def tf2nx(argv: argparse.Namespace, model_file_name: str, output_model_name: str, outputs: list, output_dir: str, scale: float, is_binary: bool, user_shapes: [None, list, np.array] = None, mean_scale_values: [dict, list] = ()): """ Convert TF GraphDef object to NetworkX representation. The resulting graph is still TF-specific and needs normalization passes to be applied. The specific TF structure assumes each GraphDef node is converted to a single NetworkX node, node id is an original TF node name, and edges go directly from one op to another op. """ meta_info = get_meta_info(argv) if argv.tensorflow_custom_layer_libraries: libraries = argv.tensorflow_custom_layer_libraries.split(',') for library in libraries: log.info('Loading library "{}" with custom operations'.format(library)) tf.load_op_library(library) graph_def, variables_values = load_tf_graph_def(graph_file_name=model_file_name, is_binary=is_binary, checkpoint=argv.input_checkpoint, user_output_node_names_list=outputs, model_dir=argv.saved_model_dir, meta_graph_file=argv.input_meta_graph, saved_model_tags=argv.saved_model_tags) try: tf.import_graph_def(graph_def, name='') except: log.warning("TensorFlow post-processing of loaded model was unsuccessful. " "This is an optional step that Model Optimizer performs for any input model but it is not usually " "required for all models." "It likely means that the original model is ill-formed. " "Model Optimizer will continue converting this model.") log.debug("Number of nodes in graph_def: {}".format(len(graph_def.node))) # pylint: disable=no-member if argv.tensorboard_logdir: tensorboard.dump_for_tensorboard(graph_def, argv.tensorboard_logdir) update_extractors_with_extensions(tf_op_extractors) try: graph = protobuf2nx(graph_def) graph.__setattr__('name', output_model_name) # 'layout' parameter change may cause an issue in EltwiseInputReshape replacer # and convert_nhwc_to_nchw(graph) graph.graph['layout'] = 'NCHW' if argv.disable_nhwc_to_nchw else 'NHWC' graph.graph['cmd_params'] = argv graph.graph['fw'] = 'tf' graph.graph['ir_version'] = 2 if argv.generate_deprecated_IR_V2 else 4 if graph.graph['ir_version'] == 2: # When the deprecated IR version was requested, # we configure only those phases that can lead to # functional regressions in the version 2. # BasicLSTMCell is one such transformation; when it is turned off, # the body of TF basic_lstm_cell is converted as-is in a decomposed form, # and should work in version 2. BasicLSTMCell.enabled = False # placeholder for request from a transformation pass to repeat the entire conversion graph.graph['repeat_conversion'] = False graph = restore_edges(graph, get_tf_edges) graph = remove_control_dependency_inputs(graph) # extract basic attributes earlier to enable some passes that relies on them before full attribute # extractor is called extract_node_attrs(graph, lambda node: (True, common_tf_fields(node))) except Exception as e: raise Error( 'Cannot pre-process TensorFlow graph after reading from model file "{}". ' \ 'File is corrupt or has unsupported format. Details: {}. ' + refer_to_faq_msg(44), model_file_name, str(e) ) from e check_empty_graph(graph, 'protobuf2nx. It may happen due to problems with loaded model') packed_user_shapes, packed_outputs, freeze_placeholder = user_data_repack(graph, user_shapes, outputs, argv.freeze_placeholder_with_value) if freeze_placeholder is not None: FreezePlaceholderValue.enabled = True FreezePlaceholderValue.replacement_dict = freeze_placeholder update_registration() GemmResolver.enabled = False inputs = list(packed_user_shapes.keys()) if packed_user_shapes is not None and isinstance(packed_user_shapes, dict) else None graph.graph['inputs'] = inputs # save user defined inputs for other extensions output_op_nodes = add_output_ops(graph, packed_outputs, inputs=packed_user_shapes) input_op_nodes = add_input_ops(graph, packed_user_shapes, True) # this call of 'graph_clean_up' removes child nodes of outputs which is useful when custom output is specified graph_clean_up_tf(graph) check_empty_graph(graph, 'add_output_ops and add_input_ops. It may happen due to absence of \'Placeholder\' layer ' 'in the model') variables_to_constants(graph, variables_values) del variables_values graph_clean_up_tf(graph) if argv.tensorflow_custom_operations_config_update: if update_custom_replacement_config_file(graph, argv.tensorflow_custom_operations_config_update): return 0 else: return 1 unsupported_ops_to_offload_to_tf = list() MAX_ITERATIONS = 5 cur_iteration = 0 while cur_iteration < MAX_ITERATIONS: graph_copy = copy.deepcopy(graph) # create a copy of graph for the case when some ops are unsupported if argv.tensorflow_subgraph_patterns is not None: csc.replace_subgraph_calls(graph, argv.tensorflow_subgraph_patterns) if argv.tensorflow_operation_patterns is not None: csc.offload_operations_to_tf(graph, argv.tensorflow_operation_patterns) if argv.offload_unsupported_operations_to_tf and len(unsupported_ops_to_offload_to_tf): csc.offload_unsupported_operations_to_tf(graph, unsupported_ops_to_offload_to_tf) extract_node_attrs(graph, lambda node: tf_op_extractor(node, check_for_duplicates(tf_op_extractors))) if argv.tensorflow_use_custom_operations_config is not None: registry = CustomReplacementRegistry() registry.add_custom_replacement_description_from_config(argv.tensorflow_use_custom_operations_config) # automatically generate sub-classes for custom replacements that replace sub-graph with a single node for replacement_desc in registry.get_all_replacements_descriptions(): if replacement_desc.has('op'): type('FrontReplacementFromConfigFileOp' + replacement_desc.op, (FrontReplacementFromConfigFileOp,), {'replacement_id': replacement_desc.id}) update_registration() override_placeholder_shapes(graph, packed_user_shapes) # the user shapes are used to convert TensorFlow Object Detection API models graph.graph['user_shapes'] = packed_user_shapes class_registration.apply_replacements(graph, class_registration.ClassType.FRONT_REPLACER) override_batch(graph, argv.batch) create_tensor_nodes(graph) graph_clean_up_tf(graph) remove_output_ops(graph) partial_infer(graph) delete_control_flow_edges(graph) replacer = AddIsCyclicAttribute() replacer.find_and_replace_pattern(graph) # TENSOR ITERATOR CREATING BEGINS if graph.graph['is_cyclic']: replacer = DeleteSelect() replacer.find_and_replace_pattern(graph) replacer = SmartInputMatcher() replacer.find_and_replace_pattern(graph) replacer = SmartOutputMatcher() replacer.find_and_replace_pattern(graph) replacer = LoopConditionMatcher() replacer.find_and_replace_pattern(graph) replacer = SimpleConditionMather() replacer.find_and_replace_pattern(graph) replacer = BackEdgesMatching() replacer.find_and_replace_pattern(graph) replacer = ConditionChecks() replacer.find_and_replace_pattern(graph) delete_not_executable(graph) graph_clean_up_tf(graph) if graph.graph['is_cyclic']: replacer = SimpleInputMatcher() replacer.find_and_replace_pattern(graph) replacer = BackEdgeSimpleInputMatcher() replacer.find_and_replace_pattern(graph) # Here will be optimizing path (ops after Enter and before body take out of body) replacer = TensorIteratorMerge() replacer.find_and_replace_pattern(graph) # TENSOR ITERATOR CREATING ENDS check_for_cycle(graph) for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf) check_empty_graph(graph, 'partial_infer') csc.prepare_tf_call_nodes(graph) graph_clean_up_tf(graph) duplicate_shared_weights(graph) input_op_nodes = add_input_ops(graph, packed_user_shapes, False) graph_clean_up_tf(graph) check_empty_graph(graph, 'add_input_ops') change_placeholders_types_to_FP32(graph) scale_input(graph, scale) add_mean_scale_values(graph, mean_scale_values) convert_dilated_convolution(graph) for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf) l2_norm_to_norm(graph) graph_clean_up_tf(graph) remove_op_nodes(graph, {'identity': True}) remove_useless_split(graph) class_registration.apply_replacements(graph, class_registration.ClassType.MIDDLE_REPLACER) mean_to_avgpool(graph) convert_nasnet(graph) fuse_pad(graph) graph_clean_up_tf(graph) convert_matmul_to_fully_connected(graph) # Mark nodes with attr 'can_be_fused': False to disable fusing for specified nodes for_graph_and_each_sub_graph_recursively(graph, lambda graph: mark_unfused_nodes(graph, argv.finegrain_fusing)) # Converting FusedBatchNorm layer to Mul->Add->Mul->Add sequence # IE doesn't support BN with 4 inputs, so we have to split it to two ScaleShift convert_batch_norm(graph) graph_clean_up_tf(graph) if not argv.disable_fusing: # Converting ScaleShift layer to Mul->Add for_graph_and_each_sub_graph_recursively(graph, convert_scale_shift_to_mul_add) for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf) # Fusing the sequences of Mul/Add operations for_graph_and_each_sub_graph_recursively(graph, fuse_mul_add_sequence) for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf) # Fusing linear operation to Convolution for_graph_and_each_sub_graph_recursively(graph, fuse_linear_ops) for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf) if not argv.disable_gfusing: grouped_convolutions_fusing(graph) graph_clean_up_tf(graph) if not argv.disable_fusing: fuse_linear_ops(graph) graph_clean_up_tf(graph) # Converting Mul->Add to ScaleShift node for_graph_and_each_sub_graph_recursively(graph, convert_muladd_to_scaleshift_or_power) for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf) for_graph_and_each_sub_graph_recursively(graph, convert_mul_add_to_power) # Need to eliminate dead nodes before doing update_fully_connected_shapes # because update_fully_connected_shapes does partial inference and dead # nodes will lead to sporadic failures. for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf) for_graph_and_each_sub_graph_recursively(graph, update_fully_connected_shapes) for_graph_and_each_sub_graph_recursively(graph, convert_mul_eltwise_to_leaky_relu) graph_clean_up_tf(graph) for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf) for_graph_and_each_sub_graph_recursively(graph, fuse_pad) for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf) for_graph_and_each_sub_graph_recursively(graph, convert_reshape) for_graph_and_each_sub_graph_recursively(graph, convert_squeeze) for_graph_and_each_sub_graph_recursively(graph, convert_add_to_scaleshift) # scale = 1 for_graph_and_each_sub_graph_recursively(graph, convert_mul_to_scaleshift) # biases = 0 if argv.reverse_input_channels: reverse_input_channels(graph) if argv.move_to_preprocess: move_scaleshift_to_preprocess(graph) graph_clean_up_tf(graph) for_graph_and_each_sub_graph_recursively(graph, fuse_sequence_of_reshapes) pattern = EltwiseInputNormalize() pattern.find_and_replace_pattern(graph) conv_flatten_concat(graph) for_graph_and_each_sub_graph_recursively(graph, apply_nhwc_to_nchw_permutation) for_graph_and_each_sub_graph_recursively(graph, merge_nodes_permutations) for_graph_and_each_sub_graph_recursively(graph, permute_data_nodes_attrs) for_graph_and_each_sub_graph_recursively(graph, permute_op_nodes_attrs) for_graph_and_each_sub_graph_recursively(graph, repack_fully_connected_weights_nhwc_to_nchw) for_graph_and_each_sub_graph_recursively(graph, transpose_fully_connected_weights) for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf) if argv.offload_unsupported_operations_to_tf: unsupported_ops_to_offload_to_tf = find_unsupported_ops(graph) if len(unsupported_ops_to_offload_to_tf) == 0: log.info('All operations are supported! Exit from the loop.') if not need_to_repeat_conversion(graph): break else: print('After {} iteration there are {} unsupported ops'.format(cur_iteration + 1, len(unsupported_ops_to_offload_to_tf))) else: if not need_to_repeat_conversion(graph): break graph = graph_copy cur_iteration += 1 class_registration.apply_replacements(graph, class_registration.ClassType.BACK_REPLACER) prepare_emit_ir(graph=graph, data_type=argv.data_type, output_dir=output_dir, output_model_name=output_model_name, meta_info=meta_info) return 0
def find_and_replace_pattern(self, graph: Graph): graph.stage = 'middle' graph.strict_mode = False create_tensor_nodes(graph) graph.strict_mode = True
def driver_R5(onnx_modelproto_bytes, precision: str, output_model_name: str, outputs: list, output_dir: str, scale: float, user_shapes: [None, list, np.array] = None, mean_scale_values: [dict, list] = ()): try: model_proto = onnx.load_from_string(bytes(onnx_modelproto_bytes)) except Exception as e: print("[python] onnx exception: ", str(e)) model_graph = model_proto.graph # pylint: disable=no-member log.debug("Number of nodes in graph_def: {}".format(len(model_graph.node))) log.debug( "Number of all input ports (not true inputs) in graph_def: {}".format( len(model_graph.input))) log.debug("Number of initializers in graph_def: {}".format( len(model_graph.initializer))) log.debug("Number of real inputs in graph_def: {}".format( len(model_graph.input) - len(model_graph.initializer))) update_extractors_with_extensions(onnx_op_extractors) try: graph = protobuf2nx(model_proto) log.debug("Number of nodes in NX graph: {}".format( graph.number_of_nodes())) graph.__setattr__( 'name', output_model_name if output_model_name else model_proto.graph.name) # pylint: disable=no-member graph.graph['layout'] = 'NCHW' graph.graph['fw'] = 'onnx' graph.graph[ 'feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3 graph.graph['ir_version'] = 4 extract_node_attrs(graph, lambda node: (True, common_onnx_fields(node))) except Exception as e: raise Error( 'Cannot pre-process ONNX graph after reading from model file "{}". ' 'File is corrupt or has unsupported format. Details: {}. ' + refer_to_faq_msg(44), model_file_name, str(e)) from e check_empty_graph( graph, 'protobuf2nx. It may happen due to problems with loaded model') packed_user_shapes, packed_outputs, _ = user_data_repack( graph, user_shapes, outputs, None) output_op_nodes = add_output_ops(graph, packed_outputs) input_op_nodes = add_input_ops(graph, packed_user_shapes, True) graph_clean_up(graph) check_empty_graph(graph, 'add_output_ops and add_input_ops') extract_node_attrs( graph, lambda node: onnx_op_extractor( node, check_for_duplicates(onnx_op_extractors))) class_registration.apply_replacements( graph, class_registration.ClassType.FRONT_REPLACER) create_tensor_nodes(graph) graph_clean_up(graph) override_placeholder_shapes(graph, packed_user_shapes) graph_clean_up(graph) remove_op_nodes(graph, {'op': 'Identity'}) graph_clean_up(graph) remove_output_ops(graph) partial_infer(graph) graph_clean_up(graph) check_empty_graph(graph, 'partial_infer') input_op_nodes = add_input_ops(graph, packed_user_shapes, False) graph_clean_up(graph) check_empty_graph(graph, 'add_input_ops') scale_input(graph, scale) add_mean_scale_values(graph, mean_scale_values) convert_dilated_convolution(graph) graph_clean_up(graph) graph_clean_up(graph) remove_op_nodes(graph, {'op': 'Identity'}) remove_useless_split(graph) class_registration.apply_replacements( graph, class_registration.ClassType.MIDDLE_REPLACER) convert_gemm_to_fully_connected(graph) NormalizeFullyConnected().find_and_replace_pattern(graph) fuse_pad(graph) graph_clean_up(graph) convert_batch_norm(graph) graph_clean_up(graph) convert_scale_shift_to_mul_add(graph) graph_clean_up(graph) fuse_mul_add_sequence(graph) graph_clean_up(graph) fuse_linear_ops(graph) graph_clean_up(graph) grouped_convolutions_fusing(graph) graph_clean_up(graph) fuse_linear_ops(graph) graph_clean_up(graph) convert_muladd_to_scaleshift_or_power(graph) graph_clean_up(graph) convert_mul_add_to_power(graph) graph_clean_up(graph) convert_reshape(graph) convert_add_to_scaleshift(graph) # scale = 1 convert_mul_to_scaleshift(graph) # biases = 0 fuse_pad(graph) graph_clean_up(graph) fuse_sequence_of_reshapes(graph) graph_clean_up(graph) pattern = EltwiseInputNormalize() pattern.find_and_replace_pattern(graph) merge_nodes_permutations(graph) permute_data_nodes_attrs(graph) permute_op_nodes_attrs(graph) class_registration.apply_replacements( graph, class_registration.ClassType.BACK_REPLACER) weights, xml_string = prepare_emit_ir(graph=graph, data_type=precision, output_dir=output_dir, output_model_name=output_model_name, meta_info={'unset': []}) return weights, xml_string
def driver(argv: argparse.Namespace, proto_file_name: str, model_file_name: str, output_model_name: str, outputs: list, output_dir: str, scale: float, user_shapes: [None, list, np.array] = None, mean_scale_values: [dict, list] = (), mean_file: str = "", mean_file_offsets: tuple = None, custom_layers_mapping_path: str = None): meta_info = get_meta_info(argv) FusePermutesSequence.enabled = False proto, model = loader.load_caffe_proto_model(proto_file_name, model_file_name) update_extractors_with_extensions( caffe_type_extractors, argv.disable_omitting_optional if hasattr( argv, 'disable_omitting_optional') else False, argv.disable_flattening_optional_params if hasattr( argv, 'disable_flattening_optional_params') else False) try: graph, original_shapes = loader.caffe_pb_to_nx(proto, model) except ValueError as e: raise Error( 'Invalid prototxt file: value error {}. ' + refer_to_faq_msg(11), str(e)) from e log.debug("After caffe_pb_to_nx") print_graph_stat(graph) check_empty_graph(graph, 'load_caffe_proto_model') graph.__setattr__('proto_path', proto_file_name) graph.__setattr__('caffemodel_path', model_file_name) graph.__setattr__('name', getattr(proto, 'name', None) or output_model_name) graph.graph['layout'] = 'NCHW' graph.graph['cmd_params'] = argv graph.graph['fw'] = 'caffe' graph.graph['ir_version'] = 2 if argv.generate_deprecated_IR_V2 else 4 extract_node_attrs(graph, lambda node: (True, common_caffe_fields(node))) log.debug("After adding specific nodes for outputs") print_graph_stat(graph) custom_layers_map = custom_layers_mapping.load_layers_xml( custom_layers_mapping_path) custom_layers_mapping.update_extractors( caffe_type_extractors, custom_layers_map, argv.disable_omitting_optional if hasattr( argv, 'disable_omitting_optional') else False, argv.enable_flattening_nested_params if hasattr( argv, 'enable_flattening_nested_params') else False) extract_node_attrs( graph, lambda node: caffe_extractor( node, check_for_duplicates(caffe_type_extractors))) log.debug("After extract_node_attr") print_graph_stat(graph) packed_user_shapes, packed_outputs, freeze_placeholder = user_data_repack( graph, user_shapes, outputs, argv.freeze_placeholder_with_value) if argv.freeze_placeholder_with_value is not None: FreezePlaceholderValue.enabled = True FreezePlaceholderValue.replacement_dict = freeze_placeholder class_registration.update_registration([FrontReplacementSubgraph]) output_op_nodes = add_output_ops(graph, packed_outputs) input_op_nodes = add_input_ops(graph, packed_user_shapes, True) override_placeholder_shapes(graph, packed_user_shapes) override_batch(graph, argv.batch) graph_clean_up(graph) check_empty_graph(graph, 'add_output_ops and add_input_ops') class_registration.apply_replacements( graph, class_registration.ClassType.FRONT_REPLACER) graph = create_tensor_nodes(graph) log.debug("After create_tensor_nodes") print_graph_stat(graph) remove_op_nodes(graph, {'op': 'Identity'}) remove_output_ops(graph) graph_clean_up(graph) log.debug("After removing specific nodes for output") print_graph_stat(graph) # you need to pass required network outputs here # but we don't have a way yet, so just passing all discovered sinks mark_outputs(graph) graph_clean_up(graph) log.debug("After graph_cleanup") print_graph_stat(graph) graph = partial_infer(graph) log.debug("After partial_infer") print_graph_stat(graph) check_empty_graph(graph, 'partial_infer') duplicate_shared_weights(graph) input_op_nodes = add_input_ops(graph, packed_user_shapes, False) graph_clean_up(graph) check_empty_graph(graph, 'add_input_ops') scale_input(graph, scale) add_mean_scale_values(graph, mean_scale_values) log.debug("Split multi input convolutions") convert_multi_input_conv(graph) graph_clean_up(graph) log.debug("After graph_cleanup") print_graph_stat(graph) remove_op_nodes(graph, {'op': 'Dropout'}) remove_op_nodes(graph, {'phase': 0}) graph_clean_up(graph) class_registration.apply_replacements( graph, class_registration.ClassType.MIDDLE_REPLACER) mean_to_avgpool(graph) # Mark nodes with attr 'can_be_fused': False to disable fusing for specified nodes mark_unfused_nodes(graph, argv.finegrain_fusing) #need this pass even without fusing to convert scale with 2 inputs convert_scale_shift_to_mul_add(graph) graph_clean_up(graph) if not argv.disable_fusing: convert_bn_to_mul_add(graph) graph_clean_up(graph) fuse_mul_add_sequence(graph) graph_clean_up(graph) fuse_linear_ops(graph) graph_clean_up(graph) if not argv.disable_resnet_optimization: stride_optimization(graph) convert_muladd_to_scaleshift_or_power(graph) convert_matmul_to_fully_connected(graph) batch_norm_fuse(graph) convert_mul_add_to_power(graph) convert_add_to_scaleshift(graph) # scale = 1 convert_mul_to_scaleshift(graph) # biases = 0 graph_clean_up(graph) log.debug("After graph_cleanup") print_graph_stat(graph) if argv.reverse_input_channels: reverse_input_channels(graph) if argv.move_to_preprocess: move_scaleshift_to_preprocess(graph) graph_clean_up(graph) fuse_sequence_of_reshapes(graph) input_names = find_inputs(graph) mf = [] try: if mean_file and len(original_shapes) == 1: mf = loader.parse_mean(mean_file, original_shapes[input_names[0]], mean_file_offsets) elif mean_file: raise Error( 'Mean file for topologies with multiple inputs is not supported. ' + refer_to_faq_msg(9)) except ValueError as e: raise Error( 'Cannot load or process mean file: value error {}. ' + refer_to_faq_msg(10), str(e)) from e class_registration.apply_replacements( graph, class_registration.ClassType.BACK_REPLACER) prepare_emit_ir(graph=graph, data_type=argv.data_type, output_dir=output_dir, output_model_name=output_model_name, mean_data=mf, input_names=input_names, meta_info=meta_info) return 0
def driver(argv: argparse.Namespace, input_model: str, output_model_name: str, outputs: list, output_dir: str, scale: float, placeholder_shapes: [None, list, np.array] = None, mean_scale_values: [dict, list] = ()): meta_info = get_meta_info(argv) try: model_nodes, model_params, model_name, iteration_number = load_symbol_def(input_model, argv.input_symbol, argv.input, argv.nd_prefix_name, argv.pretrained_model_name, argv.legacy_mxnet_model) except (ValueError, mxnet.base.MXNetError) as e: raise FrameworkError( 'The following error happened while loading mxnet model {}: {}. ' + refer_to_faq_msg(53), input_model, str(e) ) from e if argv.nd_prefix_name and argv.pretrained_model_name and argv.save_params_from_nd: save_params_file(model_name, model_params._arg_params, model_params._aux_params, iteration_number) update_extractors_with_extensions(mxnet_op_extractors) graph = symbol2nx(model_nodes, model_params, argv.input) check_empty_graph(graph, 'symbol2nx. It may happen due to problems with loaded model') graph.__setattr__('name', output_model_name) graph.graph['layout'] = 'NCHW' graph.graph['cmd_params'] = argv graph.graph['fw'] = 'mxnet' graph.graph['feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3 graph.graph['ir_version'] = 2 if argv.generate_deprecated_IR_V2 else 4 graph = extract_node_attrs(graph, mxnet_op_extractor) check_softmax_node_inputs(graph) user_shapes, packed_outputs, _ = user_data_repack(graph, placeholder_shapes, outputs, None) output_op_nodes = add_output_ops(graph, packed_outputs) input_op_nodes = add_input_ops(graph, user_shapes, True) try: override_placeholder_shapes(graph, user_shapes, argv.batch) except ValueError as err: raise Error( 'The following error happened while processing input shapes: {}. ' + refer_to_faq_msg(54), str(err) ) from err check_empty_graph(graph, 'add_output_ops and add_input_ops') class_registration.apply_replacements(graph, class_registration.ClassType.FRONT_REPLACER) add_input_data_to_prior_boxes(graph, argv.input) graph = create_tensor_nodes(graph) graph_clean_up(graph) remove_output_ops(graph) mark_outputs(graph) remove_output_ops(graph) graph_clean_up(graph) log.debug("After removing specific nodes for output") print_graph_stat(graph) graph = partial_infer(graph) graph_clean_up(graph) check_empty_graph(graph, 'partial_infer') duplicate_shared_weights(graph) scale_input(graph, scale) add_mean_scale_values(graph, mean_scale_values) remove_op_nodes(graph, {'identity': True}) graph_clean_up(graph) class_registration.apply_replacements(graph, class_registration.ClassType.MIDDLE_REPLACER) fuse_pad(graph) # Mark nodes with attr 'can_be_fused': False to disable fusing for specified nodes mark_unfused_nodes(graph, argv.finegrain_fusing) # Converting FusedBatchNorm layer to Mul->Add->Mul->Add sequence convert_batch_norm(graph) graph_clean_up(graph) if not argv.disable_fusing: # Converting ScaleShift layer to Mul->Add convert_scale_shift_to_mul_add(graph) graph_clean_up(graph) # Fusing the sequences of Mul/Add operations fuse_mul_add_sequence(graph) graph_clean_up(graph) # Fusing linear operation to Convolution fuse_linear_ops(graph) graph_clean_up(graph) if not argv.disable_resnet_optimization: stride_optimization(graph) fuse_pad(graph) # Converting Mul->Add to ScaleShift node convert_muladd_to_scaleshift_or_power(graph) graph_clean_up(graph) convert_mul_add_to_power(graph) convert_add_to_scaleshift(graph) # scale = 1 convert_mul_to_scaleshift(graph) # biases = 0 if argv.reverse_input_channels: reverse_input_channels(graph) if argv.move_to_preprocess: move_scaleshift_to_preprocess(graph) graph_clean_up(graph) pattern = EltwiseInputNormalize() pattern.find_and_replace_pattern(graph) class_registration.apply_replacements(graph, class_registration.ClassType.BACK_REPLACER) prepare_emit_ir(graph=graph, data_type=argv.data_type, output_dir=output_dir, output_model_name=output_model_name, meta_info=meta_info) return 0
def find_and_replace_pattern(self, graph: Graph): create_tensor_nodes(graph) graph.stage = 'middle'
def driver(argv, input_model, output_model_name, outputs, output_dir, scale, placeholder_shapes=None, mean_scale_values=()): meta_info = get_meta_info(argv) EltwiseChecker.enabled = False try: graph, input_shapes = load_kaldi_model(input_model) except Exception as e: raise Error('Model Optimizer is not able to read Kaldi model {}. '. format(input_model) + refer_to_faq_msg(91)) from e check_empty_graph(graph, 'load_kaldi_nnet_model') graph.graph['cmd_params'] = argv graph.graph['fw'] = 'kaldi' graph.graph['ir_version'] = 2 if argv.generate_deprecated_IR_V2 else 4 update_extractors_with_extensions(kaldi_type_extractors) extract_node_attrs(graph, lambda node: kaldi_extractor(node)) class_registration.apply_replacements( graph, class_registration.ClassType.FRONT_REPLACER) output_op_nodes = add_output_ops( graph, outputs) # TODO pass real outputs instead of None log.debug("After adding specific nodes for outputs") print_graph_stat(graph) check_empty_graph(graph, 'add_output_ops') create_tensor_nodes(graph) graph_clean_up(graph) log.debug("After removing specific nodes for output") print_graph_stat(graph) override_placeholder_shapes(graph, placeholder_shapes) override_batch(graph, argv.batch) graph_clean_up(graph) log.debug("After setting input shapes") print_graph_stat(graph) graph_clean_up(graph) remove_output_ops(graph) log.debug("After removing specific nodes for output") print_graph_stat(graph) # You need to pass required network outputs here # but we don't have a way yet, so just passing all discovered sinks mark_outputs(graph) graph_clean_up(graph) log.debug("After graph_cleanup") print_graph_stat(graph) graph = partial_infer(graph) # The order is intentional, firstly eliminate repeated, then remove redundant FuseRepeatedReshapes().find_and_replace_pattern(graph) EliminateRedundantReshape().find_and_replace_pattern(graph) check_empty_graph(graph, 'partial_infer') if argv.counts: try: counts = read_counts_file(argv.counts) except Exception as e: raise Error('Model Optimizer is not able to read counts file {}'. format(argv.counts) + refer_to_faq_msg(92)) from e apply_biases_to_last_layer(graph, counts) if argv.remove_output_softmax: RemoveLastSoftMaxPattern().find_and_replace_pattern(graph) graph_clean_up(graph) log.debug("After removing softmax") print_graph_stat(graph) # Intentionally after all transformations KaldiRemoveMemoryOutputBackReplacementPattern().find_and_replace_pattern( graph) prepare_emit_ir(graph, argv.data_type, output_dir, output_model_name, meta_info=meta_info) return 0
def driver(argv: argparse.Namespace, model_file_name: str, output_model_name: str, outputs: list, output_dir: str, scale: float, user_shapes: [None, list, np.array] = None, mean_scale_values: [dict, list] = ()): meta_info = get_meta_info(argv) model_proto = load_onnx_model(model_file_name) model_graph = model_proto.graph # pylint: disable=no-member #print(model_graph) #assert len(model_graph) == 1, "An ONNX model contains more than 1 graph: unsupported" log.debug("Number of nodes in graph_def: {}".format(len(model_graph.node))) log.debug( "Number of all input ports (not true inputs) in graph_def: {}".format( len(model_graph.input))) log.debug("Number of initializers in graph_def: {}".format( len(model_graph.initializer))) log.debug("Number of real inputs in graph_def: {}".format( len(model_graph.input) - len(model_graph.initializer))) update_extractors_with_extensions(onnx_op_extractors) try: graph = protobuf2nx(model_proto) log.debug("Number of nodes in NX graph: {}".format( graph.number_of_nodes())) graph.__setattr__( 'name', output_model_name if output_model_name else model_proto.graph.name) # pylint: disable=no-member graph.graph['layout'] = 'NCHW' graph.graph['cmd_params'] = argv graph.graph['fw'] = 'onnx' graph.graph[ 'feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3 graph.graph['ir_version'] = 2 if argv.generate_deprecated_IR_V2 else 4 # extract basic attributes earlier to enable some passes that relies on them before full attribute # extractor is called extract_node_attrs(graph, lambda node: (True, common_onnx_fields(node))) except Exception as e: raise Error( 'Cannot pre-process ONNX graph after reading from model file "{}". ' \ 'File is corrupt or has unsupported format. Details: {}. ' + refer_to_faq_msg(44), model_file_name, str(e) ) from e check_empty_graph( graph, 'protobuf2nx. It may happen due to problems with loaded model') packed_user_shapes, packed_outputs, _ = user_data_repack( graph, user_shapes, outputs, None) output_op_nodes = add_output_ops(graph, packed_outputs) input_op_nodes = add_input_ops(graph, packed_user_shapes, True) # this call of 'graph_clean_up' removes child nodes of outputs which is useful when custom output is specified graph_clean_up(graph) check_empty_graph(graph, 'add_output_ops and add_input_ops') extract_node_attrs( graph, lambda node: onnx_op_extractor( node, check_for_duplicates(onnx_op_extractors))) class_registration.apply_replacements( graph, class_registration.ClassType.FRONT_REPLACER) create_tensor_nodes(graph) graph_clean_up(graph) override_placeholder_shapes(graph, packed_user_shapes) override_batch(graph, argv.batch) graph_clean_up(graph) remove_op_nodes(graph, {'op': 'Identity'}) graph_clean_up(graph) remove_output_ops(graph) partial_infer(graph) graph_clean_up(graph) check_empty_graph(graph, 'partial_infer') input_op_nodes = add_input_ops(graph, packed_user_shapes, False) graph_clean_up(graph) check_empty_graph(graph, 'add_input_ops') #change_placeholders_types_to_FP32(graph) scale_input(graph, scale) add_mean_scale_values(graph, mean_scale_values) convert_dilated_convolution(graph) graph_clean_up(graph) graph_clean_up(graph) remove_op_nodes(graph, {'op': 'Identity'}) remove_useless_split(graph) class_registration.apply_replacements( graph, class_registration.ClassType.MIDDLE_REPLACER) convert_gemm_to_fully_connected(graph) NormalizeFullyConnected().find_and_replace_pattern(graph) fuse_pad(graph) graph_clean_up(graph) # Mark nodes with attr 'can_be_fused': False to disable fusing for specified nodes mark_unfused_nodes(graph, argv.finegrain_fusing) # Converting FusedBatchNorm layer to Mul->Add->Mul->Add sequence # IE doesn't support BN with 4 inputs, so we have to split it to two ScaleShift convert_batch_norm(graph) graph_clean_up(graph) if not argv.disable_fusing: # Converting ScaleShift layer to Mul->Add convert_scale_shift_to_mul_add(graph) graph_clean_up(graph) # Fusing the sequences of Mul/Add operations fuse_mul_add_sequence(graph) graph_clean_up(graph) # Fusing linear operation to Convolution fuse_linear_ops(graph) graph_clean_up(graph) if not argv.disable_gfusing: grouped_convolutions_fusing(graph) graph_clean_up(graph) if not argv.disable_fusing: fuse_linear_ops(graph) graph_clean_up(graph) convert_muladd_to_scaleshift_or_power(graph) graph_clean_up(graph) convert_mul_add_to_power(graph) graph_clean_up(graph) convert_reshape(graph) convert_add_to_scaleshift(graph) # scale = 1 convert_mul_to_scaleshift(graph) # biases = 0 fuse_pad(graph) graph_clean_up(graph) if argv.reverse_input_channels: reverse_input_channels(graph) if argv.move_to_preprocess: move_scaleshift_to_preprocess(graph) graph_clean_up(graph) fuse_sequence_of_reshapes(graph) graph_clean_up(graph) pattern = EltwiseInputNormalize() pattern.find_and_replace_pattern(graph) merge_nodes_permutations(graph) permute_data_nodes_attrs(graph) permute_op_nodes_attrs(graph) class_registration.apply_replacements( graph, class_registration.ClassType.BACK_REPLACER) prepare_emit_ir(graph=graph, data_type=argv.data_type, output_dir=output_dir, output_model_name=output_model_name, meta_info=meta_info) return 0