From 6056da6322dbf5f55ee01b486bf78a4bab5386f0 Mon Sep 17 00:00:00 2001
From: shingjan <yjshi03@gmail.com>
Date: Thu, 26 Mar 2020 02:19:54 -0500
Subject: [PATCH] emit structure almost finished
---
hpvm/projects/onnx/frontend/graph_builder.py | 221 ++++++++++++++++++
hpvm/projects/onnx/frontend/ir.py | 14 +-
.../onnx/frontend/{mnist.py => main.py} | 2 +-
.../frontend/onnx_frontend/graph_builder.py | 10 +-
.../onnx/frontend/onnx_frontend/util.py | 5 +-
hpvm/projects/onnx/frontend/operators.py | 66 +++++-
hpvm/projects/onnx/frontend/util.py | 62 +++++
7 files changed, 361 insertions(+), 19 deletions(-)
create mode 100644 hpvm/projects/onnx/frontend/graph_builder.py
rename hpvm/projects/onnx/frontend/{mnist.py => main.py} (93%)
create mode 100644 hpvm/projects/onnx/frontend/util.py
diff --git a/hpvm/projects/onnx/frontend/graph_builder.py b/hpvm/projects/onnx/frontend/graph_builder.py
new file mode 100644
index 0000000000..5ee60b034d
--- /dev/null
+++ b/hpvm/projects/onnx/frontend/graph_builder.py
@@ -0,0 +1,221 @@
+import sys
+import numpy as np
+import os
+
+from operators import *
+from ir import *
+
+class Node(object):
+ def __init__(self, name, shape, dtype):
+ self._name = name
+ self._shape = shape if shape else {}
+ self._dtype = dtype
+ def __str__(self):
+ return "Node: " + self._name + " with shape: " + str(self._shape) + " and data type " + str(self._dtype)
+ __repr__ = __str__
+
+class GraphBuilder(object):
+ def __init__(self, model, shape, dtype, opset):
+ self._nodes = {}
+ self._params = {}
+ self._renames = {}
+ self._num_input = 0
+ self._num_param = 0
+ self._dtype = dtype
+ self._graph = model.graph
+ self._opset = opset
+ self._shape = shape if shape else self._build_shape()
+ self.program_str = ""
+ self.dfg = DFG()
+
+ ################################################
+ # Aux functions for graph building
+ ################################################
+ def _build_shape(self):
+ shape = {}
+ for input in self._graph.input:
+ # get type of input tensor
+ tensor_type = input.type.tensor_type
+ # check if it has a shape:
+ if (tensor_type.HasField("shape")):
+ shape[input.name] = tensor_type.shape
+ return shape
+
+ def _parse_array(self, tensor_proto):
+ try:
+ from onnx.numpy_helper import to_array
+ except ImportError as e:
+ raise ImportError(
+ "Unable to import onnx which is required {}".format(e))
+ np_array = to_array(tensor_proto).reshape(tuple(tensor_proto.dims))
+ return np_array
+
+ def _parse_value_proto(self, value_proto):
+ """Parse ValueProto or raw str."""
+ try:
+ name = value_proto.name
+ except AttributeError:
+ name = value_proto
+ return name
+
+ def _parse_dtype(self, value_proto, dtype):
+ """Parse dtype."""
+ try:
+ from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
+ return TENSOR_TYPE_TO_NP_TYPE[value_proto.type.tensor_type.elem_type].name
+ except AttributeError:
+ return dtype
+
+ ################################################
+ # Emit functions for code generation
+ ################################################
+ def dump_weights(self):
+ for init_tensor in self._graph.initializer:
+ print(init_tensor)
+
+ def build_graph(self):
+ # parse init tensors
+ for init_tensor in self._graph.initializer:
+ if not init_tensor.name.strip():
+ raise ValueError("Tensor's name is required.")
+ self._params[init_tensor.name] = self._parse_array(init_tensor)
+ self._nodes[init_tensor.name] = Node(init_tensor.name,
+ 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
+ print("###############################")
+ print(op_name)
+ #print("attribute: " + str(node.attribute))
+ #print("input: " + str(node.input))
+ #print("output: " + str(node.output))
+ #print(self._nodes)
+
+ def traverse_graph(self, cur_node, visited):
+ if cur_node in visited:
+ return
+
+ if dfg.predVisited(cur_node, visited):
+ visited_nodes[cur_node.layer_name] = True
+ self.program_str += cur_node.codegen()
+ for output_node in cur_node.outputs:
+ self.codegenNode(dfg, output_node, visited)
+
+ def emit_graph(self):
+ self.build_graph()
+ visited_nodes = {}
+ self.traverse_graph(self.dfg.root, visited)
+
+ def emit_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 emit_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]
+
+ destructors = "\nllvm_hpvm_cleanupTensorRt(); \n"
+ end_main = "\nreturn 0; \n\n}\n"
+ self.program_str += destructors
+ self.program_str += end_main
+
+ def emit_batch_loop(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 emit_batch_loop_end(self):
+ end_loop_str = ""
+ end_loop_str += "\nuint32_t* labels = readLabelsBatch3(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 = computeAccuracy3(labels, " + 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 emit_program(self, src_dir):
+ f = open(src_dir + "/src.cc", "w+")
+ f.write(self.program_str)
+ f.close()
+ '''
+ Compile is a top level function to compile an onnx model into C/C++
+ program with HPVM intrinsics
+ '''
+ def codegen(self, weights_dir, test_data, test_labels):
+ if os.path.exists(weights_dir):
+ raise ValueError("Weight dir existed. Compilation interrupted!")
+ os.mkdir(weights_dir)
+ self.emit_header()
+ self.emit_weights(weights_dir)
+ self.emit_batch_loop(test_data)
+ self.emit_graph()
+ self.emit_batch_loop_end()
+ self.emit_footer(test_data)
+ self.emit_program(weights_dir)
diff --git a/hpvm/projects/onnx/frontend/ir.py b/hpvm/projects/onnx/frontend/ir.py
index 2a3bbe4a6b..6ebf1f17c6 100644
--- a/hpvm/projects/onnx/frontend/ir.py
+++ b/hpvm/projects/onnx/frontend/ir.py
@@ -1,4 +1,4 @@
-class DFG:
+class DFG(object):
root_set = False;
@@ -87,10 +87,9 @@ class DFG:
visited_nodes = {}
# Starting traversal at the DFG root node
self.traverseNode(self.root_node, visited_nodes)
-
-## This should be the place where partial evaluation happens
-class NodeFactory:
- def build(self, layer):
+
+ ## This should be the place where partial evaluation happens
+ def emitNode(self, layer):
if layer.op_type == "Conv":
pass
elif layer.op_type == "Tanh":
@@ -107,10 +106,11 @@ class NodeFactory:
pass
elif layer.op_type == "Identity":
pass
- else
+ else:
raise ValueError("Unsupported operator type!")
-class DFGNode:
+
+class DFGNode(object):
def add_output(self, output_node):
self.outputs.append(output_node)
def add_input(self, input_node):
diff --git a/hpvm/projects/onnx/frontend/mnist.py b/hpvm/projects/onnx/frontend/main.py
similarity index 93%
rename from hpvm/projects/onnx/frontend/mnist.py
rename to hpvm/projects/onnx/frontend/main.py
index 8a4f3a85d7..bd8d34f2e1 100644
--- a/hpvm/projects/onnx/frontend/mnist.py
+++ b/hpvm/projects/onnx/frontend/main.py
@@ -11,7 +11,7 @@ from onnx import numpy_helper, version_converter
from onnx_frontend.util import convert_to_hpvm
#model = onnx.load('../models/mnist/mnist.onnx')
-model = onnx.load('../models/alexnet/alexnet.onnx')
+model = onnx.load('../models/keras/alexnet.onnx')
test_data_dir = '../models/mnist/test_data_set_0'
# print('The model before conversion:\n{}'.format(model))
diff --git a/hpvm/projects/onnx/frontend/onnx_frontend/graph_builder.py b/hpvm/projects/onnx/frontend/onnx_frontend/graph_builder.py
index 5cf8644cbf..6ad0d59b6e 100644
--- a/hpvm/projects/onnx/frontend/onnx_frontend/graph_builder.py
+++ b/hpvm/projects/onnx/frontend/onnx_frontend/graph_builder.py
@@ -2,6 +2,8 @@ import sys
import numpy as np
import os
+from operators import *
+
class Node(object):
def __init__(self, name, shape, dtype):
self._name = name
@@ -12,7 +14,6 @@ class Node(object):
__repr__ = __str__
class GraphBuilder(object):
-
def __init__(self, model, shape, dtype, opset):
self._nodes = {}
self._params = {}
@@ -23,8 +24,11 @@ class GraphBuilder(object):
self._graph = model.graph
self._opset = opset
self._shape = shape if shape else self._build_shape()
+
+ def dump_weights(self):
+ print(self._graph.initializer)
- def build_cfg(self):
+ def build_graph(self):
# parse init tensors
for init_tensor in self._graph.initializer:
if not init_tensor.name.strip():
@@ -110,4 +114,4 @@ class GraphBuilder(object):
if os.path.exists(weights_dir):
raise ValueError("Weight dir existed. Compilation interrupted!")
os.mkdir(weights_dir)
- self.build_cfg()
\ No newline at end of file
+ self.build_graph()
\ No newline at end of file
diff --git a/hpvm/projects/onnx/frontend/onnx_frontend/util.py b/hpvm/projects/onnx/frontend/onnx_frontend/util.py
index 2900f2faf3..78091e7c32 100644
--- a/hpvm/projects/onnx/frontend/onnx_frontend/util.py
+++ b/hpvm/projects/onnx/frontend/onnx_frontend/util.py
@@ -1,7 +1,7 @@
import sys
import numpy as np
import os
-from .graph_builder import GraphBuilder
+from graph_builder import GraphBuilder
def check_model(onnx_model):
try:
@@ -59,4 +59,5 @@ def convert_to_hpvm(model,
opset = 1 # default opset version set to 1 if not specified
print("opset version: ", opset)
gb = GraphBuilder(model, shape, dtype, opset)
- gb.build_cfg()
\ No newline at end of file
+ gb.dump_weights()
+ gb.build_graph()
\ No newline at end of file
diff --git a/hpvm/projects/onnx/frontend/operators.py b/hpvm/projects/onnx/frontend/operators.py
index f5851e8955..3a52237307 100644
--- a/hpvm/projects/onnx/frontend/operators.py
+++ b/hpvm/projects/onnx/frontend/operators.py
@@ -1,14 +1,26 @@
-from .ir import DFGNode, ActivationNode, LogicalOpNode
+from ir import DFGNode, ActivationNode, LogicalOpNode
class AddNode(DFGNode):
+ def __init__(self, layer):
+ DFGNode.__init__(self, layer)
def codegen(self):
- pass
+ input_vars = self.getMultipleInputNames(cur_node)
+ inst_str = "void* " + out_var_name1 + " = "
+ inst_str += "tensorAdd(" + input_vars[0] + ", " + input_vars[1] + "); \n"
+ return inst_str
+
class MatMulNode(DFGNode):
+ def __init__(self, layer):
+ DFGNode.__init__(self, layer)
def codegen(self):
pass
+
class SoftMaxNode(DFGNode):
+ def __init__(self, layer):
+ DFGNode.__init__(self, layer)
def codegen(self):
pass
+
class Conv2DNode(DFGNode):
def __init__(self, layer):
DFGNode.__init__(self, layer)
@@ -20,7 +32,39 @@ class Conv2DNode(DFGNode):
print("\t", self.strides)
print("\tPadding = ", self.padding)
def codegen(self):
- pass
+ 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, "
+ if layer_type == "DepthwiseConv2D":
+ C = weights.shape[2]
+ inst_str += str(C) + "); \n"
+ else:
+ inst_str += "1); \n"
+ if strides[0] > 1 and cur_node.padding.strip() == "same":
+ print ("!ERROR: Same Padding not supported for Conv with Stride > 1")
+ print ("Use: ZeroPadding2D(padding=(" + str(padding) + "," + str(padding) + "));\n");
+ sys.exit(0)
+ return inst_str
+
class DepthwiseConv2DNode(DFGNode):
def __init__(self, layer):
DFGNode.__init__(self, layer)
@@ -63,7 +107,7 @@ class DepthwiseConv2DNode(DFGNode):
print ("!ERROR: Same Padding not supported for Conv with Stride > 1")
print ("Use: ZeroPadding2D(padding=(" + str(padding) + "," + str(padding) + "));\n");
sys.exit(0)
- return += inst_str
+ return inst_str
class DenseNode(DFGNode):
def __init__(self, layer):
@@ -88,7 +132,11 @@ class MaxPool2DNode(DFGNode):
print("\t pool_size = ", self.pool_size)
print("\t strides = ", self.strides)
def codegen(self):
- pass
+ 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"
+
class AveragePooling2DNode(DFGNode):
def __init__(self, layer):
DFGNode.__init__(self, layer)
@@ -97,7 +145,12 @@ class AveragePooling2DNode(DFGNode):
print("\t pool_size = ", self.pool_size)
print("\t strides = ", self.strides)
def codegen(self):
- pass
+ 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"
+ return insr_str
+
class ZeroPadding2DNode(DFGNode):
def __init__(self, layer):
DFGNode.__init__(self, layer)
@@ -106,6 +159,7 @@ class ZeroPadding2DNode(DFGNode):
print ("padding = ", self.padding)
def codegen(self):
pass
+
class BatchNormalizationNode(DFGNode):
def __init__(self, layer):
DFGNode.__init__(self, layer)
diff --git a/hpvm/projects/onnx/frontend/util.py b/hpvm/projects/onnx/frontend/util.py
new file mode 100644
index 0000000000..2900f2faf3
--- /dev/null
+++ b/hpvm/projects/onnx/frontend/util.py
@@ -0,0 +1,62 @@
+import sys
+import numpy as np
+import os
+from .graph_builder import GraphBuilder
+
+def check_model(onnx_model):
+ try:
+ from onnx import checker, onnx_cpp2py_export
+ if hasattr(checker, 'check_model'):
+ # try use onnx's own model checker before converting any model
+ try:
+ checker.check_model(onnx_model)
+ print("onnx model is checked valid.")
+ except onnx_cpp2py_export.checker.ValidationError as e:
+ import warnings
+ # the checker is a bit violent about errors, so simply print warnings here
+ warnings.warn(str(e))
+ except ImportError as e:
+ raise ImportError("Unable to import onnx.checker which is required {}".format(e))
+
+def convert_to_hpvm(model,
+ shape=None,
+ dtype="float32",
+ opset=None):
+ """Converting an onnx model to equivalent HPVM IR
+
+ ONNX graphs are represented as Python Protobuf objects.
+ The companion parameters will be handled automatically.
+ However, the input names from onnx graph is vague, mixing inputs and
+ network weights/bias such as "1", "2"...
+ For convenience, we rename the `real` input names to "input_0",
+ "input_1"... And renaming parameters to "param_0", "param_1"...
+
+ Parameters
+ ----------
+ model : protobuf object
+ ONNX ModelProto after ONNX v1.1.0
+
+ shape : dict of str to tuple, optional
+ The input shape to the graph
+
+ dtype : str or dict of str to str
+ The input types to the graph
+
+ opset : int, optional
+ Override to autodetected opset.
+ This can be helpful for some testing.
+
+ Returns
+ -------
+
+ """
+ check_model(model)
+ graph = model.graph
+ if opset is None:
+ try:
+ opset = model.opset_import[0].version if model.opset_import else 1
+ except AttributeError:
+ opset = 1 # default opset version set to 1 if not specified
+ print("opset version: ", opset)
+ gb = GraphBuilder(model, shape, dtype, opset)
+ gb.build_cfg()
\ No newline at end of file
--
GitLab