As a baseline for the Memento paper, replicating Section 5.2 of the VideoMem paper
Model courtesy of 3D-ResNets-PyTorch (configured as submodule).
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.
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.