a-newman/reproduce_mem_resnet3d

Training a ResNet3D on Memento

As a baseline for the Memento paper, replicating Section 5.2 of the VideoMem paper

Model courtesy of 3D-ResNets-PyTorch (configured as submodule).

Training

To train a ResNet3D-34 on Memento, run (for example):

python3 train.py --run_id "resnet3d_no_alpha" --batch_size 23 --freze_until_it 750

See train.py --help for more flags.

Evaluating

Run:

python3 evaluate.py --ckpt_path data/resnet3d_no_alpha/ckpt/ep_9__rc_0.57359067728224__total_val_loss_3625.34521484375.pth --batch_size 23

This will reproduce the number reported in the Memento paper. Output from this command is recorded in test_results.txt.

See evaluate.py --help for more flags.