I3D accuracy when finetuning on UCF-101 is very low

[ziqi-zhang]

Hi, I tried to finetune i3d on ucf-101 using pretrained models in checkpoint. But the accuracy after 10 epochs is very low (30%) on validation set. I replaced logits layer of i3d and freeze other layers. Here is my script: python -u train.py \ --dataset ucf101 \ --model i3d \ --video_path $DATASET_ROOT/jpg \ --annotation_path $DATASET_ROOT/ucfTrainTestlist/ucf101_02.json \ --batch_size 6 \ --num_classes 400 \ --finetune_num_classes 101 \ --spatial_size 224 \ --sample_duration 64 \ --learning_rate 1e-2 \ --save_dir $OUTPUT_DIR \ --dropout_keep_prob 0.5 \ --checkpoint_path checkpoints/I3D/rgb_imagenet.pth \ --finetune_prefixes logits \ --num_scales 1 \

Can you please give me some advice on finetuning on ucf-101? Thanks!

[MKowal2]

What other layers are you freezing? I replaced the logits layer but did not freeze any layers and fine tuning for 10 epochs gives about 90% accuracy on the training set and test set of UCF101_01. I have a similar batch size / learning rate too and I used rgb_imagenet.pth as my checkpoint.