search

Peterande / SAST

[CVPR 2024] SAST: Scene Adaptive Sparse Transformer for Event-based Object Detection
MIT License
28 stars 2 forks source link

epoc #8

Closed xjyisok closed 2 weeks ago

xjyisok commented 2 weeks ago

Hi how much number of iterations (it) in one epoch during training while my epoc0 dosen't stop even reach 3000 it my batch_size is 2

Peterande commented 2 weeks ago

Our training strategies are all based on iterations (it), and since we adopt a hybrid training strategy from RVT, there isn't a clear concept of epochs. You can control the evaluation cycle by changing the validation validation.val_check_interval=10000 in the launch command.

xjyisok commented 2 weeks ago

I see that in code one ckpt file will be saved after one epoch. but after training one epoch in folder output I only see some .yaml file.like this image

image

Peterande commented 2 weeks ago 
          https://github.com/Peterande/SAST/blob/b40ab1aa135f61d46dd8e3f4da1381fa733c3c58/train.py#L154

You can modify "default_root_dir" to the path.

Originally posted by @Peterande in https://github.com/Peterande/SAST/issues/9#issuecomment-2423698060

The ckpt file will be saved in "default_root_dir".