eriklindernoren / fast-neural-style-transfer Goto Github PK

Hi, you are handsome and your work are awesome! Thank you for your work.
I got a problem when I try to train a smaller 'TransformerNet' model, all input images are resize to 256x256 while training.
The 'Content' and the 'Style' reports high loss value:

And while train the model after some epoches, the model will never generate a picture:

I'm looking forward to your guidance.

i hava a video about 15 seconds, and after processing i got a nearly 1G gif. Is there anyone can tell me why it is so big ?

I am confused why there isn't a tanh function in the output of TransformerNet network. Can you please tell me your approach in the conversion of the output images?

Could you tell which dataset I should ues? And what format the dataset should be ?After reading README.md,I cannot understand how to train the network clearly.Thanks!

The param "lambda_content" and "lambda_style" seem not right. The value is too high!

Hi,
I am training your style transfer network with my own dataset and the style image as starry nights, just like yours. But even after training it for 6000 batches, I get black images as below:

Please suggest why this is happening
My dataset has the following structure:
/RGB/data/001/image_0.jpg
/RGB/data/001/image_2.jpg
...
/RGB/data/001/image_100.jpg
...
/RGB/data/084/image_8800.jpg
/RGB/data/084/image_8801.jpg

hello , thanks for your codes, can you tell me the checkpoint_model
parser.add_argument("--checkpoint_model", type=str, help="Optional path to checkpoint model")
thank you

Can u give some detail of time process on 1 image?