代码拉取完成，页面将自动刷新
同步操作将从 ahqzy/onnx_convert 强制同步，此操作会覆盖自 Fork 仓库以来所做的任何修改，且无法恢复！！！
确定后同步将在后台操作，完成时将刷新页面，请耐心等待。
# -*- coding: utf-8 -*-
import os
import onnx
from onnx import version_converter
import copy
import numpy as np
import logging
import log
import onnxruntime
import sys, getopt
import json
import argparse
import h5py
#import tensorflow as tf
import time
import fuse
from swish_convert import merge_swish_and_hard_swish
from mish_convert import merge_mish
from hard_sigmoid_convert import merge_hard_sigmiod
from gelu_fuse import merge_gelu
import bn2conv
import values
import version_check
import operation
import copy
from caffe2onnx.src.load_save_model import loadcaffemodel, saveonnxmodel
from caffe2onnx.src.caffe2onnx import Caffe2Onnx
from onnxsim.onnx_simplifier import simplify
from float16 import convert_float_to_float16
from preprocess import preproc
from postprocess import postproc
from correct_batch import correct_batch_for_opset_convert, convert_ort_type_2_np, get_data_list
#from pd2onnx import convert_pd2onnx, is_dynamic_paddle
#from pt2onnx import convert_pt2onnx
from gemm.gemm_cvt import gemm_convert
from resize_convert import merge_resize
from ln_convert import merge_layernorm
from matmul2gemm import matmul_2_gemm
from mha_optimization import mha_optimizer
from input_fp32_to_uint8 import fp32_to_uint8
from common import MXC_CONFIG
using_wheel = False
support_mish = 0
concat_result = 0
inputs_as_nchw = ''
#logging.basicConfig(level=logging.INFO, filename='./convert.log', filemode='w')
#logger = logging.getLogger("[MacaConverter]")
logger = log.getLogger('[MacaConverter]', log.INFO)
file_handler = logging.FileHandler('./convert.log')
file_handler.setLevel(logging.CRITICAL)
logger.addHandler(file_handler)
from onnx import shape_inference, TensorProto, version_converter, numpy_helper
import argparse
valid_model_type = ['caffe', 'pytorch', 'tf-h5', 'tf-ckpt', 'tf-sm', 'tf-graph', 'darknet', 'onnx', 'paddle']
optimization_op_list = ['Max', 'Min', 'Sum', 'Mean']
############ Error Code Define #################
exit_code_normal = 0
exit_code_no_caffe_cfg_file = -1
exit_code_sm2onnx = -2
exit_code_h52onnx = -3
exit_code_ckpt2onnx = -4
exit_code_pb2onnx = -5
exit_code_no_darknet_cfg_or_weights = -6
exit_code_convert_darknet2onnx = -7
exit_code_fuse_mish = -8
exit_code_check_optimization_op = -9
exit_code_check_modify_onnx2dynamic = -10
exit_code_check_convert_gap_2_ap = -11
exit_code_model_not_exist = -12
exit_code_invalid_model_type = -13
exit_code_pytorch_no_input_shape = -14
exit_code_tensorflow_no_inputs_or_outputs = -15
exit_code_extract_sub_no_inputs_or_outputs = -16
exit_code_model_type_not_onnx = -17
#########################################################
def set_using_wheel():
global using_wheel
using_wheel = True
def parse_args():
parser = argparse.ArgumentParser(description='Convert caffe/tensorflow/torch/paddle/darknet model to ONNX.')
parser.add_argument("--model_path",
type=str,
help="Input path(model file or folder)")
parser.add_argument("--model_type",
type=str,
help="Input model type(ex: caffe/pytorch/tf-h5/...)")
parser.add_argument("--output",
type=str,
help="Output path(ex: ./output.onnx)")
parser.add_argument("--concat_result",
type=int, required=False,
choices=[0, 1],
default=0,
help="When convert darknet to onnx, if set 1, the tool will concat the results")
parser.add_argument("--op_set",
type=int, required=False,
help="Set op_set version(default: 11)")
#for simplify
parser.add_argument("--simplify",
type=int, required=False,
choices=[0, 1, 2],
default=1,
help="Simplify the model(0:no simplify;1:do simplify; 2:for dynamic model)")
parser.add_argument("--simplify_hw",
type=str,
required=False,
default='',
help="When h/w is -1, you can specify h/w as you expected(together with --simplify 2)")
#force simplify
parser.add_argument("--force_simplify",
type=int, required=False,
choices=[0, 1, 2],
default=0,
help="Force simplify the model(0:no simplify;1:do simplify; 2:for dynamic model)")
#for pytorch/dynamic_paddle
parser.add_argument("--input_shape",
type=str,
required=False,
default='',
help="Input shape for pytorch/paddle(ex: [1,3,224,224] or [1,3,224,224]/[1,3,56,56])")
#######
parser.add_argument("--inputs",
type=str,
required=False,
default='',
help="When do checkpoint2ONNX/graph2ONNX/onnx_sub_graph, you should specify inputs(ex: --inputs image:0)")
parser.add_argument("--outputs",
type=str,
required=False,
default='',
help="When do checkpoint2ONNX/graph2ONNX/onnx_sub_graph, you should specify outputs(ex: --outputs predict:0)")
#for extract sub graph
parser.add_argument("--extract_sub",
type=int,
required=False,
choices=[0, 1],
default=0,
help="If set 1, the tool will extract sub graph by specify inputs/outputs")
#for dynamic batch size
parser.add_argument("--dynamic_batch",
type=int,
required=False,
default=0,
choices=[0, 1],
help="If set 1, the tool will convert batch size to -1")
#for fp32-->fp16
parser.add_argument("--fp32_to_fp16",
type=int,
required=False,
default=0,
choices=[0, 1],
help="If set 1, the tool will convert fp32 to fp16 in the model")
parser.add_argument("--support_mish",
type=int,
required=False,
default=1,
choices=[0, 1],
help="If set 1, the tool will fuse Softplus+Tanh+Mul to Mish")
#insert preproc node
parser.add_argument("--preproc_yaml",
type=str,
required=False,
default='',
help="If specify preprocess yaml file, the tool will insert preproc node in the beginning of the model")
parser.add_argument("--postproc_yaml",
type=str,
required=False,
default='',
help="If specify postprocess yaml file, the tool will insert postproc node in the ending of the model")
#for paddle dynamic model or pytorch
parser.add_argument("--model_def_file",
type=str,
required=False,
default='',
help="Paddle/pytorch model definition file location(ex: --model_def_file ./cnn.py)")
parser.add_argument("--model_weights_file",
type=str,
required=False,
default='',
help="Paddle/pytorch model weights file location(ex: --model_weights_file ./0.99667.pth)")
parser.add_argument("--model_class_name",
type=str,
required=False,
default='',
help="Paddle/pytorch model calss name(ex: --model_class_name CNN)")
parser.add_argument("--model_input_type",
type=str,
required=False,
#choices=['float', 'float32', 'float16', 'uint8', 'int8', 'uint16', 'int16', 'uint32', 'int32', 'uint64', 'int64', 'bool'],
default='',
help="Paddle/pytorch input type(default float, choice is ['float', 'float32', 'float16', 'uint8', 'int8', 'uint16', 'int16', 'uint32', 'int32', 'uint64', 'int64', 'bool'])")
parser.add_argument("--params_file",
type=str,
required=False,
default='',
help="Paddle/pytorch params declaration file location(ex: --params_file ./params.py)")
#for tensorflow
parser.add_argument("--inputs_as_nchw",
type=str,
required=False,
default='',
help="When some input of tensorflow model is nhwc, you can use it(ex: --inputs_as_nchw image:0) to convert to nchw")
#for gap-->ap
parser.add_argument("--gap_to_ap",
type=int,
required=False,
choices=[0, 1],
default=1,
help="If set 1, the tool will convert GlobalAveragePool to AveragePool for hardware acceleration")
#for pad+pool fuse
parser.add_argument("--fuse_pad_pool",
type=int,
required=False,
default=1,
choices=[0, 1],
help="If set 1, the tool will fuse Pad into Pool")
#for merge swish
parser.add_argument("--support_swish",
type=int,
required=False,
choices=[0, 1],
default=1,
help="If set 1, the tool will convert Sigmoid+Mul to Swish; HardSigmoid+Mul to HardSwish")
#for convert BN to GroupConv(1x1)
parser.add_argument("--bn_to_conv",
type=int,
required=False,
choices=[0, 1],
default=1,
help="If set 1, the tool will convert BN to group 1x1_Conv")
#for pytorch/paddle
parser.add_argument("--output_num",
type=int,
required=False,
default=1,
help="If output num of pytorch model > 1, you can specify it by --output_num")
#for pytorch
parser.add_argument("--keep_batch",
type=int,
choices=[0, 1],
required=False,
default=1,
help="For pytorch, if set 1, the tool will keep model batch size(if 0, set it to dynamic(-1))")
#for convert gemm
parser.add_argument("--gemm_optimization",
type=int,
required=False,
choices=[0, 1],
default=1,
help="If set 1, the tool will convert gemm to fc or matmul+add+mul")
#for convert Reshape+Expand+Reshape to Resize
parser.add_argument("--expand_to_resize",
type=int,
required=False,
choices=[0, 1],
default=1,
help="If set 1, the tool will convert Reshape+Expand+Reshape to Resize")
#reset model value_info(some model(batch=-1) may have wrong value info for middle node))
parser.add_argument("--reset_value_info",
type=int,
required=False,
choices=[0, 1],
default=0,
help="If set 1, the tool will try correct wrong value info")
#fuse match ops to LayerNorm
parser.add_argument("--fuse_layernorm",
type=int,
required=False,
choices=[0, 1],
default=1,
help="If set 1, the tool will fuse match ops to LayerNorm")
#convert matmul to gemm()
parser.add_argument("--matmul_to_gemm",
type=int,
required=False,
choices=[0, 1],
default=1,
help="If set 1, the tool will convert Matmul to Gemm(when A shape[0] < 32 and B is Constant)")
#reset model value_info(some model(batch=-1) may have wrong value info for middle node))
parser.add_argument("--reset_batch",
type=str,
required=False,
nargs='*',
default='', #should be 'input_batch,output_batch'
help="If set 1, the tool will try reset model batch_size")
#convert Add+Clip+Div to HardSigmoid
parser.add_argument("--fuse_hard_sigmoid",
type=int,
required=False,
default=1,
choices=[0, 1],
help="If set 1, the tool will merge Add+Cilp+Div to HardSigmoid")
#fuse Gelu
parser.add_argument("--fuse_gelu",
type=int,
required=False,
default=1,
choices=[0, 1],
help="If set 1, the tool will fuse Gelu")
#Disable all optimization
parser.add_argument("--disable_all_optimizer",
type=int,
required=False,
default=0,
choices=[0, 1],
help="If set 1, the tool will force all optimization value to 0")
#for mha optimization
parser.add_argument("--mha_optimization",
type=int,
required=False,
choices=[0, 1],
default=0,
help="If set 1, the tool will do some optimization for mha structure")
#for caffe pooling
parser.add_argument("--ceil_floor_reverse",
type=int,
required=False,
choices=[0, 1],
default=0,
help="If set 1, the tool will set ceil=1 in caffe pooling")
#fp32-->u8(for input type)
parser.add_argument("--fp32_to_u8",
type=int,
required=False,
default=0,
help="If set 1, the tool will change input type from float to uint8")
#show version
parser.add_argument('--version', '-v',
action='store_true',
default=False,
help='Show current version')
args = parser.parse_args()
return args
def get_caffe_files(model_path):
items = os.listdir(model_path)
prototxt_cnt = 0
caffemodel_cnt = 0
prototxt_file = ''
caffemodel_file = ''
for f in items:
if f.endswith(".prototxt"):
prototxt_cnt = prototxt_cnt + 1
prototxt_file = f
elif f.endswith(".caffemodel"):
caffemodel_cnt = caffemodel_cnt + 1
caffemodel_file = f
if prototxt_cnt == 1 and caffemodel_cnt == 1:
if model_path.endswith("/"):
prototxt_file = model_path + prototxt_file
caffemodel_file = model_path + caffemodel_file
else:
prototxt_file = model_path + '/' + prototxt_file
caffemodel_file = model_path + '/' + caffemodel_file
logger.info('got prototxt_file:{}, caffemodel_file:{}'.format(prototxt_file, caffemodel_file))
elif prototxt_cnt > 1 or caffemodel_cnt > 1:
prototxt_file = ''
caffemodel_file = ''
logger.error('ERROR: prototxt_cnt > 1 or caffemodel_cnt > 1')
elif prototxt_cnt == 0 or caffemodel_cnt == 0:
prototxt_file = ''
caffemodel_file = ''
logger.error('ERROR: No .prototxt file or no .caffemodel file')
return prototxt_file, caffemodel_file
def convert_caffe2onnx(model_path, output, op_set, ceil_floor_reverse):
logger.info('Begin converting caffe to onnx...')
prototxt_file, caffemodel_file = get_caffe_files(model_path)
if prototxt_file == '' or caffemodel_file == '':
sys.exit(exit_code_no_caffe_cfg_file)
onnxmodel_path = output
graph, params = loadcaffemodel(prototxt_file, caffemodel_file)
c2o = Caffe2Onnx(graph, params, onnxmodel_path, op_set, ceil_floor_reverse)
onnxmodel = c2o.createOnnxModel(op_set) #qiuzy debug
saveonnxmodel(onnxmodel, onnxmodel_path)
def convert_sm2onnx(model_path, output, op_set):
logger.info('Begin converting tf-savemodel to onnx...')
try:
import tensorflow
except Exception as e:
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@')
print(e)
print('Please install tensorflow(pip install tensorflow==2.4.0)')
print('If numpy version > 1.19.5, tensorflow version should be 2.7.4')
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@')
sys.exit(exit_code_sm2onnx)
version_check.check('tensorflow', tensorflow.__version__)
if using_wheel == False:
cmd = 'python -m tf2onnx.convert --saved-model ' + model_path + ' --opset ' + str(op_set) + ' --output ' + output
else:
cmd = 'python -m maca_converter.tf2onnx.convert --saved-model ' + model_path + ' --opset ' + str(op_set) + ' --output ' + output
if inputs_as_nchw != '':
cmd += ' --inputs-as-nchw ' + inputs_as_nchw
logger.info('convert_tfsm2onnx: {}'.format(cmd))
r = os.system(cmd)
if r != 0:
logger.error('ERROR: convert_sm2onnx failed')
sys.exit(exit_code_sm2onnx)
def convert_h52onnx(model_path, output, op_set):
logger.info('Begin converting tf-savemodel to onnx...')
try:
import tensorflow
except Exception as e:
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@')
print(e)
print('Please install tensorflow(pip install tensorflow==2.4.0)')
print('If numpy version > 1.19.5, tensorflow version should be 2.7.4')
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@')
sys.exit(exit_code_h52onnx)
version_check.check('tensorflow', tensorflow.__version__)
if using_wheel == False:
cmd = 'python -m tf2onnx.convert --keras ' + model_path + ' --opset ' + str(op_set) + ' --output ' + output
else:
cmd = 'python -m maca_converter.tf2onnx.convert --keras ' + model_path + ' --opset ' + str(op_set) + ' --output ' + output
if inputs_as_nchw != '':
cmd += ' --inputs-as-nchw ' + inputs_as_nchw
logger.info('convert_tfh52onnx: {}'.format(cmd))
r = os.system(cmd)
if r != 0:
logger.error('ERROR: convert_h52onnx failed')
sys.exit(exit_code_h52onnx)
def convert_ckpt2onnx(model_path, output, op_set, inputs, outputs):
logger.info('Begin converting tf-ckpt to onnx...')
try:
import tensorflow
except Exception as e:
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@')
print(e)
print('Please install tensorflow(pip install tensorflow==2.4.0)')
print('If numpy version > 1.19.5, tensorflow version should be 2.7.4')
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@')
sys.exit(exit_code_ckpt2onnx)
version_check.check('tensorflow', tensorflow.__version__)
if using_wheel == False:
cmd = 'python -m tf2onnx.convert --checkpoint ' + model_path + ' --opset ' + str(op_set) + ' --output ' + output \
+ ' --inputs ' + inputs + ' --outputs ' + outputs
else:
cmd = 'python -m maca_converter.tf2onnx.convert --checkpoint ' + model_path + ' --opset ' + str(op_set) + ' --output ' + output \
+ ' --inputs ' + inputs + ' --outputs ' + outputs
if inputs_as_nchw != '':
cmd += ' --inputs-as-nchw ' + inputs_as_nchw
logger.info('convert_ckpt2onnx: {}'.format(cmd))
r = os.system(cmd)
if r != 0:
logger.error('ERROR: convert_ckpt2onnx failed')
sys.exit(exit_code_ckpt2onnx)
def convert_graph2onnx(model_path, output, op_set, inputs, outputs):
logger.info('Begin converting tf-graph to onnx...')
try:
import tensorflow
except Exception as e:
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@')
print(e)
print('Please install tensorflow(pip install tensorflow==2.4.0)')
print('If numpy version > 1.19.5, tensorflow version should be 2.7.4')
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@')
sys.exit(exit_code_pb2onnx)
version_check.check('tensorflow', tensorflow.__version__)
if using_wheel == False:
cmd = 'python -m tf2onnx.convert --graphdef ' + model_path + ' --opset ' + str(op_set) + ' --output ' + output \
+ ' --inputs ' + inputs + ' --outputs ' + outputs
else:
cmd = 'python -m maca_converter.tf2onnx.convert --graphdef ' + model_path + ' --opset ' + str(op_set) + ' --output ' + output \
+ ' --inputs ' + inputs + ' --outputs ' + outputs
if inputs_as_nchw != '':
cmd += ' --inputs-as-nchw ' + inputs_as_nchw
logger.info('convert_graph2onnx: {}'.format(cmd))
r = os.system(cmd)
if r != 0:
logger.error('ERROR: convert_graph2onnx failed')
sys.exit(exit_code_pb2onnx)
def get_darknet_files(model_path):
items = os.listdir(model_path)
cfg_cnt = 0
weights_cnt = 0
cfg_file = ''
weights_file = ''
for f in items:
if f.endswith(".cfg"):
cfg_cnt = cfg_cnt + 1
cfg_file = f
elif f.endswith(".weights"):
weights_cnt = weights_cnt + 1
weights_file = f
if cfg_cnt == 1 and weights_cnt == 1:
if model_path.endswith("/"):
cfg_file = model_path + cfg_file
weights_file = model_path + weights_file
else:
cfg_file = model_path + '/' + cfg_file
weights_file = model_path + '/' + weights_file
logger.info('got cfg_file:{}, weights_file:{}'.format(cfg_file, weights_file))
elif cfg_cnt > 1 or weights_cnt > 1:
cfg_file = ''
weights_file = ''
logger.error('ERROR: cfg_cnt > 1 or weights_cnt > 1')
elif cfg_cnt == 0 or weights_cnt == 0:
cfg_file = ''
weights_file = ''
logger.error('ERROR: No .cfg file or no .weights file')
return cfg_file, weights_file
#tf.compat.v1.disable_v2_behavior()
def ckpt2h5(trained_checkpoint_prefix):
#trained_checkpoint_prefix = sys.argv[1]
export_dir = './savemodel'
graph = tf.Graph()
config=tf.compat.v1.ConfigProto(allow_soft_placement=True, log_device_placement=True)
with tf.compat.v1.Session(graph=graph, config=config) as sess:
# Restore from checkpoint
loader = tf.compat.v1.train.import_meta_graph(trained_checkpoint_prefix + '.meta')
loader.restore(sess, trained_checkpoint_prefix)
# Export checkpoint to SavedModel
builder = tf.compat.v1.saved_model.builder.SavedModelBuilder(export_dir)
builder.add_meta_graph_and_variables(sess, [tf.saved_model.TRAINING, tf.saved_model.SERVING],strip_default_attrs=False)
builder.save()
def undo_darknet_concat(output):
model = onnx.load(output)
for node in model.graph.node:
if node.name == 'outputs' and node.op_type == 'Concat':
del model.graph.output[:]
for i in node.input:
model.graph.output.extend([onnx.ValueInfoProto(name=i)])
logger.info('undo concat, new outputs: {}'.format(model.graph.output))
model.graph.node.remove(node)
onnx.save(model, output)
break
def convert_dn2onnx(model_path, output, op_set):
global support_mish
global concat_result
logger.info('Begin converting darknet to onnx...... support_mish: {}'.format(support_mish))
cfg_file, weights_file = get_darknet_files(model_path)
if cfg_file == '' or weights_file == '':
sys.exit(exit_code_no_darknet_cfg_or_weights)
if using_wheel == False:
cmd = 'python ./darknet2onnx.py --cfg_file ' + cfg_file + ' --weights_file ' + weights_file + ' --output_file ' + output + ' --support_mish ' + str(support_mish)
else:
cmd = 'python -m maca_converter.darknet2onnx --cfg_file ' + cfg_file + ' --weights_file ' + weights_file + ' --output_file ' + output + ' --support_mish ' + str(support_mish)
if '-tiny' in cfg_file or '-tiny' in weights_file:
if using_wheel == False:
cmd = 'python ./darknet2onnx.py --cfg_file ' + cfg_file + ' --weights_file ' + weights_file + ' --strides 32 16 8 ' + ' --neck FPN ' + ' --output_file ' + output + ' --support_mish ' + str(support_mish)
else:
cmd = 'python -m maca_converter.darknet2onnx --cfg_file ' + cfg_file + ' --weights_file ' + weights_file + ' --strides 32 16 8 ' + ' --neck FPN ' + ' --output_file ' + output + ' --support_mish ' + str(support_mish)
elif 'yolov3' in cfg_file or 'yolov3' in weights_file:
if using_wheel == False:
cmd = 'python ./darknet2onnx.py --cfg_file ' + cfg_file + ' --weights_file ' + weights_file + ' --strides 32 16 8 ' + ' --neck FPN ' + ' --output_file ' + output
else:
cmd = 'python -m maca_converter.darknet2onnx --cfg_file ' + cfg_file + ' --weights_file ' + weights_file + ' --strides 32 16 8 ' + ' --neck FPN ' + ' --output_file ' + output
logger.info('convert_dn2onnx: {}'.format(cmd))
r = os.system(cmd)
if r != 0:
logger.error('ERROR: convert_dn2onnx failed')
sys.exit(exit_code_convert_darknet2onnx)
if concat_result == 0:
undo_darknet_concat(output)
def convert_mish(model_path, output, op_set):
global support_mish
logger.info('Begin converting mish')
if using_wheel == False:
cmd = 'python ./mish_convert.py --onnx_file ' + model_path + ' --output_file ' + output
else:
cmd = 'python -m maca_converter.mish_convert --onnx_file ' + model_path + ' --output_file ' + output
logger.info('convert_mish: {}'.format(cmd))
r = os.system(cmd)
if r != 0:
logger.error('ERROR: convert_mish failed')
sys.exit(exit_code_fuse_mish)
def convert(model_path, model_type, output, op_set, input_shape_list, inputs, outputs,
model_def_file,
model_class_name,
model_input_type,
model_weights_file,
output_num,
keep_batch,
params_file,
ceil_floor_reverse):
if model_type == 'caffe':
convert_caffe2onnx(model_path, output, op_set, ceil_floor_reverse)
if model_type == 'tf-sm':
convert_sm2onnx(model_path, output, op_set)
if model_type == 'tf-h5':
convert_h52onnx(model_path, output, op_set)
if model_type == 'tf-ckpt':
convert_ckpt2onnx(model_path, output, op_set, inputs, outputs)
if model_type == 'tf-graph':
convert_graph2onnx(model_path, output, op_set, inputs, outputs)
if model_type == 'darknet':
convert_dn2onnx(model_path, output, op_set)
if model_type == 'pytorch':
#convert_pt2onnx(model_path, output, op_set, input_shape)
from pt2onnx import convert_pt2onnx
convert_pt2onnx(model_path, output, op_set, input_shape_list,
model_def_file, model_class_name, model_weights_file, output_num, model_input_type, keep_batch, params_file)
if model_type == 'paddle':
from pd2onnx import convert_pd2onnx
convert_pd2onnx(model_path, output, op_set, input_shape_list, model_def_file, model_class_name, model_input_type, model_weights_file)
def optimization_op(model):
#model = onnx.load(onnxfile)
delete_node_id = 0
delete = False
#export_onnx = onnxfile
for node_id, node in enumerate(model.graph.node):
#print(node_id, ", name:", node.name, ", input:", node.input, ", output:", node.output, \
# ", op:", node.op_type, ', len(input):', len(node.input))
if node.op_type in optimization_op_list and len(node.input) == 1:
delete_node_id = node_id
delete = True
break
if delete == True:
log.debug('delete: {}'.format(delete_node_id))
delete_node = model.graph.node[delete_node_id]
log.debug('delete node op: {}'.format(delete_node.op_type))
next_node = model.graph.node[delete_node_id+1]
for i, n_ in enumerate(next_node.input):
#print('next input:', n_)
if n_ == delete_node.output[0]:
log.debug('got it: {]}'.format(n_))
next_node.input[i] = delete_node.input[0]
model.graph.node.remove(delete_node)
export_onnx = onnxfile
try:
onnx.checker.check_model(model)
except onnx.checker.ValidationError as e:
logger.warning('The model cannot be saved for: {}'.format(e))
if 'No Op registered for Mish' in str(e):
logger.warning('ignore mish warning, continue saving~')
else:
logger.error('ERROR: check model failed')
sys.exit(exit_code_check_optimization_op)
else:
logger.info('---Begin saving model...')
###################
#onnx.checker.check_model(model)
#onnx.save(model, export_onnx)
return delete
def correct_output_shape(model):
for output in model.graph.output:
if len(output.type.tensor_type.shape.dim) > 0:
output_shape = output.type.tensor_type.shape.dim
output_shape = [x.dim_value for x in output_shape]
dynamic_output_shape_ = any(d==-1 or d==0 for d in output_shape)
if dynamic_output_shape_ == True:
logger.info('The model output is dynamic, output: {} {}'.format(output.name, output_shape))
output.type.tensor_type.shape.dim[0].dim_value = 1
for output in model.graph.output:
if len(output.type.tensor_type.shape.dim) > 0:
output_shape = output.type.tensor_type.shape.dim
output_shape = [x.dim_value for x in output_shape]
logger.debug('The model output is dynamic, output_shape: {}'.format(output_shape))
def reset_model_value_info(model):
model_bak = copy.deepcopy(model)
del model_bak.graph.value_info[:]
try:
new_model = onnx.shape_inference.infer_shapes(model_bak)
except BaseException as e:
logger.warning('reset_model_value_info, the model cannot be inferenced for: {}'.format(e))
return model
else:
new_model = onnx.shape_inference.infer_shapes(model_bak)
new_model = onnx.shape_inference.infer_shapes(new_model)
return new_model
def reset_batch_size(model, input_batch, output_batch):
for input_ in model.graph.input:
if len(input_.type.tensor_type.shape.dim) > 0:
dim_proto = input_.type.tensor_type.shape.dim[0]
dim_proto.dim_value = input_batch
for output_ in model.graph.output:
if len(output_.type.tensor_type.shape.dim) > 0:
dim_proto = output_.type.tensor_type.shape.dim[0]
dim_proto.dim_value = output_batch
del model.graph.value_info[:]
try:
new_model = onnx.shape_inference.infer_shapes(model)
except BaseException as e:
logger.warning('reset_batch_size, the model cannot be inferenced for: {}'.format(e))
new_model = model
else:
new_model = onnx.shape_inference.infer_shapes(new_model)
return new_model
def model_simplify(onnx_model, simplify_model, simplify_hw):
#onnx_model = onnx.load(model_path)
is_dynamic_input_shape = False
init_list = []
input_shapes_ = {}
for init in onnx_model.graph.initializer:
init_list.append(init.name)
h = -1
w = -1
for input_ in onnx_model.graph.input:
#print('graph_input_name:', input_.name)
if input_.name not in init_list:
if len(input_.type.tensor_type.shape.dim) > 0:
input_shape = input_.type.tensor_type.shape.dim
input_shape = [x.dim_value for x in input_shape]
if len(input_shape) < 2:
continue
h = input_shape[-2]
w = input_shape[-1]
dynamic_input_shape_ = any(d==-1 or d==0 for d in input_shape)
if dynamic_input_shape_ == True:
is_dynamic_input_shape = True
logger.info('The model input is dynamic, input: {} {}'.format(input_.name, input_shape))
input_shape[0] = 1
if simplify_hw != '':
hw_list = simplify_hw.split(',')
input_shape[-1] = int(hw_list[1])
input_shape[-2] = int(hw_list[0])
logger.debug('input_shape: {}'.format(input_shape))
input_shapes_[input_.name] = input_shape
logger.info('----- input_shapes_: {}'.format(input_shapes_))
skip_constant_folding_ = False
if h <= 0 or w <= 0:
skip_constant_folding_ = True
if simplify_model == 2:
if is_dynamic_input_shape == True:
model_simp, check = simplify(onnx_model, input_shapes=input_shapes_, skip_constant_folding=skip_constant_folding_)
if simplify_hw == '':
#correct_batch_for_opset_convert(model_simp)
correct_output_shape(model_simp)
try:
model_simp = reset_model_value_info(model_simp)
except Exception as e:
logger.warning('Cannot do reset_value operation~')
else:
model_simp, check = simplify(onnx_model, dynamic_input_shape=False)
else:
model_simp, check = simplify(onnx_model, dynamic_input_shape=is_dynamic_input_shape, skip_constant_folding=skip_constant_folding_)
#onnx.save(model_simp, model_path)
if model_simp.producer_version != '':
model_simp.producer_version = model_simp.producer_version + '(simplified by macaConverter)'
else:
model_simp.producer_name = model_simp.producer_name + '(simplified by macaConverter)'
return model_simp
def modify_onnx2dynamic(onnx_model):
for idx in range(len(onnx_model.graph.input)):
if len(onnx_model.graph.input[idx].type.tensor_type.shape.dim) > 0:
dim_proto_input = onnx_model.graph.input[idx].type.tensor_type.shape.dim[0]
# dim_proto_input.dim_param = 'bs'
dim_proto_input.dim_value = -1
for idx in range(len(onnx_model.graph.value_info)):
if len(onnx_model.graph.value_info[idx].type.tensor_type.shape.dim) > 0:
logger.debug('value info name: {}'.format(onnx_model.graph.value_info[idx].name))
dim_proto_input = onnx_model.graph.value_info[idx].type.tensor_type.shape.dim[0]
# dim_proto_input.dim_param = 'bs'
dim_proto_input.dim_value = -1
for idx in range(len(onnx_model.graph.output)):
if len(onnx_model.graph.output[idx].type.tensor_type.shape.dim):
dim_proto_output = onnx_model.graph.output[idx].type.tensor_type.shape.dim[0]
# dim_proto_output.dim_param = 'bs'
dim_proto_output.dim_value = -1
### for Reshape
reshape_param = []
for node_id, node in enumerate(onnx_model.graph.node):
#print(node_id, ", name:", node.name, ", input:", node.input, ", output:", node.output, \
# ", op:", node.op_type, ', len(input):', len(node.input))
if node.op_type == 'Reshape':
logger.debug('Reshape, input: {}'.format(node.input))
if node.input[1] not in reshape_param:
reshape_param.append(node.input[1])
for n in reshape_param:
for init in onnx_model.graph.initializer:
logger.info('loop init.name: {}'.format(init.name))
if n == init.name:
logger.info('got it in initializer: {} {}'.format(n, init.int64_data))
#init.int64_data[0] = -1
dtype = init.data_type
np_dtype = convert_ort_type_2_np(dtype)
if init.raw_data:
params_list = np.fromstring(init.raw_data, dtype=np_dtype)
logger.debug('len(params_list): {}'.format(len(params_list)))
adjust = True
for val in params_list:
if val == -1:
adjust = False
if adjust == True and params_list[0] != -1:
params_list[0] = -1
init.raw_data = params_list.tostring()
else:
data_list = get_data_list(dtype, init)
adjust = True
logger.debug('len(data_list): {}'.format(len(data_list)))
for val in data_list:
if val == -1:
adjust = False
if adjust == True and len(data_list) > 0 and data_list[0] != -1:
data_list[0] = -1
############# for constant node
for n in reshape_param:
for node in onnx_model.graph.node:
if node.op_type == 'Constant':
if node.output[0] == n:
logger.info('got constant output: {}'.format(node.output))
attributes = node.attribute
for attr in attributes:
if attr.name == 'value':
v = values.get_tensor_value(attr.t)
#print('got type v:', type(v))
adjust = True
for val in v:
if val == -1:
adjust = False
break
if adjust == True:
v[0] = -1
vv = [v_ for v_ in v]
#print('-----new vv:', vv, type(vv))
if isinstance(v, np.ndarray) == True:
values.set_tensor_value(attr.t, v)
else:
values.set_tensor_value(attr.t, vv)
########################
#onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
try:
onnx.checker.check_model(onnx_model)
except onnx.checker.ValidationError as e:
print('*** The model cannot be modified for: %s' % e)
if 'No Op registered for Mish' in str(e):
logger.warning('ignore mish warning, continue saving~')
else:
logger.error('ERROR: check model failed in modify_onnx2dynamic')
#sys.exit(exit_code_check_modify_onnx2dynamic)
else:
logger.info('*** The model is modified!')
return onnx_model
def convert_gap_2_ap(model):
#model = onnx.load(onnxfile)
node_list = []
has_global_average_pool = False
need_convert = False
for node_id, node in enumerate(model.graph.node):
#print(node_id, ", name:", node.name, ", input:", node.input, ", output:", node.output, \
# ", op:", node.op_type)
if node.op_type == 'GlobalAveragePool':
has_global_average_pool = True
dict={'id':node_id, 'input':node.input, 'output':node.output, 'op':node.op_type}
node_list.append(dict)
if has_global_average_pool == True:
node_list2 = []
for v in model.graph.value_info:
input_shape = v.type.tensor_type.shape.dim
input_shape = [x.dim_value for x in input_shape]
if len(input_shape) >= 4:
dict2 = {'name':v.name, 'shape':input_shape}
node_list2.append(dict2)
#print("+++++++++++ name:", v.name, input_shape)
for d in node_list:
if d['op'] == 'GlobalAveragePool':
logger.debug('op id: {}, op input: {}'.format(d['id'], d['input']))
for v in node_list2:
#print('v.name:', v['name'])
if d['input'][0] == v['name']:
logger.debug('got GlobalAveragePool, shape: {} {}'.format(v['shape'], v['name']))
if v['shape'][2] <= 15 and v['shape'][3] <= 15:
need_convert = True
logger.info('GlobalAveragePool===>AveragePool......')
old_node = model.graph.node[d['id']]
model.graph.node.remove(old_node)
new_node = onnx.helper.make_node(
name = '',
op_type="AveragePool",
inputs=d['input'],
outputs=d['output'],
kernel_shape=[v['shape'][2], v['shape'][3]],
)
model.graph.node.insert(d['id'], new_node)
if need_convert == True:
#onnx.checker.check_model(model)
try:
onnx.checker.check_model(model)
except onnx.checker.ValidationError as e:
print('The model cannot be saved for: %s' % e)
if 'No Op registered for Mish' in str(e):
logger.warning('ignore mish warning, continue saving~')
else:
logger.error('ERROR: check model failed in convert_gap_2_ap')
sys.exit(exit_code_check_convert_gap_2_ap)
else:
logger.info('+++ Begin saving model...')
#onnx.save(model, onnxfile)
return need_convert
def post_process(new_model, inference_success, gap_to_ap):
start_time = time.time()
debug_print = False
delete = optimization_op(new_model)
while delete == True:
debug_print = True
delete = optimization_op(new_model)
end_time1 = time.time()
if debug_print == True:
logger.info('optimization_op cost {} seconds'.format(end_time1 - start_time))
debug_print = False
if gap_to_ap == 1:
if inference_success == True:
debug_print = convert_gap_2_ap(new_model)
else:
logger.warning('Cannot do inference, so skip global_average_pool-->average_pool')
end_time2 = time.time()
if debug_print == True:
logger.info('convert_gap_2_ap cost {} seconds'.format(end_time2 - end_time1))
def my_extract_model(
input_path, # type: Text
output_path, # type: Text
input_names, # type: List[Text]
output_names # type: List[Text]
): # type: (...) -> None
"""Extracts sub-model from an ONNX model.
The sub-model is defined by the names of the input and output tensors *exactly*.
Note: For control-flow operators, e.g. If and Loop, the _boundary of sub-model_,
which is defined by the input and output tensors, should not _cut through_ the
subgraph that is connected to the _main graph_ as attributes of these operators.
Arguments:
input_path (string): The path to original ONNX model.
output_path (string): The path to save the extracted ONNX model.
input_names (list of string): The names of the input tensors that to be extracted.
output_names (list of string): The names of the output tensors that to be extracted.
"""
if not os.path.exists(input_path):
raise ValueError("Invalid input model path: %s" % input_path)
if not output_path:
raise ValueError("Output model path shall not be empty!")
if not output_names:
raise ValueError("Output tensor names shall not be empty!")
#onnx.checker.check_model(input_path)
try:
onnx.checker.check_model(input_path)
except onnx.checker.ValidationError as e:
print('Extract warning:: %s' % e)
else:
logger.info('~~~~ Begin extracting model...')
model = onnx.load(input_path)
e = onnx.utils.Extractor(model)
extracted = e.extract_model(input_names, output_names)
onnx.save(extracted, output_path)
#onnx.checker.check_model(output_path)
try:
onnx.checker.check_model(output_path)
except onnx.checker.ValidationError as e:
print('Extracted warning: %s' % e)
else:
logger.info('^^^^ Finish extracting model...')
return True
def extract_sub_graph(input_path, output_path, input_names, output_names):
logger.info('input_names: {}, output_names: {}'.format(input_names, output_names))
input_list = input_names.split(',')
output_list = output_names.split(',')
#onnx.utils.extract_model(input_path, output_path, input_list, output_list)
return my_extract_model(input_path, output_path, input_list, output_list)
'''
try:
onnx.utils.extract_model(input_path, output_path, input_list, output_list)
except BaseException as e:
print('The model cannot be extracted for: %s' % e)
return False
else:
print('Inference success---')
return True
'''
def process(args):
global support_mish
global inputs_as_nchw
global concat_result
model_path = args.model_path
model_type = args.model_type
output = args.output
concat_result = args.concat_result
op_set = args.op_set
input_shape = args.input_shape
inputs = args.inputs
outputs = args.outputs
simplify_model = args.simplify
extract_sub = args.extract_sub
dynamic_batch = args.dynamic_batch
fp32_to_fp16 = args.fp32_to_fp16
support_mish = args.support_mish
preproc_yaml = args.preproc_yaml
postproc_yaml = args.postproc_yaml
model_def_file = args.model_def_file
model_class_name = args.model_class_name
model_input_type = args.model_input_type
model_weights_file = args.model_weights_file
inputs_as_nchw = args.inputs_as_nchw
gap_to_ap = args.gap_to_ap
fuse_pad_pool = args.fuse_pad_pool
support_swish = args.support_swish
bn_to_conv = args.bn_to_conv
output_num = args.output_num
keep_batch = args.keep_batch
params_file = args.params_file
simplify_hw = args.simplify_hw
force_simplify = args.force_simplify
gemm_optimization = args.gemm_optimization
expand_to_resize = args.expand_to_resize
reset_value_info = args.reset_value_info
fuse_layernorm = args.fuse_layernorm
matmul_to_gemm = args.matmul_to_gemm
reset_batch = args.reset_batch
fuse_hard_sigmoid = args.fuse_hard_sigmoid
fuse_gelu = args.fuse_gelu
disable_all_optimizer = args.disable_all_optimizer
mha_optimization = args.mha_optimization
fp32_to_u8 = args.fp32_to_u8
ceil_floor_reverse = args.ceil_floor_reverse
if args.version:
print('maca_converter version:', MXC_CONFIG.VERSION)
print('last modified:', MXC_CONFIG.LAST_MODIFIED)
exit(0)
if model_path == None or model_type == None or output == None:
print('WARNING: model_path/model_type/output COULD NOT be null')
exit(-1)
if disable_all_optimizer == 1:
print('------- disable all optimazation')
support_mish = 0
gap_to_ap = 0
fuse_pad_pool = 0
support_swish = 0
bn_to_conv = 0
gemm_optimization = 0
expand_to_resize = 0
fuse_layernorm = 0
matmul_to_gemm = 0
fuse_hard_sigmoid = 0
fuse_gelu = 0
ceil_floor_reverse = 0
logger.info('model_path:{}, model_type:{}, output:{}'.format(model_path, model_type, output))
if model_type == 'tf-ckpt' or model_type == 'tf-graph' :
logger.debug('checkpoint: {} {}'.format(inputs, outputs))
logger.info('---input_shape: {}'.format(input_shape))
input_shape_list = input_shape.split('/')
logger.info('---input_shape_list: {}'.format(input_shape_list))
#input_shape = input_shape_list[0]
dynamic_paddle = False
if model_type == 'paddle':
from pd2onnx import is_dynamic_paddle
dynamic_paddle = is_dynamic_paddle(input_shape_list, model_def_file, model_class_name, model_weights_file)
#if dynamic_paddle == True and model_input_type == '':
# model_input_type = 'float32'
can_ignore_model_path = False
if model_type == 'pytorch':
#if '.' in model_class_name:
if model_weights_file != '':
can_ignore_model_path = True
if dynamic_paddle == False and can_ignore_model_path == False and not os.path.exists(model_path):
logger.error('ERROR: {} is not exist'.format(model_path))
sys.exit(exit_code_model_not_exist)
if model_type not in valid_model_type:
logger.error('Valid mode type is {}'.format(valid_model_type))
logger.error('ERROR: {} is not valid mode type'.format(model_type))
sys.exit(exit_code_invalid_model_type)
op_set_default = 11
if op_set != None and op_set < op_set_default:
op_set_default = op_set
if model_type == 'pytorch' and args.input_shape == '':
logger.warning('WARNNIG: when converting pytorch model, you must tell the input shape(ex: --input_shape [1, 3, 32, 32])')
logger.warning('WARNNIG: also, you should provide model definition file')
sys.exit(exit_code_pytorch_no_input_shape)
if (model_type == 'tf-ckpt' or model_type == 'tf-graph') and (args.inputs == '' or args.outputs == ''):
logger.warning('WARNNIG: When converting checkpoint/graph, you must tell the inputs(ex: --inputs input0:0,input1:0) and outputs(ex: --outputs output0:0)')
sys.exit(exit_code_tensorflow_no_inputs_or_outputs)
if extract_sub == 1:
if args.inputs == '' or args.outputs == '':
logger.warning('WARNNIG: When extract sub graph, you must tell the inputs(ex: --inputs input0:0,input1:0) and outputs(ex: --outputs output0:0)')
sys.exit(exit_code_extract_sub_no_inputs_or_outputs)
if model_type != 'onnx':
logger.warning('WARNNING: only onnx model supports extracting...')
sys.exit(exit_code_model_type_not_onnx)
r = extract_sub_graph(model_path, output, inputs, outputs)
if r == True:
logger.critical('Convert Success!')
sys.exit(exit_code_normal)
logger.info('begin convert..')
begin_time = time.time()
if model_type != 'onnx':
convert(model_path,
model_type,
output,
op_set_default,
input_shape_list,
inputs,
outputs,
model_def_file,
model_class_name,
model_input_type,
model_weights_file,
output_num,
keep_batch,
params_file,
ceil_floor_reverse)
end_time1 = time.time()
logger.info('finish convert, it cost {} seconds'.format(end_time1 - begin_time))
if model_type != 'onnx':
model = onnx.load(output)
else:
model = onnx.load(model_path)
if op_set != None :
if model_type == 'onnx':
logger.info('ONNX, add_value_info_for_constants...')
correct_batch_for_opset_convert(model)
operation.add_value_info_for_constants(model)
model = version_converter.convert_version(model, op_set)
elif op_set != op_set_default:
correct_batch_for_opset_convert(model)
operation.add_value_info_for_constants(model)
model = version_converter.convert_version(model, op_set)
operation.eliminate_unused_input_initializer(model)
inference_success = False
new_model = model
#new_model = onnx.shape_inference.infer_shapes(model)
try:
new_model = onnx.shape_inference.infer_shapes(model)
except BaseException as e:
print('The model cannot be inferenced for: %s' % e)
new_model = model
else:
logger.info('Inference success---')
inference_success = True
#onnx.checker.check_model(new_model)
try:
onnx.checker.check_model(new_model)
except BaseException as e: #onnx.checker.ValidationError as e:
logger.warning('ignore warning(check_model), continue saving~')
else:
logger.info('### Begin saving model...')
#onnx.save(new_model, output)
if dynamic_batch == 1:
logger.info('modify model to dynamic batch...')
new_model = modify_onnx2dynamic(new_model)
end_time2 = time.time()
logger.info('generate inference shape model, it cost {} seconds'.format(end_time2 - end_time1))
#post_process(new_model, inference_success, gap_to_ap)
if simplify_model == 1 or simplify_model == 2:
logger.info('begin doing simplify...')
producer_name = new_model.producer_name
producer_version = new_model.producer_version
simplify_flag = '(simplified by macaConverter)'
if simplify_flag not in producer_name and simplify_flag not in producer_version:
new_model = model_simplify(new_model, simplify_model, simplify_hw)
else:
if force_simplify != 0:
simplify_model = force_simplify
new_model = model_simplify(new_model, simplify_model, simplify_hw)
else:
logger.info('The model has been simplified by macaConverter, ignore this operation~~')
post_process(new_model, inference_success, gap_to_ap)
if mha_optimization == 1:
new_model = mha_optimizer(new_model)
if reset_batch != '':
batchs = reset_batch #.split(' ')
input_batch = int(batchs[0])
output_batch = input_batch
if len(batchs) >= 2:
output_batch = int(batchs[1])
if input_batch == 0:
input_batch = -1
if output_batch == 0:
output_batch = -1
logger.info('got batchs: {}'.format(batchs))
new_model = reset_batch_size(new_model, input_batch, output_batch)
if fuse_pad_pool == 1:
logger.info('begin doing fuse_pad_to_pool...')
new_model = fuse.fuse_pad_to_pool(new_model)
'''
if fp32_to_fp16 == 1:
print('begin doing fp32-->fp16...')
new_model = convert_float_to_float16(new_model, keep_io_types=True)
'''
#if model_type == 'onnx' and support_mish == 1:
if support_mish == 1:
new_model = merge_mish(new_model)
#if model_type == 'onnx' and support_swish == 1:
if support_swish == 1:
new_model = merge_swish_and_hard_swish(new_model)
#if model_type == 'onnx' and gemm_optimization == 1:
if gemm_optimization == 1:
new_model = gemm_convert(new_model)
if model_type == 'onnx' and preproc_yaml != '':
if os.path.exists(preproc_yaml):
new_model = preproc(new_model, preproc_yaml)
else:
logger.warning('pre_proc yaml file {} is not exist'.format(preproc_yaml))
if model_type == 'onnx' and postproc_yaml != '':
if os.path.exists(postproc_yaml):
new_model = postproc(new_model, postproc_yaml)
else:
logger.warning('post_proc yaml file {} is not exist'.format(postproc_yaml))
if bn_to_conv == 1 and simplify_model != 0:
new_model = bn2conv.bn2conv(new_model)
#if model_type == 'onnx' and expand_to_resize == 1:
if expand_to_resize == 1:
new_model = merge_resize(new_model)
if model_type == 'onnx' and reset_value_info == 1:
new_model = reset_model_value_info(new_model)
#if model_type == 'onnx' and fuse_layernorm == 1:
if fuse_layernorm == 1:
new_model = merge_layernorm(new_model)
#if model_type == 'onnx' and matmul_to_gemm == 1:
if matmul_to_gemm == 1:
new_model = matmul_2_gemm(new_model)
if fuse_hard_sigmoid == 1:
new_model = merge_hard_sigmiod(new_model)
if fuse_gelu== 1:
new_model = merge_gelu(new_model)
if fp32_to_u8 == 1:
new_model = fp32_to_uint8(new_model)
if fp32_to_fp16 == 1:
logger.info('begin doing fp32-->fp16...')
new_model = convert_float_to_float16(new_model, keep_io_types=True)
delete = operation.eliminate_redundant_reshape(new_model)
while delete == True:
delete = operation.eliminate_redundant_reshape(new_model)
operation.eliminate_unused_input_initializer(new_model)
operation.eliminate_unused_constant_node(new_model)
operation.remove_unused_initializer(new_model)
onnx.save(new_model, output)
end_time3 = time.time()
logger.info('The whole progress cost {} seconds'.format(end_time3 - begin_time))
logger.critical('Convert Success!')
def usage():
print('python model_convert.py --model_path ./my_model --model_type caffe --output ./c2o.onnx')
print('or')
print('python model_convert.py --model_path ./my_model \
--model_type tf-h5 \
--output ./t2o.onnx \
--op_set 12 \
--datatype_convert fp32_to_fp16 \
--q_dataset_file ./quantization.npy \
--q_onnx_file ./quantization.onnx')
def main(args):
#clear log file
with open("./convert.log", 'r+') as file:
file.truncate(0)
process(args)
if __name__ == "__main__":
args = parse_args()
main(args)
此处可能存在不合适展示的内容，页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容，可点击提交进行申诉，我们将尽快为您处理。
马建仓 AI 助手