Myujth

I am creating simple code that creates an RGB image in grayscale. Even if this does not work properly, I hope the code will be executed. I have a question about thread usage. Below is the code.

with tf.variable_scope("color"): -> make variable(similar to VGG16)
def conv_layer(x, weights, biases, stride, name="convlayer", padding='SAME'):
 return tf.nn.relu(tf.nn.conv2d(x, weights, strides=stride, padding=padding) + biases, name=name)


 def read_my_file_format(filename_queue, randomize=False):
 reader = tf.WholeFileReader()
 key, file = reader.read(filename_queue)
 uint8image = tf.image.decode_jpeg(file, channels=3)
 uint8image = tf.random_crop(uint8image, (224, 224, 3))
 if randomize:
 uint8image = tf.image.random_flip_left_right(uint8image)
 uint8image = tf.image.random_flip_up_down(uint8image, seed=None)
 float_image = tf.div(tf.cast(uint8image, tf.float32), 255)
 return float_image

 def input_pipeline(filenames, batch_size, num_epochs=None):
 filename_queue = tf.train.string_input_producer(
 filenames, num_epochs=num_epochs, shuffle=False)
 example = read_my_file_format(filename_queue, randomize=False)
 min_after_dequeue = 5
 capacity = min_after_dequeue + 3 * batch_size
 example_batch = tf.train.shuffle_batch(
 [example], batch_size=batch_size, capacity=capacity,
 min_after_dequeue=min_after_dequeue)
 return example_batch

 with tf.name_scope("images_setting"):
 filenames = sorted(glob.glob("C:/example/*.jpg"))
 # filenames = ['C:/example/000005.jpg', 'C:/example/000007.jpg ~~~~']
 batch_size = 2
 num_epochs = 100

 colorimage = input_pipeline(filenames, batch_size, num_epochs=num_epochs)
 grayscale = tf.image.rgb_to_grayscale(colorimage)

 with tf.name_scope("layer_explain"):
 expand = tf.image.grayscale_to_rgb(grayscale)
 conv1_1 = conv_layer(expand, conv1_1_weights, conv1_1_biases, stride1, 'conv1_1')
 conv1_2 = conv_layer(conv1_1, conv1_2_weights, conv1_2_biases, stride1, 'conv1_2')

 conv2_1 = conv_layer(conv1_2, conv2_1_weights, conv2_1_biases, stride1, 'conv2_1')
 conv2_2 = conv_layer(conv2_1, conv2_2_weights, conv2_2_biases, stride1, 'conv2_2')

 conv3_1 = conv_layer(conv2_2, conv3_1_weights, conv3_1_biases, stride1, 'conv3_1')
 conv3_2 = conv_layer(conv3_1, conv3_2_weights, conv3_2_biases, stride1, 'conv3_2')
 conv3_3 = conv_layer(conv3_2, conv3_3_weights, conv3_3_biases, stride1, 'conv3_3')

 conv4_1 = conv_layer(conv3_3, conv4_1_weights, conv4_1_biases, stride1, 'conv4_1')
 conv4_2 = conv_layer(conv4_1, conv4_2_weights, conv4_2_biases, stride1, 'conv4_2')
 conv4_3 = conv_layer(conv4_2, conv4_3_weights, conv4_3_biases, stride1, 'conv4_3')

 conv5_1 = conv_layer(conv4_3, conv5_1_weights, conv5_1_biases, stride1, 'conv5_1')
 conv5_2 = conv_layer(conv5_1, conv5_2_weights, conv5_2_biases, stride1, 'conv5_2')
 conv5_3 = conv_layer(conv5_2, conv5_3_weights, conv5_3_biases, stride1, 'conv5_3')


 print("conv5_3: ", conv5_3)
 print("colorimage: ", colorimage)
 loss = tf.reduce_mean(tf.square(conv5_3 - colorimage))
 optimizer = tf.train.GradientDescentOptimizer(0.001)
 opt = optimizer.minimize(loss)


 init_global = tf.global_variables_initializer()
 init_local = tf.local_variables_initializer()
 sess = tf.Session()
 sess.run(init_global)
 sess.run(init_local)

 # Start input enqueue threads.
 coord = tf.train.Coordinator()
 threads = tf.train.start_queue_runners(sess=sess, coord=coord)

 print("expand: ", expand)
 print("conv1_1: ", conv1_1)

 print("grayscale: ", grayscale)
 print(filenames, '**********************')
 try:
 while not coord.should_stop():
 training_opt = sess.run(opt)

 for i in range(10):
 loss = sess.run(loss)
 print("cost: ".format(loss))
 except Exception as ex:
 print(ex)
 print("Done training -- epoch limit reached")
 finally:
 coord.request_stop()
 coord.join(threads)
 sess.close()

Error message:

(cost: 0.2219611406326294)

Fetch argument 0.22196114 has invalid type
, must be a string or Tensor. (Can not convert
a float32 into a Tensor or Operation.)
--> This is error... I think the message means something wrong at "loss funcion"

(Done training -- epoch limit reached)

The issue is the following line:

 loss = sess.run(loss)

The first time it runs, loss is a Tensor, so when session.run returns its value, the python variable loss is not a python float, which you cannot pass to session.run.

Do instead something like

 loss_value = sess.run(loss)

and you'll be fine.