This is the source code for our paper End-to-End Dense Video Captioning with Masked Transformer. It mainly supports dense video captioning on generated segments. To generate captions on GT segments, please refer to our new GVD repo and also our notes.
1) Miniconda3 for Python 3.6
2) CUDA 9.2 and CUDNN v7.1
3) PyTorch 0.4.0. Follow the instructions to install pytorch and torchvision.
4) Install other required modules (e.g., torchtext)
pip install -r requirements.txt
Optional: If you would like to use visdom to track training do pip install visdom
Optional: If you would like to use spacy tokenizer do pip install spacy
Note: The code has been tested on a variety of GPUs, including 1080 Ti, Titan Xp, P100, V100 etc. However, for the latest RTX GPUs (e.g., 2080 Ti), CUDA 10.0 and hence PyTorch 1.0 are required. The code needs to be upgraded to PyTorch 1.0.
For ActivityNet, download the re-formatted annotation files from here, decompress and place under directory data. The frame-wise appearance (with suffix _resnet.npy) and motion (with suffix _bn.npy) feature files for each spilt are available [train (27.7GB), val (13.7GB), test (13.6GB)] and should be decompressed and placed under your dataset directory (refer to as feature_root in the configuration files).
Similarly for YouCook2, the annotation files are available here and should be placed under data. The feature files are [train (9.6GB), val (3.2GB), test (1.5GB)].
You could also extract the feature on your own with this code. Note that ActivityNet is processed with an older version of the repo while YouCook2 is processed with the latest code which had a minor change regarding the sampling approach. This accounts for the difference in the formulation of frame_to_second conversion.
Download the dense video captioning evaluation scripts and place it under the tools directory. Make sure you recursively clone the repo. Our code is equavalent to the official evaluation code from ActivityNet 2017 Challenge, but faster. Note that the current evaluation scripts had a few major bugs fixed towards ActivityNet 2018 Challenge.
The evaluate script for event proposal can be found under tools.
First, set the paths in configuration files (under cfgs) to your own data and feature directories. Create new directories log and results under the root directory to save log and result files.
The example command on running a 4-GPU distributed data parallel job (for ActivityNet):
For Masked Transformer:
CUDA_VISIBLE_DEVICES=0 python3 scripts/train.py --dist_url $dist_url --cfgs_file $cfgs_file \
--checkpoint_path ./checkpoint/$id --batch_size $batch_size --world_size 4 \
--cuda --sent_weight $sent_weight | tee log/$id-0 &
CUDA_VISIBLE_DEVICES=1 python3 scripts/train.py --dist_url $dist_url --cfgs_file $cfgs_file \
--checkpoint_path ./checkpoint/$id --batch_size $batch_size --world_size 4 \
--cuda --sent_weight $sent_weight | tee log/$id-1 &
CUDA_VISIBLE_DEVICES=2 python3 scripts/train.py --dist_url $dist_url --cfgs_file $cfgs_file \
--checkpoint_path ./checkpoint/$id --batch_size $batch_size --world_size 4 \
--cuda --sent_weight $sent_weight | tee log/$id-2 &
CUDA_VISIBLE_DEVICES=3 python3 scripts/train.py --dist_url $dist_url --cfgs_file $cfgs_file \
--checkpoint_path ./checkpoint/$id --batch_size $batch_size --world_size 4 \
--cuda --sent_weight $sent_weight | tee log/$id-3For End-to-End Masked Transformer:
CUDA_VISIBLE_DEVICES=0 python3 scripts/train.py --dist_url $dist_url --cfgs_file $cfgs_file \
--checkpoint_path ./checkpoint/$id --batch_size $batch_size --world_size 4 \
--cuda --sent_weight $sent_weight --mask_weight $mask_weight --gated_mask | tee log/$id-0 &
CUDA_VISIBLE_DEVICES=1 python3 scripts/train.py --dist_url $dist_url --cfgs_file $cfgs_file \
--checkpoint_path ./checkpoint/$id --batch_size $batch_size --world_size 4 \
--cuda --sent_weight $sent_weight --mask_weight $mask_weight --gated_mask | tee log/$id-1 &
CUDA_VISIBLE_DEVICES=2 python3 scripts/train.py --dist_url $dist_url --cfgs_file $cfgs_file \
--checkpoint_path ./checkpoint/$id --batch_size $batch_size --world_size 4 \
--cuda --sent_weight $sent_weight --mask_weight $mask_weight --gated_mask | tee log/$id-2 &
CUDA_VISIBLE_DEVICES=3 python3 scripts/train.py --dist_url $dist_url --cfgs_file $cfgs_file \
--checkpoint_path ./checkpoint/$id --batch_size $batch_size --world_size 4 \
--cuda --sent_weight $sent_weight --mask_weight $mask_weight --gated_mask | tee log/$id-3Arguments: batch_size=14, mask_weight=1.0, sent_weight=0.25, cfgs_file='cfgs/anet.yml', dist_url='file:///home/luozhou/nonexistent_file' (replace with your directory), id indicates the model name.
For YouCook2 dataset, you can simply replace cfgs/anet.yml with cfgs/yc2.yml. To monitor the training (e.g., training & validation losses), start the visdom server with visdom in the background (e.g., tmux). Then, add --enable_visdom as a command argument.
Note that at least 15 GB of free RAM is required for the training. The nonexistent_file will normally be cleaned up automatically, but might need a manual delete if otherwise. More about distributed data parallel see here (0.4.0). You can also run the code with a single GPU by setting world_size=1.
Due to legacy reasons, we store the feature files as individual .npy files, which causes latency in data loading and hence instability during distributed model training. By default, we set the value of num_workers to 1. It could be set up to 6 for a faster data loading. However, if encouter any data parallel issue, try setting it to 0.
The pre-trained models can be downloaded from here (1GB). Make sure you uncompress the file under the checkpoint directory (create one under the root directory if not exists).
For Masked Transformer (id=anet-2L-gt-mask):
python3 scripts/test.py --cfgs_file $cfgs_file --densecap_eval_file ./tools/densevid_eval/evaluate.py \
--batch_size 1 --start_from ./checkpoint/$id/model_epoch_$epoch.t7 --id $id-$epoch \
--val_data_folder $split --cuda | tee log/eval-$id-epoch$epochFor End-to-End Masked Transformer (id=anet-2L-e2e-mask):
python3 scripts/test.py --cfgs_file $cfgs_file --densecap_eval_file ./tools/densevid_eval/evaluate.py \
--batch_size 1 --start_from ./checkpoint/$id/model_epoch_$epoch.t7 --id $id-$epoch \
--val_data_folder $split --learn_mask --gated_mask --cuda | tee log/eval-$id-epoch$epochArguments: epoch=19, split='validation', cfgs_file='cfgs/anet.yml'
This gives you the language evaluation results on the validation set. You need at least 8GB of free GPU memory for the evaluation. The current evaluation script only supports batch_size=1 and is slow (1hr for yc2 and 4hr for anet). We actively welcome pull requests.
The official evaluation servers are available under ActivityNet and YouCook2. Note that the NEW evaluation scripts from ActivityNet 2018 Challenge are used in both cases.
We use a different code base for captioning-only models (dense captioning on GT segments). Please contact luozhou@umich.edu for details. Note that it can potentially work with this code base if you feed in GT segments into the captioning module rather than the generated segments. However, there is no guarantee on reproducing the results from the paper. You can also refer to this implementation where you need to config --att_model to 'transformer'.
@inproceedings{zhou2018end,
title={End-to-End Dense Video Captioning with Masked Transformer},
author={Zhou, Luowei and Zhou, Yingbo and Corso, Jason J and Socher, Richard and Xiong, Caiming},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
pages={8739--8748},
year={2018}
}