_6_b);
void* var_22 = tensorRelu(var_21);
void* var_24 = tensorConvolution(var_22, conv2d_7_w, 1, 1, 1, 1, 1, 0);
void* var_25 = tensorAdd(var_24, conv2d_7_b);
void* var_26 = tensorRelu(var_25);
void* var_27 = tensorPooling(var_26,0,2,2,0,0,2,2);
void* var_28 = tensorConvolution(var_27, conv2d_8_w, 1, 1, 1, 1, 1, 0);
void* var_29 = tensorAdd(var_28, conv2d_8_b);
void* var_30 = tensorRelu(var_29);
void* var_32 = tensorConvolution(var_30, conv2d_9_w, 1, 1, 1, 1, 1, 0);
void* var_33 = tensorAdd(var_32, conv2d_9_b);
void* var_34 = tensorRelu(var_33);
void* var_36 = tensorConvolution(var_34, conv2d_10_w, 1, 1, 1, 1, 1, 0);
void* var_37 = tensorAdd(var_36, conv2d_10_b);
void* var_38 = tensorRelu(var_37);
void* var_39 = tensorPooling(var_38,0,2,2,0,0,2,2);
void* var_40 = tensorConvolution(var_39, conv2d_11_w, 1, 1, 1, 1, 1, 0);
void* var_41 = tensorAdd(var_40, conv2d_11_b);
void* var_42 = tensorRelu(var_41);
void* var_44 = tensorConvolution(var_42, conv2d_12_w, 1, 1, 1, 1, 1, 0);
void* var_45 = tensorAdd(var_44, conv2d_12_b);
void* var_46 = tensorRelu(var_45);
void* var_48 = tensorConvolution(var_46, conv2d_13_w, 1, 1, 1, 1, 1, 0);
void* var_49 = tensorAdd(var_48, conv2d_13_b);
void* var_50 = tensorRelu(var_49);
void* var_51 = tensorPooling(var_50,0,2,2,0,0,2,2);
void* var_54 = tensorGemmGPU(var_51, dense_1_w);
void* var_55 = tensorAdd(var_54, dense_1_b);
void* var_56 = tensorRelu(var_55);
void* var_58 = tensorGemmGPU(var_56, dense_2_w);
void* var_59 = tensorAdd(var_58, dense_2_b);
void* var_60 = tensorSoftmax(var_59);
uint8_t* labels = readLabelsBatch(labels_path.c_str(),start,end);
float accuracy = computeAccuracy2(labels, batch_size, var_60);
final_accuracy += accuracy;
freeBatchMemory();
}
final_accuracy = final_accuracy / batch_count;
dumpFinalAccuracy(final_accuracy);
llvm_hpvm_cleanupTensorRt();
return 0;
}
hpvm/projects/onnx/keras_ref/keras_environment.yml deleted 100644 → 0
+ 0
− 322
View file @ 6491cfe3
name: approxhpvm_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
- 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
- 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
- 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
hpvm/projects/onnx/keras_ref/keras_frontend/__init__.py deleted 100644 → 0
+ 0
− 0
View file @ 6491cfe3
hpvm/projects/onnx/keras_ref/keras_frontend/setup.py deleted 100644 → 0
+ 0
− 12
View file @ 6491cfe3
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=[],
)
hpvm/projects/onnx/keras_ref/setup.py deleted 100644 → 0
+ 0
− 12
View file @ 6491cfe3
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=[],
)
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment