ResNet18 performs much better than expected!

[ashra-main]

I ran this repo with ResNet18 using 4 GPUs and with the latest Pytorch version and I got 95.67% instead of 93.02% reported in the ReadMe table.
I'm wondering whether anything improved in Pytorch that could explain this major improvement?

My requirements list:
numpy==1.22.3
torch==1.11.0+cu115
torchvision==0.12.0+cu115
-f https://download.pytorch.org/whl/torch_stable.html

[seuretm]

Hello,

TL;DR: the longer you train, the higher the result, but it's not a test result, but a validation result.

This code has a big issue, which I already raised: it uses the test set as validation set. What you get is the highest validation accuracy, not a test accuracy. As there is always a bit of randomness on the validation results, the longer you train, the higher the chance that due to this randomness you get a higher result.

Yet, this does not reflect what you would get in a never-seen test set, and any scientific publication using this methodology should be immediately retracted.

The right way of training the network would be to use some samples from the training set as validation, and use the test set only once, at the very end of the training - instead of using the test set in the optimization process.

[AminJun]

@seuretm All points you mentioned are true. However, even if you don't use the test data during training, and only use it once (after the training is done), you get about 95% acc, where the repo reports 93% acc. So my experience is that the point @arashash is making here is valid.

[zaccharieramzi]

In a different setup, using a different implementation of the same model (ResNet-18 in CIFAR configuration), and a different code to perform the optimization, I also find that without using the test set during the evaluation I can reach 95.4 - 95.5% (I don't have a script to share since I am doing it as part of a bigger benchmark, but the gist of it is here). I am therefore concurring that this is not due to early stopping on the test set, or retaining the best test accuracy.

However it is also true that this specific setup (e.g. in particular the weight decay and learning rate values) might have been tuned on the test set, which would be problematic. But I think this is another topic.

As to why there is an improvement compared to the numbers reported in the repo, my guess is that with newer models, new training strategies were implemented and 93% is what you would get with an old strategy.
To back that guess a little bit, here is the version of main.py when the resnet (and the 93% result) was added.
You can see that the learning rate and the weight decay are much lower.

[skrbnv]

This is not original ResNet18 network. That's the reason why accuracy is so high: #136 (comment)

[zaccharieramzi]

@skrbnv the differences pointed out in this comment are just the differences between the CIFAR-10 and the ImageNet versions.
In the original ResNet paper, there were indeed 2 different versions : one for CIFAR-10 (which has smaller res images) and one for ImageNet. The one for CIFAR-10 doesn't have striding in the first conv, and the initial MaxPooling.

If you use a ResNet with striding in the first conv, and the initial MaxPooling, you will not obtain accuracies above 91% (I tried). The main reason is that you lose too much information at the very beginning of the network.

The plain/residual architectures follow the form in Fig. 3
(middle/right). The network inputs are 32×32 images, with
the per-pixel mean subtracted. The first layer is 3×3 convolutions. Then we use a stack of 6n layers with 3×3 convolutions on the feature maps of sizes {32, 16, 8} respectively,
with 2n layers for each feature map size. The numbers of
filters are {16, 32, 64} respectively. The subsampling is performed by convolutions with a stride of 2. The network ends
with a global average pooling, a 10-way fully-connected
layer, and softmax. There are totally 6n+2 stacked weighted
layers