Hi,
Thank you for the great work.

I've been trying to replicate the results of this paper "https://arxiv.org/pdf/1508.06576.pdf" using your implementation, by setting "random_init=true, gram_ratio=0.5, content_ratio=0.5". However, all I get is noise.

Can you give any insight that clarifies the issue I'm having?

Thank you again.
Regards,

In the function stylenet_core.py : get_style_cost_gram,

layer_gram_lost = l2_norm_cost(G1 - G2) / (size ** 2)

the code in function l2_norm_cost() already divides the result by (size ** 2), and in this line, the result is divided by (size ** 2) again. Is there a problem?

patches map finished 0 3.54528403282
finished joining patches 5.4882171154
Traceback (most recent call last):
File "stylenet_patch.py", line 342, in
"./images/husky_paint_region.jpg", "./test_data/husky_real_region.jpg", height=500)
File "stylenet_patch.py", line 330, in render_gen
output_file=output_file)
File "stylenet_patch.py", line 141, in render
blur_mapping=blur_mapping)
File "/Users/hcai/work/ai/stylenet/stylenet_core.py", line 240, in get_style_cost_patch2
np.save(full_patch_file, full_patch_out)
File "/Library/Python/2.7/site-packages/numpy/lib/npyio.py", line 477, in save
fid = open(file, "wb")
IOError: [Errno 2] No such file or directory: './train/husky_paint-husky_real-style_map_3_full.npy'

Looks like the following line of code in stylenet_core.py is causing an exception when using a large height value
filters = tf.constant(filters_out, tf.float32)

Is there any fix for this? I looked around on google for a bit and found something about using feed_dict but how to do this was not very clear to me.

Hey, so I started using your code and I have encountered quite high memory usage. I have GPU-enabled Tensorflow, CUDA works well and it stays around 2.5GB in GPU memory for the default settings. My issue is with RAM - it keeps rising up to around 9GB. Is this normal behaviour or is there something wrong with my Tensorflow installation? Do you have maybe some tips for other settings I might use in your code?

Thanks

I am trying to understand the MRF paper and after I read your code, one part confuses me a little bit, which is get_style_cost_patch2() in stylenet_core.py. I am wondering if you could help me out?

From my understanding, function get_style_cost_patch2() is trying to calculate MRF cost, defined as Eq. (2) in the paper. And within it, calculating Eq.(3) is done by this line of code:
max_arg_out, patches = _create_patch(sess, content_input, style_input, content_regions, style_regions, blur_mapping)
It seems that this line of code finds the maximized matching patches between the style image and the content image. However in the paper, Eq (3) aims to find the maximized matching patches between the style image and the synthesized image.

I am wondering if I miss something from your implementation, or you modify the algorithm on purpose?

I also tried to modify this part back according to the paper and write a simply modified function get_style_cost_patch(). However TF reports an error saying Attempting to use uninitialized value Variable. I am wondering if you have some insight about why the error happens? Thanks,
My modified code is below:

def get_style_cost_patch(sess, var_input, style_input, save_file,
                          content_regions=None, style_regions=None,
                          load_saved_mapping=True, blur_mapping=False):
    dim = var_input.get_shape().as_list()
    h, w, d = dim[1], dim[2], dim[3]
    ph, pw = h - 2, w - 2
    pn = ph * pw
    full_patch_file = save_file + "_full.npy"

    full_patch_out = None
    if load_saved_mapping:
        try:
            full_patch_out = np.load(full_patch_file)
        except:
            print("saved full patch not found")
    if full_patch_out is None:
        with tf.device("/cpu:0"):
            max_arg_out, patches = _create_patch(sess, var_input, style_input,
                                                 content_regions, style_regions, blur_mapping)
            assert patches.get_shape().as_list() == [3, 3, d, pn]

            max_arg_out = np.reshape(max_arg_out, [ph, pw])
            # print max_arg_out
            full_patch = _join_patches_np(sess, patches, max_arg_out)

            assert full_patch.get_shape() == var_input.get_shape()
            full_patch_out = sess.run(full_patch)
            np.save(full_patch_file, full_patch_out)

    full_patch = tf.constant(full_patch_out)

    cost = l2_norm_cost(var_input - full_patch)

    return cost

this work is amazing! But is there any arxiv paper for this stylenet? i want to know if there are some more details about the two modifications. Thanks.

Hi,@machrisaa

Thank you for the released code.The link for Tensorflow in the "#Requirement" in readme is broken.And Ido not Know where the "tensoflow_vgg" in the file "custom_vgg19.py" is .How can I get it?

Looking forward to receiving your early reply.