pradeeplam / anime-sketch-coloring-with-swish-gated-residual-unet Goto Github PK

This is bad. Fix this.

• Batch size?
• In up layers, how does swish blocks upsample? Unpooling or conv2d transpose? (conv2d transpose)

This is bad. Fix this.

code:
https://colab.research.google.com/drive/1_ITjS2r-OJzbNlPAlMKf9L7XevMWuV3Q?usp=sharing

Is the cause of this error a mismatch in the number of images in the image_bw and image_rgb folder contents?

datasets
https://drive.google.com/drive/folders/1GmEwRcu9zK3hQnqk7bTPyU273g7X4vyy?usp=sharing

Tried to install as is and got

After dropping the version of opencv-python and some others build has failed

Ubuntu 20.04 LTS
pip 20.0.2 from /usr/lib/python3/dist-packages/pip (python 3.8)

Readme needs to incorporate:

• Result images
• Instructions on how to run predictions on new images using evaluate.py script

In the network structure diagram of the paper, the resolution of the feature map is divided into 6 levels, but in model.py, there are only 5 resolution levels. Actually, when the number of channels of the feature map is 512, the resolution of the feature map is the smallest, which is still obtained by downsampling, and it has no horizontal connection like the right branch.
In model.py, in the final output of the right branch, the output of the horizontal swish layer of the left branch is not acquired.

This leads to the asymmetry of the left and right branches of the network, misaligned connections.

Or fix it

I have an idea that perhaps a small number of verification sets should be build that can verify the effect of the model after each training eopch. Instead of choosing the last epoch of training or the epoch with the best convergence. It may prevent over-training (over-fitting), resulting in inconsistent colors in a certain area of the generated picture. For example, the hair of the person in the final epoch of training will show multiple colors, and even the left and right colors of the clothing are not symmetrically consistent.
In the dataset given in the paper, there are dozens of numbered pictures at the end for the verification phase. It seems to be about 60 sketches.

We need to upload our pretrained checkpoints and dataset used to the release section.

The two diagrams below don't seem consistent. I mentioned this in the slack, but I'm putting it here so this conversation will be easier to access.


Essentially the SGB marked as (c) doesn't seem to match with the SGB in the first image (brown dashed box).

If the dimension of the input image is not evenly divisible by 32 (the network cuts it in half 5 times) there will most likely be a wonky concatenation issue due to mismatching tensor dimensions. The referenced line performs a hack that prevents this by resizing all images to 128x128. A better solution would be to recreate a dataset where all images are 224 or 256 or something.

What GPU did you use?