def train_upscale_cnn_gpu(datasets, batch_size, input_shape, output_shape, channel_size, sess):
iterator = tf.compat.v1.data.make_one_shot_iterator(datasets)
dataset = iterator.get_next()
parsed_dataset = tf.io.parse_example(dataset, features={
'filename': tf.io.FixedLenFeature([], tf.string),
"x_image": tf.io.FixedLenFeature([], tf.string),
"y_image": tf.io.FixedLenFeature([], tf.string)})
x_s = tf.cast([tf.image.decode_jpeg(parsed_dataset['x_image'][index]) for index in range(0, batch_size)], tf.float32)
y_s = tf.cast([tf.image.decode_jpeg(parsed_dataset['y_image'][index]) for index in range(0, batch_size)], tf.float32)
# 七層CNN 同尺寸 低解 -> 高解
# with tf.device('/job:localhost/replica:0/task:0/device:XLA_GPU:0'):
# stage 1
y1 = cnn.add_deconv_layer(x_s, [3, 3, 128, 3], [batch_size, output_shape[0], output_shape[1], 128])
y2 = cnn.add_cnn_layer(y1, [1, 1, 128, 32])
y3 = cnn.add_cnn_layer(y2, [3, 3, 32, 3])
loss = tf.reduce_mean(tf.square(y3 - y_s))
optimizer = tf.compat.v1.train.AdamOptimizer(1e-4)
train_step = optimizer.minimize(loss)
if os.path.isfile("trained_parameters/upscale_128_3_layer_" + str(batch_size) + "p.index"):
saver = tf.compat.v1.train.Saver()
saver.restore(sess=sess, save_path="trained_parameters/upscale_128_3_layer_" + str(batch_size) + "p")
else:
saver = tf.compat.v1.train.Saver()
sess.run(tf.compat.v1.global_variables_initializer())
while True:
_, loss_val = sess.run([train_step, loss])
saver.save(sess=sess, save_path="trained_parameters/upscale_128_3_layer_" + str(batch_size) + "p")
print(loss_val)
def two_times_srcnn_model(self):
with self.g.as_default():
iterator = tf.compat.v1.data.make_one_shot_iterator(self.datasets)
dataset = iterator.get_next()
parsed_dataset = tf.io.parse_example(dataset, features={
'filename': tf.io.FixedLenFeature([], tf.string),
"x_image": tf.io.FixedLenFeature([], tf.string),
"y_image": tf.io.FixedLenFeature([], tf.string)})
x_s = tf.cast([tf.image.decode_jpeg(parsed_dataset['x_image'][index]) for index in range(0, self.batch_size)], tf.float32)
y_s = tf.cast([tf.image.decode_jpeg(parsed_dataset['y_image'][index]) for index in range(0, self.batch_size)], tf.float32)
y1 = cnn.add_deconv_layer(x_s, [3, 3, 128, 3], [self.batch_size, self.output_shape[0], self.output_shape[1], 128])
y2 = cnn.add_cnn_layer(y1, [1, 1, 128, 32])
self.y3 = cnn.add_cnn_layer(y2, [3, 3, 32, 3])
self.loss = tf.reduce_mean(tf.square(self.y3 - y_s))
optimizer = tf.compat.v1.train.AdamOptimizer(1e-4)
self.train_step = optimizer.minimize(self.loss)
self.saver = tf.compat.v1.train.Saver()
self.sess = tf.Session(graph=self.g)
if os.path.isfile("trained_parameters/srcnn_2x.index"):
self.saver.restore(sess=self.sess, save_path="trained_parameters/srcnn_2x")
else:
self.sess.run(tf.compat.v1.global_variables_initializer())
def cnn_with_batch_norm(x_s, k_size, input_features, output_features, strides):
y1 = cnn.add_cnn_layer(x_s, [k_size[0], k_size[1], input_features, output_features], strides=strides)
mean, variance = tf.nn.moments(y1, [0,1,2] if input_features > 1 else [0])
A_1_bn = tf.nn.batch_normalization(y1, mean=mean, variance=variance, offset=None, scale=None, variance_epsilon=1e-4)
y1_act = tf.nn.relu(A_1_bn)
A_1_pool = tf.nn.max_pool(y1_act, ksize=(1, 3, 3, 1), strides=(1, 1, 1, 1), padding='SAME')
return A_1_pool
def two_times_srcnn_predit(self, image, input_shape, channel_size):
x_s = tf.placeholder(tf.float32, [None, input_shape[0], input_shape[1], channel_size], "x_test")
y1 = cnn.add_deconv_layer(x_s, [3, 3, 128, 3], [batch_size, input_shape[0]*2, output_shape[1]*2, 128])
y2 = cnn.add_cnn_layer(y1, [1, 1, 128, 32])
y3 = cnn.add_cnn_layer(y2, [3, 3, 32, 3])
pred = tf.cast(y3, tf.uint8)
saver = tf.compat.v1.train.Saver()
sess = tf.Session()
if os.path.isfile("trained_parameters/srcnn_2x.index"):
saver.restore(sess=sess, save_path="trained_parameters/srcnn_2x")
else:
raise '請先完成訓練，或將指定權重放在正確的資料夾下'
out_put = sess.run(pred, feed_dict={x_s: image})
return out_put[0]
[np.shape(test_dataset)[0], shape_size]) test_labels = im_creator.test_labels() # define placeholder for inputs to network x_s = tf.placeholder(tf.float32, [None, 2352]) # 28x28 y_s = tf.placeholder(tf.float32, [None, 10]) keep_prob = tf.placeholder(tf.float32) x_image = tf.reshape(x_s, [-1, 28, 28, 3]) # print(x_image.shape) # [n_samples, 28,28,1] ## conv1 layer ## # W_conv1 = weight_variable([5,5, 3,32]) # patch 5x5, in size 1, out size 32 # b_conv1 = bias_variable([32]) # h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) # output size 28x28x32 # h_pool1 = max_pool_2x2(h_conv1) # output size 14x14x32 y1 = cnn.add_cnn_layer(x_image, tf.shape(x_image)[0], [5, 5, 3, 32]) h_pool1 = cnn.add_pooling_layer(y1) ## conv2 layer ## # W_conv2 = weight_variable([5,5, 32, 64]) # patch 5x5, in size 32, out size 64 # b_conv2 = bias_variable([64]) # h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) # output size 14x14x64 # h_pool2 = max_pool_2x2(h_conv2) # output size 7x7x64 ## fc1 layer ## # W_fc1 = weight_variable([7*7*64, 1024]) # b_fc1 = bias_variable([1024]) # # [n_samples, 7, 7, 64] ->> [n_samples, 7*7*64] # h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64]) # h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) # h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
def train_srcnn_gpu(datasets, batch_size, input_shape, output_shape, channel_size, sess):
iterator = tf.compat.v1.data.make_one_shot_iterator(datasets)
dataset = iterator.get_next()
parsed_dataset = tf.io.parse_example(dataset, features={
'filename': tf.io.FixedLenFeature([], tf.string),
"x_image": tf.io.FixedLenFeature([], tf.string),
"y_image": tf.io.FixedLenFeature([], tf.string)})
x_s = tf.cast([tf.image.decode_jpeg(parsed_dataset['x_image'][index]) for index in range(0, batch_size)], tf.float32)
y_s = tf.cast([tf.image.decode_jpeg(parsed_dataset['y_image'][index]) for index in range(0, batch_size)], tf.float32)
# deconv_3 = add_deconv_layer(x_s, [3, 3, 32, 3], [batch_size, input_shape[0] * 2, input_shape[1] * 2, 32], stride=2)
# 七層CNN 同尺寸 低解 -> 高解
# with tf.device('/job:localhost/replica:0/task:0/device:XLA_GPU:0'):
# stage 1
y1 = cnn.add_cnn_layer(x_s, [3, 3, 3, 64])
y2 = cnn.add_cnn_layer(y1, [1, 1, 64, 64])
y3 = cnn.add_cnn_layer(y2, [3, 3, 64, 64])
y4 = cnn.add_cnn_layer(y3, [3, 3, 64, 64])
y5 = cnn.add_cnn_layer(y4, [3, 3, 64, 128])
x_ = cnn.add_cnn_layer(x_s, [3, 3, 3, 128])
y5_act = tf.nn.relu(tf.add(y5, x_))
y6 = cnn.add_cnn_layer(y5_act, [3, 3, 128, 3])
# y1 = cnn_with_batch_norm(x_s, [3, 3], 3, 64, 1)
# y2 = cnn_with_batch_norm(y1, [1, 1], 64, 64, 1)
# y3 = cnn_with_batch_norm(y2, [3, 3], 64, 64, 1)
# y4 = cnn_with_batch_norm(y3, [3, 3], 64, 64, 1)
# y5 = cnn_with_batch_norm(y4, [3, 3], 64, 128, 1)
# stage 1 x
# x_ = cnn_with_batch_norm(x_s, [1,1], 3, 128, 1)
# mean, variance = tf.nn.moments(x_, [0,1,2])
# x_bn = tf.nn.batch_normalization(x_, mean=mean, variance=variance, offset=None, scale=None, variance_epsilon=1e-4)
# y5_act = tf.nn.relu(tf.add(y5, x_bn))
# y6 = cnn_with_batch_norm(y5_act, [3,3], 128, 3, 1)
# pool = tf.nn.avg_pool(y6, ksize=(1, 3, 3, 1), strides=(1, 1, 1, 1), padding='SAME')
loss = tf.reduce_mean(tf.square(y6 - y_s))
optimizer = tf.compat.v1.train.AdamOptimizer(1e-4)
train_step = optimizer.minimize(loss)
if os.path.isfile("trained_parameters/srcnn_f33" + str(batch_size) + "p.index"):
saver = tf.compat.v1.train.Saver()
saver.restore(sess=sess, save_path="trained_parameters/srcnn_f33" + str(batch_size) + "p")
else:
saver = tf.compat.v1.train.Saver()
sess.run(tf.compat.v1.global_variables_initializer())
while True:
_, loss_val = sess.run([train_step, loss])
saver.save(sess=sess, save_path="trained_parameters/srcnn_f33" + str(batch_size) + "p")
print(loss_val)