Configs for Image Classification (cifar10)

[keroro824]

Thanks for the great work! I have a question regarding the hyperparams for training cifar10. I used the setting in this repo and replaced several hyperparams (eg n_layers n_heads etc) with the ones reported in the paper, but the best testing acc I got was 0.36:

import ml_collections

NUM_EPOCHS = 200 TRAIN_EXAMPLES = 45000 VALID_EXAMPLES = 10000

def get_config(): """Get the default hyperparameter configuration.""" config = ml_collections.ConfigDict() config.batch_size = 256 config.eval_frequency = TRAIN_EXAMPLES // config.batch_size config.num_train_steps = (TRAIN_EXAMPLES // config.batch_size) * NUM_EPOCHS config.num_eval_steps = VALID_EXAMPLES // config.batch_size config.weight_decay = 0. config.grad_clip_norm = None

config.save_checkpoints = True config.restore_checkpoints = False config.checkpoint_freq = (TRAIN_EXAMPLES // config.batch_size) * NUM_EPOCHS // 2 config.random_seed = 0

config.learning_rate = .0005 (if using 0.01 from the paper, the loss is not going down) config.factors = 'constant linear_warmup cosine_decay' config.warmup = (TRAIN_EXAMPLES // config.batch_size) 1 config.steps_per_cycle = (TRAIN_EXAMPLES // config.batch_size) NUM_EPOCHS

model params

config.model = ml_collections.ConfigDict() config.model.emb_dim = 32 config.model.num_heads = 4 config.model.num_layers = 3 config.model.qkv_dim = 64 config.model.mlp_dim = 128 config.model.dropout_rate = 0.3 config.model.attention_dropout_rate = 0.2 config.model.classifier_pool = 'CLS' config.model.learn_pos_emb = True

config.trial = 0 # dummy for repeated runs. return config

Could you point out which params I could adjust to match the accuracy (this is for full attention).

[vanzytay]

@MostafaDehghani for clarity on image tasks.

Note: we might take awhile to reply due to the upcoming deadlines. Thanks

[keroro824]

@vanzytay Thanks for the quick response! @MostafaDehghani Could you help me check this when you got time and let me know any possible ones I could try (we're also trying for the upcoming deadline :)) Thanks!

[MostafaDehghani]

Thank you @keroro824 for the question.

So if I understood correctly, you are looking for the configs for the vanilla transformer to reproduce its results on CIFAR10 dataset in LRA. For that, you can use the followings as the model hparams:

  # model params
  config.model = ml_collections.ConfigDict()
  config.model.emb_dim = 128
  config.model.num_heads = 8
  config.model.num_layers = 1
  config.model.qkv_dim = 64
  config.model.mlp_dim = 128
  config.model.dropout_rate = 0.3
  config.model.attention_dropout_rate = 0.2
  config.model.classifier_pool = 'CLS'
  config.model.learn_pos_emb = True

We are planning to release the code for all models and the best performing configurations, as soon as possible. In the meantime, please let us know if you had any questions :)

[keroro824]

@MostafaDehghani Thank you !!! I can replicate it now!

[MostafaDehghani]

No problem at all! Perfect! ... and good luck with the deadline :)

[alexmathfb]

The above comment states 1 layer and leaves learning rate unspecified. This means learning rate will be 0.0005 inherited from base_cifar10_config.

The arxiv paper states: 3 layers, learning rate 0.01.

The openreview paper states: 3 layers, learning rate 0.01.

Notably, the config file still contains nothing.

Currently, the code in this repository is inconsistent with the published articles. Do you plan on fixing these inconsistencies? Or did you abandon this project?

[vanzytay]

IIRC config files takes precedence over the paper hparams. We will update the readme here to state this.

[MostafaDehghani]

The best results in the paper are all reproducible from the code in the repo. Have you tried the configs that are shared here? Many people reproduced the results without any issue after our last update.

LRA is a living benchmark. We tried our best to tune hyper-parameters of each model we had in the paper and some of the authors of those models reached out to help us find better hyper parameters. The codebase has the most updated version of those and it can be used for reproducing the results.

Notably, the config file still contains nothing.

If you read the code carefully, you can see that the config file you are referring to is inheriting from the base config!

[alexmathfb]

IIRC config files takes precedence over the paper hparams. We will update the readme here to state this.

This was not clear to me, I apologize for the misunderstanding.

If you read the code carefully, you can see that the config file you are referring to is inheriting from the base config!

I meant that the file was empty, so learning rate was inherited as 0.0005 from base config file while article reported learning rate 0.01. I was under the impression that the article hyperparams would be used, but as vanzytay clarified this is not the case.

[vanzytay]

It is a good reminder to us that an update of the paper is due to ask researchers to defer to the codebase to reproduce the results.

In our 2nd update, we ran all the cifar results again to make sure they were reproducible. So the code configs should be good. Do give it a try and let us know if you run into any other issues. Thanks!