def generator_forward(config,
noise=None,
scope="generator",
name=None,
reuse=False,
num_samples=-1):
with tf.variable_scope(scope, name, reuse=reuse):
if noise is None:
noise = tf.random_normal(
[config.batch_size if num_samples == -1 else num_samples, 128],
name="noise")
output = fully_connected(noise, 4 * 4 * 4 * config.dim)
output = batch_normalization(output)
output = tf.nn.relu(output)
output = tf.reshape(output, [-1, 4, 4, 4 * config.dim])
output = conv_2d_transpose(output,
2 * config.dim,
5, [8, 8],
strides=2)
output = output[:, :7, :7, :]
output = conv_2d_transpose(output, config.dim, 5, [14, 14], strides=2)
output = tf.nn.relu(output)
output = conv_2d_transpose(output, 1, 5, [28, 28], strides=2)
output = tf.tanh(output)
return output
def Network():
network = input_data(shape=[None, 128, 128, 1], name='input')
network = conv_2d(network, 3, 3, strides=1, activation='relu')
layer1 = max_pool_2d(network, 3)
network = conv_2d(layer1, 64, 3, strides=1, activation='relu')
network = conv_2d(network, 64, 3, strides=1, activation='relu')
layer2 = max_pool_2d(network, 3)
network = conv_2d(layer2, 128, 3, strides=1, activation='relu')
network = conv_2d(network, 128, 3, strides=1, activation='relu')
network = conv_2d(network, 128, 3, strides=1, activation='relu')
network = conv_2d(network, 128, 3, strides=2, activation='relu')
layer3 = max_pool_2d(network, 3)
network = conv_2d(layer3, 256, 3, strides=1, activation='relu')
network = conv_2d(network, 256, 3, strides=1, activation='relu')
network = conv_2d(network, 256, 3, strides=1, activation='relu')
network = conv_2d(network, 256, 3, strides=2, activation='relu')
layer4 = max_pool_2d(network, 3, name="layer4")
network = conv_2d(layer4, 512, 3, strides=1, activation='relu')
network = conv_2d(network, 512, 3, strides=1, activation='relu')
network = conv_2d(network, 512, 3, strides=1, activation='relu')
network = conv_2d(network, 512, 3, strides=2, activation='relu')
layer5 = max_pool_2d(network, 3, name="layer5")
network = conv_2d(layer5, 256, 1, strides=1, activation='relu')
network = batch_normalization(network)
network = conv_2d_transpose(network,
256,
1, [32, 32],
strides=2,
activation='relu')
layer4 = batch_normalization(layer4)
network = tflearn.layers.merge_ops.merge([layer4, network],
mode="elemwise_sum")
network = conv_2d(network, 128, 3, strides=1, activation='relu')
network = conv_2d_transpose(network,
128,
1, [64, 64],
strides=2,
activation='relu')
layer3 = batch_normalization(layer3)
network = tflearn.layers.merge_ops.merge([layer3, network],
mode="elemwise_sum",
axis=1)
network = conv_2d(network, 64, 3, strides=1, activation='relu')
network = conv_2d_transpose(network,
64,
1, [128, 128],
strides=2,
activation='relu')
layer2 = batch_normalization(layer2)
network = tflearn.layers.merge_ops.merge([layer2, network],
mode="elemwise_sum",
axis=1)
network = conv_2d(network, 3, 3, strides=1, activation='relu')
layer1 = batch_normalization(layer1)
network = tflearn.layers.merge_ops.merge([layer1, network],
mode="elemwise_sum",
axis=1)
network = conv_2d(network, 3, 3, strides=1, activation='relu')
network = conv_2d(network, 2, 3, strides=1, activation='softmax')
return network
def model():
# loading saved numpy files with training data
INPATH = 'training_input_cropped.npy'
OUTPATH = 'training_output_cropped.npy'
if not os.path.exists(INPATH) or not os.path.exists(OUTPATH):
training_input, training_output = fetch_data.create_train_data()
else:
training_input = np.load(INPATH)
training_output = np.load(OUTPATH)
#split into training and validation set
X = training_input[:88000].reshape(-1, 100, 100, 3)
Y = training_output[:88000].reshape(-1, 100, 100, 3)
test_x = training_input[12000:].reshape(-1, 100, 100, 3)
test_y = training_output[12000:].reshape(-1, 100, 100, 3)
convnet = input_data(shape=[None, 100, 100, 3], name='input')
convnet = conv_2d(convnet, 32, 5, activation='relu')
convnet = conv_2d(convnet, 64, 5, activation='relu')
convnet = conv_2d(convnet, 128, 5, activation='relu')
convnet = conv_2d(convnet, 256, 5, activation='relu')
convnet = conv_2d(convnet, 512, 5, activation='relu')
convnet = conv_2d_transpose(convnet, 256, 5, [100, 100], activation='relu')
convnet = conv_2d_transpose(convnet, 128, 5, [100, 100], activation='relu')
convnet = conv_2d(convnet, 32, 3, activation='relu')
convnet = dropout(convnet, 0.8)
convnet = conv_2d(convnet, 3, 5, activation='linear')
# Adam optizer used to minimize mean square loss between pixel values
convnet = regression(convnet,
optimizer='adam',
learning_rate=0.001,
loss='mean_square',
name='targets')
model = tflearn.DNN(convnet, tensorboard_dir='log')
if os.path.exists('model3'):
model.load('model3\sketch1')
print('model loaded')
else:
model.fit({'input': X}, {'targets': Y},
n_epoch=6,
validation_set=({
'input': test_x
}, {
'targets': test_y
}),
snapshot_step=500,
show_metric=True)
model.save('model3\sketch1')
return model
def generator_forward(config,
noise=None,
scope="generator",
name=None,
reuse=False,
num_samples=-1):
with tf.variable_scope(scope, name, reuse=reuse):
if noise is None:
noise = tf.random_normal(
[config.batch_size if num_samples == -1 else num_samples, 128],
name="noise")
output = fully_connected(noise,
4 * 4 * 8 * config.gen_dim,
name="input")
output = tf.reshape(output, [-1, 4, 4, 8 * config.gen_dim])
output = batch_normalization(output)
output = relu(output)
output = conv_2d_transpose(output,
4 * config.gen_dim,
5, [8, 8],
name="conv1",
strides=2)
output = batch_normalization(output)
output = relu(output)
output = conv_2d_transpose(output,
2 * config.gen_dim,
5, [16, 16],
name="conv2",
strides=2)
output = batch_normalization(output)
output = relu(output)
output = conv_2d_transpose(output,
config.gen_dim,
5, [32, 32],
name="conv3",
strides=2)
output = batch_normalization(output)
output = relu(output)
output = conv_2d_transpose(output,
3,
5, [64, 64],
name="conv4",
strides=2)
output = tf.tanh(output)
return output
def deconv(h_0,
filters,
kernel_size,
strides,
output_shape=[],
activation='relu',
trainable=True):
if not output_shape:
h = int(((h_0.shape[1] - 1) * strides) + kernel_size) - 1
w = int(((h_0.shape[2] - 1) * strides) + kernel_size) - 1
else:
h = int(output_shape[1])
w = int(output_shape[2])
init = tflearn.initializations.truncated_normal(shape=None,
mean=0.0,
stddev=0.01,
dtype=tf.float32,
seed=None)
h1 = conv_2d_transpose(h_0,
nb_filter=filters,
filter_size=kernel_size,
strides=strides,
output_shape=[h, w],
weights_init=init,
trainable=trainable)
h1_bn = batch_normalization(h1, trainable=trainable)
if activation == 'relu':
h1_o = tf.nn.relu(h1_bn)
elif activation == 'none':
h1_o = h1_bn
print(h1_o.shape)
return h1_o
def conv_2d_transpose_layer(input, n_filters, stride, output_shape):
return conv.conv_2d_transpose(input,
n_filters,
3,
output_shape=output_shape,
strides=stride,
padding='same',
activation='elu',
bias_init='zeros',
scope=None,
name='Conv3D')
def default_generator(net):
net = fully_connected(net, 7 * 7 * 16, scope='fc')
net = tf.reshape(net, [tf.shape(net)[0], 7, 7, 16])
net = conv_2d_transpose(net, 1, 5, [28, 28], scope="l0", strides=4, bias=True)
return net
conv4_1 = conv_2d(pool3, nb_filter = 512, filter_size = 3, strides=1, padding = 'same' ,activation='relu') conv4_2 = conv_2d(conv4_1, nb_filter = 512, filter_size = 3, strides=1, padding = 'same' ,activation='relu') conv4_3 = conv_2d(conv4_2, nb_filter = 512, filter_size = 3, strides=1, padding = 'same' ,activation='relu') pool4 = max_pool_2d(conv4_3, kernel_size = 2, strides=2) conv5_1 = conv_2d(pool4, nb_filter = 512, filter_size = 3, strides=1, padding = 'same' ,activation='relu') conv5_2 = conv_2d(conv5_1, nb_filter = 512, filter_size = 3, strides=1, padding = 'same' ,activation='relu') conv5_3 = conv_2d(conv5_2, nb_filter = 512, filter_size = 3, strides=1, padding = 'same' ,activation='relu') # DSN conv 1 score_dsn1_up = conv_2d(conv1_2, nb_filter = 1 , filter_size = 1, strides=1, padding = 'same' ,activation='relu') dsn1_loss = regression(score_dsn1_up, optimizer='rmsprop',loss='categorical_crossentropy',learning_rate=0.0001) # DSN conv 2 score_dsn2 = conv_2d(conv2_2, nb_filter = 1 , filter_size = 1, strides = 1, padding = 'same' ,activation='relu') score_dsn2_up = conv_2d_transpose(score_dsn2, nb_filter = 1 , filter_size = 4,output_shape=[224,224], strides = 2, padding = 'same' ,activation='relu') dsn2_loss = regression(score_dsn2_up, optimizer='rmsprop',loss='categorical_crossentropy',learning_rate=0.0001) # DSN conv 3 score_dsn3 = conv_2d(conv3_3, nb_filter = 1 , filter_size = 1, strides = 1, padding = 'same' ,activation='relu') score_dsn3_up = conv_2d_transpose(score_dsn3, nb_filter = 1 , filter_size = 8, output_shape=[224,224],strides = 4, padding = 'same' ,activation='relu') dsn3_loss = regression(score_dsn3_up, optimizer='rmsprop',loss='categorical_crossentropy',learning_rate=0.0001) # DSN conv 4 score_dsn4 = conv_2d(conv4_3, nb_filter = 1 , filter_size = 1, strides = 1, padding = 'same' ,activation='relu') score_dsn4_up = conv_2d_transpose(score_dsn4, nb_filter = 1 , filter_size = 16, output_shape=[224,224],strides = 8, padding = 'same' ,activation='relu') dsn4_loss = regression(score_dsn4_up, optimizer='rmsprop',loss='categorical_crossentropy',learning_rate=0.0001) # DSN conv 5 score_dsn5 = conv_2d(conv5_3, nb_filter = 1 , filter_size = 1, strides = 1, padding = 'same' ,activation='relu') score_dsn5_up = conv_2d_transpose(score_dsn5, nb_filter = 1 , filter_size = 32, output_shape=[224,224],strides = 16, padding = 'same' ,activation='relu')
def build_tiramisu(network):
n = 5
Ni = 8
#Pool1
network_1 = conv_2d(network, Ni, 3, regularizer='L2',
weight_decay=0.0001) # 256
network_1 = residual_block(network_1, n, Ni)
network_1 = residual_block(network_1, 1, Ni)
pool_1 = max_pool_2d(network_1, 2) # downsampling 2x - 128
#Pool2
network_2 = residual_block(pool_1, n - 1, 2 * Ni)
network_2 = residual_block(network_2, 1, 2 * Ni)
pool_2 = max_pool_2d(network_2, 2) # downsampling 4x - 64
#Pool3
network_3 = residual_block(pool_2, n - 1, 4 * Ni)
network_3 = residual_block(network_3, 1, 4 * Ni)
pool_3 = max_pool_2d(network_3, 2) # downsampling 8x - 32
#Pool4
network_4 = residual_block(pool_3, n - 1, 8 * Ni)
network_4 = residual_block(network_4, 1, 8 * Ni)
pool_4 = max_pool_2d(network_4, 2) # downsampling 16x - 16
#Pool5
network_5 = residual_block(pool_4, n - 1, 16 * Ni)
network_5 = residual_block(network_5, 1, 16 * Ni)
# /////////////////////////////////////////////////////////////////////////////////
Unpool1 = conv_2d_transpose(network_5,
8 * Ni,
3,
strides=2,
output_shape=[HEIGHT // 8, WIDTH // 8, 8 * Ni])
merge1 = merge([Unpool1, network_4], mode='concat', axis=3) # merge
merge1 = conv_2d(merge1, 8 * Ni, 3, activation='relu')
merge1 = conv_2d(merge1, 8 * Ni, 3, activation='relu') #
Unpool2 = conv_2d_transpose(merge1,
4 * Ni,
3,
strides=2,
output_shape=[HEIGHT // 4, WIDTH // 4, 4 * Ni])
merge2 = merge([Unpool2, network_3], mode='concat', axis=3) # merge
merge2 = conv_2d(merge2, 4 * Ni, 3, activation='relu')
merge2 = conv_2d(merge2, 4 * Ni, 3, activation='relu')
Unpool3 = conv_2d_transpose(merge2,
2 * Ni,
3,
strides=2,
output_shape=[HEIGHT // 2, WIDTH // 2, 2 * Ni])
merge3 = merge([Unpool3, network_2], mode='concat', axis=3) # merge
merge3 = conv_2d(merge3, 2 * Ni, 3, activation='relu')
merge3 = conv_2d(merge3, 2 * Ni, 3, activation='relu')
Unpool4 = conv_2d_transpose(merge3,
Ni,
3,
strides=2,
output_shape=[HEIGHT, WIDTH, Ni])
merge4 = merge([Unpool4, network_1], mode='concat', axis=3) # merge
merge4 = conv_2d(merge4, Ni, 3, activation='relu')
merge4 = conv_2d(merge4, Ni, 3, activation='relu')
final_merge = conv_2d(merge4, 3, 1, activation='relu')
network = tflearn.regression(final_merge,
optimizer='adam',
loss='mean_square')
return network
def build_unet(network):
Ni = 8
#Pool1
network_1 = conv_2d(network, Ni, 3, activation='relu')
network_1 = conv_2d(network_1, Ni, 3, activation='relu')
pool1 = max_pool_2d(network_1, 2) # downsampling 2x
#Pool2
network_2 = conv_2d(pool1, 2 * Ni, 3, activation='relu')
network_2 = conv_2d(network_2, 2 * Ni, 3, activation='relu')
pool2 = max_pool_2d(network_2, 2) # downsampling 4x
#Pool3
network_3 = conv_2d(pool2, 4 * Ni, 3, activation='relu')
network_3 = conv_2d(network_3, 4 * Ni, 3, activation='relu')
pool3 = max_pool_2d(network_3, 2) # downsampling 8x
#Pool4
network_4 = conv_2d(pool3, 8 * Ni, 3, activation='relu')
network_4 = conv_2d(network_4, 8 * Ni, 3, activation='relu')
pool4 = max_pool_2d(network_4, 2) # downsampling 16x
#Pool5
network_5 = conv_2d(pool4, 16 * Ni, 3, activation='relu')
network_5 = conv_2d(network_5, 16 * Ni, 3, activation='relu')
Unpool1 = conv_2d_transpose(network_5,
8 * Ni,
3,
strides=2,
output_shape=[HEIGHT // 8, WIDTH // 8, 8 * Ni])
merge1 = merge([Unpool1, network_4], mode='concat', axis=3) # merge
merge1 = conv_2d(merge1, 8 * Ni, 3, activation='relu')
merge1 = conv_2d(merge1, 8 * Ni, 3, activation='relu')
Unpool2 = conv_2d_transpose(merge1,
4 * Ni,
3,
strides=2,
output_shape=[HEIGHT // 4, WIDTH // 4, 4 * Ni])
merge1 = merge([Unpool2, network_3], mode='concat', axis=3) # merge
merge1 = conv_2d(merge1, 4 * Ni, 3, activation='relu')
merge1 = conv_2d(merge1, 4 * Ni, 3, activation='relu')
Unpool3 = conv_2d_transpose(merge1,
2 * Ni,
3,
strides=2,
output_shape=[HEIGHT // 2, WIDTH // 2, 2 * Ni])
merge1 = merge([Unpool3, network_2], mode='concat', axis=3) # merge
merge1 = conv_2d(merge1, 2 * Ni, 3, activation='relu')
merge1 = conv_2d(merge1, 2 * Ni, 3, activation='relu')
Unpool4 = conv_2d_transpose(merge1,
Ni,
3,
strides=2,
output_shape=[HEIGHT, WIDTH, Ni])
merge1 = merge([Unpool4, network_1], mode='concat', axis=3) # merge
merge1 = conv_2d(merge1, Ni, 3, activation='relu')
merge1 = conv_2d(merge1, Ni, 3, activation='relu')
merge1 = conv_2d(merge1, 3, 1, activation='relu')
network = tflearn.regression(merge1, optimizer='adam', loss='mean_square')
return network
def build_segnet_half(network):
#Pool1
network_1 = conv_2d(network, 8, 3,
activation='relu') #output 2x_downsampled
network_1 = conv_2d(network_1, 8, 3,
activation='relu') #output 2x_downsampled
pool1 = max_pool_2d(network_1, 2)
#Pool2
network_2 = conv_2d(pool1, 16, 3,
activation='relu') #output 4x_downsampled
network_2 = conv_2d(network_2, 16, 3,
activation='relu') #output 4x_downsampled
pool2 = max_pool_2d(network_2, 2)
#Pool3
network_3 = conv_2d(pool2, 32, 3,
activation='relu') #output 8x_downsampled
network_3 = conv_2d(network_3, 32, 3,
activation='relu') #output 8x_downsampled
pool3 = max_pool_2d(network_3, 2)
#Pool4
network_3 = conv_2d(pool3, 64, 3,
activation='relu') #output 16x_downsampled
network_3 = conv_2d(network_3, 64, 3,
activation='relu') #output 16x_downsampled
pool4 = max_pool_2d(network_3, 2)
# ----- decoder -----
decoder = conv_2d_transpose(pool4,
64,
3,
strides=4,
output_shape=[
HEIGHT // 4, WIDTH // 4, 64
]) # 16x downsample to 4x downsample
decoder = conv_2d(decoder, 64, 3, activation='relu')
pool5 = conv_2d(decoder, 64, 3, activation='relu')
decoder = conv_2d_transpose(pool3,
32,
3,
strides=2,
output_shape=[
HEIGHT // 4, WIDTH // 4, 32
]) # 8x downsample to 4x downsample
decoder = conv_2d(decoder, 32, 3, activation='relu')
pool6 = conv_2d(decoder, 32, 3, activation='relu')
pool6 = merge([pool6, pool5, pool2], mode='concat',
axis=3) #merge all 4x downsampled layers
decoder = conv_2d_transpose(pool6,
16,
3,
strides=4,
output_shape=[HEIGHT, WIDTH, 16])
decoder = conv_2d(decoder, 16, 3, activation='relu')
pool6 = conv_2d(decoder, 16, 3, activation='relu')
decoder = conv_2d(pool6, CHANNELS, 1)
network = tflearn.regression(decoder, optimizer='adam', loss='mean_square')
return network
def build_autoencoder(network):
# encoder
encoder = conv_2d(network, 16, 7, activation='relu')
encoder = conv_2d(encoder, 16, 7, activation='relu')
encoder = max_pool_2d(encoder, 2)
encoder = conv_2d(encoder, 32, 5, activation='relu')
encoder = conv_2d(encoder, 32, 5, activation='relu')
encoder = max_pool_2d(encoder, 2)
encoder = conv_2d(encoder, 64, 3, activation='relu')
encoder = conv_2d(encoder, 64, 3, activation='relu')
encoder = max_pool_2d(encoder, 2)
# decoder
decoder = conv_2d_transpose(encoder,
64,
3,
strides=2,
output_shape=[HEIGHT // 4, WIDTH // 4, 64])
decoder = conv_2d(decoder, 64, 3, activation='relu')
decoder = conv_2d(decoder, 64, 3, activation='relu')
decoder = conv_2d_transpose(decoder,
32,
5,
strides=2,
output_shape=[HEIGHT // 2, WIDTH // 2, 32])
decoder = conv_2d(decoder, 32, 5, activation='relu')
decoder = conv_2d(decoder, 32, 5, activation='relu')
decoder = conv_2d_transpose(decoder,
16,
7,
strides=2,
output_shape=[HEIGHT, WIDTH, 16])
decoder = conv_2d(decoder, 16, 7, activation='relu')
decoder = conv_2d(decoder, 16, 7, activation='relu')
# decoder = upsample_2d(decoder, 2)
decoder = conv_2d(decoder, CHANNELS, 1)
def my_loss(y_pred, y_true):
return tflearn.objectives.weak_cross_entropy_2d(y_pred,
y_true,
num_classes=3)
def my_metric(y_pred, y_true):
return tflean.metrics.Top_k(k=3)
network = regression(
decoder,
optimizer='adam',
#loss='mean_square',
loss='categorical_crossentropy',
#loss='weak_cross_entropy_2d',
#loss=my_loss,
#learning_rate=0.00005,
#learning_rate=0.0005,
#metric=my_metric
)
return network