# Copyright 2016 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """ResNet model. Related papers: https://arxiv.org/pdf/1603.05027v2.pdf https://arxiv.org/pdf/1512.03385v1.pdf https://arxiv.org/pdf/1605.07146v1.pdf """ from __future__ import absolute_import from __future__ import division from __future__ import print_function from __future__ import unicode_literals import numpy as np import six import tensorflow as tf from .discriminator import Discriminator from collections import namedtuple from functools import partial HParams = namedtuple('HParams', 'batch_size, num_classes, min_lrn_rate, lrn_rate, ' 'num_residual_units, use_bottleneck, weight_decay_rate, ' 'relu_leakiness, optimizer') class ResNet(Discriminator): """ResNet model.""" def __init__(self, hps, dims): """ResNet constructor. Args: @hps: Hyperparameters. @mode: One of 'train' and 'eval'. """ self.hps = hps self._extra_train_ops = [] super(ResNet, self).__init__(dims=dims) def _get_logits_op(self, X, n_classes, train=True, reuse=False, **kwargs): """Build the core model within the graph.""" with tf.variable_scope('init'): x = self._conv('init_conv', X, 3, 3, 16, self._stride_arr(1)) strides = [1, 2, 2] activate_before_residual = [True, False, False] if self.hps.use_bottleneck: res_func = partial(self._bottleneck_residual, train=train, reuse=reuse) filters = [16, 64, 128, 256] else: res_func = partial(self._residual, train=train, reuse=reuse) filters = [16, 16, 32, 64] # Uncomment the following codes to use w28-10 wide residual network. # It is more memory efficient than very deep residual network and has # comparably good performance. # https://arxiv.org/pdf/1605.07146v1.pdf # filters = [16, 160, 320, 640] # Update hps.num_residual_units to 9 with tf.variable_scope('unit_1_0'): x = res_func(x, filters[0], filters[1], self._stride_arr(strides[0]), activate_before_residual[0]) for i in six.moves.range(1, self.hps.num_residual_units): with tf.variable_scope('unit_1_%d' % i): x = res_func(x, filters[1], filters[1], self._stride_arr(1), False) with tf.variable_scope('unit_2_0'): x = res_func(x, filters[1], filters[2], self._stride_arr(strides[1]), activate_before_residual[1]) for i in six.moves.range(1, self.hps.num_residual_units): with tf.variable_scope('unit_2_%d' % i): x = res_func(x, filters[2], filters[2], self._stride_arr(1), False) with tf.variable_scope('unit_3_0'): x = res_func(x, filters[2], filters[3], self._stride_arr(strides[2]), activate_before_residual[2]) for i in six.moves.range(1, self.hps.num_residual_units): with tf.variable_scope('unit_3_%d' % i): x = res_func(x, filters[3], filters[3], self._stride_arr(1), False) with tf.variable_scope('unit_last'): x = self._batch_norm('final_bn', x, train=train, reuse=reuse) x = self._relu(x, self.hps.relu_leakiness) x = self._global_avg_pool(x) with tf.variable_scope('logit'): logits = self._fully_connected(x, n_classes) return logits def _stride_arr(self, stride): """Map a stride scalar to the stride array for tf.nn.conv2d.""" return [1, stride, stride, 1] def _batch_norm(self, name, x, train=True, reuse=False): # Note we replace the batch norm from the tensorflow/models code with the # contrib.layers implementation here for better scoping behavior... return tf.contrib.layers.batch_norm( x, decay=0.9, center=True, # I.e. use beta scale=True, # I.e. use gamma epsilon=1e-5, # Note: important to leave this unset! # updates_collections=None, variables_collections=[self.bn_vars_collection], is_training=train, reuse=reuse, scope=name, trainable=True ) def _residual(self, x, in_filter, out_filter, stride, activate_before_residual=False, train=True, reuse=False): """Residual unit with 2 sub layers.""" if activate_before_residual: with tf.variable_scope('shared_activation'): x = self._batch_norm('init_bn', x, train=train, reuse=reuse) x = self._relu(x, self.hps.relu_leakiness) orig_x = x else: with tf.variable_scope('residual_only_activation'): orig_x = x x = self._batch_norm('init_bn', x, train=train, reuse=reuse) x = self._relu(x, self.hps.relu_leakiness) with tf.variable_scope('sub1'): x = self._conv('conv1', x, 3, in_filter, out_filter, stride) with tf.variable_scope('sub2'): x = self._batch_norm('bn2', x, train=train, reuse=reuse) x = self._relu(x, self.hps.relu_leakiness) x = self._conv('conv2', x, 3, out_filter, out_filter, [1, 1, 1, 1]) with tf.variable_scope('sub_add'): if in_filter != out_filter: orig_x = tf.nn.avg_pool(orig_x, stride, stride, 'VALID') orig_x = tf.pad( orig_x, [[0, 0], [0, 0], [0, 0], [(out_filter-in_filter)//2, (out_filter-in_filter)//2]]) x += orig_x tf.logging.debug('image after unit %s', x.get_shape()) return x def _bottleneck_residual(self, x, in_filter, out_filter, stride, activate_before_residual=False, train=True, reuse=False): """Bottleneck residual unit with 3 sub layers.""" if activate_before_residual: with tf.variable_scope('common_bn_relu'): x = self._batch_norm('init_bn', x, train=train, reuse=reuse) x = self._relu(x, self.hps.relu_leakiness) orig_x = x else: with tf.variable_scope('residual_bn_relu'): orig_x = x x = self._batch_norm('init_bn', x, train=train, reuse=reuse) x = self._relu(x, self.hps.relu_leakiness) with tf.variable_scope('sub1'): x = self._conv('conv1', x, 1, in_filter, out_filter/4, stride) with tf.variable_scope('sub2'): x = self._batch_norm('bn2', x, train=train, reuse=reuse) x = self._relu(x, self.hps.relu_leakiness) x = self._conv('conv2', x, 3, out_filter/4, out_filter/4, [1, 1, 1, 1]) with tf.variable_scope('sub3'): x = self._batch_norm('bn3', x, train=train, reuse=reuse) x = self._relu(x, self.hps.relu_leakiness) x = self._conv('conv3', x, 1, out_filter/4, out_filter, [1, 1, 1, 1]) with tf.variable_scope('sub_add'): if in_filter != out_filter: orig_x = self._conv('project', orig_x, 1, in_filter, out_filter, stride) x += orig_x tf.logging.info('image after unit %s', x.get_shape()) return x def get_weight_decay_op(self): """L2 weight decay loss.""" costs = [] for var in tf.trainable_variables(): if var.op.name.find(r'DW') > 0: costs.append(tf.nn.l2_loss(var)) return tf.add_n(costs) def _conv(self, name, x, filter_size, in_filters, out_filters, strides): """Convolution.""" with tf.variable_scope(name): n = filter_size * filter_size * out_filters kernel = tf.get_variable( 'DW', [filter_size, filter_size, in_filters, out_filters], tf.float32, initializer=tf.random_normal_initializer( stddev=np.sqrt(2.0/n))) return tf.nn.conv2d(x, kernel, strides, padding='SAME') def _relu(self, x, leakiness=0.0): """Relu, with optional leaky support.""" return tf.where(tf.less(x, 0.0), leakiness * x, x, name='leaky_relu') def _fully_connected(self, x, out_dim): """FullyConnected layer for final output.""" w = tf.get_variable( 'DW', [x.get_shape()[1], out_dim], initializer=tf.uniform_unit_scaling_initializer(factor=1.0)) b = tf.get_variable('biases', [out_dim], initializer=tf.constant_initializer()) return tf.nn.xw_plus_b(x, w, b) def _global_avg_pool(self, x): assert x.get_shape().ndims == 4 return tf.reduce_mean(x, [1, 2]) hps_default = HParams(batch_size=100, num_classes=10, min_lrn_rate=0.0001, lrn_rate=0.1, num_residual_units=9, use_bottleneck=False, weight_decay_rate=0.0002, relu_leakiness=0.1, optimizer='mom') ResNetDefault = partial(ResNet, hps=hps_default)