gchaperon/alexnet

My replication code for the AlexNet paper.

alexnet

My replication code for the AlexNet paper.

tldr: go to Results

Data

The data used is ImageNet 2012. I downloaded it from kaggle. The paper also experiments with ImageNet 2010 but I couldn't find this dataset. The authors also experiment with pretraining on ImageNet Fall 2011 which isn't even available anymore. The closest would be current ImageNet21K, but I don't have enough compute for that.

Download and extract ImangeNet

mkdir -p data/imagenet && cd data/imagenet
kaggle competitions download -c imagenet-object-localization-challenge
unzip imagenet-object-localization-challenge.zip

Instructions on how to configure kaggle are here

Also, if you want to train on tiny-imagenet download the data as follows.

mkdir -p data && cd data
wget https://image-net.org/data/tiny-imagenet-200.zip
unzip tiny-imagenet-200.zip
cd ..

I coded a couple more tasks, but those are downloaded automatically.

User guide

Install the deps found in requirements.txt. I used python 3.9 and pytorch 1.12. You should modify the cuda version according to your hardware.

These are the command line options

$ python -m alexnet --help
Usage: python -m alexnet [OPTIONS]

Options:
  --task [mnist|fashion-mnist|cifar10|cifar100|tiny-imagenet|imagenet]
                                  [required]
  --batch-size INTEGER            [default: 128]
  --dropout FLOAT                 [default: 0.5]
  --learn-rate FLOAT              [default: 0.0001]
  --seed INTEGER                  [default: 12331]
  --extra-logging                 Whether to log histograms of parameters and
                                  grads.
  --fast-dev-run                  Run only a couple of steps, to check if
                                  everything is working properly.
  --help                          Show this message and exit.

The available tasks can be seen above. Default hparams where chosen according to the paper and my own experimentation. To run a setup pretty close to the one on the paper simply run

python -m alexnet --task imagenet

Results

See the training curves in the tensorboard.dev experiment.

All experiments were done using an RTX3090. ImageNet training took ~3 days to reach 100 epochs (hardware has come a long way since 2012, compare with the 5 days it took on an RTX580).

Below are the results on different tasks.

	imagenet (my experiment)	imagenet (paper)	mnist	fashion-mnist	cifar10	cifar100	tiny-imagenet
error@1	0.48	0.40	0.01	0.09	0.13	0.41	0.57
error@5	0.25	0.18	0.00	0.00	0.01	0.16	0.32

Results aren't that close, but also not that far off. As we can see in the training curves, we might have been able to obtain a better results with more training time, since val error seemed to be going down still when training stoped. For this scope however it's good enough.

Summary of features / techniques used in the paper

• for rect images, first rescale so that shorter side is of 256 pix, then crop longer side to 256 pix, always square images.
• substract "mean activity" from each pixel, apparently this is mean activity per channel
• ReLU non linearity everywhere
• training on multiple gpus (see section 3.2 and fig 2). I wil not do this, instead one can leverage the groups option in nn.Conv2d so emulate the behaviour
• local response norm, pytorch implementation divides alpha by n, so in order to replicate paper alpha should be multiplied by n
• net description taken from section 3.5 and figure 2
• augmentation: at train time, extract random 224 x 224 patches and horizontal reflection
• augmentation: at test time, extract 10 224 x 224 patches (corners + center) + h reflections, results is averaged prediction over the 10 patches
• augmentation: PCA color augmentation, see paper section 4.1, extra resources:
  • https://github.com/koshian2/PCAColorAugmentation
  • https://github.com/aparico/pca-color-augment
  • https://groups.google.com/g/lasagne-users/c/meCDNeA9Ud4?pli=1
• dropout

Training details

• optim SGD, see exact update rule in paper section 5 (not used see, notes below)
• batch size 128
• dropout of 0.5
• momentum 0.9 (see notes about optimization below)
• weight decay 0.0005 (see notes about optimzation below)
• weight init is guassian with mean 0, std 0.01. Bias in 2, 4, 5 conv layers and also fc layers with constant 1, bias in other layers init with 0
• same lr for al layers, start at 0.01 (close, but see notes)
• decay learn rate by factor of 10 when val error stops improving. This results in 3 reductions during training.
• 90 epochs, which took 5/6 days.

Notes

• Optimizer changed, other user also saw poor performance using optim.SGD. I even implemented the actual optimizer step described in the paper, since it's a little bit different than pytorch's algortihm, but I saw no improvement. I kept Adam optim and used a learn rate of 1e-4, and also lowered it on plateau.
• Apparently the net convergence is super sensitive to param initialization. Out of three seed values I tried only one made the net learn in the imagenet experiment (using the exact same hparams elsewhere). The current default seed is the one I found.