search

tobyperrett / few-shot-action-recognition

Implementations of some few-shot action recognition methods.
MIT License
41 stars 6 forks source link

A considerable gap from the SOTA #3

Open Brave-peng opened 2 years ago

Brave-peng commented 2 years ago

I am trying to reimplement the OTAM and using the default arguments, this is options and the result in ssv2。

Options: Namespace( dataset='data/ssv2_256x256q5', learning_rate=0.001, tasks_per_batch=16, checkpoint_dir='checkpoint_dir', test_model_name='checkpoint_best_val.pt', training_iterations=60000, resume_from_checkpoint=False, way=5, shot=5, query_per_class=5, query_per_class_test=1, val_iters=[], num_val_tasks=1000, num_test_tasks=10000, print_freq=100, seq_len=8, num_workers=8, backbone='resnet50', opt='sgd', save_freq=5000, img_size=224, num_gpus=4, sch=[1000000], method='otam', pretrained_backbone=None, val_on_test=False, trans_linear_in_dim=2048)

test accuracies: data/ssv2_256x256q5: 44.5+/-0.4

This is a considerable gap from the SOTA. Instead, I also testing in the kinetics and using the TRX, only get the 63.2% accuracy(5-way-1-shot). Could you tell me what might be causing this?

tobyperrett commented 2 years ago

It's been a while since I've used this, but I seem to remember that OTAM works best with tasks_per_batch=1 (they don't mention gradient accumulation in their paper). Others who've used this code to replicate the OTAM results used that and it seemed to work OK. Hope that helps.

tobyperrett commented 2 years ago

In fact, that's what the setting is in scripts/examples.md. So just try running that command.

Brave-peng commented 2 years ago

Thank you very much for your quick reply, which is very helpful to me