confused about the full structure about neural_style HOT 6 OPEN

Hi,
First, make sure you pulled the very last version of the project.

This project has pretty complicated structure because i've left all the work i made to facilitate my iteration on different models.

I don't really know about the "frozen effect" but this looks like "RGB <-> BGR" problem. Have you changed something to the code?

from neural_style.

  you are right, it is the problem about "RGB <-> BGR" and I have solved this problem. 

I got another question about ConvolutionTranspose2D.py in model's directory.
For the using of "tf.nn.conv2d_transpose", exactly the using of tf.pack to outshape,

it leads to the input image size must be the same as traing dataset when doing forward only. For example, if I used an image of 640_480 to do style transforming, but 512_512 for training, it will send errors like "InvalidArgumentError: Conv2DCustomBackpropInput: Size of out_backprop doesn't match computed: actual = 188, computed = 250". Is there any method to solve this problem or it just can process image size equal to training dataSet

from neural_style.

Since the models are fully convolutionnal, you should be able to train and predict on different image sizes:
Also the input_shape for prediction is built dynamically from the image_size. Make sure the input_shape is well defined.

Also, i've never tested with rectangular shapes (non squared), there could be a bug there. Check with a squared image to see if you still have the problem.

from neural_style.

Thanks for your reply. perphase my statement is not very ckearly. I used "export_keras_model.py" to convert tensorflow's chkp and meta model files to pb format, and do forward using the pb format's model file. Just like you do in mobile_app directory. some code as follows:

def create_graph(pbModelPath):
with tf.gfile.FastGFile(pbModelPath, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(graph_def, name='')

def forward_test(imagePath, pbModelPath,imgOutPath):
img_src = Image.open(imagePath).convert('RGB')
img = np.asarray(img_src, dtype=np.float32)
batch = img.reshape((1,) + img.shape)

device_ = '/cpu:0'
with tf.device(device_):
    create_graph(pbModelPath)
    with tf.Session() as sess:
        out_tensor = sess.graph.get_tensor_by_name("mul:0")         # "add_35" could handle any input image size, "mul"  can't
        in_tensor = sess.graph.get_tensor_by_name("input_node:0")
        outputs = sess.run(out_tensor, feed_dict={in_tensor:batch}) 
        print outputs.shape
    outputs = outputs.reshape((outputs.shape[1:]))
    imgOut = Image.fromarray(np.uint8(outputs))
    imgOut.save(imgOutPath)

when I set "out_tensor = sess.graph.get_tensor_by_name("add_35:0")", it can handle any input image size. but when i set it to "out_tensor = sess.graph.get_tensor_by_name("mul:0")", which should be the correct output tensor, it strictly need input image size == training using image size. I almostly sure the problem is th.nn.conv2d_transpose 's outshape, could you give some solutions

from neural_style.

Hum,
maybe something changed in keras or tensorflow which my layer implementation of conv2d_transpose doesn't handle anymore.

I'm sorry i won't be able to dig on this issue but if you found out what is happening, i'll be glad to merge a PR on your issue 👍🏻

from neural_style.

I suspect that is the case as well

from neural_style.