Densecap in Pytorch

A simplified pytorch version of jcjohnson/densecap

Issue or Contact me directly by [email protected]

What is densecap?

• Detect regions from an image and describe each region with a short sentence.
• See jcjohnson's repo or their paper

What about this implementation?

• Things are up to date: Python 3.7 and pytorch 1.4.0
• Our model code inherits GeneralizedRCNN from torchvision directly and try to keep it easy to understand and use (wish to be). So the region detector part is not the same as the original paper.
• Should be trainable on a single GTX 1080Ti GPU with 12GB memory.

Requirments

• python 3.7
• pytorch 1.4.0
• torchvision 0.5.0
• numpy
• PIL
• h5py
• tqdm
• tensorboard
• Prefetch version of DataLoader: https://github.com/IgorSusmelj/pytorch-styleguide/issues/5
• apex: https://github.com/NVIDIA/apex (to speed up training)
• nlgeval*: https://github.com/Maluuba/nlg-eval

Note: We use nlgeval to calculate Meteor, you can modify evaluate.py to use other methods like coco-eval.

Pre-preprocess

1. mkdir ./data and place Visual Genome Dataset in a sub directory visual-genome
   • Images are in two directories ./data/visual-genome/VG_100K and ./data/visual-genome/VG_100K_2
   • Annotation files are ./data/visual-genome/region_descriptions.json and ./data/visual-genome/image_data.json
2. python preprocess.py
3. Now we get ./data/VG-regions-dicts-lite.pkl and ./data/VG-regions-lite.h5. See preprocess.py for more details.
4. After preprocessing, file structures are listed below:

    data
    ├── visual-genome
    │   ├── VG_100K
    │   ├── VG_100K_2
    │   ├── region_descriptions.json
    │   └── image_data.json
    ├── VG-regions-dicts-lite.pkl
    └── VG-regions-lite.h5

Start Training

1. mkdir model_params
2. python train.py
3. change settings by set_args() and global variables

Evaluating

Since we need to combine metrics from vision and language, we rewrite the Lua version and the code is in ./model/evaluator.py.

• We provide functions to evaluate in the script evaluate.py.
• You can use describe.py to do inference on your images.
• Visualize bounding boxes and their descriptions in vg_dataset_visualization.ipynb

Trained model

We provide a playable checkpoint, Here is the link - OneDrive or BaiduYun (code is lysh)

• It is initiated from the pretrained torchvision Faster R-CNN and trained for at most 10 epochs on the training set.
• Place the checkpoint (.pth.tar) and the directory under ./model_params

Performance on the val set (train_all_val_all_bz_2_epoch_10_inject_init.pth.tar):

{
  "map": 0.09381764795879523,
  "ap_breakdown": {
    "iou_0.3_meteor_0": 0.27547998070716856,
    "iou_0.3_meteor_0.05": 0.20472119629383087,
    "iou_0.3_meteor_0.1": 0.1440167324244976,
    "iou_0.3_meteor_0.15": 0.07862544938921928,
    "iou_0.3_meteor_0.2": 0.042336766347289084,
    "iou_0.3_meteor_0.25": 0.023921287432312966,
    "iou_0.4_meteor_0": 0.2457992273569107,
    "iou_0.4_meteor_0.05": 0.18552129060029984,
    "iou_0.4_meteor_0.1": 0.13265080079436303,
    "iou_0.4_meteor_0.15": 0.07398858115077019,
    "iou_0.4_meteor_0.2": 0.04003382280468941,
    "iou_0.4_meteor_0.25": 0.02333026934415102,
    "iou_0.5_meteor_0": 0.20341708421707153,
    "iou_0.5_meteor_0.05": 0.15629490286111833,
    "iou_0.5_meteor_0.1": 0.11364261746406555,
    "iou_0.5_meteor_0.15": 0.06471541225910186,
    "iou_0.5_meteor_0.2": 0.035920637771487234,
    "iou_0.5_meteor_0.25": 0.021687612012028692,
    "iou_0.6_meteor_0": 0.15223818764090538,
    "iou_0.6_meteor_0.05": 0.11921414405107499,
    "iou_0.6_meteor_0.1": 0.08984904080629348,
    "iou_0.6_meteor_0.15": 0.052889608442783356,
    "iou_0.6_meteor_0.2": 0.03046290695667267,
    "iou_0.6_meteor_0.25": 0.018970464691519737,
    "iou_0.7_meteor_0": 0.09079854160547257,
    "iou_0.7_meteor_0.05": 0.07260968565940856,
    "iou_0.7_meteor_0.1": 0.056333212852478026,
    "iou_0.7_meteor_0.15": 0.03415838725864887,
    "iou_0.7_meteor_0.2": 0.01916669186204672,
    "iou_0.7_meteor_0.25": 0.011734895706176758
  },
  "detmap": 0.25295563289523126,
  "det_breakdown": {
    "iou_0.3": 0.37301586389541624,
    "iou_0.4": 0.3269594985246658,
    "iou_0.5": 0.26560761243104936,
    "iou_0.6": 0.18992780923843383,
    "iou_0.7": 0.10926738038659095
  }
}

Developer

• soloist97
• fuyun1107