model-optimizer/mo/pipeline/onnx.py

   1 """
   2  Copyright (c) 2018 Intel Corporation
   3
   4  Licensed under the Apache License, Version 2.0 (the "License");
   5  you may not use this file except in compliance with the License.
   6  You may obtain a copy of the License at
   7
   8       http://www.apache.org/licenses/LICENSE-2.0
   9
  10  Unless required by applicable law or agreed to in writing, software
  11  distributed under the License is distributed on an "AS IS" BASIS,
  12  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  See the License for the specific language governing permissions and
  14  limitations under the License.
  15 """
  16
  17 from __future__ import absolute_import
  18 from __future__ import division
  19 from __future__ import print_function
  20 from __future__ import unicode_literals
  21
  22 import argparse
  23 import copy
  24 import logging as log
  25
  26 import onnx
  27 import os
  28
  29 import numpy as np
  30
  31 from mo.front.common.custom_replacement_registry import CustomReplacementRegistry
  32 from mo.front.common.find_unsupported_ops import find_unsupported_ops
  33 from mo.front.common.register_custom_ops import check_for_duplicates
  34 from mo.front.common.register_custom_ops import update_extractors_with_extensions
  35 from mo.front.extractor import restore_edges, add_output_ops, add_input_ops, \
  36     extract_node_attrs, create_tensor_nodes, remove_output_ops, user_data_repack
  37 from mo.front.onnx.extractor import common_onnx_fields, onnx_op_extractor, onnx_op_extractors
  38 from mo.front.onnx.loader import load_onnx_model, protobuf2nx
  39 from mo.middle.passes.conv import convert_add_to_scaleshift, \
  40     convert_weights_yxio_to_oiyx, convert_weights_yxio_to_goiyx, convert_gemm_to_fully_connected, \
  41     convert_muladd_to_scaleshift_or_power, fuse_pad, transpose_fully_connected_weights, \
  42     convert_dilated_convolution, convert_mul_to_scaleshift, convert_nasnet
  43 from mo.middle.passes.fusing.decomposition import convert_batch_norm, convert_scale_shift_to_mul_add
  44 from mo.middle.passes.eliminate import graph_clean_up, remove_op_nodes, remove_useless_split
  45 from mo.middle.passes.fusing.fuse_linear_ops import fuse_linear_ops
  46 from mo.middle.passes.fusing.fuse_linear_seq import fuse_mul_add_sequence
  47 from mo.middle.passes.fusing.fuse_grouped_conv import grouped_convolutions_fusing
  48 from mo.middle.passes.fusing.mark_unfused_nodes import mark_unfused_nodes
  49 from mo.middle.passes.mean_scale_values import move_scaleshift_to_preprocess
  50 from mo.middle.passes.infer import scale_input, override_placeholder_shapes, partial_infer, convert_mul_add_to_power, \
  51     update_fully_connected_shapes, add_mean_scale_values, override_batch
  52 from mo.middle.passes.l2normalization import l2_norm_to_norm
  53 from mo.middle.passes.pool import mean_to_avgpool
  54 from mo.middle.passes.shape import convert_squeeze, convert_reshape, convert_nhwc_to_nchw, reverse_input_channels, \
  55     conv_flatten_concat, fuse_sequence_of_reshapes
  56 from mo.utils import class_registration
  57 from mo.pipeline.common import prepare_emit_ir
  58 from mo.utils.custom_replacement_config import update_custom_replacement_config_file
  59 from mo.utils.error import Error
  60 from mo.utils.utils import refer_to_faq_msg
  61
  62
  63 def driver(argv: argparse.Namespace, model_file_name: str, output_model_name: str, outputs: list, output_dir: str,
  64           scale: float,
  65           user_shapes: [None, list, np.array] = None,
  66           mean_scale_values: [dict, list] = ()):
  67
  68     model_proto = load_onnx_model(model_file_name)
  69     model_graph = model_proto.graph
  70     #print(model_graph)
  71     #assert len(model_graph) == 1, "An ONNX model contains more than 1 graph: unsupported"
  72     log.debug("Number of nodes in graph_def: {}".format(len(model_graph.node)))
  73     log.debug("Number of all input ports (not true inputs) in graph_def: {}".format(len(model_graph.input)))
  74     log.debug("Number of initializers in graph_def: {}".format(len(model_graph.initializer)))
  75     log.debug("Number of real inputs in graph_def: {}".format(len(model_graph.input) - len(model_graph.initializer)))
  76     update_extractors_with_extensions(onnx_op_extractors)
  77
  78     try:
  79         graph = protobuf2nx(model_proto)
  80         log.debug("Number of nodes in NX graph: {}".format(graph.number_of_nodes()))
  81         graph.__setattr__('name', output_model_name if output_model_name else model_proto.graph.name)
  82         graph.graph['layout'] = 'NCHW'
  83         graph.graph['cmd_params'] = argv
  84         graph.graph['fw'] = 'onnx'
  85         graph.graph['feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3
  86         # extract basic attributes earlier to enable some passes that relies on them before full attribute
  87         # extractor is called
  88         extract_node_attrs(graph, lambda node: (True, common_onnx_fields(node)))
  89     except Exception as e:
  90         raise Error(
  91             'Cannot pre-process ONNX graph after reading from model file "{}". ' \
  92             'File is corrupt or has unsupported format. Details: {}. ' +
  93             refer_to_faq_msg(44),
  94             model_file_name,
  95             str(e)
  96         ) from e
  97
  98     user_shapes, outputs, _ = user_data_repack(graph, user_shapes, outputs, None)
  99
 100     graph, output_op_nodes = add_output_ops(graph, outputs)
 101     graph, input_op_nodes = add_input_ops(graph, user_shapes, True)
 102
 103     # this call of 'graph_clean_up' removes child nodes of outputs which is useful when custom output is specified
 104     graph_clean_up(graph)
 105
 106     extract_node_attrs(graph, lambda node: onnx_op_extractor(node, check_for_duplicates(onnx_op_extractors)))
 107
 108     class_registration.apply_replacements(graph, class_registration.ClassType.FRONT_REPLACER)
 109
 110     create_tensor_nodes(graph)
 111     graph_clean_up(graph)
 112
 113     override_placeholder_shapes(graph, user_shapes)
 114     override_batch(graph, argv.batch)
 115
 116     graph_clean_up(graph)
 117     remove_op_nodes(graph, {'op': 'Identity'})
 118
 119     graph_clean_up(graph)
 120
 121     remove_output_ops(graph)
 122
 123     partial_infer(graph)
 124     graph_clean_up(graph)
 125
 126
 127     graph, input_op_nodes = add_input_ops(graph, user_shapes, False)
 128     graph_clean_up(graph)
 129
 130     #change_placeholders_types_to_FP32(graph)
 131
 132     scale_input(graph, scale)
 133     add_mean_scale_values(graph, mean_scale_values)
 134
 135     convert_dilated_convolution(graph)
 136     graph_clean_up(graph)
 137
 138     graph_clean_up(graph)
 139
 140     remove_op_nodes(graph, {'op': 'Identity'})
 141     remove_useless_split(graph)
 142
 143     class_registration.apply_replacements(graph, class_registration.ClassType.MIDDLE_REPLACER)
 144
 145     convert_gemm_to_fully_connected(graph)
 146
 147     fuse_pad(graph)
 148     graph_clean_up(graph)
 149
 150     # Mark nodes with attr 'can_be_fused': False to disable fusing for specified nodes
 151     mark_unfused_nodes(graph, argv.finegrain_fusing)
 152
 153     # Converting FusedBatchNorm layer to Mul->Add->Mul->Add sequence
 154     # IE doesn't support BN with 4 inputs, so we have to split it to two ScaleShift
 155     convert_batch_norm(graph)
 156     graph_clean_up(graph)
 157
 158     if not argv.disable_fusing:
 159         # Converting ScaleShift layer to Mul->Add
 160         convert_scale_shift_to_mul_add(graph)
 161         graph_clean_up(graph)
 162
 163         # Fusing the sequences of Mul/Add operations
 164         fuse_mul_add_sequence(graph)
 165         graph_clean_up(graph)
 166
 167         # Fusing linear operation to Convolution
 168         fuse_linear_ops(graph)
 169         graph_clean_up(graph)
 170
 171     if not argv.disable_gfusing:
 172         grouped_convolutions_fusing(graph)
 173         graph_clean_up(graph)
 174         if not argv.disable_fusing:
 175             fuse_linear_ops(graph)
 176             graph_clean_up(graph)
 177
 178     convert_muladd_to_scaleshift_or_power(graph)
 179     graph_clean_up(graph)
 180
 181     convert_mul_add_to_power(graph)
 182
 183     # Need to eliminate dead nodes before doing update_fully_connected_shapes
 184     # because update_fully_connected_shapes does partial inference and dead
 185     # nodes will lead to sporadic failures.
 186     graph_clean_up(graph)
 187     update_fully_connected_shapes(graph)
 188
 189     convert_reshape(graph)
 190     convert_add_to_scaleshift(graph)  # scale = 1
 191     convert_mul_to_scaleshift(graph)  # biases = 0
 192
 193     fuse_pad(graph)
 194     graph_clean_up(graph)
 195
 196     if argv.reverse_input_channels:
 197         reverse_input_channels(graph)
 198
 199     if argv.move_to_preprocess:
 200         move_scaleshift_to_preprocess(graph)
 201         graph_clean_up(graph)
 202
 203     fuse_sequence_of_reshapes(graph)
 204     graph_clean_up(graph)
 205
 206     class_registration.apply_replacements(graph, class_registration.ClassType.BACK_REPLACER)
 207
 208     prepare_emit_ir(graph=graph, data_type=argv.data_type, output_dir=output_dir, output_model_name=output_model_name)
 209
 210     return 0