How to evaluate or do inference?

Generate predictions for nocaps val or nocaps test using a pretrained checkpoint as follows.

python scripts/inference.py \
    --config /path/to/config.yaml \
    --checkpoint-path /path/to/checkpoint.pth \
    --output-path /path/to/save/predictions.json \
    --gpu-ids 0

Add --evalai-submit flag if you wish to submit the predictions directly to EvalAI and get results.

Using Constrained Beam Search

To use Constrained Beam Search during inference, add –config-override MODEL.USE_CBS True or specify it in the config file. This is only valid for checkpoints trained with frozen GloVe embeddings (MODEL.EMBEDDING_SIZE 300).