stacyyang / mxnet-gluon-style-transfer Goto Github PK

Hi, I tried the following command as you provide:

python main.py eval --content-image images/content/venice-boat.jpg --style-image images/styles/candy.jpg --model models/21styles.params --content-size 1024

but got the following error:

Traceback (most recent call last):
File "main.py", line 214, in
main()
File "main.py", line 203, in main
evaluate(args)
File "main.py", line 129, in evaluate
style_model.load_params(args.model, ctx=ctx)
File "/usr/local/lib/python2.7/dist-packages/mxnet/gluon/block.py", line 317, in load_params
self.prefix)
File "/usr/local/lib/python2.7/dist-packages/mxnet/gluon/parameter.py", line 669, in load
"Parameter %s is missing in file %s"%(name[lprefix:], filename)
AssertionError: Parameter conv0_weight is missing in file models/21styles.params

I have checked the names of params in 21styles.params,
they are something like model.2.conv_block.2.weight, model.4.conv_block.2.bias and so on.
It seems that the net and the params are not matching.
Can you check that?
Thank you

Hi Hang,
In the paper, you have proposed an upsampled convolution to replace traditional deconvolution. However, in the code, it seems that upsampling + convolution is used instead of your proposed upsampled conv. May I ask how to implement your version of upsampled convolution?

Best,
Ed

I was trying to put a symbol into the model and save it as a symbolic network, but many of the Hybrid things only implement forward and not hybrid_forward.

I am struggling to do this myself.

do you know if anybody else has already done this as an exercise.

I am stuck on Inspiration(HybridBlock)

I was able to do Reflectance padding on my own after reading this document:

http://gluon.mxnet.io/chapter07_distributed-learning/hybridize.html

Any help would be apprieciated.

It would be a textbook example of being able to train in Python and deploy in C++

Sam

@zhanghang1989
Thanks for sharing this nice project.

I am trying to use only the evaluate method.
What value of args.ngf should I give ?
Rest all args are handled, but for this I am not able to find the default value.

hi:
I have tried to run

python main.py optim --content-image images/content/venice-boat.jpg --style-image images/styles/candy.jpg

but I got an error:

LocalFileSystem: fail to open "models\mxvgg.params"

I have checked the models dir, which there is no mxvgg.params file, but it has an download_model.sh script,
then, but it only download 21styles.params no mxvgg.params.
where can i find mxvgg.params?

Line 145-148 in main.py
I think content_image is obtained in RBG fomat.
However, utils.subtract_imagenet_mean_preprocess_batch take img in BGR fomat.
Is that right?

#main.py
    content_image = utils.tensor_load_rgbimage(args.content_image,ctx, size=args.content_size, keep_asp=True)
    content_image = utils.subtract_imagenet_mean_preprocess_batch(content_image)
    style_image = utils.tensor_load_rgbimage(args.style_image, ctx, size=args.style_size)
    style_image = utils.subtract_imagenet_mean_preprocess_batch(style_image)

#utils.py
def subtract_imagenet_mean_preprocess_batch(batch):
    """Subtract ImageNet mean pixel-wise from a BGR image."""
    batch = F.swapaxes(batch,0, 1)
    (r, g, b) = F.split(batch, num_outputs=3, axis=0)

And I think a better understanding way to transform RGB to BGR or vise versa is to use:
F.swapaxes(batch,0, 1) is not that obvious.

rgb_img = bgr_img[::-1,,] # C*H*W
bgr_img = rgb_img[::-1,,] # C*H*W

It may be helpful to keep this code up to date with the code in the mx-net repository for example in main.py load_params has been deprecated and replaced by load_parameters. Alternatively, this repository could be archived with all further development occuring in the mx net repository.