Model performance about nfnets_pytorch

[benjs]

The accuracy is the result of taking the pretrained weights from the original repo and running eval.py.

When training from scratch using this code, there are multiple things missing for best results.
train.py should reach the baseline accuracy as described in the paper, however, this is not tested as I don't have the computing resources or money to train such large models.

[guozhiyao]

Then why the results are inconsistent with Top-1 accuracy Brock et al.?

[benjs]

They preprocess the input images differently. I tried to implement their method but it was a bit worse than simple resizing.

With such small differences, there can be many things that cause them (resizing algorithm, ...)

[guozhiyao]

ok. Thanks a lot.

[guozhiyao]

I calculated the #Params and found that they were inconsistent with those reported in the paper. The calculation method

size = 0
for p in net.parameters():
    size += p.numel()
print(size*4/1024/1024)

[benjs]

What was the result of your calculation?

[guozhiyao]

I test all NFNets and set the num_classes=1000, stochdepth_rate=0.25, alpha=0.2, se_ratio=0.5, activation='gelu', and the results are: 272.7MB, 506.0MB, 739.2MB, 972.5MB, 1205.7MB, 1439.0MB, 1672.2MB, 1905.5MB.

[benjs]

I guess the correct parameter count is in the #Params column in table 3 of the paper.

Your code counts the number of bytes, assuming that one parameter is of type float32 and thus has 4 bytes.

The paper however simply provides the number of parameter, not the number of bytes.

When you change your print statement to
print(size/1000/1000), you'll see that the count is correct.

[guozhiyao]

Yes, you are right. Thank you for your reply.