search

antoine77340 / howto100m

Code for the HowTo100M paper
Apache License 2.0
250 stars 38 forks source link

Embd_dim #3

Open Petr-Byv opened 4 years ago

Petr-Byv commented 4 years ago

Hi, thank you for sharing the model. For the evaluation the line suggest to use 6144 for the embedding: python eval.py --eval_msrvtt=1 --eval_youcook=1 --eval_lsmdc=1 --num_thread_reader=8 --embd_dim=6144 --pretrain_path=model/howto100m_pt_model.pth.

Is that indeed so, and how is it different to the 4096 recommended for the training: python train.py --num_thread_reader=8 --epochs=15 --batch_size=32 --n_pair=64 --n_display=100 --embd_dim=4096 --checkpoint_dir=ckpt --features_path_2D=folder_for_2d_features --features_path_3D=folder_for_3d_features --train_csv=train.csv ?

Thank you.

zhenliulv commented 4 years ago

Hi, i have the same question, can you give us some idea ? Thank you.