Transformer_model

• A pytorch implementation of Attention Is All You Need (Transformer) for image captioning.
• Add SCST training from "Self-critical Sequence Training for Image Captioning".
• Clear and easy to learn.

Environment

• Python 3.7
• Pytorch 1.3.1

Method

1. Architecture

2. Main Process

Usage

1. Preprocessing

Extract image features by ResNet-101 (denoted as grid-based features) and process coco captions data (from Karpathy splits) through preprocess.py. Need to adjust the parameters, where resnet101_file comes from here. Image features can also be obtained from here or extracted using ezeli/bottom_up_features_extract repository (from Bottom-Up Attention paper, using fixed 36 features per image, denoted as region-based features).

This project is not limited to the MSCOCO dataset, but you need to process your data according to the data format in the preprocess.py file.

2. Training

• First adjust the parameters in opt.py:
  • train_mode: 'xe' for pre-training, 'rl' for fine-tuning (+SCST).
  • learning_rate: '4e-4' for xe, '4e-5' for rl.
  • resume: resume training from this checkpoint. required for rl.
  • N_enc/N_dec: number of encoder/decoder layers, the default is 4.
  • other parameters can be modified as needed.
• Run:
  • python train.py
  • checkpoint save in checkpoint dir, test result save in result dir.

3. Test

• python test.py -t model.pth -i image.jpg
• only applicable to the model trained by grid-based features.
• for region-based features, you can first extract the image feature through ezeli/bottom_up_features_extract repository, and then simply modify the test.py file to use.

Result

Evaluation metrics

Evaluation tool: ezeli/caption_eval

XE represents Cross-Entropy loss, and +SCST means using reinforcement learning to fine-tune the model (using CIDEr reward).

features	training	Bleu-1	Bleu-2	Bleu-3	Bleu-4	METEOR	ROUGE_L	CIDEr	SPICE
grid-based	XE	74.3	57.8	44.4	34.0	27.3	55.2	109.5	20.3
grid-based	+SCST	78.5	62.8	48.5	36.8	27.9	57.0	121.3	21.8
region-based	XE	76.3	60.2	46.3	35.5	27.7	56.2	113.4	20.7
region-based	+SCST	79.9	64.4	50.0	38.3	28.7	58.3	127.3	22.4

Examples

a man riding a motorcycle down a dirt road.
a bunch of knives sitting on top of a wooden cutting board.