mmf_convert_hm --zip_file=data.zip --password=1 --bypass_checksum=1

`mmf_run config=projects/hateful_memes/configs/

mmf_run config=projects/hateful_memes/configs/vilbert/vitencbert.yaml \
    model=vitencbert \
    dataset=hateful_memes \
    run_type=train_val

Code Changes

• Slight bugfix for wandb logger
• Added code for the hateful memes dataset
• Adjust preprocessing code

mmf_run config=projects/hateful_memes/configs/vilbert/vitbert.yaml \
    model=vitbert \
    dataset=hateful_memes \
    run_type=train_val 

• Copy over embeddings
• Alignment of attention layers
  • There are 12 layers each for text and image What was done

mmf_run config=projects/hateful_memes/configs/vilbert/vitbert.yaml \
    model=vitbert \
    dataset=hateful_memes \
    run_type=train_val 

mmf_predict config=projects/hateful_memes/configs/vilbert/vitencbert.yaml
model=vitencbert
dataset=hateful_memes
run_type=test
checkpoint.resume_file=