def roipooling_infer(node: Node): """ Sets shape of output node according specified parameters input blobs and node Sets number from the first input blob, channels from the second one, height and width are specified Parameters ---------- node """ shapes = [node.in_node(i).shape for i in range(len(node.in_nodes()))] if any(s is None for s in shapes): return if len(node.in_nodes()) == 4: # TensorFlow case of CropAndResize operation crop_size = node.in_node(3).value if crop_size is None: log.error('The ROIPooling size is not known for node {}'.format( node.soft_get('name'))) return if not isinstance(crop_size, np.ndarray) or len(crop_size) != 2: log.error( 'The ROIPooling size is should have 2 elements for node {}'. format(node.soft_get('name'))) node.pooled_h = crop_size[0] node.pooled_w = crop_size[1] node.graph.remove_edge(node.in_node(3).id, node.id) node.graph.remove_edge(node.in_node(2).id, node.id) layout = node.graph.graph['layout'] assert len(layout) == 4 node.out_node().shape = shape_for_layout( layout, batch=shapes[1][get_batch_dim(layout, 4)], features=shapes[0][get_features_dim(layout, 4)], height=node.pooled_h, width=node.pooled_w)
def infer(node: Node): in_shape = node.in_port(0).data.get_shape() if in_shape.size != 4: raise Error( 'TensorFlow DepthToSpace operation is supported for 4D \'NHWC\' input layout only. ' 'Current input shape is \'{}\''.format(in_shape)) layout = node.graph.graph['layout'] N = in_shape[get_batch_dim(layout, 4)] H = in_shape[get_height_dim(layout, 4)] W = in_shape[get_width_dim(layout, 4)] C = in_shape[get_features_dim(layout, 4)] block_size = node['block_size'] if C is not dynamic_dimension and C % (block_size**2): raise Error( 'Feature dimensions of input tensor of DepthToSpace operation have to be divisible by square ' 'of DepthToSpace \'block_size\' parameter. Input tensor shape = {}. Feature dimension = {}. ' 'block_size = {}'.format(in_shape, C, block_size)) out_shape = shape_for_layout(layout, batch=N, features=C // (block_size * block_size), height=H * block_size, width=W * block_size) if is_fully_defined(in_shape) and is_fully_defined( out_shape) and np.prod(in_shape) != np.prod(out_shape): raise Error( 'Number of input elements "{}" is not equal to number of output elements "" for node "{}"' ''.format(in_shape, out_shape, node.soft_get('name', node.id))) node.out_port(0).data.set_shape(out_shape)
def infer(node: Node): in_shape = node.in_node().shape if in_shape.size != 4: raise Error( 'TensorFlow SpaceToDepth operation is supported for 4D \'NHWC\' input layout only. ' 'Current input shape is \'{}\''.format(in_shape)) layout = node.graph.graph['layout'] N = in_shape[get_batch_dim(layout, 4)] H = in_shape[get_height_dim(layout, 4)] W = in_shape[get_width_dim(layout, 4)] C = in_shape[get_features_dim(layout, 4)] block_size = node['block_size'] if (H is not dynamic_dimension and H % block_size) or (W is not dynamic_dimension and W % block_size): raise Error( 'Spatial dimensions of input tensor of SpaceToDepth operation have to be divisible by ' 'SpaceToDepth \'block_size\' parameter. Input tensor shape = {}. Spatial dimensions = {},{}. ' 'block_size = {}'.format(in_shape, H, W, block_size)) out_shape = shape_for_layout(layout, batch=N, features=C * (block_size**2), height=H // block_size, width=W // block_size) node.out_port(0).data.set_shape(out_shape)
def upsample_infer(node: Node): layout = node.graph.graph['layout'] assert len(layout) == 4 input_shape = node.in_node(0).shape if input_shape is None: return if len(node.in_nodes()) == 1: in_height = input_shape[get_height_dim(layout, 4)] in_width = input_shape[get_width_dim(layout, 4)] assert node.has('width_scale') is not None and node.has( 'height_scale') is not None out_height = math.floor(in_height * node.height_scale) out_width = math.floor(in_width * node.width_scale) node.out_node().shape = shape_for_layout( layout, batch=input_shape[get_batch_dim(layout, 4)], features=input_shape[get_features_dim(layout, 4)], height=out_height, width=out_width) else: assert node.in_node(1).value is not None eps = 1e-5 # This is to make rounding in case of very close number to round to closest instead of down # generic output shape calculation to support 5D input shape case node.out_node().shape = np.array( (input_shape + eps) * node.in_node(1).value).astype(np.int64)
def upsample_infer(node: Node): layout = node.graph.graph['layout'] assert len(layout) == 4 input_shape = node.in_node(0).shape if input_shape is None: return if len(node.in_nodes()) == 1: in_height = input_shape[get_height_dim(layout, 4)] in_width = input_shape[get_width_dim(layout, 4)] assert node.has('width_scale') is not None and node.has( 'height_scale') is not None out_height = math.floor(in_height * node.height_scale) out_width = math.floor(in_width * node.width_scale) node.out_node().shape = shape_for_layout( layout, batch=input_shape[get_batch_dim(layout, 4)], features=input_shape[get_features_dim(layout, 4)], height=out_height, width=out_width) else: assert node.in_node(1).value is not None # generic output shape calculation to support 5D input shape case node.out_node().shape = input_shape * node.in_node(1).value
def find_and_replace_pattern(self, graph: Graph): layout = graph.graph['layout'] for eltwise_op_node in graph.get_op_nodes(is_eltwise=True): out_shape = eltwise_op_node.out_port().data.get_shape() if 4 <= len(out_shape) <= 5: out_features = out_shape[get_features_dim( layout, len(out_shape))] for port, node in eltwise_op_node.in_nodes().items(): if len(node.shape) != len(out_shape) and len( node.shape) == 1 and out_features == node.shape[0]: new_shape = shape_for_layout( layout, batch=1, features=out_features, height=1, width=1, depth=1 if len(out_shape) == 5 else None) dim_const = Const(graph, { 'value': new_shape, 'name': node.id + '/Dim' }).create_node() reshape_op = Reshape(graph, attrs={ 'dim': new_shape, 'name': node.id + '/Broadcast' }).create_node() eltwise_op_node.in_port(port).get_source( ).node.out_port(0).get_connection().set_destination( reshape_op.in_port(0)) reshape_op.in_port(1).connect(dim_const.out_port(0)) reshape_op.out_port(0).connect( eltwise_op_node.in_port(port))
def replace_pattern(self, graph: Graph, match: dict): bias_add = match['BiasAdd'] # Replace BiasAdd by Add operation new_add = Add(graph, {'name': bias_add.id + '/Add'}).create_node() bias_add.in_port(0).get_connection().set_destination( new_add.in_port(0)) bias_add.in_port(1).get_connection().set_destination( new_add.in_port(1)) bias_add.out_port(0).get_connection().set_source(new_add.out_port(0)) if bias_add.data_format != 'NCHW': return input_shape = new_add.in_port(0).data.get_shape() bias_shape = new_add.in_port(1).data.get_shape() assert len(bias_shape) == 1 unsqueeze_dims = np.arange(len(input_shape)) channel_dim = get_features_dim('NCHW', len(input_shape)) unsqueeze_dims = np.delete(unsqueeze_dims, channel_dim, 0) unsqueeze_node = Unsqueeze(graph, { 'name': new_add.id + '/BiasUnsqueeze' }).create_node() unsqueeze_dims_node = Const(graph, { 'name': new_add.id + '/Dims', 'value': unsqueeze_dims }).create_node() # Reconnecting nodes unsqueeze_node.in_port(1).connect(unsqueeze_dims_node.out_port(0)) unsqueeze_node['override_output_shape'] = True new_add.in_port(1).get_connection().insert_node(unsqueeze_node)
def pad_op_transform(graph: Graph, match: dict): op = match['op'] pad_op = match['pad_op'] input_data = pad_op.in_node(0) pads = pad_op.in_node(1).value if len( pad_op.in_nodes()) == 2 else pad_op.pads if pad_op.mode != 'constant': log.info( 'The pad node "{}" with pad mode "{}" cannot be fused.'.format( pad_op.soft_get('name'), pad_op.mode)) return if pad_op.mode == 'constant' and pad_op.fill_value != 0.0: log.info('The pad node "{}" with non-zero fill value cannot be fused.'. format(pad_op.soft_get('name'))) return input_tensor_dims = len(match['pad_output'].shape) if np.any(pads[get_features_dim(op.graph.graph['layout'], input_tensor_dims)] != 0) or \ np.any(pads[get_batch_dim(op.graph.graph['layout'], input_tensor_dims)] != 0): log.info( 'The pad node "{}" with padding over feature/batch dimension cannot be fused.' .format(pad_op.soft_get('name'))) return op.pad += pads op.pad_spatial_shape = op.pad[op.spatial_dims] op['auto_pad'] = None if op.type == 'Pooling': op['exclude_pad'] = False assert (graph[match['pad_output'].node][match['op'].node][0]['in'] == 0) edge_attrs = graph.get_edge_data(match['pad_output'].id, match['op'].id)[0] graph.remove_edge(match['pad_output'].id, match['op'].id) graph.add_edge(input_data.id, match['op'].id, **{'in': 0, **edge_attrs})
def infer(node: Node): in_shape = node.in_node().shape if in_shape.size != 4: raise Error( 'TensorFlow SpaceToDepth operation is supported for 4D \'NHWC\' input layout only. ' 'Current input shape is \'{}\''.format(in_shape)) layout = node.graph.graph['layout'] N = in_shape[get_batch_dim(layout, 4)] H = in_shape[get_height_dim(layout, 4)] W = in_shape[get_width_dim(layout, 4)] C = in_shape[get_features_dim(layout, 4)] block_size = node['block_size'] if H % block_size or W % block_size: raise Error( 'Spatial dimensions of input tensor of SpaceToDepth operation have to be divisible by ' 'SpaceToDepth \'block_size\' parameter. Input tensor shape = {}. Spatial dimensions = {},{}. ' 'block_size = {}'.format(in_shape, H, W, block_size)) out_shape = shape_for_layout(layout, batch=N, features=int(C * (block_size**2)), height=int(H / block_size), width=int(W / block_size)) assert np.prod(in_shape) == np.prod(out_shape) node.out_node().shape = int64_array(out_shape)
def infer(node: Node): in_shape = node.in_node().shape if in_shape.size != 4: raise Error( 'TensorFlow DepthToSpace operation is supported for 4D \'NHWC\' input layout only. ' 'Current input shape is \'{}\''.format(in_shape)) layout = node.graph.graph['layout'] N = in_shape[get_batch_dim(layout, 4)] H = in_shape[get_height_dim(layout, 4)] W = in_shape[get_width_dim(layout, 4)] C = in_shape[get_features_dim(layout, 4)] block_size = node['block_size'] if C % (block_size**2): raise Error( 'Feature dimensions of input tensor of DepthToSpace operation have to be divisible by square ' 'of DepthToSpace \'block_size\' parameter. Input tensor shape = {}. Feature dimension = {}. ' 'block_size = {}'.format(in_shape, C, block_size)) out_shape = shape_for_layout(layout, batch=N, features=int(C / (block_size**2)), height=int(H * block_size), width=int(W * block_size)) assert np.prod(in_shape) == np.prod(out_shape) node.out_node().shape = int64_array(out_shape)
def infer(node): layout = node.graph.graph['layout'] node_name = node.soft_get('name', node.id) assert len([port for port in node.in_ports().values() if not port.disconnected()]) == 3, \ 'The node "{}" must 3 inputs'.format(node_name) assert node.has_valid('pooled_w'), '"pooled_w" attribute is not set for node "{}"'.format(node_name) assert node.has_valid('pooled_h'), '"pooled_h" attribute is not set for node "{}"'.format(node_name) assert node.has_valid('mode'), '"mode" attribute is not set for node "{}"'.format(node_name) assert node.mode in ['avg', 'max'], \ '"mode" attribute range of values is ["avg", "max"], got {} for node "{}"'.format(node.mode, node_name) input_shape = node.in_port(0).data.get_shape() rois_shape = node.in_port(1).data.get_shape() indices_shape = node.in_port(2).data.get_shape() assert input_shape is not None and rois_shape is not None and indices_shape is not None, \ 'The node "{}" input shape is None'.format(node_name) assert rois_shape[0] == indices_shape[0], 'The number of batch indices does not correspond to number of ROIs ' \ 'for node "{}"'.format(node_name) assert rois_shape[1] == 4, 'The size of ROI element must be 4 for node "{}"'.format(node_name) assert len(input_shape) == 4, 'The rank of port 0 input tensor of node "{}" must be 4.'.format(node_name) node.out_port(0).data.set_shape( shape_for_layout(layout, batch=rois_shape[0], features=input_shape[get_features_dim(layout, 4)], height=node.pooled_h, width=node.pooled_w) )
def infer(node: None): input_shape = node.in_node(0).shape name = node.soft_get('name', node.id) if node.axes is not None and node.across_channels is not None: raise Error('Either axes or across_channels can be set for the MVN in node "{}".'.format(name)) if node.across_channels is None: if node.axes is not None: # normalizing (replacing -1 with actual index) axes_data_value = node.axes axes = [axes_data_value.item()] if axes_data_value.size == 1 else axes_data_value axes = [get_canonical_axis_index(input_shape, a) for a in axes] # deduce across_channels from the axes, e.g. if the first axis is included (assuming batch is zero axis) feature_dim = get_features_dim(node.graph.graph['layout'], len(input_shape)) \ if (4 <= len(input_shape) <= 5) \ else 1 node.across_channels = int(feature_dim in axes) if 0 in axes: raise Error('Reduction over the batch dimension in node "{}" ' 'is not supported by the backend.'.format(name)) for i in range(2, len(input_shape)): if i not in axes: raise Error( 'Reduction over spatial dimensions in node "{}" ' 'is obligatory for the backend.'.format(name)) else: node.across_channels = 0 # default copy_shape_infer(node)
def insert_pre_processing(graph: Graph, input_node: Node, node_mean_scale_values: np.array, preprocessing_name: str): assert preprocessing_name in ['scale', 'mean'] if node_mean_scale_values.get(preprocessing_name) is None: return user_value = node_mean_scale_values[preprocessing_name] value = 1 / user_value if preprocessing_name == 'scale' else user_value * (-1) optimize_value = int(preprocessing_name == 'scale') op = Mul if preprocessing_name == 'scale' else Add if all([x == optimize_value for x in value]): return assert input_node.has_valid('shape') features_dim_idx = get_features_dim(graph.graph['layout'], len(input_node.shape)) assert value.size == input_node.shape[features_dim_idx] or value.size == 1 shape = np.ones(len(input_node.shape), dtype=np.int64) shape[features_dim_idx] = value.size value = value.reshape(shape) name = input_node.soft_get('name', input_node.id) + '/' + preprocessing_name preprocessing = create_op_with_const_inputs(graph, op=op, port_value_dict={1: value}, op_attrs={'name': name}) for dst in input_node.out_port(0).get_destinations(): if dst.node.soft_get('type') != 'ShapeOf': dst.get_connection().set_source(preprocessing.out_port(0)) input_node.out_port(0).connect(preprocessing.in_port(0))
def infer(node: Node): assert [port.idx for port in node.in_ports().values() if not port.disconnected()] == [0], \ 'Wrong input nodes number for node {} with type ExtractImagePatches'.format(node.soft_get('name', node.id)) input_shape = node.in_port(0).data.get_shape() name = node.soft_get('name', node.id) assert input_shape is not None, 'Input shape is not set for node {} with type ExtractImagePatches'.format(name) assert len(input_shape) == 4, 'ExtractImagePatches operation supports only 4D tensors' layout = node.graph.graph['layout'] N = input_shape[get_batch_dim(layout, 4)] C = input_shape[get_features_dim(layout, 4)] size_spatial = int64_array(node.sizes)[node.spatial_dims] input_spatial_shape = input_shape[node.spatial_dims] stride_spatial_shape = node.strides[node.spatial_dims] size_extent = node.rates[node.spatial_dims] * (size_spatial - 1) + 1 pad_spatial_shape, output_spatial_shape = tf_window_op_pad_infer(input_spatial_shape, size_extent, stride_spatial_shape, node.auto_pad, False) out_shape = shape_for_layout(layout, batch=N, features=C * np.prod(size_spatial), height=output_spatial_shape[0], width=output_spatial_shape[1]) node.out_port(0).data.set_shape(int64_array(out_shape))
def infer(node: Node): layout = node.graph.graph['layout'] assert len(layout) == 4 assert len( [p for p in node.in_ports().values() if not p.disconnected()]) assert node.has_valid('mode') assert node.has_valid('axes') src_shape = node.in_port(0).data.get_shape() assert src_shape is not None dst_shape = node.in_port(1).data.get_value() assert dst_shape is not None out_height = dst_shape[0] out_width = dst_shape[1] node.out_node().shape = shape_for_layout( layout, batch=src_shape[get_batch_dim(layout, 4)], features=src_shape[get_features_dim(layout, 4)], height=out_height, width=out_width) PermuteAttrs.create_permute_attrs(node, attrs=[('axes', 'input:0')])
def insert_pre_processing(graph: Graph, input_node: Node, node_mean_scale_values: np.array, preprocessing_name: str): assert preprocessing_name in ['scale', 'mean'] if node_mean_scale_values.get(preprocessing_name) is None: return user_value = node_mean_scale_values[preprocessing_name] value = 1 / user_value if preprocessing_name == 'scale' else user_value * (-1) optimize_value = int(preprocessing_name == 'scale') op = Mul if preprocessing_name == 'scale' else Add if all([x == optimize_value for x in value]): return assert input_node.has_valid('shape') features_dim_idx = get_features_dim(graph.graph['layout'], len(input_node.shape)) assert value.size == input_node.shape[features_dim_idx] or value.size == 1 shape = np.ones(len(input_node.shape), dtype=np.int64) shape[features_dim_idx] = value.size value = value.reshape(shape) name = input_node.soft_get('name', input_node.id) + '/' + preprocessing_name preprocessing = create_op_with_const_inputs(graph, op=op, port_value_dict={1: value}, op_attrs={'name': name}) for dst in input_node.out_port(0).get_destinations(): if dst.node.soft_get('type') != 'ShapeOf': # After the insertion of additional operations model optimizer # should keep the link to the input layer. Parameter node in framework # should map to parameter node in IR. # For this reason 'fw_tensor_debug_info' should be kept in data node. dst.get_connection().set_source(preprocessing.out_port(0), "source") input_node.out_port(0).connect(preprocessing.in_port(0))
def resample_infer(node: Node): layout = node.graph.graph['layout'] assert len(layout) == 4 input_shape = node.in_node(0).shape if input_shape is None: return in_height = input_shape[get_height_dim(layout, 4)] in_width = input_shape[get_width_dim(layout, 4)] if node.has('fw') and node.fw == 'tf': dst_shape = node.in_node(1).value if dst_shape is None or len(input_shape) != 4 or len( dst_shape) != 2: log.error( 'Node {} with op {} cannot be converted to Resample layer because there is no enough info about ' 'src/dst shapes: src_shape = {}, dst_shape = {}'.format( node.name, node.op, input_shape, dst_shape)) node.type = None # prevent translation to a valid IE layer return out_height = dst_shape[0] out_width = dst_shape[1] node.graph.remove_edge(node.in_node(1).id, node.id) else: if len(node.in_nodes()) == 1: if node.has('width') and node.has('height'): out_height = node.height out_width = node.width else: out_height = node.factor * in_height out_width = node.factor * in_width else: out_height = node.in_node(1).shape[get_height_dim(layout, 4)] out_width = node.in_node(1).shape[get_width_dim(layout, 4)] node.factor = factor_update( node.factor, [float(out_height) / in_height, float(out_width) / in_width], [in_height, in_width], [out_height, out_width], node.soft_get('name')) node.out_node().shape = shape_for_layout( layout, batch=input_shape[get_batch_dim(layout, 4)], features=input_shape[get_features_dim(layout, 4)], height=out_height, width=out_width)
def find_and_replace_pattern(self, graph: Graph): layout = graph.graph['layout'] for n in list(graph.nodes()): if 'type' in graph.node[n] and graph.node[n]['type'] == 'Eltwise' and get_value_id(Node(graph, n)) is None: eltwise_op_node = Node(graph, n) out_shape = eltwise_op_node.out_node().shape if 4 <= len(out_shape) <= 5: out_features = out_shape[get_features_dim(layout, len(out_shape))] for port, node in eltwise_op_node.in_nodes().items(): if len(node.shape) != len(out_shape) and len(node.shape) == 1 and out_features == node.shape[0]: in_atts = deepcopy(graph.get_edge_data(node.id, n)[0]) graph.remove_edge(node.id, n) new_shape = shape_for_layout(layout, batch=1, features=out_features, height=1, width=1, depth=1 if len(out_shape) == 5 else None) reshape_data_op = Reshape(graph, attrs={'dim': new_shape, 'name': node.id + '/Broadcast'}) reshape_data_node = reshape_data_op.create_node_with_data([node]) graph.add_edge(reshape_data_node.id, eltwise_op_node.id, **in_atts)
def upsample_infer(node: Node): node_name = node.soft_get('name', node.id) layout = node.graph.graph['layout'] assert len( layout ) == 4, 'Input tensor rank must be equal to 4 for node "{}"'.format( node_name) input_shape = node.in_port(0).data.get_shape() if len(node.in_nodes()) == 1: in_height = input_shape[get_height_dim(layout, 4)] in_width = input_shape[get_width_dim(layout, 4)] assert node.has('width_scale') is not None and node.has( 'height_scale') is not None if in_height is not dynamic_dimension: out_height = math.floor(in_height * node.height_scale) else: out_height = dynamic_dimension if in_width is not dynamic_dimension: out_width = math.floor(in_width * node.width_scale) else: out_width = dynamic_dimension node.out_port(0).data.set_shape( shape_for_layout(layout, batch=input_shape[get_batch_dim(layout, 4)], features=input_shape[get_features_dim( layout, 4)], height=out_height, width=out_width)) else: scales = node.in_port(1).data.get_value() assert scales is not None, 'The input with scales for node "{}" is not constant'.format( node_name) eps = 1e-5 # This is to make rounding in case of very close number to round to closest instead of down # generic output shape calculation to support 5D input shape case output_shape = shape_array( [dynamic_dimension for _ in range(len(input_shape))]) for idx in range(len(output_shape)): if input_shape[idx] is not dynamic_dimension: output_shape[idx] = int( (input_shape[idx] + eps) * scales[idx]) else: output_shape[idx] = dynamic_dimension_value node.out_port(0).data.set_shape(output_shape)
def pad_op_transform(graph: Graph, match: dict): op = match['op'] pad_op = match['pad_op'] input_data = pad_op.in_node(0) if pad_op.mode != 'constant': log.info( 'The pad node "{}" with pad mode "{}" cannot be fused.'.format( pad_op.soft_get('name'), pad_op.mode)) return if op.type == 'Pooling' and op.pool_method == 'max': return if pad_op.mode == 'constant': fill_value = pad_op.in_port(3).data.get_value() if fill_value is None or fill_value != 0.0: log.info( 'The pad node "{}" with non-zero fill value cannot be fused.'. format(pad_op.soft_get('name'))) return input_tensor_dims = len(match['pad_output'].shape) for in_port in [1, 2]: pads = pad_op.in_port(in_port).data.get_value() if pads[get_features_dim(op.graph.graph['layout'], input_tensor_dims)] != 0 or \ pads[get_batch_dim(op.graph.graph['layout'], input_tensor_dims)] != 0: log.info( 'The pad node "{}" with padding over feature/batch dimension cannot be fused.' .format(pad_op.soft_get('name'))) return op.pad += np.concatenate([ pad_op.in_port(1).data.get_value().reshape([-1, 1]), pad_op.in_port(2).data.get_value().reshape([-1, 1]) ], axis=1) op.pad_spatial_shape = op.pad[op.spatial_dims] op['auto_pad'] = None if op.type == 'Pooling': op['exclude_pad'] = False assert (graph[match['pad_output'].node][match['op'].node][0]['in'] == 0) edge_attrs = graph.get_edge_data(match['pad_output'].id, match['op'].id)[0] graph.remove_edge(match['pad_output'].id, match['op'].id) graph.add_edge(input_data.id, match['op'].id, **{'in': 0, **edge_attrs})
def replace_pattern(graph: Graph, match: dict): reshape = match['reshape'] assert len(reshape.in_nodes()) > 0 if graph.graph['layout'] == 'NCHW' or reshape.has_and_set('nchw_layout') or\ reshape.soft_get('correct_data_layout') is True: return input_node = reshape.in_node() output_node = reshape.out_node() input_shape = input_node.shape output_shape = output_node.shape if len(input_shape) >= 4 and len(output_shape) == 3: # Check that we will permute some shapes in this Reshape by our permutation pass layout = 'NCHW' c_idx = get_features_dim(layout, len(input_shape)) hw_idx = [ get_width_dim(layout, len(input_shape)), get_height_dim(layout, len(input_shape)) ] if input_shape[c_idx] != 1 and np.any( input_shape[hw_idx] != [1, 1]): # then nhwc -> nchw permutation can change shapes significantly # We need to wrap up node with NCHW -> NHWC permutes and don't touch it later permutation = PermuteAttrs.get_nchw_to_nhwc_permutation( len(input_shape)) permutation_back = PermuteAttrs.get_nchw_to_nhwc_permutation( len(input_shape)) # 1. Insert input Permute # This Permute will permute input from original input layout to operation layout edge_attrs = graph.get_edge_data(input_node.id, reshape.id)[0] graph.remove_edge(input_node.id, reshape.id) permute_op = Permute(graph, { 'order': permutation.perm, 'name': reshape.name + '/Permute_' }) permute_data_node = permute_op.create_node_with_data( [input_node]) graph.add_edge(permute_data_node.id, reshape.id, **edge_attrs)
def replace_pattern(self, graph: Graph, match: dict): reshape1 = match['reshape1'] reshape2 = match['reshape2'] transpose = match['transpose'] # Check that Reshape->Transpose->Reshape shuffle only feature channel input_shape = np.array(reshape1.in_node(0).shape) reshape1_shape = np.array(reshape1.out_node().shape) output_shape = np.array(reshape2.out_node().shape) # Check that input shape is 4D if len(input_shape) != 4: log.warning( 'Can\'t convert Reshape->Transpose({})->Reshape sequence due to input shape should be 4D ' '(instead of {}D)'.format(transpose.name, len(input_shape))) return # Check that output shape the same as input shape if not np.prod(input_shape) == np.prod(output_shape): log.warning( 'Can\'t convert Reshape->Transpose({})->Reshape sequence due to output shape should be equal ' 'to input shape: {} and {}'.format(transpose.name, input_shape, output_shape)) return # Input shapes can be either NCHW or NHWC, so in case of channel split, feature channel can be splited as # follows in comments below # So feature_dims_split list contains possible dims responsible for feature dim layout = graph.graph['layout'] feature_dim = get_features_dim(layout, len(input_shape)) spatial_dims = [ get_height_dim(layout, len(input_shape)), get_width_dim(layout, len(input_shape)) ] if layout == 'NCHW': # NC1C2HW or NC1C2(H*W) feature_dims_split = np.array([feature_dim, feature_dim + 1]) else: # NHWC1C2 or N(H*W)C1C2 or (N*H*W)C1C2 feature_dims_split = np.array( [len(reshape1_shape) - 2, len(reshape1_shape) - 1]) # Check that feature_dims_split suits reshape layer shape for dim in feature_dims_split: if dim < 0 or dim >= len(reshape1_shape): log.warning( 'Can\'t convert Reshape({}:{})->Transpose->Reshape sequence. Can\'t detect feature shuffle.' ''.format(reshape1.shape, reshape1_shape)) return if not np.prod(np.delete(reshape1_shape, feature_dims_split)) == np.prod( np.delete(input_shape, feature_dim)): log.warning( 'Can\'t convert Reshape->Transpose->Reshape sequence. Can\'t detect feature shuffle. {} ' 'should be equal to {}'.format( np.prod(np.delete(reshape1_shape, feature_dims_split)), np.prod(np.delete(input_shape, feature_dim)))) return # Check transpose order if not np.array_equal(feature_dims_split[::-1], transpose.order[feature_dims_split]): log.warning( 'Can\'t convert Reshape->Transpose({})->Reshape sequence. Transpose operation should witch ' 'feature order (given order: {})'.format( transpose.name, transpose.order)) return # Now we are sure that Reshape->Transpose->Reshape shuffle feature dims # So, then we change Reshape and Transpose attrs to suite NCHW layout # The resulting shape for Reshape1 layer : [N,C1,C2,(H*W)] new_reshape1_shape = np.concatenate( (np.array([input_shape[0]]), np.array(reshape1_shape[feature_dims_split]), np.array([np.prod(input_shape[spatial_dims])]))) new_transpose_order = np.array([0, 2, 1, 3]) new_transpose_shape = np.array(new_reshape1_shape[new_transpose_order]) reshape1.out_node().shape = new_reshape1_shape reshape1.dim = np.copy(new_reshape1_shape) transpose.order = new_transpose_order transpose.out_node().shape = new_transpose_shape # Preserve layers from conversion to NCHW (in case of NHWC topology layout) reshape1['nchw_layout'] = True reshape1.out_node()['nchw_layout'] = True transpose['nchw_layout'] = True transpose.out_node()['nchw_layout'] = True
def replace_pattern(self, graph: Graph, match: dict): layout = graph.graph['layout'] if layout != 'NHWC': return reshape1 = match['reshape1'] softmax = match['softmax'] # Check that Reshape->Softmax->Reshape shuffle only feature channel input_shape = np.array(reshape1.in_node(0).shape) reshape1_shape = np.array(reshape1.out_node().shape) # Check that input shape is 4D if len(input_shape) != 4: log.warning( 'Can\'t convert Reshape({})->Softmax->Reshape sequence due to input shape should be 4D ' '(instead of {}D {})'.format(reshape1.name, len(input_shape), input_shape)) return if len(reshape1_shape) != 2: log.warning( 'This pass expect 2D output tensor for first Reshape {} layer (given shape: {})' ''.format(reshape1.name, reshape1_shape)) return # Define feature dim feature_dim = get_features_dim(layout, len(input_shape)) spatial_dims = [ get_height_dim(layout, len(input_shape)), get_width_dim(layout, len(input_shape)) ] # Skip transform in case if spatial dims in input shape are equal to [1,1] if np.array_equal(input_shape[spatial_dims], np.array([1, 1])): log.info('Skip this transformation due to spatial dims are [1,1]') return # Check that Reshape1 has out dims [-1, feature_dims] if not (reshape1_shape[-1] == input_shape[-1] and reshape1_shape[0] == np.prod(np.delete(input_shape, feature_dim))): log.warning( 'Output shape for Reshape operation should be [{},{}] instead of {}' .format(np.prod(np.delete(input_shape, feature_dim)), input_shape[-1], reshape1_shape)) return # Now we are sure that Reshape->Softmax suits for this transformation # The resulting shape for Reshape1 layer : [N,C,(H*W)] new_reshape1_shape = np.concatenate( (np.array([input_shape[0]]), np.array([reshape1_shape[-1]]), np.array([np.prod(input_shape[spatial_dims])]))) # update 'dim' attribute but preserve batch dimension size which could be -1 reshape1.dim = int64_array([reshape1.dim[0], *new_reshape1_shape[1:]]) old_shape = np.array(reshape1.out_node().shape) reshape1.out_node().shape = new_reshape1_shape softmax.out_node().shape = new_reshape1_shape # Preserve layers from conversion to NCHW (in case of NHWC topology layout) reshape1['nchw_layout'] = True reshape1.out_node()['nchw_layout'] = True softmax['nchw_layout'] = True softmax.out_node()['nchw_layout'] = True # Create final Reshape to keep original shape for softmax output if softmax is not the last node softmax_out_data = softmax.out_node() if len(softmax_out_data.out_nodes()) != 0: next_operation = softmax_out_data.out_node() # Save edge attributes & remove edge edge_attrs = graph.get_edge_data(softmax_out_data.id, next_operation.id)[0] graph.remove_edge(softmax_out_data.id, next_operation.id) reshape_op = Reshape( graph, dict(name=softmax.id + "/Reshape", dim=np.array(old_shape), nchw_layout=True)) reshape_out_data = reshape_op.create_node_with_data( inputs=[softmax_out_data]) graph.add_edges_from([(reshape_out_data.id, next_operation.id, edge_attrs)])
def replace_pattern(self, graph: Graph, match: dict): y = match['maximum'].in_port(0).data.get_value() if y is None: y = match['maximum'].in_port(1).data.get_value() if y is None or y.shape != (): log.debug( 'The value of the "maximum_y_data" is not defined or is not constant' ) return # We need to check axes which performed reduction because IE supports only 2D, 3D, 4D inputs and # reduction only along spatial and channel dimensions. input_rank = len(match['sum'].in_port(0).data.get_shape()) if input_rank not in [2, 3, 4]: log.debug( 'IE supports L2 normalization only for 2D, 3D and 4D tensors.') return axes = match['sum'].in_port(1).data.get_value() axes = int64_array(axes) if axes.shape == (): axes = int64_array([axes]) axes = int64_array( [axis if axis >= 0 else axis + input_rank for axis in axes]) axes.sort() transformation_applicable = False # check for case C + all spatial dims. Works for 2D (NC), 3D (NCH) and 4D (NCHW and NHWC) if len(axes) + 1 == input_rank and np.array_equal( axes, int64_array(np.arange(start=1, stop=input_rank))): transformation_applicable = True # check for pure C channel normalization if len(axes) == 1 and ((input_rank == 4 and get_features_dim( graph.graph['layout'], input_rank) == axes[0]) or (input_rank != 4 and axes[0] == 1)): transformation_applicable = True if not transformation_applicable: log.debug( 'IE doesn\'t support l2 normalization with reduction along axes {}.' .format(axes)) return output_name = match['l2_normalize'].soft_get('name', match['l2_normalize'].id) normalize_node = create_op_node_with_second_input( graph, NormalizeL2Op, axes, { 'name': output_name, 'eps_mode': 'max', 'eps': y }) match['square'].in_port(0).get_source().connect( normalize_node.in_port(0)) match['square'].in_port(0).disconnect() if match['l2_normalize'].in_port( 0).get_source().node.id == match['rsqrt'].id: match['l2_normalize'].in_port(1).disconnect() else: match['l2_normalize'].in_port(0).disconnect() match['l2_normalize'].out_port(0).get_connection().set_source( normalize_node.out_port(0)) rename_nodes([(match['l2_normalize'], output_name + "/TBR"), (normalize_node, output_name)])
def interp_infer(node: Node): layout = node.graph.graph['layout'] assert len(layout) == 4 if len(node.in_nodes()) == 2: src_shape = node.in_node(0).shape dst_shape = node.in_node(1).shape # in Caffe can be 2 inputs too, but shape should be got from shape of the second input if node.parse_2nd_input == 'shape': dst_shape = [ dst_shape[get_height_dim(layout, 4)], dst_shape[get_width_dim(layout, 4)] ] else: # it is TF case dst_shape = node.in_node(1).value if src_shape is None or dst_shape is None or len( src_shape) != 4 or len(dst_shape) != 2: log.error( 'Node {} with op {} cannot be converted to Resample layer because there is no enough info about ' 'src/dst shapes: src_shape = {}, dst_shape = {}'.format( node.name, node.op, src_shape, dst_shape)) node.type = None # prevent translation to a valid IE layer return in_height = src_shape[get_height_dim(layout, 4)] in_width = src_shape[get_width_dim(layout, 4)] out_height = dst_shape[0] out_width = dst_shape[1] node.factor = factor_update( node.factor, [float(out_height) / in_height, float(out_width) / in_width], [in_height, in_width], [out_height, out_width], node.soft_get('name')) if node.factor is None: node['width'] = out_width node['height'] = out_height node.out_node().shape = shape_for_layout( layout, batch=src_shape[get_batch_dim(layout, 4)], features=src_shape[get_features_dim(layout, 4)], height=out_height, width=out_width) node.graph.remove_edge(node.in_node(1).id, node.id) else: outn = node.out_node(0) in_shape = node.in_node(0) num_ = in_shape.shape[get_batch_dim(layout, 4)] channels_ = in_shape.shape[get_features_dim(layout, 4)] height_in_ = in_shape.shape[get_height_dim(layout, 4)] width_in_ = in_shape.shape[get_width_dim(layout, 4)] height_out_ = height_in_ + node.pad_beg + node.pad_end width_out_ = width_in_ + node.pad_beg + node.pad_end if node.shrink_factor != 1 and node.zoom_factor == 1: shrink_factor = node.shrink_factor if shrink_factor < 1: log.error( 'Shrink factor should be positive in node {}'.format( node.id)) return None height_out_ = (height_out_ - 1) / shrink_factor + 1 width_out_ = (width_out_ - 1) / shrink_factor + 1 elif node.shrink_factor == 1 and node.zoom_factor != 1: zoom_factor = node.zoom_factor if zoom_factor < 1: log.error( 'Zoom factor should be positive in node {}'.format( node.id)) return None node['debug_message'] = 'Interp layer shape inference function may be wrong, please, try to update ' \ 'layer shape inference function in the file (extensions/ops/interp.op at the ' \ 'line {}).'.format(inspect.currentframe().f_lineno) + refer_to_faq_msg(100) # Reshape methods can be different in some cases # Commented out section represents reshape that used in deeplab-caffe # Uncomment the following lines, if your model was trained with deeplab-caffe # or have the same reshape method # height_out_ = height_out_ + (height_out_ - 1) * (zoom_factor - 1) # width_out_ = width_out_ + (width_out_ - 1) * (zoom_factor - 1) # Comment out the following lines if you use the reshape method from previous section height_out_ = height_out_ * zoom_factor width_out_ = width_out_ * zoom_factor elif node.width != 0 and node.height != 0: height_out_ = node.height width_out_ = node.width elif node.shrink_factor != 1 and node.zoom_factor != 1: shrink_factor = node.shrink_factor zoom_factor = node.zoom_factor if shrink_factor < 1: log.error( 'Shrink factor should be positive in node {}'.format( node.id)) return None if zoom_factor < 1: log.error( 'Zoom factor should be positive in node {}'.format( node.id)) return None height_out_ = (height_out_ - 1) / shrink_factor + 1 width_out_ = (width_out_ - 1) / shrink_factor + 1 height_out_ = height_out_ + (height_out_ - 1) * (zoom_factor - 1) width_out_ = width_out_ + (width_out_ - 1) * (zoom_factor - 1) outn.shape = shape_for_layout(layout, batch=num_, features=channels_, height=height_out_, width=width_out_)
def replace_pattern(self, graph: nx.MultiDiGraph, match: dict): node = match['reduce'] if not node.has_valid('reduce_type') or node.reduce_type.lower() not in self.supported_reduce_types: log.error("Reduce type {} is not supported for node {}".format(node.soft_get('reduce_type'), node.id)) return reduce_type = node.reduce_type.lower() if reduce_type not in self.pool_method_map: log.error("Reduce type {} is not included in pool_method_map. Please update pool_method_map with new key " "{}".format(reduce_type, reduce_type)) return input_data = node.in_node() output_data = node.out_node() input_shape = node.in_node().shape output_shape = node.out_node().shape # normalize node.axis to exclude negative indices node.axis = [get_canonical_axis_index(input_shape, a) for a in node.axis] axis = node.axis # Check that values in axis list are consecutive for idx in range(1, len(axis)): if axis[idx] != (axis[idx - 1] + 1): log.error("Reduce with not consecutive axes {} is not supported ".format(axis)) return layout = graph.graph['layout'] # So now we are sure that we can convert Reduce to appropriate operation # 1. Calculate shape that will be used in reduction reduction_dim = np.prod([input_shape[idx] for idx in axis]) begin_dims = np.array([input_shape[idx] for idx in range(axis[0])]) end_dim = np.prod([input_shape[idx] for idx in range(axis[-1] + 1, len(input_shape))]) # 2. Create reshape with appropriate shape if layout == 'NCHW': if len(begin_dims) > 2: begin_dims = np.array([np.prod(begin_dims[0:-1]), begin_dims[-1]], dtype=np.int64) else: # Expand begin_dims to 2 begin_dims = np.array(np.append(begin_dims, [1] * (2 - len(begin_dims))), dtype=np.int64) reshape_shape = np.array([*begin_dims, reduction_dim, end_dim], dtype=np.int64) pool_window = np.array([1, 1, reduction_dim, 1], dtype=np.int64) elif layout == 'NHWC': begin_dims = np.prod(begin_dims) reshape_shape = np.array([begin_dims, reduction_dim, 1, end_dim], dtype=np.int64) pool_window = np.array([1, reduction_dim, 1, 1], dtype=np.int64) else: log.error('{} layout currently is not supported'.format(layout)) return # 3. Reduce => Reshape->Pooling->Reshape reshape_op = Reshape(graph, {'name': node.id + '/Reshape', 'dim': reshape_shape}) final_reshape_op = Reshape(graph, {'name': node.id + '/FinalReshape', 'dim': output_shape}) pooling_op = Pooling(graph, dict(name=node.id + '/Pool', window=pool_window, output_spatial_shape=None, batch_dims=np.array([get_batch_dim(layout, 4)], dtype=np.int64), channel_dims=np.array([get_features_dim(layout, 4)], dtype=np.int64), exclude_pad='false', pool_method=self.pool_method_map[reduce_type])) graph.remove_edge(input_data.id, node.id) graph.remove_edge(node.id, output_data.id) final_reshape_op.create_node_with_data( inputs=[pooling_op.create_node_with_data( inputs=[reshape_op.create_node_with_data( inputs=[input_data] )] )], data_nodes=output_data) # 4. If it is reduction with summation, we need to multiply by size of the reduction slice with Mul op if reduce_type == 'sum': output_data.in_node().insert_node_with_data_after( output_data, Power, {'name': node.name + '/Mul', 'scale': float(reduction_dim)} )
def test_get_features_dim_NDHWC(self): self.assertEqual(get_features_dim('NHWC', 5), 4)
def test_get_features_dim_NCDHW(self): self.assertEqual(get_features_dim('NCHW', 5), 1)