out_frame_count=1,
shuffle=False,
)
inputs = Input(shape=(11, 771))
fc1 = Flatten()(inputs)
fc1 = Dropout(0.3)(fc1)
fc1 = Dense(2048)(fc1)
fc1 = BatchNormalization(momentum=0.999)(fc1)
fc1 = PReLU()(fc1)
fc1 = Dropout(0.3)(fc1)
fc2 = Dense(2048)(fc1)
fc2 = BatchNormalization(momentum=0.999)(fc2)
fc2 = PReLU()(fc2)
fc2 = Dropout(0.3)(fc2)
out = Dense(257)(fc2)
model = Model(inputs=inputs, outputs=out)
adam = Adam(lr=0.0001, decay=1e-8)
model.compile(optimizer=adam, loss='mse')
for epoch in range(100):
train_loss = model.fit_generator(train_loader.batchify(), 5299)
test_loss = model.evaluate_generator(test_loader.batchify(), 1600)
print("Epoch", epoch)
print("Test loss:", test_loss)
layer = tf.layers.dropout(layer, rate=0.3, training=training)
out = tf.layers.dense(layer, 257, kernel_initializer=tf.random_normal_initializer(0, 0.02))
loss = tf.losses.mean_squared_error(out_frames, out)
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(0.0001, global_step, 1e4, 0.95)
train = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(100):
print("Epoch", epoch)
train_loss = 0
for in_frame_batch, out_frame_batch in train_loader.batchify():
fd = {in_frames: in_frame_batch, out_frames: out_frame_batch, training: True}
batch_loss, _ = sess.run([loss, train], fd)
train_loss += batch_loss
print("Train loss:", train_loss / 5300)
test_loss = 0
for in_frame_batch, out_frame_batch in dev_loader.batchify():
fd = {in_frames: in_frame_batch, out_frames: out_frame_batch, training: False}
test_loss += sess.run(loss, fd)
print("Test loss:", test_loss / 1612)
buffer_size=100,
context=5,
out_frame_count=1,
shuffle=False,
)
model = Sequential([
Reshape(target_shape=(11, 257, 1), input_shape=(11, 257)),
Conv2D(64, 7, padding='same', activation='relu'),
MaxPooling2D((2, 3), padding='same'),
Conv2D(96, 3, padding='same', activation='relu'),
MaxPooling2D((2, 3), padding='same'),
Conv2D(128, 3, padding='same', activation='relu'),
Flatten(),
Dense(2048, activation='relu'),
Dropout(0.3),
Dense(2048, activation='relu'),
Dropout(0.3),
Dense(257),
])
adam = Adam(lr=0.0001)
model.compile(optimizer=adam, loss='mse')
for epoch in range(100):
model.fit_generator(train_loader.batchify(), 21200)
test_loss = model.evaluate_generator(test_loader.batchify(), 6410)
print("Epoch", epoch)
print("Test loss:", test_loss)
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(0.0001, global_step, 1e4, 0.95)
train = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step)
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
min_loss = 1
for epoch in range(100):
print("Epoch", epoch)
train_loss = 0
count = 0
for in_frame_batch, out_frame_batch in train_loader.batchify(shuffle_batches=True, include_deltas=False):
fd = {in_frames: in_frame_batch, out_frames: out_frame_batch, training: True}
batch_loss, _ = sess.run([loss, train], fd)
train_loss += batch_loss
count += 1
print("Train loss:", train_loss / count)
test_loss = 0
count = 0
for in_frame_batch, out_frame_batch in dev_loader.batchify(include_deltas=False):
fd = {in_frames: in_frame_batch, out_frames: out_frame_batch, training: False}
test_loss += sess.run(loss, fd)
count += 1
test_loss /= count
loss, global_step=global_step)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
if args.load_dir:
saver.restore(sess, tf.train.latest_checkpoint(args.load_dir))
min_loss = 1
for epoch in range(0, 100):
print("Epoch", epoch)
train_loss = 0
count = 0
for noisy, clean in train_loader.batchify():
noisy = noisy[:, :, :257]
fd = {
noisy_frames: noisy,
clean_frames: clean,
resnet.training: True
}
batch_loss, _ = sess.run([loss, train], fd)
train_loss += batch_loss
count += 1
print("Train loss:", train_loss / count)
test_loss = 0
count = 0
for noisy, clean in dev_loader.batchify():