shkarupa-alex / tfswin Goto Github PK

Hi and thank you for this amazing work. I was wondering if the models are compatible with Colab TPU since the training fails with error " Socket closed".

I am trying to use this repo and tried Evaluation code(https://github.com/shkarupa-alex/tfswin#evaluation), written on readme, for 1st step. However, the result returns accuracy=0.001 as follows.

1250/1250 [==============================] - 172s 127ms/step - loss: nan - accuracy: 0.0010 - sparse_top_k_categorical_accuracy: 0.0000e+00
[nan, 0.0010000000474974513, 0.0]

I read model.py and find "load wights" so I thought that Evaluation code returns around 70% acc., like table attached on Evaluation section. And I found imagenet_v2 download is success, I can see some jpeg img on imagenet_V2 dir, created by Evaluation code.

imagenet2 = tfds.load('imagenet_v2', split='test', shuffle_files=True)

My question is, how to get the result like the table on evaluation section? (and what is the point I was wrong with getting acc=0.001? )

Here is a Dockerfile I created and used for this repo (sry idk how to attach Dockerfile so it change to txt file)
Dockerfile_tfswin.txt

docker run \ --gpus all \ -it -e \ TF_FORCE_GPU_ALLOW_GROWTH=true \ -v $MOUNT_DATA:/root/data \ -v $MOUNT_DIR:/root/scripts \ -v /tmp/.X11-unix:/tmp/.X11-unix \ -p 6060:6060 \ --env="DISPLAY" \ --name tfswin \ --rm tfswin \ bash

and Evaluation code I used (copy&paste from evaluation section on readme, and just change model name) (for same reason with Dockerfile, I change py to txt)
test_tfswin.txt

Thank you!

Hi, does the SwinTransformer v2 do SimMiM pretraining? This is shown in the paper:
https://arxiv.org/pdf/2111.09883.pdf

If not, any plans to add/how difficult would it be to port?

I tried to use different input size such as 256 x 256 x 11 and use the pretrained weights. It did not work. Any help on this.