Awesome

imprinted-weights

This is an unofficial pytorch implementation of Low-Shot Learning with Imprinted Weights.

Requirements

Python 3.5+
PyTorch 0.4.1
torchvision
pandas (0.23.4)
progress
matplotlib
numpy

Major Difference

Paper: InceptionV1 + RMSProp

This implementation: ResNet-50 + SGD

Preparation

Download CUB_200_2011 Dataset.

Unzip and locate it in this directory.

The whole directory should be look like this:

imprinted-weights
│   README.md
│   pretrain.py
│   models.py
│   loader.py
│   imprint.py
│   imprint_ft.py
│   alljoint.py
|
└───utils
│   
└───CUB_200_2011
    │   images.txt
    │   image_class_labels.txt
    │   train_test_split.txt
    │
    └───images
        │   001.Black_footed_Albatross
        │   002.Laysan_Albatross
        │   ...

Usage

Pretrain models

Train the model on the first 100 classes of CUB_200_2011.

python pretrain.py

Trained models will be saved at pretrain_checkpoint.

Imprint weights

Use N novel exemplar from the training split to imprint weights.

python imprint.py --model pretrain_checkpoint/model_best.pth.tar --num-sample N

Trained models will be saved at imprint_checkpoint.

For more details and parameters, please refer to --help option.

All w/o FT results of Table 1 and Table 2 in the paper can be reproduced by this script.

Imprint weights + FT

Apply fine-tuning to the imprinting model.

python imprint_ft.py --model pretrain_checkpoint/model_best.pth.tar --num-sample N

Trained models will be saved at imprint_ft_checkpoint.

All w/ FT results of Table 1 and Table 2 in the paper can be reproduced by this script.

All class joint

Train model for base and novel classes jointly without a separate low-shot learning phase.

python alljoint.py --num-sample N

Trained models will be saved at alljoint_checkpoint.

Results

200-way top-1 accuracy for novel-class examples in CUB-200-2011

w/o FT

n =	1	2	5	10	20
Rand-noFT (paper)	0.17	0.17	0.17	0.17	0.17
Imprinting (paper)	21.26	28.69	39.52	45.77	49.32
Imprinting + Aug (paper)	21.40	30.03	39.35	46.35	49.80
Rand-noFT	0.00	0.00	0.00	0.00	0.00
Imprinting	28.77	37.61	50.20	56.31	60.58
Imprinting + Aug	28.81	38.02	49.90	56.18	60.44

w/ FT

n =	1	2	5	10	20
Rand + FT (paper)	5.25	13.41	34.95	54.33	65.60
Imprinting + FT (paper)	18.67	30.17	46.08	59.39	68.77
AllClassJoint (paper)	3.89	10.82	33.00	50.24	64.88
Rand + FT	2.39	13.00	36.11	56.83	69.90
Imprinting + FT	26.14	34.81	54.40	63.41	73.41
AllClassJoint	7.34	18.09	45.67	60.58	73.41

200-way top-1 accuracy measured across examples in all classes of CUB-200-2011

w/o FT

n =	1	2	5	10	20
Rand-noFT (paper)	37.36	37.36	37.36	37.36	37.36
Imprinting (paper)	44.75	48.21	52.95	55.99	57.47
Imprinting + Aug (paper)	44.60	48.48	52.78	56.51	57.84
Rand-noFT	41.39	41.39	41.39	41.39	41.39
Imprinting	53.50	57.35	62.65	65.38	67.09
Imprinting + Aug	53.40	57.47	62.56	65.21	67.26

w/ FT

n =	1	2	5	10	20
Rand + FT (paper)	39.26	43.36	53.69	63.17	68.75
Imprinting + FT (paper)	45.81	50.41	59.15	64.65	68.73
AllClassJoint (paper)	38.02	41.89	52.24	61.11	68.31
Rand + FT	41.27	46.41	57.71	67.79	74.72
Imprinting + FT	53.02	57.82	67.21	71.73	76.25
AllClassJoint	42.73	48.43	62.20	69.95	75.89

References

[1]: H. Qi, M. Brown and D. Lowe. "Low-Shot Learning with Imprinted Weights", in CVPR, 2018.