few_shot_image

Vision Tasks

Running the code

Datasets

miniImageNet (courtesy of Spyros Gidaris)
tieredImageNet (courtesy of Kwonjoon Lee)
domainNet
ImageNet-800

result

standard FT

Backbone	dataset	shot	Accuracy
CLIP-ViT_B32	mini	1	69.39 +- 0.30
		5	88.39 +- 0.15
	tiered	1	61.20 +- 0.37
		5	80.65 +- 0.27
	domain	1	47.72 +- 0.37
		5	72.82 +- 0.29
CLIP-RN50	mini	1	63.15 +- 0.31
		5	83.45 +- 0.17
	tiered	1	55.77 +- 0.35
		5	75.28 +- 0.29
	domain	1	43.77 +- 0.38
		5	67.30 +- 0.31
DINOv2-ViT	mini	1	91.07 +- 0.19
		5	97.32 +- 0.06
	tiered	1	82.40 +- 0.30
		5	93.07 +- 0.16
	domain	1	61.84 +- 0.39
		5	79.63 +- 0.25
torchVis-ViT	mini	1	95.28 +- 0.13
		5	98.33 +- 0.04
	tiered	1	86.44 +- 0.27
		5	94.91 +- 0.12
	domain	1	30.93 +- 0.31
		5	52.14 +- 0.29
torchVis-RN50	mini	1	84.10 +- 0.22
		5	94.81 +- 0.09
	tiered	1	74.48 +- 0.33
		5	88.35 +- 0.19
	domain	1	34.10 +- 0.31
		5	55.08 +- 0.29

5-way accuracy (%) on mini-ImageNet

Backbone	Task(M)	shot images	query images	image per task(m)	Accuracy
CLIP-ViT_B32	0	None	None	None	83.03 +- 0.24
	200	1	7	40	88.53 +- 0.22
		5	35	200	89.50 +- 0.20
		25	175	1000	89.93 +- 0.20
	1000	1	7	40	89.37 +- 0.20
		5	35	200	90.81 +- 0.19
		25	175	1000	90.97 +- 0.19
	5000	1	7	40	89.95 +- 0.20
		5	35	200	90.94 +- 0.19
		25	175	1000	91.16 +- 0.18

M vs m trend

15-way accuracy (%) on mini-ImageNet

Backbone	Task(M)	shot images	query images	image per task(m)	Accuracy
CLIP-ViT_B32	0	None	None	None	68.41 +- 0.54
	200	1	7	120	78.42 +- 0.15(78.62 +-0.15)
	200	1	9	150	78.51 +- 0.15
		2	18	300	80.26 +- 0.15
		3	27	450	80.59 +- 0.14
		4	36	600	80.67 +- 0.14
	400	1	9	150	78.95 +- 0.15
		2	18	300	80.85 +- 0.15
		3	27	450	81.31 +- 0.14
		4	36	600	81.51 +- 0.14
	600	1	9	150	79.44 +- 0.15
		2	18	300	80.87 +- 0.15
		3	27	450	81.38 +- 0.14
		4	36	600	81.65 +- 0.14
	800	1	9	150	79.87 +- 0.15
		2	18	300	81.06 +- 0.15
		3	27	450	81.44 +- 0.14
		4	36	600	81.65 +- 0.14
CLIP-RN50	0	None	None	None	61.31 +- 0.31
	200	1	9	150	67.32 +- 0.16
		2	18	300	67.19 +- 0.16
		3	27	450	67.12 +- 0.16
		4	36	600	-
	400	1	9	150	67.25 +- 0.16
		2	18	300	67.16 +- 0.16
		3	27	450	67.14 +- 0.16
		4	36	600	-
	600	1	9	150	67.16 +- 0.16
		2	18	300	67.38 +- 0.16
		3	27	450	67.32 +- 0.16
		4	36	600	-
	800	1	9	150	67.47 +- 0.16
		2	18	300	67.31 +- 0.16
		3	27	450	-
		4	36	600	-
DinoV2-ViT_b14	0	None	None	None	90.61 +- 0.19
	200	1	9	150	92.77 +- 0.18

5-shot in testing

Backbone	Task(M)	shot images	query images	image per task(m)	Accuracy
CLIP-ViT_B32	0	None	None	None	88.67 +- 0.14
	200	1	9	150	93.22 +- 0.11
CLIP-RN50	0	None	None	None	82.03 +- 0.18
	200	1	9	150	85.09 +- 0.17
DinoV2-ViT_b14	0	None	None	None	97.20 +- 0.06
	200	1	9	150	97.68 +- 0.06

15-way accuracy (%) on tiered-ImageNet

1-shot in testing

Backbone	Task(M)	shot images	query images	image per task(m)	Accuracy
CLIP-ViT_B32	0	None	None	None	59.55 +- 0.21
	200	1	9	150	68.57 +- 0.37
		2	18	300	70.51 +- 0.37
		3	27	450	71.14 +- 0.36
		4	36	600	71.30 +- 0.36
	400	1	9	150	70.19 +- 0.53
		2	18	300	71.63 +- 0.36
		3	27	450	72.33 +- 0.36
		4	36	600	72.44 +- 0.36
	600	1	9	150	70.54 +- 0.36
		2	18	300	72.33 +- 0.36
		3	27	450	72.64 +- 0.35
		4	36	600	72.81 +- 0.36
	800	1	9	150	71.44 +- 0.36
		2	18	300	72.28 +- 0.35
		3	27	450	72.78 +- 0.36
		4	36	600	72.93 +- 0.35
CLIP-RN50	0	None	None	None	51.76 +- 0.36
	200	1	9	150	57.56 +- 0.36
		2	18	300	57.64 +- 0.36
		3	27	450	57.74 +- 0.36
		4	36	600	58.90 +- 0.35
	400	1	9	150	58.19 +- 0.36
		2	18	300	58.33 +- 0.36
		3	27	450	58.40 +- 0.36
		4	36	600	59.24 +- 0.36
	600	1	9	150	58.40 +- 0.50
		2	18	300	58.63 +- 0.36
		3	27	450	58.72 +- 0.36
		4	36	600	59.63 +- 0.36
	800	1	9	150	58.86 +- 0.35
		2	18	300	58.95 +- 0.36
		3	27	450	59.73 +- 0.36
		4	36	600	59.91 +- 0.36
DinoV2-ViT_b14	0	None	None	None	82.33 +- 0.30
	200	1	9	150	84.74 +- 0.30

5-shot in testing

Backbone	Task(M)	shot images	query images	image per task(m)	Accuracy
CLIP-ViT_B32	0	None	None	None	79.51 +- 0.27
	200	1	9	150	84.79 +- 0.22
CLIP-RN50	0	None	None	None	71.40 +- 0.30
	200	1	9	150	75.80 +- 0.28
DinoV2-ViT_b14	0	None	None	None	92.90 +- 0.16
	200	1	9	150	93.65 +- 0.16

15-way accuracy (%) on domain-Net

1-shot in testing

Backbone	Task(M)	shot images	query images	image per task(m)	Accuracy
CLIP-ViT_B32	0	None	None	None	46.48 +- 0.37
	200	1	9	150	64.97 +- 0.39
CLIP-RN50	0	None	None	None	40.55 +- 0.36
	200	1	9	150	52.67 +- 0.39
DinoV2-ViT_b14	0	None	None	None	61.65 +- 0.41
	200	1	9	150	68.22 +- 0.40

5-shot in testing

Backbone	Task(M)	shot images	query images	image per task(m)	Accuracy
CLIP-ViT_B32	0	None	None	None	72.01 +- 0.29
	200	1	9	150	80.49 +- 0.25
CLIP-RN50	0	None	None	None	64.90 +- 0.31
	200	1	9	150	72.19 +- 0.30
DinoV2-ViT_b14	0	None	None	None	79.34 +- 0.25
	200	1	9	150	82.62 +- 0.24

Vary num shot

15-way accuracy (%) on mini-ImageNet

Backbone	Task(M)	shot images	query images	image per task(m)	Accuracy
CLIP-ViT_B32	0	None	None	None	68.41 +- 0.54
	200	1	19	300	-
		5	15	300	-
		10	10	300	-
		15	5	300	-
	400	1	19	300	-
		5	15	300	-
		10	10	300	-
		15	5	300	-
	400	1	39	600	-
		5	35	600	-
		10	30	600	-
		15	25	600	-
CLIP-RN50	0	None	None	None	-
	200	1	19	300	-
		5	15	300	-
		10	10	300	-
		15	5	300	-
	400	1	19	300	-
		5	15	300	-
		10	10	300	-
		15	5	300	-
	400	1	39	600	-
		5	35	600	-
		10	30	600	-
		15	25	600	-

15-way accuracy (%) on tiered-ImageNet

Backbone	Task(M)	shot images	query images	image per task(m)	Accuracy
CLIP-ViT_B32	0	None	None	None	59.55 +- 0.21
	200	1	19	300	69.10 +- 0.37
		5	15	300	71.27 +- 0.36
		10	10	300	71.10 +- 0.37
		15	5	300	71.01 +- 0.36
	400	1	19	300	70.84 +- 0.36
		5	15	300	72.43 +- 0.36
		10	10	300	72.42 +- 0.36
		15	5	300	72.11 +- 0.35
	400	1	39	600	70.89 +- 0.36
		5	35	600	72.45 +- 0.36
		10	30	600	72.42 +- 0.36
		15	25	600	72.20 +- 0.36
CLIP-RN50	0	None	None	None	51.76 +- 0.36
	200	1	19	300	59.38 +- 0.36
		5	15	300	59.27 +- 0.36
		10	10	300	58.86 +- 0.36
		15	5	300	58.64 +- 0.36
	400	1	19	300	60.47 +- 0.36
		5	15	300	59.89 +- 0.36
		10	10	300	59.77 +- 0.36
		15	5	300	59.57 +- 0.36
	400	1	39	600	60.44 +- 0.37
		5	35	600	59.72 +- 0.36
		10	30	600	59.54 +- 0.36
		15	25	600	59.48 +- 0.36

Examples

python command

python finetune.py \
    --config=configs/clip/mini-imagenet/finetune_ViT.yaml \
    --n_batch_train 200 \
    --n_shot 1 \
    --sample_per_task 150
    
python test.py --config=configs/clip/mini-imagenet/test.yaml

modify path in test.yaml as saving model path or add --path [path] in command python test.py

run one job

sbatch script.sh

Modify script.sh with different python command.

run multiple jobs with varying one inputs

sbatch --array=1-3 jobArrayScript.sh

Modify jobArrayScript.sh and/or input_file_list.txt, input_path_list.txt.

run multiple jobs with varying several inputs

bash multi_jobs.sh

Modify script.sh, multi_jobs.sh.

Local data directory

datasets/
├── mini-imagenet
│   ├── miniImageNet_category_split_meta_train_limited100.pickle
│   ├── miniImageNet_category_split_meta_train_limited10.pickle
│   ├── miniImageNet_category_split_meta_train_limited20.pickle
│   ├── miniImageNet_category_split_meta_train_limited50.pickle
│   ├── miniImageNet_category_split_meta_train_limited_class16.pickle
│   ├── miniImageNet_category_split_meta_train_limited_class32.pickle
│   ├── miniImageNet_category_split_meta_train_limited_class8.pickle
│   ├── miniImageNet_category_split_test.pickle
│   ├── miniImageNet_category_split_train_phase_test.pickle
│   ├── miniImageNet_category_split_train_phase_train.pickle
│   ├── miniImageNet_category_split_train_phase_val.pickle
│   └── miniImageNet_category_split_val.pickle

Name		Name	Last commit message	Last commit date
Latest commit History 118 Commits
configs		configs
datasets		datasets
input_files_jobarray		input_files_jobarray
models		models
notebook		notebook
scripts		scripts
tools		tools
utils		utils
.gitignore		.gitignore
README.md		README.md
VisionLM.md		VisionLM.md
backup_finetune_before_real_maml.py		backup_finetune_before_real_maml.py
debug.py		debug.py
finetune.py		finetune.py
finetune_backup.py		finetune_backup.py
finetune_maml.py		finetune_maml.py
input_file_list.txt		input_file_list.txt
input_path_list.txt		input_path_list.txt
interactive.sh		interactive.sh
jobArrayScript.sh		jobArrayScript.sh
lp_finetune.py		lp_finetune.py
main.ipynb		main.ipynb
multi_jobs.sh		multi_jobs.sh
partial_result.txt		partial_result.txt
requirements.txt		requirements.txt
run.py		run.py
runVisionLM.py		runVisionLM.py
script.sh		script.sh
sub.py		sub.py
test.py		test.py
visionLM.py		visionLM.py

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

few_shot_image

Vision Tasks

Running the code

result

standard FT

5-way accuracy (%) on mini-ImageNet

M vs m trend

15-way accuracy (%) on mini-ImageNet

15-way accuracy (%) on tiered-ImageNet

15-way accuracy (%) on domain-Net

Vary num shot

15-way accuracy (%) on mini-ImageNet

15-way accuracy (%) on tiered-ImageNet

Examples

Local data directory

About

Releases

Packages

Contributors 2

Languages

zhuoyan-xu/few_shot_image

Folders and files

Latest commit

History

Repository files navigation

few_shot_image

Vision Tasks

Running the code

result

standard FT

5-way accuracy (%) on mini-ImageNet

M vs m trend

15-way accuracy (%) on mini-ImageNet

15-way accuracy (%) on tiered-ImageNet

15-way accuracy (%) on domain-Net

Vary num shot

15-way accuracy (%) on mini-ImageNet

15-way accuracy (%) on tiered-ImageNet

Examples

Local data directory

About

Resources

Stars

Watchers

Forks

Releases

Packages 0

Contributors 2

Languages

Packages