# Copyright 2017 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. # ============================================================================== """PNASNet Faster R-CNN implementation. Based on PNASNet model: https://arxiv.org/abs/1712.00559 """ import tensorflow as tf from object_detection.meta_architectures import faster_rcnn_meta_arch from nets.nasnet import nasnet_utils from nets.nasnet import pnasnet arg_scope = tf.contrib.framework.arg_scope slim = tf.contrib.slim def pnasnet_large_arg_scope_for_detection(is_batch_norm_training=False): """Defines the default arg scope for the PNASNet Large for object detection. This provides a small edit to switch batch norm training on and off. Args: is_batch_norm_training: Boolean indicating whether to train with batch norm. Returns: An `arg_scope` to use for the PNASNet Large Model. """ imagenet_scope = pnasnet.pnasnet_large_arg_scope() with arg_scope(imagenet_scope): with arg_scope([slim.batch_norm], is_training=is_batch_norm_training) as sc: return sc def _filter_scaling(reduction_indices, start_cell_num): """Compute the expected filter scaling at given PNASNet cell start_cell_num. In the pnasnet.py code, filter_scaling starts at 1.0. We instead adapt filter scaling to depend on the starting cell. At first cells, before any reduction, filter_scalling is 1.0. With passing any reduction cell, the filter_scaling is multiplied by 2. Args: reduction_indices: list of int indices. start_cell_num: int. Returns: filter_scaling: float. """ filter_scaling = 1.0 for ind in reduction_indices: if ind < start_cell_num: filter_scaling *= 2.0 return filter_scaling # Note: This is largely a copy of _build_pnasnet_base inside pnasnet.py but # with special edits to remove instantiation of the stem and the special # ability to receive as input a pair of hidden states. It constructs only # a sub-network from the original PNASNet model, starting from the # start_cell_num cell and with modified final layer. def _build_pnasnet_base( hidden_previous, hidden, normal_cell, hparams, true_cell_num, start_cell_num): """Constructs a PNASNet image model for proposal classifier features.""" # Find where to place the reduction cells or stride normal cells reduction_indices = nasnet_utils.calc_reduction_layers( hparams.num_cells, hparams.num_reduction_layers) filter_scaling = _filter_scaling(reduction_indices, start_cell_num) # Note: The None is prepended to match the behavior of _imagenet_stem() cell_outputs = [None, hidden_previous, hidden] net = hidden # Run the cells for cell_num in range(start_cell_num, hparams.num_cells): is_reduction = cell_num in reduction_indices stride = 2 if is_reduction else 1 if is_reduction: filter_scaling *= hparams.filter_scaling_rate prev_layer = cell_outputs[-2] net = normal_cell( net, scope='cell_{}'.format(cell_num), filter_scaling=filter_scaling, stride=stride, prev_layer=prev_layer, cell_num=true_cell_num) true_cell_num += 1 cell_outputs.append(net) # Final nonlinearity. # Note that we have dropped the final pooling, dropout and softmax layers # from the default pnasnet version. with tf.variable_scope('final_layer'): net = tf.nn.relu(net) return net # TODO(shlens): Only fixed_shape_resizer is currently supported for PNASNet # featurization. The reason for this is that pnasnet.py only supports # inputs with fully known shapes. We need to update pnasnet.py to handle # shapes not known at compile time. class FasterRCNNPNASFeatureExtractor( faster_rcnn_meta_arch.FasterRCNNFeatureExtractor): """Faster R-CNN with PNASNet feature extractor implementation.""" def __init__(self, is_training, first_stage_features_stride, batch_norm_trainable=False, reuse_weights=None, weight_decay=0.0): """Constructor. Args: is_training: See base class. first_stage_features_stride: See base class. batch_norm_trainable: See base class. reuse_weights: See base class. weight_decay: See base class. Raises: ValueError: If `first_stage_features_stride` is not 16. """ if first_stage_features_stride != 16: raise ValueError('`first_stage_features_stride` must be 16.') super(FasterRCNNPNASFeatureExtractor, self).__init__( is_training, first_stage_features_stride, batch_norm_trainable, reuse_weights, weight_decay) def preprocess(self, resized_inputs): """Faster R-CNN with PNAS preprocessing. Maps pixel values to the range [-1, 1]. Args: resized_inputs: A [batch, height_in, width_in, channels] float32 tensor representing a batch of images with values between 0 and 255.0. Returns: preprocessed_inputs: A [batch, height_out, width_out, channels] float32 tensor representing a batch of images. """ return (2.0 / 255.0) * resized_inputs - 1.0 def _extract_proposal_features(self, preprocessed_inputs, scope): """Extracts first stage RPN features. Extracts features using the first half of the PNASNet network. We construct the network in `align_feature_maps=True` mode, which means that all VALID paddings in the network are changed to SAME padding so that the feature maps are aligned. Args: preprocessed_inputs: A [batch, height, width, channels] float32 tensor representing a batch of images. scope: A scope name. Returns: rpn_feature_map: A tensor with shape [batch, height, width, depth] end_points: A dictionary mapping feature extractor tensor names to tensors Raises: ValueError: If the created network is missing the required activation. """ del scope if len(preprocessed_inputs.get_shape().as_list()) != 4: raise ValueError('`preprocessed_inputs` must be 4 dimensional, got a ' 'tensor of shape %s' % preprocessed_inputs.get_shape()) with slim.arg_scope(pnasnet_large_arg_scope_for_detection( is_batch_norm_training=self._train_batch_norm)): with arg_scope([slim.conv2d, slim.batch_norm, slim.separable_conv2d], reuse=self._reuse_weights): _, end_points = pnasnet.build_pnasnet_large( preprocessed_inputs, num_classes=None, is_training=self._is_training, final_endpoint='Cell_7') # Note that both 'Cell_6' and 'Cell_7' have equal depth = 2160. # Cell_7 is the last cell before second reduction. rpn_feature_map = tf.concat([end_points['Cell_6'], end_points['Cell_7']], 3) # pnasnet.py does not maintain the batch size in the first dimension. # This work around permits us retaining the batch for below. batch = preprocessed_inputs.get_shape().as_list()[0] shape_without_batch = rpn_feature_map.get_shape().as_list()[1:] rpn_feature_map_shape = [batch] + shape_without_batch rpn_feature_map.set_shape(rpn_feature_map_shape) return rpn_feature_map, end_points def _extract_box_classifier_features(self, proposal_feature_maps, scope): """Extracts second stage box classifier features. This function reconstructs the "second half" of the PNASNet network after the part defined in `_extract_proposal_features`. Args: proposal_feature_maps: A 4-D float tensor with shape [batch_size * self.max_num_proposals, crop_height, crop_width, depth] representing the feature map cropped to each proposal. scope: A scope name. Returns: proposal_classifier_features: A 4-D float tensor with shape [batch_size * self.max_num_proposals, height, width, depth] representing box classifier features for each proposal. """ del scope # Number of used stem cells. num_stem_cells = 2 # Note that we always feed into 2 layers of equal depth # where the first N channels corresponds to previous hidden layer # and the second N channels correspond to the final hidden layer. hidden_previous, hidden = tf.split(proposal_feature_maps, 2, axis=3) # Note that what follows is largely a copy of build_pnasnet_large() within # pnasnet.py. We are copying to minimize code pollution in slim. # TODO(shlens,skornblith): Determine the appropriate drop path schedule. # For now the schedule is the default (1.0->0.7 over 250,000 train steps). hparams = pnasnet.large_imagenet_config() if not self._is_training: hparams.set_hparam('drop_path_keep_prob', 1.0) # Calculate the total number of cells in the network total_num_cells = hparams.num_cells + num_stem_cells normal_cell = pnasnet.PNasNetNormalCell( hparams.num_conv_filters, hparams.drop_path_keep_prob, total_num_cells, hparams.total_training_steps) with arg_scope([slim.dropout, nasnet_utils.drop_path], is_training=self._is_training): with arg_scope([slim.batch_norm], is_training=self._train_batch_norm): with arg_scope([slim.avg_pool2d, slim.max_pool2d, slim.conv2d, slim.batch_norm, slim.separable_conv2d, nasnet_utils.factorized_reduction, nasnet_utils.global_avg_pool, nasnet_utils.get_channel_index, nasnet_utils.get_channel_dim], data_format=hparams.data_format): # This corresponds to the cell number just past 'Cell_7' used by # _extract_proposal_features(). start_cell_num = 8 true_cell_num = start_cell_num + num_stem_cells with slim.arg_scope(pnasnet.pnasnet_large_arg_scope()): net = _build_pnasnet_base( hidden_previous, hidden, normal_cell=normal_cell, hparams=hparams, true_cell_num=true_cell_num, start_cell_num=start_cell_num) proposal_classifier_features = net return proposal_classifier_features def restore_from_classification_checkpoint_fn( self, first_stage_feature_extractor_scope, second_stage_feature_extractor_scope): """Returns a map of variables to load from a foreign checkpoint. Note that this overrides the default implementation in faster_rcnn_meta_arch.FasterRCNNFeatureExtractor which does not work for PNASNet checkpoints. Args: first_stage_feature_extractor_scope: A scope name for the first stage feature extractor. second_stage_feature_extractor_scope: A scope name for the second stage feature extractor. Returns: A dict mapping variable names (to load from a checkpoint) to variables in the model graph. """ variables_to_restore = {} for variable in tf.global_variables(): if variable.op.name.startswith( first_stage_feature_extractor_scope): var_name = variable.op.name.replace( first_stage_feature_extractor_scope + '/', '') var_name += '/ExponentialMovingAverage' variables_to_restore[var_name] = variable if variable.op.name.startswith( second_stage_feature_extractor_scope): var_name = variable.op.name.replace( second_stage_feature_extractor_scope + '/', '') var_name += '/ExponentialMovingAverage' variables_to_restore[var_name] = variable return variables_to_restore