Created
March 6, 2018 01:44
-
-
Save oraoto/7656c7593b5bc7e481118e8c167d4d13 to your computer and use it in GitHub Desktop.
modified TensorFlowParser.py
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| # Copyright 2017 Intel Corporation. | |
| # The source code, information and material ("Material") contained herein is | |
| # owned by Intel Corporation or its suppliers or licensors, and title to such | |
| # Material remains with Intel Corporation or its suppliers or licensors. | |
| # The Material contains proprietary information of Intel or its suppliers and | |
| # licensors. The Material is protected by worldwide copyright laws and treaty | |
| # provisions. | |
| # No part of the Material may be used, copied, reproduced, modified, published, | |
| # uploaded, posted, transmitted, distributed or disclosed in any way without | |
| # Intel's prior express written permission. No license under any patent, | |
| # copyright or other intellectual property rights in the Material is granted to | |
| # or conferred upon you, either expressly, by implication, inducement, estoppel | |
| # or otherwise. | |
| # Any license under such intellectual property rights must be express and | |
| # approved by Intel in writing. | |
| import sys | |
| import tensorflow as tf | |
| import google.protobuf as proto | |
| import numpy as np | |
| import math | |
| import re | |
| from Models.Network import * | |
| from Models.NetworkStage import * | |
| from Models.EnumDeclarations import * | |
| from tensorflow.core.framework import graph_pb2 | |
| from tensorflow.python.framework import ops | |
| from Controllers.TensorFlowPreproc import TFPreprocessor | |
| from Controllers.TensorFlowPreproc import PatternType | |
| sys.setrecursionlimit(5000) | |
| placeholder_dict = {} | |
| const_dict = {} | |
| node_dict = {} | |
| variable_dict = {} | |
| concat_tracker = [] | |
| reshape_tracker = [] | |
| identity_tracker = [] | |
| padding_tracker = [] | |
| inputnode = 'input' | |
| os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2" | |
| def apply_padding(pad_type, in_dim, kernel_dim, stride_dim): | |
| if pad_type == b'SAME': | |
| return same_padding(in_dim, kernel_dim, stride_dim) | |
| elif pad_type == b'VALID': | |
| return valid_padding(in_dim, kernel_dim, stride_dim) | |
| else: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "No Such Pad Type Supported." + | |
| str(pad_type)) | |
| def same_padding(in_dim, kernel_dim, stride_dim): | |
| """ | |
| Calculates the output dimension and also the padding required for that dimension. | |
| :param in_dim: Width/Height of Input | |
| :param kernel_dim: Width/Height of Kernel | |
| :param stride_dim: Vertical/Horizontal Stride | |
| """ | |
| output_dim = math.ceil(float(in_dim) / float(stride_dim)) | |
| pad = ((output_dim - 1) * stride_dim + kernel_dim - in_dim) / 2 | |
| return output_dim, pad | |
| def valid_padding(in_dim, kernel_dim, stride_dim): | |
| output_dim = math.ceil(float(in_dim - kernel_dim + 1) / float(stride_dim)) | |
| pad = 0 | |
| return output_dim, pad | |
| def get_deconv_padding(input_shape, output_shape, kernel_shape, stride): | |
| """ | |
| input_shape: N,H,W,C | |
| output_shape: N,H,W,C | |
| kernel_shape: kh,kw | |
| stride: 1,sh,sw,sc | |
| """ | |
| pady = 0 | |
| padx = 0 | |
| # 2Xpad = stride X (input - 1) + kernel - out | |
| pady = stride[1] * (input_shape[1] - 1) + kernel_shape[0] - output_shape[1] | |
| padx = stride[2] * (input_shape[2] - 1) + kernel_shape[1] - output_shape[2] | |
| if (pady % 2 == 1): | |
| pady = -1 | |
| else: | |
| pady = pady / 2 | |
| if (padx % 2 == 1): | |
| padx = -1 | |
| else: | |
| padx = padx / 2 | |
| return int(pady), int(padx) | |
| def get_input(name, fail=True): | |
| global placeholder_dict | |
| global const_dict | |
| global concat_tracker | |
| global reshape_tracker | |
| global identity_tracker | |
| if len(concat_tracker) != 0: | |
| # We track all previous concats, as we may have a many-to-many | |
| # connection | |
| for concat in concat_tracker: | |
| if concat[0] == name: | |
| # If the next layer will try to attach to the non-existant | |
| # concat intermediary node. | |
| ret = [] | |
| for l in concat[1]: | |
| a = get_input(l) | |
| if isinstance(a[0], list): | |
| for a1 in a[0]: | |
| ret.append(a1) | |
| else: | |
| ret.append(a[0]) | |
| return [ret] | |
| if len(reshape_tracker) != 0: | |
| for reshape in reshape_tracker: | |
| if reshape[0] == name: | |
| return get_input(reshape[1]) | |
| if len(identity_tracker) != 0: | |
| for idn in identity_tracker: | |
| if idn[0] == name: | |
| if idn[1] == inputnode: | |
| return None | |
| if get_input(idn[1], False) == 0: | |
| return 0 | |
| return get_input(idn[1]) | |
| if name == inputnode: | |
| return None | |
| if name in node_dict.keys(): | |
| return [node_dict[name].unprocessed_name] | |
| if not fail: | |
| return 0 | |
| if name in const_dict.keys(): | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Top Not Supported - Constants " + | |
| str(name)) | |
| else: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Top Not Found " + str(name)) | |
| def get_padding_input(name): | |
| global padding_tracker | |
| for padding in padding_tracker: | |
| if padding[0] == name: | |
| return padding[1], padding[2] | |
| return None, None | |
| def have_first_input(name): | |
| if name == inputnode: | |
| return True | |
| if len(identity_tracker) != 0: | |
| for idn in identity_tracker: | |
| if idn[0] == name: | |
| if idn[1] == inputnode: | |
| return True | |
| return False | |
| def strip_tensor_id(word): | |
| return word.replace(':0', '').replace(':','#') | |
| # Count how many times we need this as an input | |
| def count_inputs(t): | |
| graph = tf.get_default_graph() | |
| count = 0 | |
| for node in graph.get_operations(): | |
| for a in node.inputs: | |
| if a.name == t: | |
| count = count + 1 | |
| return count | |
| def parse_tensor(arguments, myriad_conf, preprocess=True, debug=True, file_gen=False): | |
| global const_dict | |
| global placeholder_dict | |
| global node_dict | |
| global concat_tracker | |
| global identity_tracker | |
| global reshape_tracker | |
| global padding_tracker | |
| global inputnode | |
| path = arguments.net_description | |
| image = arguments.image | |
| output_node_name = arguments.output_node_name | |
| input_node_name = arguments.input_node_name | |
| filename = arguments.outputs_name | |
| if input_node_name is not None: | |
| inputnode = input_node_name | |
| # debug = True | |
| with tf.Session() as sess: | |
| filetype = path.split(".")[-1] | |
| if filetype == 'pb': | |
| graph_def = graph_pb2.GraphDef() | |
| with open(path, 'rb') as f: | |
| graph_def.ParseFromString(f.read()) | |
| tf.import_graph_def(graph_def, name="") | |
| else: | |
| saver = tf.train.import_meta_graph(path, clear_devices=True) | |
| if saver is not None: | |
| weight_filename = arguments.net_weights | |
| if weight_filename == None: | |
| weight_filename = path[:path.rfind('.')] | |
| saver.restore(sess, weight_filename) | |
| graph = tf.get_default_graph() | |
| preprocessor = None | |
| if preprocess: | |
| preprocessor = TFPreprocessor() | |
| preprocessor.preprocess(graph) | |
| inputTensor = graph.get_tensor_by_name(inputnode + ':0') | |
| if output_node_name is None: | |
| output_node_name = 'output' | |
| try: | |
| outputTensor = graph.get_tensor_by_name(output_node_name + ':0') | |
| except: | |
| throw_error(ErrorTable.NoOutputNode, output_node_name + ':0') | |
| shape = inputTensor.get_shape() | |
| if shape.dims is None: | |
| # If the network does not give an input shape, assume this | |
| # we will need to add this as a parameter | |
| shape = [1, 224, 224, 3] | |
| if arguments.input_size: | |
| shape = [1, | |
| arguments.input_size[1], | |
| arguments.input_size[0], | |
| 3] | |
| if isinstance(shape, tf.TensorShape): | |
| shape_list = shape.as_list() | |
| # Tensorflow can have None in the batch size field of the | |
| # input shape, if that is the case then set it to 1 | |
| if None == shape_list[0]: | |
| shape_list[0] = 1 | |
| shape = shape_list | |
| inputTensor.set_shape(shape) | |
| elif None in shape: | |
| throw_error(ErrorTable.TFNotEvaluated) | |
| if image is None or image == "Debug": | |
| input_data = np.random.uniform(0, 1, shape) | |
| if debug: | |
| print("Input image shape", shape) | |
| else: | |
| input_data = parse_img(image, | |
| [int(shape[0]), | |
| int(shape[3]), | |
| int(shape[1]), | |
| int(shape[2])], | |
| raw_scale=arguments.raw_scale, | |
| mean=arguments.mean, | |
| channel_swap=arguments.channel_swap) | |
| input_data = input_data.transpose([0, 2, 3, 1]) | |
| network = Network("TensorFlow Network", input_data) | |
| arguments.network = network | |
| res = outputTensor.eval(feed_dict={inputnode + ':0' : input_data}) | |
| prev_node = None | |
| prev_node_label = None | |
| cnt = 0 | |
| inputfound = False | |
| for idx, node in enumerate(graph.get_operations()): | |
| if debug: | |
| print(" ", idx, node.type, node.name) | |
| for a in node.inputs: | |
| print(" IN:", a.name) | |
| for a in node.outputs: | |
| print(" OUT:", a.name) | |
| if not inputfound: | |
| if have_first_input(strip_tensor_id(node.outputs[0].name)): | |
| inputfound = True | |
| if debug: | |
| print('Starting to process') | |
| continue | |
| # Each layer can have a placeholder for the batch size. | |
| for output_item in node.outputs: | |
| item_shape = output_item.shape.as_list() | |
| if len(item_shape) > 0 and \ | |
| item_shape[0] == None: | |
| item_shape[0] = 1 | |
| output_item.set_shape(item_shape) | |
| if preprocessor: | |
| pattern_found, current_pattern = preprocessor.pattern_found(node) | |
| if pattern_found: | |
| if current_pattern.get_type() == PatternType.Completed: | |
| continue | |
| else: | |
| if current_pattern.get_type() == PatternType.LeakyReLU: | |
| if debug: | |
| print("LeakyRelu") | |
| if len(current_pattern.get_input_shape()) == 4: | |
| node.outputs[0].set_shape([current_pattern.get_input_shape()[0], | |
| current_pattern.get_input_shape()[1], | |
| current_pattern.get_input_shape()[2], | |
| current_pattern.get_output_shape()[3]]) | |
| elif len(current_pattern.get_input_shape()) == 2: | |
| node.outputs[0].set_shape([current_pattern.get_input_shape()[0], | |
| current_pattern.get_input_shape()[1]]) | |
| else: | |
| throw_error(ErrorTable.StageDetailsNotSupported, "Unsupported LeakyRelu Dimensions") | |
| prev_node.postOp = StageType.leaky_relu | |
| prev_node.post_param1 = current_pattern.get_param(0) | |
| prev_node.changeName(current_pattern.get_name()) | |
| prev_node_label = strip_tensor_id(current_pattern.get_prev_name()) | |
| node_dict[prev_node_label] = prev_node | |
| else: | |
| throw_error(ErrorTable.StageDetailsNotSupported, | |
| "Pattern not supported " + str(current_pattern.get_type().name)) | |
| if node.type == "Const": | |
| const_dict[node.name] = node.outputs[0].get_shape() | |
| elif node.type == "Placeholder": | |
| placeholder_dict[node.name] = node.outputs[0].get_shape() | |
| elif node.type == 'Variable' or node.type == 'VariableV2': | |
| variable_dict[node.name] = node.outputs[0].get_shape() | |
| elif node.type == "Conv2D": | |
| if debug: | |
| print("Conv2D") | |
| inputs = node.inputs[0] | |
| input_shape = inputs.get_shape() | |
| #print(input_shape) | |
| # If the network does not have predetermined input shape, take | |
| # if from input | |
| if input_shape.dims is None and inputs.name == input_node_name + ':0': | |
| input_shape = input_data.shape | |
| if arguments.input_size: | |
| input_shape = [1, | |
| arguments.input_size[1], | |
| arguments.input_size[0], | |
| 3] | |
| taps = node.inputs[1] | |
| taps_shape = node.inputs[1].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| ksize = taps_shape[0] | |
| stride = node.get_attr("strides") | |
| output_size = [input_shape[0], | |
| apply_padding(node.get_attr("padding"), | |
| int(input_shape[1]), | |
| int(ksize), | |
| stride[1])[0], | |
| apply_padding(node.get_attr("padding"), | |
| int(input_shape[2]), | |
| int(ksize), | |
| stride[2])[0], | |
| outputs[3]] | |
| node.outputs[0].set_shape(output_size) | |
| top, padding = get_padding_input(strip_tensor_id(inputs.name)) | |
| padx = 0 | |
| pady = 0 | |
| padstyle = PadStyle.tfsame if node.get_attr( | |
| "padding") == b'SAME' else PadStyle.tfvalid | |
| if top is not None: | |
| if top == inputnode: | |
| top = None | |
| else: | |
| top = [top] | |
| pady = padding[1][0] | |
| padx = padding[2][0] | |
| padstyle = PadStyle.caffe | |
| input_shape = [ | |
| input_shape[0], | |
| input_shape[1] - 2 * pady, | |
| input_shape[2] - 2 * padx, | |
| input_shape[3]] | |
| else: | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| pady, | |
| padx, | |
| padstyle, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.convolution, | |
| int(taps_shape[0]), | |
| int(taps_shape[1]), | |
| stride[1], | |
| stride[2], | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| int(taps_shape[0]), | |
| int(taps_shape[1]), | |
| int(taps_shape[3]), | |
| np.array(taps.eval()), | |
| TapsOrder.orderHWCK, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| # node_dict | |
| cnt += 1 | |
| elif node.type == 'DepthwiseConv2dNative': | |
| if debug: | |
| print("DepthwiseConv2dNative") | |
| inputs = node.inputs[0] | |
| input_shape = inputs.get_shape() | |
| # If the network does not have predetermined input shape, take | |
| # if from input | |
| if input_shape.dims is None and inputs.name == input_node_name + ':0': | |
| input_shape = input_data.shape | |
| if arguments.input_size: | |
| input_shape = [1, | |
| arguments.input_size[1], | |
| arguments.input_size[0], | |
| 3] | |
| taps = node.inputs[1] | |
| taps_shape = node.inputs[1].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| ksize = taps_shape[0] | |
| stride = node.get_attr("strides") | |
| output_size = [input_shape[0], | |
| apply_padding(node.get_attr("padding"), | |
| int(input_shape[1]), | |
| int(ksize), | |
| stride[1])[0], | |
| apply_padding(node.get_attr("padding"), | |
| int(input_shape[2]), | |
| int(ksize), | |
| stride[2])[0], | |
| outputs[3]] | |
| if debug: | |
| print(output_size) | |
| node.outputs[0].set_shape(output_size) | |
| top, padding = get_padding_input(strip_tensor_id(inputs.name)) | |
| padx = 0 | |
| pady = 0 | |
| padstyle = PadStyle.tfsame if node.get_attr( | |
| "padding") == b'SAME' else PadStyle.tfvalid | |
| if top is not None: | |
| if top == inputnode: | |
| top = None | |
| else: | |
| top = [top] | |
| pady = padding[1][0] | |
| padx = padding[2][0] | |
| padstyle = PadStyle.caffe | |
| input_shape = [ | |
| input_shape[0], | |
| input_shape[1] - 2 * pady, | |
| input_shape[2] - 2 * padx, | |
| input_shape[3]] | |
| else: | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| taps2 = np.array(taps.eval()) | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| pady, | |
| padx, | |
| padstyle, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.depthwise_convolution, | |
| int(taps_shape[0]), | |
| int(taps_shape[1]), | |
| stride[1], | |
| stride[2], | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| int(taps_shape[0]), | |
| int(taps_shape[1]), | |
| int(taps_shape[2]) * int(taps_shape[3]), | |
| taps2, | |
| TapsOrder.orderHWCK, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| # node_dict | |
| cnt += 1 | |
| elif node.type == "Conv2DBackpropInput": | |
| inputs = node.inputs[2] | |
| input_shape = inputs.get_shape() | |
| # If the network does not have predetermined input shape, take | |
| # if from input | |
| if input_shape.dims is None and inputs.name == input_node_name + ':0': | |
| input_shape = input_data.shape | |
| if arguments.input_size: | |
| input_shape = [1, | |
| arguments.input_size[1], | |
| arguments.input_size[0], | |
| 3] | |
| taps = node.inputs[1] | |
| taps_shape = node.inputs[1].get_shape().as_list() | |
| outputs = node.outputs[0].get_shape() | |
| ksize = [taps_shape[0], taps_shape[1]] | |
| stride = node.get_attr("strides") | |
| output_size = node.inputs[0].eval() | |
| node.outputs[0].set_shape(output_size) | |
| top, padding = get_padding_input(strip_tensor_id(inputs.name)) | |
| pady, padx = get_deconv_padding(input_shape.as_list(), output_size, ksize, stride) | |
| if pady < 0 or padx < 0: | |
| throw_error(ErrorTable.StageDetailsNotSupported, "Wrong deconvolution output shape.") | |
| padstyle = PadStyle.caffe | |
| if top is not None: | |
| if top == inputnode: | |
| top = None | |
| else: | |
| top = [top] | |
| pady = padding[1][0] | |
| padx = padding[2][0] | |
| input_shape = [ | |
| input_shape[0], | |
| input_shape[1] - 2 * pady, | |
| input_shape[2] - 2 * padx, | |
| input_shape[3]] | |
| else: | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| tapval = taps.eval() | |
| tapval = np.swapaxes(tapval, 2, 3) | |
| tapval = tapval[::-1,::-1,:,:] | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| pady, | |
| padx, | |
| padstyle, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.deconvolution, | |
| int(taps_shape[0]), | |
| int(taps_shape[1]), | |
| stride[1], | |
| stride[2], | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| int(taps_shape[0]), | |
| int(taps_shape[1]), | |
| int(taps_shape[2]), | |
| np.array(tapval), | |
| TapsOrder.orderHWCK, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| # node_dict | |
| cnt += 1 | |
| elif (node.type == 'BiasAdd' or node.type == 'Add' and | |
| get_input(strip_tensor_id(node.inputs[0].name), False) != 0 and | |
| get_input(strip_tensor_id(node.inputs[1].name), False) == 0 and | |
| len(node.inputs[0].get_shape()) != 1): | |
| if debug: | |
| print("BiasAdd") | |
| inputs = node.inputs[0].get_shape() | |
| bias_data = node.inputs[1] | |
| outputs = node.outputs[0].get_shape() | |
| if(len(inputs) == 4): | |
| node.outputs[0].set_shape( | |
| [inputs[0], inputs[1], inputs[2], outputs[3]]) | |
| elif(len(inputs) == 2): | |
| node.outputs[0].set_shape([inputs[0], inputs[1]]) | |
| else: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Unsupported Bias Dimensions") | |
| prev_node = network.search(strip_tensor_id(node.inputs[0].name)) | |
| bias = np.array(bias_data.eval()) | |
| if bias.size == 1: | |
| #bias is a constant, transform it to a vector | |
| prev_node.addBias( (np.ones(outputs) * bias).astype(np.float16) ) | |
| else: | |
| prev_node.addBias( bias.astype(np.float16) ) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| prev_node.changeName(prev_node_label) | |
| node_dict[prev_node_label] = prev_node | |
| elif node.type == "MaxPool": | |
| if debug: | |
| print("MaxPool") | |
| inputs = node.inputs[0] | |
| input_shape = node.inputs[0].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| ksize = node.get_attr("ksize") | |
| stride = node.get_attr("strides") | |
| pad = 0 | |
| output_size = [input_shape[0], | |
| apply_padding(node.get_attr("padding"), | |
| int(input_shape[1]), | |
| int(ksize[1]), | |
| stride[1])[0], | |
| apply_padding(node.get_attr("padding"), | |
| int(input_shape[2]), | |
| int(ksize[2]), | |
| stride[2])[0], | |
| outputs[3]] | |
| node.outputs[0].set_shape(output_size) | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| if len(input_shape) == 4: | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| else: | |
| xyz = (1, 1, int(input_shape[1])) | |
| prev_node = NetworkStage( | |
| strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.tfsame if node.get_attr("padding") == b'SAME' else PadStyle.tfvalid, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.max_pooling, | |
| ksize[1], | |
| ksize[2], | |
| stride[1], | |
| stride[2], | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| ksize[1], | |
| ksize[2], | |
| int(output_size[3]), | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif node.type == "Relu": | |
| if debug: | |
| print("ReLU") | |
| inputs = node.inputs[0].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| if(len(inputs) == 4): | |
| node.outputs[0].set_shape( | |
| [inputs[0], inputs[1], inputs[2], outputs[3]]) | |
| elif(len(inputs) == 2): | |
| node.outputs[0].set_shape([inputs[0], inputs[1]]) | |
| else: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Unsupported ReLU Dimensions") | |
| prev_node.postOp = StageType.relu | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| prev_node.changeName(prev_node_label) | |
| node_dict[prev_node_label] = prev_node | |
| elif node.type == "Relu6": | |
| if debug: | |
| print("ReLU6") | |
| inputs = node.inputs[0].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| if(len(inputs) == 4): | |
| node.outputs[0].set_shape( | |
| [inputs[0], inputs[1], inputs[2], outputs[3]]) | |
| elif(len(inputs) == 2): | |
| node.outputs[0].set_shape([inputs[0], inputs[1]]) | |
| else: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Unsupported ReLU Dimensions") | |
| prev_node.postOp = StageType.relu_x | |
| prev_node.post_param1 = 6.0 | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| prev_node.changeName(prev_node_label) | |
| node_dict[prev_node_label] = prev_node | |
| elif node.type == "LRN": | |
| if debug: | |
| print("LRN") | |
| inputs = node.inputs[0] | |
| input_shape = node.inputs[0].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| node.outputs[0].set_shape( | |
| [input_shape[0], input_shape[1], input_shape[2], outputs[3]]) | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| bias = np.array([node.get_attr("bias"), | |
| node.get_attr("alpha") * (2 * node.get_attr("depth_radius") + 1), | |
| node.get_attr("beta"), | |
| 0], dtype=np.float16) | |
| prev_node = NetworkStage( | |
| strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.LRN, | |
| 0, | |
| 2 * node.get_attr("depth_radius") + 1, | |
| 1, | |
| 1, | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| 0, | |
| 0, | |
| xyz[2], | |
| None, | |
| None, | |
| bias, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif node.type == "MatMul": # Note: Assuming MatMul and FCL are the same. | |
| if debug: | |
| print("FCL / MatMul") | |
| inputs = node.inputs | |
| input_shape = node.inputs[0].get_shape() | |
| taps = node.inputs[1] | |
| taps_shape = node.inputs[1].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| node.outputs[0].set_shape([node.inputs[0].get_shape()[0], | |
| node.inputs[1].get_shape()[1]]) | |
| top = get_input(strip_tensor_id(inputs[0].name)) | |
| if len(input_shape) == 2: | |
| xyz = (0, int(input_shape[0]), int(input_shape[1])) #0 is special flag for saying the input is 2D | |
| else: | |
| xyz = (1, 1, int(input_shape[1])) | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.fully_connected_layer, | |
| 1, | |
| 1, | |
| 1, | |
| 1, | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| 1, | |
| 1, | |
| int(taps_shape[1]), | |
| np.array(taps.eval()).astype(np.float16), | |
| TapsOrder.orderHWCK, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif node.type == "Softmax" or node.type == "Sigmoid" or node.type == "Tanh": | |
| if debug: | |
| print(node.type) | |
| inputs = node.inputs[0] | |
| input_shape = node.inputs[0].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| if(len(input_shape) == 4): | |
| node.outputs[0].set_shape( | |
| [input_shape[0], input_shape[1], input_shape[2], outputs[3]]) | |
| elif(len(input_shape) == 2): | |
| node.outputs[0].set_shape([input_shape[0], input_shape[1]]) | |
| else: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Unsupported " + node.type + " dimensions") | |
| taps_shape = [1, 1, 1, 1] | |
| stride = [1, 1, 1, 1] | |
| pad = 0 | |
| opParams=None | |
| if node.type == "Softmax": | |
| stagetype = StageType.soft_max | |
| opParams = np.array([1], dtype=np.dtype("<i4")) # softmax would be performed on C - axis | |
| elif node.type == "Sigmoid": | |
| stagetype = StageType.sigmoid | |
| else: | |
| stagetype = StageType.tanh | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| if len(input_shape) == 4: | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| else: | |
| xyz = (1, int(input_shape[0]), int(input_shape[1])) | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| stagetype, | |
| int(taps_shape[0]), | |
| int(taps_shape[0]), | |
| stride[1], | |
| stride[2], | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| int(taps_shape[0]), | |
| int(taps_shape[0]), | |
| int(input_shape[1]), | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments, | |
| opParams=opParams) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif node.type == "AvgPool": | |
| if debug: | |
| print("Avg Pool") | |
| inputs = node.inputs[0] | |
| input_shape = node.inputs[0].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| ksize = node.get_attr("ksize") | |
| stride = node.get_attr("strides") | |
| pad = 0 | |
| output_size = [input_shape[0], | |
| apply_padding(node.get_attr("padding"), | |
| int(input_shape[1]), | |
| int(ksize[1]), | |
| stride[1])[0], | |
| apply_padding(node.get_attr("padding"), | |
| int(input_shape[2]), | |
| int(ksize[2]), | |
| stride[2])[0], | |
| outputs[3]] | |
| node.outputs[0].set_shape(output_size) | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| if len(input_shape) == 4: | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| else: | |
| xyz = (1, 1, int(input_shape[1])) | |
| prev_node = NetworkStage( | |
| strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.tfsame if node.get_attr("padding") == b'SAME' else PadStyle.tfvalid, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.average_pooling, | |
| ksize[1], | |
| ksize[2], | |
| stride[1], | |
| stride[2], | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| ksize[1], | |
| ksize[2], | |
| int(output_size[3]), | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif node.type == "Mean": | |
| if debug: | |
| print("Mean") | |
| inputs = node.inputs[0] | |
| input_shape = node.inputs[0].get_shape() | |
| dimensions = node.inputs[1].eval() | |
| if dimensions[0] != 1 or dimensions[1] != 2: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Unsupported Mean operation") | |
| outputs = node.outputs[0].get_shape() | |
| ksize = [0, int(input_shape[1]), int(input_shape[2]), 0] | |
| stride = [1, 1, 1, 1] | |
| output_size = [input_shape[0], 1, 1, outputs[3]] | |
| node.outputs[0].set_shape(output_size) | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| if len(input_shape) == 4: | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| else: | |
| xyz = (1, 1, int(input_shape[1])) | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.tfvalid, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.average_pooling, | |
| ksize[1], | |
| ksize[2], | |
| stride[1], | |
| stride[2], | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| ksize[1], | |
| ksize[2], | |
| int(output_size[3]), | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif node.type == "Reshape": | |
| if debug: | |
| print("Reshape") | |
| inputs = node.inputs | |
| input_shape = node.inputs[0].get_shape() | |
| desired_shape = node.inputs[1].eval(feed_dict={inputnode + ':0' : input_data}) | |
| # Check for -1 in desired_shape | |
| if -1 in desired_shape: | |
| if desired_shape[desired_shape == -1].size > 1: | |
| throw( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Illegal Reshape dimension") | |
| # desired_shape = np.reshape(input_shape.as_list(), desired_shape)[0] | |
| input_size = input_shape.num_elements() | |
| desired_size = np.product(desired_shape[desired_shape >= 0]) | |
| negative_index = np.argmin(desired_shape) | |
| desired_shape[negative_index] = input_size / desired_size | |
| node.outputs[0].set_shape(desired_shape) | |
| reshape_tracker += [(strip_tensor_id(node.outputs[0].name), | |
| strip_tensor_id(inputs[0].name))] | |
| elif node.type == "Shape": | |
| if debug: | |
| print(node.type, len(node.inputs[0].get_shape())) | |
| elif node.type == "Squeeze": | |
| if debug: | |
| print("Squeeze") | |
| identity_tracker += [(strip_tensor_id(node.outputs[0].name), | |
| strip_tensor_id(node.inputs[0].name))] | |
| elif node.type == "Identity": | |
| if debug: | |
| print("Identity") | |
| inputs = node.inputs | |
| input_shape = node.inputs[0].get_shape() | |
| node.outputs[0].set_shape(node.inputs[0].get_shape()) | |
| identity_tracker += [(strip_tensor_id(node.outputs[0].name), | |
| strip_tensor_id(inputs[0].name))] | |
| elif node.type == "NoOp": | |
| if debug: | |
| print("No OP") | |
| pass | |
| elif (node.type == "Concat" or node.type == "ConcatV2") and not node.inputs[0].dtype.is_integer: | |
| explicit = False | |
| #If at least one input is const, do explicit concat | |
| for i in range(0, len(node.inputs)-1): | |
| if get_input(strip_tensor_id(node.inputs[i].name), False) == 0: | |
| explicit = True | |
| break | |
| if explicit or arguments.explicit_concat: | |
| if debug: | |
| print('Explicit concat') | |
| if node.type == 'Concat': | |
| axis_select = node.inputs[0].eval() | |
| input_indices = range(1, len(node.inputs)) | |
| else: | |
| axis_select = node.inputs[-1].eval() | |
| input_indices = range(0, len(node.inputs) - 1) | |
| outstride = 0 | |
| for idx in input_indices: | |
| outstride += int(node.inputs[idx].get_shape()[axis_select]) | |
| outstride *= 2 | |
| for idx in input_indices: | |
| inshape = list(node.inputs[idx].get_shape()) | |
| while len(inshape) < 4: | |
| inshape.insert(0,1) | |
| in_node = get_input(strip_tensor_id(node.inputs[idx].name), False) | |
| taps = None | |
| if in_node == 0: | |
| #Node not found, we assume it's a constant | |
| taps = np.array(node.inputs[idx].eval()) | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[0].name) + '_' + str(idx), [in_node] if in_node!=0 and in_node is not None else None, StorageOrder.orderYXZ, | |
| 0, 0, PadStyle.none, | |
| DataType.fp16, DataType.fp16, | |
| StageType.copy, # op_type | |
| 1, 1, # op_x, op_y, | |
| 1, 1, # sx, sy, | |
| int(inshape[1]), int(inshape[2]), int(inshape[3]), # X, Y, Z | |
| 0, 0, int(inshape[3]), | |
| taps, TapsOrder.orderHWCK, None, # taps, taps_order, bias, | |
| None, # Pre Op | |
| StageType.none, # Post Op | |
| None, # Post Op Param 1 | |
| 0, # Post Op StrideX | |
| 0, # Post Op StrideX | |
| myriad_config=myriad_conf, args=arguments) | |
| network.attach(prev_node) | |
| #What follows will work only if axis_select is on the last index | |
| if idx == 0: | |
| outputPointer, outputIndex = prev_node.setoutput(outstride) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| else: | |
| prev_node.setoutput(outstride, outputPointer, outputIndex) | |
| outputPointer = outputPointer + 2 * int(inshape[3]) | |
| cnt += 1 | |
| continue | |
| if debug: | |
| print("Concat") | |
| concat_channel_size = 0 | |
| inputs = node.inputs | |
| for src in inputs: | |
| dim = len(src.get_shape())-1 | |
| if dim > 0: | |
| concat_channel_size += int(src.get_shape()[dim]) | |
| a_input = node.inputs[1].get_shape() | |
| if len(a_input) == 2: | |
| node.outputs[0].set_shape([a_input[0], concat_channel_size]) | |
| else: | |
| node.outputs[0].set_shape([a_input[0], a_input[1], a_input[2], concat_channel_size]) | |
| rep_arr = [] | |
| if node.type == 'Concat': | |
| for inp in inputs[1:]: | |
| rep_arr.append(strip_tensor_id(inp.name)) | |
| else: | |
| for inp in inputs[:-1]: | |
| rep_arr.append(strip_tensor_id(inp.name)) | |
| concat_tracker += [(strip_tensor_id(node.outputs[0].name), rep_arr)] | |
| elif ((node.type == 'Add' or node.type == 'Mul' or node.type == 'Maximum') and | |
| get_input(strip_tensor_id(node.inputs[0].name), False) != 0 and | |
| get_input(strip_tensor_id(node.inputs[1].name), False) != 0): | |
| # Elementwise operations of the outputs of two existing nodes | |
| if debug: | |
| print(node.type) | |
| top = [get_input(strip_tensor_id(node.inputs[0].name))[0], | |
| get_input(strip_tensor_id(node.inputs[1].name))[0]] | |
| input_shape = node.inputs[0].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| if len(input_shape) == 4: | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| else: | |
| xyz = (1, int(input_shape[0]), int(input_shape[1])) | |
| if node.type == 'Add': | |
| op = StageType.eltwise_sum | |
| elif node.type == 'Mul' and node.inputs[1].shape[-1] == 1: | |
| op = StageType.scale_with_scalar | |
| elif node.type == 'Mul': | |
| op = StageType.eltwise_prod | |
| else: | |
| op = StageType.eltwise_max | |
| prev_node = NetworkStage( | |
| strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| op, | |
| 1, | |
| 1, | |
| 1, | |
| 1, | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif ((node.type == 'Mul' or node.type == 'Div' or node.type == 'RealDiv') and prev_node_label is not None and | |
| (strip_tensor_id(node.inputs[0].name) == prev_node_label or strip_tensor_id(node.inputs[1].name) == prev_node_label)): | |
| # We are probably multiplying with a constant, try | |
| iidx = 1 if strip_tensor_id(node.inputs[1].name) == prev_node_label else 0 | |
| # Check if absorption is possible into a convolution, to be possible, we should use | |
| # the convolution output only once, here | |
| if prev_node.op == StageType.convolution and count_inputs( | |
| prev_node_label + ":0") == 1: | |
| if debug: | |
| print('Mul with constant absorbed into convolution') | |
| if node.type == 'Mul': | |
| prev_node.taps = np.multiply( | |
| prev_node.taps, node.inputs[1 - iidx].eval()) # Eval may fail | |
| if prev_node.bias is not None: | |
| prev_node.bias = np.multiply( | |
| prev_node.bias, node.inputs[1 - iidx].eval()) # Eval may fail | |
| else: | |
| prev_node.taps = np.divide( | |
| prev_node.taps, node.inputs[1 - iidx].eval()) # Eval may fail | |
| if prev_node.bias is not None: | |
| prev_node.bias = np.divide( | |
| prev_node.bias, node.inputs[1 - iidx].eval()) # Eval may fail | |
| node_dict[node.name] = node_dict[prev_node_label] | |
| prev_node_label = node.name | |
| prev_node.name = prev_node.unprocessed_name + '/' + strip_tensor_id(node.outputs[0].name) | |
| prev_node.changeName(prev_node.name) | |
| prev_node.alias.append(prev_node.unprocessed_name) | |
| prev_node.alias.append(strip_tensor_id(node.outputs[0].name)) | |
| else: | |
| if debug: | |
| print('Mul with constant') | |
| inputs = node.inputs[iidx] | |
| input_shape = node.inputs[iidx].get_shape() | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| if len(input_shape) == 4: | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| else: | |
| xyz = (1, 1, int(input_shape[1])) | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.scale, | |
| 0, | |
| 0, | |
| 1, | |
| 1, | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| 0, | |
| 0, | |
| xyz[2], | |
| node.inputs[1 - iidx].eval(), | |
| TapsOrder.orderHWCK, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif (node.type == 'Add' or node.type == 'Sub') and prev_node_label is not None \ | |
| and strip_tensor_id(node.inputs[0].name) == prev_node_label: | |
| # We are probaby adding a constant bias, try | |
| if debug: | |
| print('Add (bias)') | |
| inputs = node.inputs[0].get_shape() | |
| bias_data = None | |
| bias_data = node.inputs[1].eval() # This eval may fail | |
| outputs = node.outputs[0].get_shape() | |
| if(len(inputs) == 4): | |
| node.outputs[0].set_shape( | |
| [inputs[0], inputs[1], inputs[2], outputs[3]]) | |
| elif(len(inputs) == 2): | |
| node.outputs[0].set_shape([inputs[0], inputs[1]]) | |
| else: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Unsupported Bias Dimensions") | |
| # Add scalar | |
| if bias_data.ndim == 1 and bias_data.shape[0] == 1: | |
| # Populate bias array with data | |
| value = bias_data[0] | |
| bias_data = np.empty([int(outputs[3])]) | |
| bias_data.fill(value) | |
| if node.type == 'Add': | |
| prev_node.addBias(np.array(bias_data).astype(np.float16)) | |
| else: | |
| prev_node.addBias(np.array(-bias_data).astype(np.float16)) | |
| prev_node.changeName(node.name) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| prev_node.changeName(prev_node_label) | |
| node_dict[prev_node_label] = prev_node | |
| elif node.type == "Maximum": | |
| if debug: | |
| print(node.type) | |
| if prev_node_label == None: | |
| iidx = 0 | |
| else: | |
| iidx = 1 if node.inputs[1].name == prev_node_label + ":0" else 0 | |
| inputs = node.inputs[iidx] | |
| input_shape = inputs.get_shape() | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| if len(input_shape) == 4: | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| else: | |
| xyz = (1, 1, int(input_shape[1])) | |
| prev_node = NetworkStage(node.name, | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.max_with_const, | |
| 0, | |
| 0, | |
| 1, | |
| 1, | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| 0, | |
| 0, | |
| xyz[2], | |
| node.inputs[1 - iidx].eval(), | |
| TapsOrder.orderHWCK, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif (node.type == "Square" or node.type == "Rsqrt") and \ | |
| (get_input(strip_tensor_id(node.inputs[0].name), False) != 0): | |
| if debug: | |
| print(node.type) | |
| inputs = node.inputs[0] | |
| input_shape = node.inputs[0].get_shape() | |
| outputs = node.outputs[0].get_shape() | |
| if(len(input_shape) == 4): | |
| node.outputs[0].set_shape( | |
| [input_shape[0], input_shape[1], input_shape[2], outputs[3]]) | |
| elif(len(input_shape) == 2): | |
| node.outputs[0].set_shape([input_shape[0], input_shape[1]]) | |
| else: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Unsupported " + node.type + " dimensions") | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| if len(input_shape) == 4: | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| else: | |
| xyz = (1, 1, int(input_shape[1])) | |
| if node.type == "Square": | |
| op_type = StageType.square | |
| else: | |
| op_type = StageType.rsqrt | |
| prev_node = NetworkStage(node.name, | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| op_type, | |
| 0, | |
| 0, | |
| 1, | |
| 1, | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| 0, | |
| 0, | |
| int(input_shape[1]), | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif node.type == "Sum": | |
| if debug: | |
| print(node.type) | |
| inputs = node.inputs[0] | |
| input_shape = node.inputs[0].get_shape() | |
| output_shape = node.outputs[0].get_shape() | |
| axis = node.inputs[1].eval() | |
| if axis != len(output_shape) - 1: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Unsupported " + node.type + " axis") | |
| axis_param = np.array([axis, 0], dtype=np.float16) | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| if len(input_shape) == 4: | |
| xyz = (int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3])) | |
| else: | |
| xyz = (1, 1, int(input_shape[1])) | |
| prev_node = NetworkStage(node.name, | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.sum_reduce, | |
| 0, | |
| 0, | |
| 1, | |
| 1, | |
| xyz[0], | |
| xyz[1], | |
| xyz[2], | |
| 0, | |
| 0, | |
| 1, | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments, | |
| opParams=axis_param) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif (node.type == 'FusedBatchNorm' and prev_node_label is not None and | |
| len(node.inputs) == 5): | |
| if debug: | |
| print('FusedBatchNorm') | |
| # Fold the batchnorm into the weights | |
| if prev_node.op == StageType.convolution or \ | |
| prev_node.op == StageType.fully_connected_layer or \ | |
| prev_node.op == StageType.depthwise_convolution: | |
| if debug: | |
| print('FusedBatchNorm absorbed into convolution') | |
| eps = node.get_attr('epsilon') | |
| scale_param = node.inputs[1].eval() | |
| offset = node.inputs[2].eval() | |
| mean = node.inputs[3].eval() | |
| var = node.inputs[4].eval() | |
| if len(mean) == 0 or len(var) == 0: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "FusedBatchNorm inputs mean and variance are not defined. The graph is not created for inference.") | |
| variance = var + eps | |
| scale = np.reciprocal(np.sqrt(variance)) * scale_param | |
| bias = offset - (mean * scale) | |
| scale = np.reshape(scale, [1, 1, 1, -1]) | |
| bias = np.reshape(bias, [1, 1, 1, -1]) | |
| if prev_node.op == StageType.depthwise_convolution: | |
| scale = np.swapaxes(scale,2,3) | |
| bias = np.swapaxes(bias,2,3) | |
| prev_node.taps = prev_node.taps * scale | |
| if prev_node.bias is not None: | |
| if bias is not None: | |
| prev_node.bias = prev_node.bias * scale + bias | |
| else: | |
| prev_node.bias = prev_node.bias * scale | |
| else: | |
| if bias is not None: | |
| prev_node.addBias(np.array(bias).astype(np.float16)) | |
| node_dict[node.name] = node_dict[prev_node_label] | |
| prev_node_label = node.name | |
| prev_node.name = prev_node.unprocessed_name + '/' + node.name | |
| prev_node.changeName(prev_node.name) | |
| prev_node.alias.append(prev_node.unprocessed_name) | |
| prev_node.alias.append(node.name) | |
| else: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "FusedBatchNorm must be preceded by convolution or fully connected layer") | |
| elif node.type == 'Slice' and not node.inputs[0].dtype.is_integer: | |
| if debug: | |
| print('Slice') | |
| input_shape = node.inputs[0].get_shape() | |
| slicingbegin = node.inputs[1].eval() | |
| slicingsize = node.inputs[2].eval() | |
| if (len(input_shape) != 4 or len(slicingbegin) != 4 or len(slicingsize) != 4 or | |
| slicingbegin[0] != 0 or slicingbegin[1] != 0 or slicingbegin[2] != 0 or | |
| input_shape[0] != slicingsize[0] or input_shape[1] != slicingsize[1] or input_shape[2] != slicingsize[2]): | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Slice type not supported") | |
| top = get_input(strip_tensor_id(node.inputs[0].name)) | |
| curslicing = [] | |
| curslicing.append( | |
| (top, int( | |
| slicingbegin[3]), int( | |
| slicingbegin[3] + slicingsize[3]))) | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.copy, | |
| 1, | |
| 1, | |
| 1, | |
| 1, | |
| int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3]), | |
| 1, | |
| 1, | |
| slicingsize[3], | |
| None, | |
| TapsOrder.orderKCHW, | |
| None, | |
| None, | |
| StageType.none, | |
| None, | |
| 0, | |
| 0, | |
| curslicing, | |
| myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif node.type == 'ExtractImagePatches': | |
| if debug: | |
| print("ExtractImagePatches") | |
| # Currently not supported, will be interpreted as reorg for yolo-v2 | |
| throw_error(ErrorTable.StageDetailsNotSupported, node.type) | |
| ''' | |
| inputs = node.inputs[0] | |
| input_shape = node.inputs[0].get_shape() | |
| stride = node.get_attr("stride") | |
| output_shape = [input_shape[0], tf.Dimension(int(input_shape[1]) / stride), | |
| tf.Dimension(int(input_shape[2]) / stride), | |
| tf.Dimension(int(input_shape[3]) * stride * stride)] | |
| node.outputs[0].set_shape(output_shape) | |
| top = get_input(strip_tensor_id(inputs.name)) | |
| op_node = NetworkStage( | |
| node.name + '_op', | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.reorg, | |
| 0, | |
| 0, | |
| 0, | |
| 0, | |
| int(input_shape[1]), | |
| int(input_shape[2]), | |
| int(input_shape[3]), | |
| int(output_shape[1]), | |
| int(output_shape[2]), | |
| int(output_shape[3]), | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments, | |
| opParams=np.array([stride], dtype=np.int32)) | |
| network.attach(op_node) | |
| prev_node_label = node.name + '_op' | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| prev_node = NetworkStage( | |
| node.name, | |
| [node.name + '_op'], | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.copy, | |
| 1, | |
| 1, | |
| 1, | |
| 1, | |
| int(output_shape[1]), | |
| int(output_shape[2]), | |
| int(output_shape[3]), | |
| int(output_shape[1]), | |
| int(output_shape[2]), | |
| int(output_shape[3]), | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| None, | |
| 0, | |
| 0, | |
| myriad_config=myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = node.name | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| ''' | |
| elif (node.type == 'StridedSlice') and not node.inputs[0].dtype.is_integer: | |
| if debug: | |
| print('StridedSlice') | |
| input_shape = node.inputs[0].get_shape() | |
| slicingbegin = node.inputs[1].eval() | |
| slicingsize = node.inputs[2].eval() - slicingbegin | |
| slicingstride = node.inputs[3].eval() | |
| begin_mask = node.get_attr('begin_mask') | |
| end_mask = node.get_attr('end_mask') | |
| ellipsis_mask = node.get_attr('ellipsis_mask') | |
| new_axis_mask = node.get_attr('new_axis_mask') | |
| shrink_axis_mask = node.get_attr('shrink_axis_mask') | |
| if begin_mask != 5 or end_mask != 5 or ellipsis_mask != 0 or new_axis_mask != 0 or shrink_axis_mask != 2: | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "StridedSlice attributes not supported") | |
| if (len(input_shape) != 3 or len(slicingbegin) != 3 or len(slicingsize) != 3 or len(slicingstride) != 3 or | |
| slicingsize[1] != 1 or slicingstride[0] != 1 or slicingstride[1] != 1 or slicingstride[2] != 1): | |
| throw_error( | |
| ErrorTable.StageDetailsNotSupported, | |
| "Slice type not supported") | |
| top = get_input(strip_tensor_id(node.inputs[0].name)) | |
| curslicing = [] | |
| curslicing.append( | |
| (top, int(slicingbegin[1]) * int(input_shape[2]), (int(slicingbegin[1])+1) * int(input_shape[2])) | |
| ) | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[0].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.copy, | |
| 1, | |
| 1, | |
| 1, | |
| 1, | |
| int(input_shape[0]), | |
| int(input_shape[1]), | |
| int(input_shape[2]), | |
| 1, | |
| 1, | |
| int(input_shape[2]), | |
| None, | |
| TapsOrder.orderKCHW, | |
| None, | |
| None, | |
| StageType.none, | |
| None, | |
| 0, | |
| 0, | |
| curslicing, | |
| myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[0].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif node.type == 'Split' and node.inputs[0].dtype.is_integer: | |
| if debug: | |
| print('Split') | |
| n = len(node.outputs) | |
| dim = node.inputs[0].eval() | |
| input_shape = node.inputs[1].get_shape() | |
| size = int(input_shape[dim]) | |
| dim += 1 | |
| while dim < len(input_shape): | |
| size *= int(input_shape[dim]) | |
| dim += 1 | |
| top = get_input(strip_tensor_id(node.inputs[1].name)) | |
| for i in range(n): | |
| curslicing = [] | |
| curslicing.append((top[0] if top is not None else None, size//n * i, size//n * (i+1))) | |
| prev_node = NetworkStage(strip_tensor_id(node.outputs[i].name), | |
| top, | |
| StorageOrder.orderYXZ, | |
| 0, | |
| 0, | |
| PadStyle.none, | |
| DataType.fp16, | |
| DataType.fp16, | |
| StageType.copy, | |
| 1, | |
| 1, | |
| 1, | |
| 1, | |
| 1, | |
| int(input_shape[0]), | |
| int(input_shape[1]), | |
| 1, | |
| 1, | |
| size//n, | |
| None, | |
| TapsOrder.orderKCHW, | |
| None, | |
| None, | |
| StageType.none, | |
| None, | |
| 0, | |
| 0, | |
| curslicing, | |
| myriad_conf, | |
| args=arguments) | |
| network.attach(prev_node) | |
| prev_node_label = strip_tensor_id(node.outputs[i].name) | |
| node_dict[prev_node_label] = prev_node | |
| cnt += 1 | |
| elif node.type == 'Pad': | |
| if debug: | |
| print('Pad') | |
| padding_tracker += [(strip_tensor_id(node.outputs[0].name), | |
| strip_tensor_id(node.inputs[0].name), | |
| node.inputs[1].eval())] | |
| elif (node.type == 'TruncatedNormal' or node.type == 'Assign' or | |
| node.type == 'RandomUniform' or node.type == 'Div' or node.type == 'RealDiv' or node.type == 'Mul' or | |
| node.type == 'Floor' or node.type == 'Add' or node.type == 'Sub' or | |
| node.type == 'Rsqrt' or node.type == 'RandomStandardNormal' or | |
| node.type == 'L2Loss' or node.type == 'Pack' or node.type == 'Slice' or | |
| node.type == 'Prod' or node.type == 'ExpandDims' or node.type == 'ConcatV2' or | |
| node.type == 'StridedSlice' or node.type == 'Fill'or node.type == 'StringJoin' or node.type == 'SaveV2' or | |
| node.type == 'RestoreV2' or node.type == 'ShardedFilename' or | |
| node.type == 'MergeV2Checkpoints'): | |
| pass | |
| else: | |
| throw_error(ErrorTable.StageDetailsNotSupported, node.type) | |
| if len(node.outputs) > 0 and strip_tensor_id(node.outputs[0].name) == output_node_name: | |
| if node.type == 'Concat' or node.type == 'ConcatV2': | |
| nodes = network.search_several(get_input(node.name)[0]) | |
| NetworkStage.concat(nodes) | |
| break | |
| if len(res.shape) == 4: | |
| network.outputTensor = ( | |
| res.shape[0], | |
| res.shape[1], | |
| res.shape[2], | |
| res.shape[3]) | |
| else: | |
| network.outputTensor = (res.shape[0], 1, 1, res.shape[1]) | |
| if file_gen: | |
| try: | |
| np.save(filename + "_expected.npy", res) | |
| except BaseException: | |
| throw_error(ErrorTable.NoOutputNode, extra=net.blob.keys()) | |
| return network |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment