joeyy5588 / p2c

Official Implementation of the paper "Paraphrasing is all you need for Novel Object Captioning", accepted by NeurIPS 2022

Paraphrasing Is All You Need for Novel Object Captioning

This repository contains the official PyTorch implementation for the paper. paper link

Yang, C. F., Tsai, Y. H. H., Fan, W. C., Salakhutdinov, R. R., Morency, L. P., & Wang, F. (2022). Paraphrasing Is All You Need for Novel Object Captioning. Advances in Neural Information Processing Systems, 35, 6492-6504.

Installation

• Follow the instruction from here: link to set up the environment of our codebase, and replace the directory oscar/ with p2c/
• Install CLIP from the offical repo: link, and place the CLIP directory under the root directory
• Download the pre-trained CLIP model and save it using the following code: torch.save(_use_new_zipfile_serialization=False) Then replace line 381 in CLIP/clip/model.py with this line: for attr in [*[f"{s}_proj_weight" for s in ["in"]], "in_proj_bias"]:
• Install the object detection model TSD from here: link

Data

• It is recommended to download large files with AzCopy for faster speed

path/to/azcopy copy <link> <target folder> --recursive

• For COCO dataset, download the region features and detection tags from here: link
• Download the testing data for nocaps from here: link Note that this link only provides data in validation set. For testing set, please extract them from the data below.
• For Open Images Dataset, download the region features from here: link
• For Open Images Dataset, use the TSD model to generate the detection tags.
• Parse the region features and detection tags into the same format as the COCO dataset, and place the generated files in p2c/data. Codes for generating the tsv file can be found here: link

Training

P2C:

• To perform VIVO pre-training

$ bash scripts/vivo.sh

• To perform our stage 1 training (Describing novel objects with linguistic fluency)

$ bash scripts/stage1.sh

• To perform our stage 2 training (Learning novel object captions with fidelity and adequacy)

$ bash scripts/stage2.sh

Reproducing VinVL+VIVO:

• To perform VIVO pre-training

$ bash scripts/vivo.sh

• To perform cross-entropy optimization

$ bash scripts/xe.sh

• To perform SCST optimization

$ bash scripts/scst.sh

Testing

To evaluate the result on the testing set, replace val.yaml with test.yaml

• To perform greedy decoding

$ bash scripts/inference.sh --num_beams 1 --test_yaml data/val.yaml

• To perform beam search

$ bash scripts/inference.sh --num_beams 5 --test_yaml data/val.yaml

• To perform Constrained Beam Search (CBS)

$ bash scripts/cbs.sh --test_yaml data/val.yaml