def build_architecture(self):
x = self.placeholders['shape_batch']
num_classes = self.num_classes # Chair/table classification
x = layers.conv3d(x, 64, 3, strides=2, padding='same', name='conv1', reuse=self.reuse)
x = tf.layers.batch_normalization(x, training=self.is_training, name='conv1_batch_norm',
reuse=self.reuse)
x = layers.relu(x, name='conv1_relu')
x = layers.conv3d(x, 128, 3, strides=2, padding='same', name='conv2', reuse=self.reuse)
x = tf.layers.batch_normalization(x, training=self.is_training, name='conv2_batch_norm',
reuse=self.reuse)
x = layers.relu(x, name='conv2_relu')
x = layers.conv3d(x, 256, 3, strides=2, padding='same', name='conv3', reuse=self.reuse)
x = tf.layers.batch_normalization(x, training=self.is_training, name='conv3_batch_norm',
reuse=self.reuse)
x = layers.relu(x, name='conv3_relu')
x = layers.avg_pooling3d(x, name='avg_pool4')
x = layers.dense(x, 128, name='fc5', reuse=self.reuse)
encoder_output = x
x = layers.dense(x, num_classes, name='fc6', reuse=self.reuse)
prob = layers.softmax(x, name='softmax_layer')
output_dict = {
'logits': x,
'probabilities': prob,
'encoder_output': encoder_output,
}
return output_dict
def build_architecture(self, inputs_dict):
x = inputs_dict['shape_batch']
if cfg.CONST.DATASET == 'shapenet':
num_classes = 2 # Chair/table classification
elif cfg.CONST.DATASET == 'primitives':
train_inputs_dict = open_pickle(cfg.DIR.PRIMITIVES_TRAIN_DATA_PATH)
val_inputs_dict = open_pickle(cfg.DIR.PRIMITIVES_VAL_DATA_PATH)
test_inputs_dict = open_pickle(cfg.DIR.PRIMITIVES_TEST_DATA_PATH)
f = lambda inputs_dict: list(inputs_dict['category_matches'].keys())
categories = f(train_inputs_dict) + f(val_inputs_dict) + f(test_inputs_dict)
categories = list(set(categories))
num_classes = len(categories)
else:
raise ValueError('Please select a valid dataset')
x = layers.conv3d(x, 64, 3, strides=2, padding='same', name='conv1', reuse=self.reuse)
x = tf.layers.batch_normalization(x, training=self.is_training, name='conv1_batch_norm',
reuse=self.reuse)
x = layers.relu(x, name='conv1_relu')
x = layers.conv3d(x, 128, 3, strides=2, padding='same', name='conv2', reuse=self.reuse)
x = tf.layers.batch_normalization(x, training=self.is_training, name='conv2_batch_norm',
reuse=self.reuse)
x = layers.relu(x, name='conv2_relu')
x = layers.conv3d(x, 256, 3, strides=2, padding='same', name='conv3', reuse=self.reuse)
x = tf.layers.batch_normalization(x, training=self.is_training, name='conv3_batch_norm',
reuse=self.reuse)
x = layers.relu(x, name='conv3_relu')
# get intermediate results
intermediate_output = x
x = layers.avg_pooling3d(x, name='avg_pool4')
x = layers.dense(x, 128, name='fc5', reuse=self.reuse)
encoder_output = x
x = layers.dense(x, num_classes, name='fc6', reuse=self.reuse)
prob = layers.softmax(x, name='softmax_layer')
output_dict = {
'logits': x,
'probabilities': prob,
'encoder_output': encoder_output,
'intermediate_output': intermediate_output
}
return output_dict
def build_architecture(self, inputs_dict):
print('--> building 3D GAN discriminator architecture')
leaky_relu_fn = layers.leaky_relu(leak=0.2)
with tf.variable_scope('architecture', reuse=self.reuse):
x = inputs_dict['shape_batch']
print('\t\tinput', x.get_shape())
x = layers.conv3d(x,
filters=64,
kernel_size=4,
strides=2,
padding='SAME',
activation=None,
name='conv1',
reuse=self.reuse)
x = leaky_relu_fn(x, name='conv1_lrelu')
x = layers.conv3d(x,
filters=128,
kernel_size=4,
strides=2,
padding='SAME',
activation=None,
name='conv2',
reuse=self.reuse)
x = leaky_relu_fn(x, name='conv2_lrelu')
x = layers.conv3d(x,
filters=256,
kernel_size=4,
strides=2,
padding='SAME',
activation=None,
name='conv3',
reuse=self.reuse)
x = leaky_relu_fn(x, name='conv3_lrelu')
# Text embedding input
embedding_fc_dim = 256
# Add FC layer
embedding_batch = inputs_dict['text_encoding_without_noise']
fc_embedding_output = layers.dense(embedding_batch,
embedding_fc_dim,
activation=None,
name='fc_embedding_1')
fc_embedding_output = leaky_relu_fn(fc_embedding_output,
name='fc_embedding_1_lrelu')
# Add FC layer
fc_embedding_output = layers.dense(fc_embedding_output,
embedding_fc_dim,
activation=None,
name='fc_embedding_2')
fc_embedding_output = leaky_relu_fn(fc_embedding_output,
name='fc_embedding_2_lrelu')
x = layers.conv3d(x,
filters=512,
kernel_size=4,
strides=2,
padding='SAME',
activation=None,
name='conv4',
reuse=self.reuse)
x = leaky_relu_fn(x, name='conv4_lrelu')
x = layers.conv3d(x,
filters=256,
kernel_size=2,
strides=2,
padding='SAME',
activation=None,
name='conv5',
reuse=self.reuse)
x = leaky_relu_fn(x, name='conv5_lrelu')
x = layers.reshape(x, [cfg.CONST.BATCH_SIZE, -1],
scope='reshape_to_concat')
x = layers.concat([x, fc_embedding_output],
axis=1,
name='concat_text_shape')
# Add FC layer
x = layers.dense(x, 128, activation=None, name='fc6')
x = leaky_relu_fn(x, name='fc6_lrelu')
# Add FC layer
x = layers.dense(x, 64, activation=None, name='fc7')
x = leaky_relu_fn(x, name='fc7_lrelu')
# Add FC layer
logits = layers.dense(x, 1, activation=None, name='fc8')
sigmoid_output = tf.sigmoid(logits)
return {'sigmoid_output': sigmoid_output, 'logits': logits}
def build_architecture(self, inputs_dict, last_activation=tf.sigmoid):
with tf.variable_scope('architecture'):
x = inputs_dict['text_encoding_with_noise']
print('\t\tinput', x.get_shape())
# Conv1
x = layers.dense(x,
units=512 * 4 * 4 * 4,
activation=None,
name='fc1',
reuse=self.reuse)
x = tf.layers.batch_normalization(x,
training=self.is_training,
name='fc1_batch_norm',
reuse=self.reuse)
x = layers.relu(x, name='fc1_relu')
x = layers.reshape(x,
shape=[-1, 4, 4, 4, 512],
scope='reshape_fc1')
# Conv2
x = layers.conv3d_transpose(x,
num_output_channels=512,
filter_size=4,
stride=1,
activation_fn=None,
scope='conv_transpose2',
reuse=self.reuse)
x = tf.layers.batch_normalization(x,
training=self.is_training,
name='conv_tranpose2_batch_norm',
reuse=self.reuse)
x = layers.relu(x, name='conv_transpose2_relu')
# Conv3
x = layers.conv3d_transpose(x,
num_output_channels=256,
filter_size=4,
stride=2,
activation_fn=None,
scope='conv_transpose3',
reuse=self.reuse)
x = tf.layers.batch_normalization(x,
training=self.is_training,
name='conv_tranpose3_batch_norm',
reuse=self.reuse)
x = layers.relu(x, name='conv_transpose3_relu')
# Conv4
x = layers.conv3d_transpose(x,
num_output_channels=128,
filter_size=4,
stride=2,
activation_fn=None,
scope='conv_transpose4',
reuse=self.reuse)
x = tf.layers.batch_normalization(
x,
training=self.is_training,
name='conv_transpose4_batch_norm',
reuse=self.reuse)
x = layers.relu(x, name='conv_transpose4_relu')
# Conv5
logits = layers.conv3d_transpose(
x,
num_output_channels=inputs_dict['num_output_channels'],
filter_size=4,
stride=2,
padding='SAME',
activation_fn=None,
scope='conv_transpose6',
reuse=self.reuse)
sigmoid_output = last_activation(logits)
return {'sigmoid_output': sigmoid_output, 'logits': logits}