facebookresearch/mmbt

Supervised Multimodal Bitransformers for Classifying Images and Text

mmbt's Issues

train.jsonl for foo101

Downloaded UPMC_Food101, and trying to run the code for the food101 experiment. It seems the code needs json files for the dataset (which is not provided by dataset itself). How can I get them?

$ python mmbt/train.py --batch_sz 4 --gradient_accumulation_steps 40  --savedir mmbt/ --name mmbt_model_run  --data_path mmbt/data/  --task food101 --task_type classification  --model mmbt --num_image_embeds 3 --freeze_txt 5 --freeze_img 3   --patience 5 --dropout 0.1 --lr 5e-05 --warmup 0.1 --max_epochs 100 --seed 1
Traceback (most recent call last):
  File "mmbt/train.py", line 280, in <module>
    cli_main()
  File "mmbt/train.py", line 272, in cli_main
    train(args)
  File "mmbt/train.py", line 183, in train
    train_loader, val_loader, test_loaders = get_data_loaders(args)
  File "/home/fb/mmbt/data/helpers.py", line 113, in get_data_loaders
    os.path.join(args.data_path, args.task, "train.jsonl")
  File "/home/fb/mmbt/data/helpers.py", line 40, in get_labels_and_frequencies
    data_labels = [json.loads(line)["label"] for line in open(path)]
FileNotFoundError: [Errno 2] No such file or directory: 'mmbt/data/food101/train.jsonl'

Run on GQA dataset

Hello
Do you have an instructions to run MMBT on the GQA dataset?
I've tried to make changes in the repo but it doesn't seem to work.

I've tried to run with this CMD:
python mmbt/train.py --batch_sz 4 --gradient_accumulation_steps 40 --savedir /users/yonatan/mmbt/savedir --name mmbt_model_run --data_path /users/yonatan/gqa_data_and_images_dir --task food101 --task_type classification --model mmbt --num_image_embeds 3 --freeze_txt 5 --freeze_img 3 --patience 5 --dropout 0.1 --lr 5e-05 --warmup 0.1 --max_epochs 100 --seed 1
Thank you

Hello, can you provide a link to the dataset? for example，MM-IMDB，FOOD101,V-SNLI

Order of text and image embeddings concatenation does not match the comment

In the MultimodalBertEncoder, the text and image embeddings order of concatentation does not match the comment:

encoder_input = torch.cat([img_embed_out, txt_embed_out], 1) # Bx(TEXT+IMG)xHID

Pretrained models?

Hi,

Can you please provide pretrained models for the different models/baselines used in the paper?

Can you please teach me how to prepare "test_hard_gt.jsonl"

Hi,

I downloaded UPMC_Food101, and ran food_101.py. And I'm trying to run the code "train.py" for the food101 datasets.
It seems that the code need jsonl file named "test_hard_gt.jsonl".

Error is follows, please could you help me prepare "test_hard_gt.jsonl".

python mmbt/train.py --batch_sz 4 --gradient_accumulation_steps 40 --savedir savedir --name mmbt_model_run --data_path datasets --task food101 --task_type classification --model mmbt --num_image_embeds 3 --freeze_txt 5 --freeze_img 3 --patience 5 --dropout 0.1 --lr 5e-05 --warmup 0.1 --max_epochs 100 --seed 1
Traceback (most recent call last):
  File "mmbt/train.py", line 280, in <module>
    cli_main()
  File "mmbt/train.py", line 272, in cli_main
    train(args)
  File "mmbt/train.py", line 183, in train
    train_loader, val_loader, test_loaders = get_data_loaders(args)
  File "E:\mmbt\mmbt\data\helpers.py", line 197, in get_data_loaders
    args,
  File "E:\mmbt\mmbt\data\dataset.py", line 23, in __init__
    self.data = [json.loads(l) for l in open(data_path)]
FileNotFoundError: [Errno 2] No such file or directory: 'datasets\\food101\\test_hard_gt.jsonl'