alinlab / psco

Unsupervised Meta-Learning via Few-shot Pseudo-supervised Contrastive Learning (ICLR 2023)

Unsupervised Meta-learning via Few-shot Pseudo-supervised Contrastive Learning

PyTorch implementation for "Unsupervised Meta-learning via Few-shot Pseudo-supervised Contrastive Learning" (accepted Spotlight presentation in ICLR 2023)

TL;DR: Constructing online pseudo-tasks via momentum representations and applying contrastive learning improves the pseudo-labeling strategy progressively for meta-learning.

Install

conda create -n unsup_meta python=3.9
conda activate unsup_meta
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
conda install ignite -c pytorch
pip install packaging tensorboard sklearn

Download datasets

• omniglot: images_background.zip and images_evaluation.zip
• miniimagenet
• cross-domain

Meta-Training PsCo

Omniglot

python train.py --model psco --backbone conv4 --prediction --num-shots 1 \
    --dataset omniglot --datadir DATADIR \
    --logdir logs/omniglot/psco

miniImageNet

python train.py --model psco --backbone conv5 --prediction --num-shots 4 \
    --dataset miniimagenet --datadir DATADIR \
    --logdir logs/miniimagenet/psco

Meta-Testing PsCo

Standard few-shot classification (Table 1)

• For Omniglot

python test.py --model psco --backbone conv4 --prediction --num-shots 1 \
    --ckpt logs/omniglot/psco/last.pth \
    --pretrained-dataset omniglot \
    --dataset omniglot --datadir [DATADIR] \
    --N 5 --K 1 --num-tasks 2000 \
    --eval-fewshot-metric supcon

• For miniImageNet

python test.py --model psco --backbone conv5 --prediction --num-shots 4 \
    --ckpt logs/miniimagenet/psco/last.pth \
    --pretrained-dataset miniimagenet \
    --dataset miniimagenet --datadir [DATADIR] \
    --N 5 --K 1 --num-tasks 2000 \
    --eval-fewshot-metric supcon

Cross-domain few-shot classification with miniImageNet pretrained (Table 2)

• miniImageNet to [DATASET]

python test.py --model psco --backbone conv5 --prediction --num-shots 4 \
    --ckpt logs/miniimagenet/psco/last.pth \
    --pretrained-dataset miniimagenet \
    --dataset [DATASET] --datadir [DATADIR] \
    --N 5 --K 5 --num-tasks 2000 \
    --eval-fewshot-metric ft-supcon

• [DATASET] list
  • cub200 (For CUB200)
  • cars (For Cars)
  • places (For Places)
  • plantae (For Plantae)
  • cropdiseases (For CropDiseases)
  • eurosat (For EuroSAT)
  • isic (For ISIC)
  • chestx (For ChestX)

psco's Issues

What is vinyals?

Trying to train using the omniglot dataset right now, getting an error because I can't find vinyals_x_.json as specified in

        with open(os.path.join(self.root, f'vinyals_{self.split}_labels.json'), mode='r') as f:
            dir_list = json.load(f)

What is this supposed to be?

No such file or directory: '***/omniglot/vinyals_train_labels.json'

Hi，thanks for your great work!
When i run your code , I meet a problem:"No such file or directory: '***/omniglot/vinyals_train_labels.json'.But I can't find this file.And could you please share the link of downloading miniimagenet and other cross domain dataset?
Thanks a lot!

AttributeError: module 'torch.ao' has no attribute 'nn'

Thank you for good work.
(unsup_meta) /mnt/workspace/code/ICLR2023/PsCo> CUDA_DEVICES_VISIBLE="4,5" python train.py --model psco --backbone conv4 --prediction --num-shots 1 --dataset omniglot --datadir /mnt/workspace/code/ICLR2023/PsCo/data --logdir logs/omniglot/psco
/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: '/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torchvision/image.so: undefined symbol: _ZN3c104impl8GPUTrace13gpuTraceStateE'If you don't plan on using image functionality from torchvision.io, you can ignore this warning. Otherwise, there might be something wrong with your environment. Did you have libjpeg or libpng installed before building torchvision from source?
warn(
Traceback (most recent call last):
File "/mnt/workspace/code/ICLR2023/PsCo/train.py", line 12, in
import utils
File "/mnt/workspace/code/ICLR2023/PsCo/utils.py", line 17, in
import models
File "/mnt/workspace/code/ICLR2023/PsCo/models.py", line 4, in
import torchvision
File "/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torchvision/init.py", line 6, in
from torchvision import datasets, io, models, ops, transforms, utils
File "/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torchvision/models/init.py", line 17, in
from . import detection, optical_flow, quantization, segmentation, video
File "/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torchvision/models/quantization/init.py", line 3, in
from .mobilenet import *
File "/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torchvision/models/quantization/mobilenet.py", line 1, in
from .mobilenetv2 import * # noqa: F401, F403
File "/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torchvision/models/quantization/mobilenetv2.py", line 5, in
from torch.ao.quantization import DeQuantStub, QuantStub
File "/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torch/ao/quantization/init.py", line 3, in
from .fake_quantize import * # noqa: F403
File "/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torch/ao/quantization/fake_quantize.py", line 8, in
from torch.ao.quantization.observer import (
File "/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torch/ao/quantization/observer.py", line 15, in
from torch.ao.quantization.utils import (
File "/mnt/workspace/workgroup/yfchen/anaconda3/envs/unsup_meta/lib/python3.9/site-packages/torch/ao/quantization/utils.py", line 655, in
) -> torch.ao.nn.quantizable.LSTM:
AttributeError: module 'torch.ao' has no attribute 'nn'