1# Copyright 2019 The TensorFlow Authors. All Rights Reserved. 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================== 15"""Test configs for transpose_conv.""" 16from __future__ import absolute_import 17from __future__ import division 18from __future__ import print_function 19 20import numpy as np 21import tensorflow.compat.v1 as tf 22from tensorflow.lite.testing.zip_test_utils import create_tensor_data 23from tensorflow.lite.testing.zip_test_utils import make_zip_of_tests 24from tensorflow.lite.testing.zip_test_utils import register_make_test_function 25 26 27# Since compute output_shape is fairly complicated for 28# tf.nn.conv2d_transpose input_sizes argument, so we here first perform a 29# "conv2d" operation to get the output, then we use the output to feed in 30# tf.nn.conv2d_backprop_input. 31# This test will depend on the "conv2d" operation's correctness. 32@register_make_test_function() 33def make_transpose_conv_tests(options): 34 """Make a set of tests to do transpose_conv.""" 35 36 # Tensorflow only supports equal strides 37 test_parameters = [ 38 { 39 "input_shape": [[1, 3, 4, 1], [1, 10, 10, 3], [3, 20, 20, 1]], 40 "filter_size": [[1, 1], [1, 2], [3, 3]], 41 "has_bias": [False], 42 "strides": [[1, 1, 1, 1], [1, 3, 3, 1]], 43 "padding": ["SAME", "VALID"], 44 "data_format": ["NHWC"], 45 "channel_multiplier": [1, 2], 46 "output_shape": [[]], 47 "fully_quantize": [False], 48 "const_weight_bias": [False] 49 }, 50 # TODO(yunluli): Adding simple tests for now to unblock edgetpu debugging. 51 # Need to add more test cases. 52 { 53 "input_shape": [[1, 3, 3, 1]], 54 "filter_size": [[3, 3, 2, 1]], 55 "has_bias": [False], 56 "strides": [[1, 1, 1, 1]], 57 "padding": ["SAME"], 58 "data_format": ["NHWC"], 59 "channel_multiplier": [1], 60 "output_shape": [[1, 3, 3, 2]], 61 "fully_quantize": [True], 62 "const_weight_bias": [True] 63 }, 64 { 65 "input_shape": [[1, 3, 3, 1]], 66 "filter_size": [[3, 3, 2, 1]], 67 "has_bias": [False], 68 "strides": [[1, 1, 1, 1]], 69 "padding": ["SAME"], 70 "data_format": ["NHWC"], 71 "channel_multiplier": [1], 72 "output_shape": [[1, 3, 3, 2]], 73 "fully_quantize": [False], 74 "const_weight_bias": [True] 75 }, 76 { 77 "input_shape": [[1, 3, 3, 1]], 78 "filter_size": [[3, 3, 2, 1]], 79 "has_bias": [False], 80 "strides": [[1, 2, 2, 1]], 81 "padding": ["SAME"], 82 "data_format": ["NHWC"], 83 "channel_multiplier": [1], 84 "output_shape": [[1, 6, 6, 2]], 85 "fully_quantize": [True], 86 "const_weight_bias": [True] 87 }, 88 { 89 "input_shape": [[1, 4, 3, 1]], 90 "filter_size": [[3, 3, 2, 1]], 91 "has_bias": [False], 92 "strides": [[1, 2, 2, 1]], 93 "padding": ["SAME"], 94 "data_format": ["NHWC"], 95 "channel_multiplier": [1], 96 "output_shape": [[1, 8, 6, 2]], 97 "fully_quantize": [True], 98 "const_weight_bias": [True] 99 }, 100 { 101 "input_shape": [[1, 3, 3, 1]], 102 "filter_size": [[3, 3, 2, 1]], 103 "has_bias": [True], 104 "strides": [[1, 1, 1, 1]], 105 "padding": ["SAME"], 106 "data_format": ["NHWC"], 107 "channel_multiplier": [1], 108 "output_shape": [[1, 3, 3, 2]], 109 "fully_quantize": [True], 110 "const_weight_bias": [True] 111 }, 112 ] 113 114 def get_tensor_shapes(parameters): 115 input_shape = parameters["input_shape"] 116 filter_size = parameters["filter_size"] 117 if not parameters["const_weight_bias"]: 118 filter_shape = filter_size + [ 119 input_shape[3], parameters["channel_multiplier"] 120 ] 121 return [input_shape, filter_shape] 122 return [input_shape, filter_size] 123 124 def build_graph(parameters): 125 """Build a transpose_conv graph given `parameters`.""" 126 input_shape, filter_shape = get_tensor_shapes(parameters) 127 input_tensor = tf.compat.v1.placeholder( 128 dtype=tf.float32, name="input", shape=input_shape) 129 130 filter_input = tf.compat.v1.placeholder( 131 dtype=tf.float32, name="filter", shape=filter_shape) 132 133 if not parameters["const_weight_bias"]: 134 input_tensors = [input_tensor, filter_input] 135 conv_outputs = tf.nn.conv2d( 136 input_tensor, 137 filter_input, 138 strides=parameters["strides"], 139 padding=parameters["padding"], 140 data_format=parameters["data_format"]) 141 out = tf.compat.v1.nn.conv2d_backprop_input( 142 input_shape, 143 filter_input, 144 conv_outputs, 145 strides=parameters["strides"], 146 padding=parameters["padding"], 147 data_format=parameters["data_format"]) 148 else: 149 input_tensors = [input_tensor] 150 if parameters["fully_quantize"]: 151 filter_input = create_tensor_data( 152 np.float32, filter_shape, min_value=-1, max_value=1) 153 else: 154 filter_input = create_tensor_data(np.float32, filter_shape) 155 out = tf.nn.conv2d_transpose( 156 input_tensor, 157 filter_input, 158 parameters["output_shape"], 159 strides=parameters["strides"], 160 padding=parameters["padding"], 161 data_format=parameters["data_format"]) 162 if parameters["has_bias"]: 163 if parameters["fully_quantize"]: 164 bias_input = create_tensor_data( 165 np.float32, (parameters["output_shape"][-1],), 166 min_value=-1, 167 max_value=1) 168 else: 169 bias_input = create_tensor_data(np.float32, 170 (parameters["output_shape"][-1],)) 171 out = tf.nn.bias_add( 172 out, bias_input, data_format=parameters["data_format"]) 173 174 mul_data = create_tensor_data(np.float32, 175 (parameters["output_shape"][-1],)) 176 out = tf.math.multiply(out, mul_data) 177 178 return input_tensors, [out] 179 180 def build_inputs(parameters, sess, inputs, outputs): 181 input_shape, filter_shape = get_tensor_shapes(parameters) 182 if not parameters["const_weight_bias"]: 183 values = [ 184 create_tensor_data(np.float32, input_shape), 185 create_tensor_data(np.float32, filter_shape) 186 ] 187 else: 188 if parameters["fully_quantize"]: 189 values = [ 190 create_tensor_data( 191 np.float32, input_shape, min_value=-1, max_value=1), 192 ] 193 else: 194 values = [create_tensor_data(np.float32, input_shape),] 195 196 return values, sess.run(outputs, feed_dict=dict(zip(inputs, values))) 197 198 make_zip_of_tests(options, test_parameters, build_graph, build_inputs) 199