model-optimizer/mo/pipeline/mx.py

   1 """
   2  Copyright (c) 2018-2019 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 from extensions.back.CreateConstNodes import CreateConstNodesReplacement
  17 from extensions.front.restore_ports import RestorePorts
  18 from mo.middle.pattern_match import for_graph_and_each_sub_graph_recursively
  19 from mo.utils.error import Error, FrameworkError
  20 from mo.utils.utils import refer_to_faq_msg
  21
  22 try:
  23     import mxnet
  24 except ImportError:
  25     raise Error('Module mxnet was not found. Please install appropriate version of mxnet via install_prerequisites '
  26                 'script.' + refer_to_faq_msg(52))
  27
  28 import argparse
  29
  30 from mo.front.extractor import extract_node_attrs, remove_output_ops
  31 from mo.front.mxnet.extractor import mxnet_op_extractor
  32 from mo.front.mxnet.loader import symbol2nx, load_symbol_def
  33 from mo.middle.passes.fusing.decomposition import convert_batch_norm, convert_scale_shift_to_mul_add
  34 from mo.middle.passes.conv import convert_muladd_to_scaleshift_or_power, \
  35     convert_add_or_mul_to_scaleshift, fuse_pad
  36 from mo.middle.passes.eliminate import graph_clean_up, remove_const_ops
  37 from mo.middle.passes.fusing.fuse_linear_ops import fuse_linear_ops
  38 from mo.middle.passes.fusing.fuse_linear_seq import fuse_mul_add_sequence
  39 from mo.middle.passes.fusing.mark_unfused_nodes import mark_unfused_nodes
  40 from mo.middle.passes.fusing.resnet_optimization import stride_optimization
  41 from mo.middle.passes.infer import convert_mul_add_to_power
  42 from mo.middle.passes.mean_scale_values import move_scaleshift_to_preprocess
  43 from mo.middle.passes.shape import reverse_input_channels
  44 from mo.pipeline.common import prepare_emit_ir
  45 from mo.front.mxnet.nd_to_params import save_params_file
  46 from mo.front.common.register_custom_ops import update_extractors_with_extensions
  47 from mo.front.mxnet.extractor import mxnet_op_extractors
  48 from mo.utils import class_registration
  49 from mo.utils.cli_parser import get_meta_info
  50 from extensions.middle.EltwiseInputNormalization import EltwiseInputNormalize
  51
  52
  53 def driver(argv: argparse.Namespace, input_model: str, output_model_name: str, output_dir: str):
  54     meta_info = get_meta_info(argv)
  55
  56     try:
  57         model_nodes, model_params, model_name, iteration_number = load_symbol_def(input_model, argv.input_symbol,
  58                                                                                   argv.input,
  59                                                                                   argv.nd_prefix_name,
  60                                                                                   argv.pretrained_model_name,
  61                                                                                   argv.legacy_mxnet_model)
  62     except (ValueError, mxnet.base.MXNetError) as e:
  63         raise FrameworkError(
  64             'The following error happened while loading mxnet model {}: {}. ' +
  65             refer_to_faq_msg(53),
  66             input_model,
  67             str(e)
  68         ) from e
  69
  70     if argv.nd_prefix_name and argv.pretrained_model_name and argv.save_params_from_nd:
  71         save_params_file(model_name, model_params._arg_params, model_params._aux_params, iteration_number)
  72
  73     update_extractors_with_extensions(mxnet_op_extractors)
  74     graph = symbol2nx(model_nodes, model_params, argv.input)
  75     graph.check_empty_graph('symbol2nx. It may happen due to problems with loaded model')
  76
  77     graph.__setattr__('name', output_model_name)
  78     graph.graph['layout'] = 'NCHW'
  79     graph.graph['cmd_params'] = argv
  80     graph.graph['fw'] = 'mxnet'
  81     graph.graph['feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3
  82     graph.graph['ir_version'] = 2 if argv.generate_deprecated_IR_V2 else 5
  83     extract_node_attrs(graph, mxnet_op_extractor)
  84
  85     # --------------------------------- LOAD END ------------------------------------------------------
  86     class_registration.apply_replacements(graph, class_registration.ClassType.FRONT_REPLACER)
  87     class_registration.apply_replacements(graph, class_registration.ClassType.MIDDLE_REPLACER)
  88
  89     fuse_pad(graph)
  90
  91     # Mark nodes with attr 'can_be_fused': False to disable fusing for specified nodes
  92     mark_unfused_nodes(graph, argv.finegrain_fusing)
  93
  94     # Converting FusedBatchNorm layer to Mul->Add->Mul->Add sequence
  95     convert_batch_norm(graph)
  96     graph_clean_up(graph)
  97
  98     if not argv.disable_fusing:
  99         # Converting ScaleShift layer to Mul->Add
 100         convert_scale_shift_to_mul_add(graph)
 101         graph_clean_up(graph)
 102
 103         # Fusing the sequences of Mul/Add operations
 104         fuse_mul_add_sequence(graph)
 105         graph_clean_up(graph)
 106
 107         # Fusing linear operation to Convolution
 108         fuse_linear_ops(graph)
 109         graph_clean_up(graph)
 110
 111     if not argv.disable_resnet_optimization:
 112         stride_optimization(graph)
 113
 114     fuse_pad(graph)
 115
 116     # Converting Mul->Add to ScaleShift node
 117     convert_muladd_to_scaleshift_or_power(graph)
 118     graph_clean_up(graph)
 119
 120     convert_mul_add_to_power(graph)
 121     graph_clean_up(graph)
 122     convert_add_or_mul_to_scaleshift(graph)  # scale = 1
 123     graph_clean_up(graph)
 124
 125     if argv.reverse_input_channels:
 126         reverse_input_channels(graph)
 127
 128     if argv.move_to_preprocess:
 129         move_scaleshift_to_preprocess(graph)
 130         graph_clean_up(graph)
 131
 132     pattern = EltwiseInputNormalize()
 133     pattern.find_and_replace_pattern(graph)
 134
 135     class_registration.apply_replacements(graph, class_registration.ClassType.BACK_REPLACER)
 136
 137     for_graph_and_each_sub_graph_recursively(graph, remove_const_ops)
 138     CreateConstNodesReplacement().find_and_replace_pattern(graph)
 139
 140     for_graph_and_each_sub_graph_recursively(graph, remove_output_ops)
 141
 142     prepare_emit_ir(graph=graph, data_type=argv.data_type, output_dir=output_dir, output_model_name=output_model_name,
 143                     meta_info=meta_info)
 144     return 0