erezposner / fast_dense_feature_extraction Goto Github PK

Hi,

Could you check typo in "Expected verbose"?

Total time for C_P => C_I ?

Averaged time for C_I => C_P ?

And why you do not compare Total time between two models?

Thank you.

Hi,

First, this is a great work that largely improves the processing speed for patch-based CNN. After reading the paper, I found the transpose-reshape operation might be replaced by the pixel shuffle that is provided by the PyTorch framework. Is it correct?

I'm looking forward to your reply.

Best,

Xiaohong

Hello,

Would it be possible if you can give an example where the original network is an architecture like Resnet? This would be extremely helpful.

Thanks!

the SlimNet has a 3-dimensions output with shape (-1, imH, imW) now.
but if we have a batch of images for input, should the net have output with 4-dimensions ? like(batch_size, -1, imH, imW)?

thanks.

Running method: allPatches on device: cuda:0
Total time for C_I: 0.01324005126953125sec
Averaged time for C_P per Patch without warm up: 0.0012916574080009014sec
------- Comparison between a base_net over all patches output and slim_net -------
aggregated difference percentage = 0.0000458457%
Traceback (most recent call last):
File "d:/user/FE_pj/Fast_Dense_Feature_Extraction/sample_code.py", line 175, in
yi, xi = np.unravel_index(index, (imH, imW))
File "<array_function internals>", line 6, in unravel_index
ValueError: index 108153647 is out of bounds for array with size 1228800

Hi,

I have a concern about the equality of the proposed method and the commonly used one. For example, if the image size is 240 * 240 and the patch size is 24 * 24, the receptive field of the commonly used method (i.e., extract all patches one by one from the whole image) would not exceed the patch size due to the fact that all patches are fed into the network independently for feature extraction. However, the receptive field of this FDFE seems to consider the region larger than 24 * 24 if the pooling operation is used.

Could you please explain why they are equal if we require the latent code (the results from feature extraction) is only generated from each fixed patch size (e.g., 24 * 24)?

Thanks,

Xiaohong