krishnabits001 / domain_specific_cl Goto Github PK

@krishnabits001
Hello, Thanks for this wonderful work. I have checked your code for pretrain the decoder.
it was more than 800 lines within the same instance method, which is very hard to understand. Is it possible to have some words to summarize it a bit?
In the paper, the descriptions are not clear too.

I did not see the MMWHS on create_cropped_imgs.py. Does that mean there is no need to pre-process on MMWHS?

Thanks!

Thank you for providing the code. Can I use it on the coco dataset

The code you provided is based on TensorFlow, which is difficult to implement. Could you provide the code of PyTorch?

Hi Krishna,

Great work! I am very new to the topic of self-supervised learning. May I ask what is the intuition that pretrained initialization would help the segmentation performance? If I understand your paper correctly, the network learned the global and local discriminative features for some datasets with your self-supervised learning scheme, how would that relate to the segmentation. Is there any paper or material you can refer to?

Best regards,

How can I fix the following error while pre-training the decoder?

Traceback (most recent call last):
  File "pretr_decoder_local_contrastive_loss.py", line 318, in <module>
    tmp_op=new_var.assign(var_val)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/resource_variable_ops.py", line 902, in assign
    (tensor_name, self._shape, value_tensor.shape))
ValueError: Cannot assign to variable enc_c1_a/W:0 due to variable shape (3, 3, 1, 16) and value shape (122,) are incompatible

There is some question about fintuning the segmentation model on "mmwhs":

1. The training loss is hard to converge. We set "--no_of_tr_imgs tr8 --comb_tr_imgs c1" but the training loss only decrease to 0.5x, and the val_score is 0 in all epoches.
2. We found the augmentation will change the original value in label array, thus we use a function to re-map the value:
   def ReMod(x, dataset = 'mmwhs'):
   '''
   for tf-onehot
   '''
   if dataset == 'mmwhs':
   mapp_v = np.array([ 0, 205, 420, 500, 550, 600, 820, 850])
   x = np.repeat(np.expand_dims(x, axis=-1), len(mapp_v), -1)
   x = np.argmin(np.abs(x - mapp_v),axis=-1)
   return x

Dear author:
You propose an effective augmentation way about 3D dataset,if I use the augmentation on 3D CT what should I attention?