""" Authors: Neal Jean, Michael Xie, Chris Yeh, Anthony Perez """ from .base_model import BaseModel from . import layers from .init_from_pretrained_utils import init_vggf_from_numpy import tensorflow as tf class VGGF(BaseModel): '''VGGF Model Adapted from what Neal and Michael used in the Science paper http://science.sciencemag.org/content/353/6301/790.long - add regularization losses to the regularization losses collection, not "losses" - add option to use dilated convolution in 1st layer - take (224, 224) input images, instead of (400, 400), which requires removing the pool7 layer Based on original VGG-F model https://arxiv.org/abs/1405.3531 ''' def __init__(self, inputs, is_training, fc_reg, conv_reg=0.0005, device='/gpu:0', reuse=False, use_dilated_conv_in_first_layer=False): ''' Args: inputs: tf.Tensor with shape [batch_size, img_height, img_width, img_depth], dtype tf.float32 is_training: bool fc_reg: float, regularization for weights in the fully-connected layer conv_reg: float, regularization for weights in the conv layers device: str, '/cpu:i' or '/gpu:i' where i is nonnegative int reuse: bool (reuse variable weights or create new weights) use_dilated_conv_in_first_layer: bool ''' super(VGGF, self).__init__( inputs=inputs, is_training=is_training, fc_reg=fc_reg, conv_reg=conv_reg) self.device = device self.update_ops_collection_key = tf.GraphKeys.UPDATE_OPS with tf.variable_scope('vggf'): self.set_up_graph(images=inputs, is_training=is_training, reuse=reuse, reg=conv_reg, use_dilation=use_dilated_conv_in_first_layer) def init_from_numpy(self, path, sess, hs_weight_init='random'): ''' Args: path: str, path to saved weights sess: tf.Session hs_weight_init: str, one of ['random', 'same'] ''' init_vggf_from_numpy(path, sess, hs_weight_init=hs_weight_init) def set_up_graph(self, images, is_training, reuse, reg, use_dilation): """ Builds computation graph. :param images: Tensor :param is_training: bool, whether weights are trainable and to use dropout :param reuse: bool (reuse variable weights or create new weights) :param reg: float, regularization for weights in the conv layers :param use_dilation: bool, whether to use dilated convolution in first layer """ with tf.device(self.device): self.images = images if use_dilation: # In shape: [B, 224, 224, 9]. Out shape: [B, 224, 224, 64], [B, 54, 54, 64] self.conv1 = layers.dilated_conv( self.images, 11, 64, 4, stride_pad="VALID", bias_init=tf.constant_initializer(value=0.1), l2loss=reg, reuse=reuse, trainable=is_training, name='conv1') else: # In shape: [B, 224, 224, C]. Out shape: [B, 54, 54, 64] self.conv1 = layers.conv2d( self.images, 11, 64, 4, pad='VALID', bias_init=tf.constant_initializer(value=0.1), l2loss=reg, reuse=reuse, trainable=is_training, name='conv1') self.norm1 = layers.lrn( self.conv1, local_size=5, alpha=0.0005, beta=0.75, bias=2.0, name='norm1') self.pool1 = layers.max_pool( self.norm1, 3, 2, pad='SAME', name='pool1') self.conv2 = layers.conv2d( self.pool1, 5, 256, 1, pad='SAME', bias_init=tf.constant_initializer(value=0.1), l2loss=reg, reuse=reuse, trainable=is_training, name='conv2') self.norm2 = layers.lrn( self.conv2, local_size=5, alpha=0.0005, beta=0.75, bias=2.0, name='norm2') self.pool2 = layers.max_pool( self.norm2, 3, 2, pad='VALID', name='pool2') self.conv3 = layers.conv2d( self.pool2, 3, 256, 1, pad='SAME', bias_init=tf.constant_initializer(value=0.1), l2loss=reg, reuse=reuse, trainable=is_training, name='conv3') self.conv4 = layers.conv2d( self.conv3, 3, 256, 1, pad='SAME', bias_init=tf.constant_initializer(value=0.1), l2loss=reg, reuse=reuse, trainable=is_training, name='conv4') self.conv5 = layers.conv2d( self.conv4, 3, 256, 1, pad='SAME', bias_init=tf.constant_initializer(value=0.1), l2loss=reg, reuse=reuse, trainable=is_training, name='conv5') self.pool5 = layers.max_pool( self.conv5, 3, 2, pad='SAME', name='pool5') self.conv6 = layers.conv2d( self.pool5, 6, 4096, 6, pad='VALID', bias_init=tf.constant_initializer(value=0.1), l2loss=reg, reuse=reuse, trainable=is_training, name='conv6') # after conv6: # - if images are (400, 400), then self.conv6 is (2, 2, 4096) # - if images are (224, 224), then self.conv6 is (1, 1, 4096) if is_training: self.dropout6 = layers.dropout( self.conv6, prob=0.5, name='dropout6') self.conv7 = layers.conv2d( self.dropout6, 1, 4096, 1, pad='VALID', bias_init=tf.constant_initializer(value=0.1), l2loss=reg, reuse=reuse, trainable=is_training, name='conv7') else: self.conv7 = layers.conv2d( self.conv6, 1, 4096, 1, pad='VALID', bias_init=tf.constant_initializer(value=0.1), l2loss=reg, reuse=reuse, trainable=is_training, name='conv7') # Extract features self.features_layer = tf.squeeze( self.conv7, squeeze_dims=[1, 2], name='features') # self.pool7 = layers.average_pool( # self.conv7, 2, 1, pad='VALID', name='pool7') # self.features = tf.squeeze( # self.pool7, squeeze_dims=[1, 2], name='features')