def test_shape_for_layout_unknown_parameter(self): with self.assertRaises(Error): shape_for_layout('NHWC', batch=2, features=3, height=4, width=5, unknown_parameter=123)
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: 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 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 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 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): 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): 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 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 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 test_shape_for_layout_NHWC(self): self.assertListEqual([2, 4, 5, 3], list( shape_for_layout('NHWC', batch=2, features=3, height=4, width=5)))
def test_shape_for_layout_NCHW(self): self.assertListEqual([2, 3, 4, 5], list( shape_for_layout('NCHW', batch=2, features=3, height=4, width=5)))
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 psroipooling_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 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=node.output_dim, height=node.group_size, width=node.group_size)
def regionyolo_infer(node: Node): input_shape = node.in_node(0).shape if input_shape is None: return axis = get_canonical_axis_index(input_shape, node.axis) end_axis = get_canonical_axis_index(input_shape, node.end_axis) node.axis = axis node.end_axis = end_axis if node.do_softmax: flat_dim = np.prod(input_shape[axis: end_axis + 1]) node.out_node().shape = np.array([*input_shape[:axis], flat_dim, *input_shape[end_axis + 1:]]) else: layout = node.graph.graph['layout'] assert len(layout) == 4 node.out_node().shape = shape_for_layout(layout, batch=input_shape[get_batch_dim(layout, 4)], features=(node.classes + node.coords + 1) * len(node.mask), height=input_shape[get_height_dim(layout, 4)], width=input_shape[get_width_dim(layout, 4)])
def regionyolo_infer(node: Node): input_shape = node.in_port(0).data.get_shape() axis = get_canonical_axis_index(input_shape, node.axis) end_axis = get_canonical_axis_index(input_shape, node.end_axis) node.axis = axis node.end_axis = end_axis if node.do_softmax: dims_to_flatten = input_shape[axis: end_axis + 1] if is_fully_defined(dims_to_flatten): flat_dim = np.ma.prod(dims_to_flatten) else: flat_dim = dynamic_dimension_value node.out_port(0).data.set_shape([*input_shape[:axis], flat_dim, *input_shape[end_axis + 1:]]) else: layout = node.graph.graph['layout'] assert len(layout) == 4 node.out_port(0).data.set_shape(shape_for_layout(layout, batch=input_shape[get_batch_dim(layout, 4)], features=(node.classes + node.coords + 1) * len(node.mask), height=input_shape[get_height_dim(layout, 4)], width=input_shape[get_width_dim(layout, 4)]))
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 test_shape_for_layout_missing_features(self): with self.assertRaises(Error): shape_for_layout('NCHW', batch=2, height=4, width=5)
def test_shape_for_layout_missing_width(self): with self.assertRaises(Error): shape_for_layout('NHWC', batch=2, features=3, height=4)