/simclr

SimCLRv2 - Big Self-Supervised Models are Strong Semi-Supervised Learners

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

SimCLR - A Simple Framework for Contrastive Learning of Visual Representations

SimCLR Illustration
An illustration of SimCLR (from our blog here).



News! We updated the code to reflect some changes in SimCLRv2. The SimCLRv2 checkpoints and accompanying colabs are also released (see below).

Pre-trained models for SimCLRv2

Open In Colab

We opensourced total 65 pretrained models here, corresponding to those in Table 1 of the SimCLRv2 paper:

Depth Width SK Param (M) F-T (1%) F-T(10%) F-T(100%) Linear eval Supervised
50 1X False 24 57.9 68.4 76.3 71.7 76.6
50 1X True 35 64.5 72.1 78.7 74.6 78.5
50 2X False 94 66.3 73.9 79.1 75.6 77.8
50 2X True 140 70.6 77.0 81.3 77.7 79.3
101 1X False 43 62.1 71.4 78.2 73.6 78.0
101 1X True 65 68.3 75.1 80.6 76.3 79.6
101 2X False 170 69.1 75.8 80.7 77.0 78.9
101 2X True 257 73.2 78.8 82.4 79.0 80.1
152 1X False 58 64.0 73.0 79.3 74.5 78.3
152 1X True 89 70.0 76.5 81.3 77.2 79.9
152 2X False 233 70.2 76.6 81.1 77.4 79.1
152 2X True 354 74.2 79.4 82.9 79.4 80.4
152 3X True 795 74.9 80.1 83.1 79.8 80.5

These checkpoints are stored in Google Cloud Storage:

We also provide examples on how to use the checkpoints in colabs/ folder.

Pre-trained models for SimCLRv1

The pre-trained models (base network with linear classifier layer) can be found below. Note that for these SimCLRv1 checkpoints, the projection head is not available.

Model checkpoint and hub-module ImageNet Top-1
ResNet50 (1x) 69.1
ResNet50 (2x) 74.2
ResNet50 (4x) 76.6

A note on the signatures of the TensorFlow Hub module: default is the representation output of the base network; logits_sup is the supervised classification logits for ImageNet 1000 categories. Others (e.g. initial_max_pool, block_group1) are middle layers of ResNet; refer to resnet.py for the specifics. See this tutorial for additional information regarding use of TensorFlow Hub modules.

Enviroment setup

Our models are trained with TPUs. It is recommended to run distributed training with TPUs when using our code for pretraining.

Our code can also run on a single GPU. It does not support multi-GPUs, for reasons such as global BatchNorm and contrastive loss across cores.

The code is compatible with both TensorFlow v1 and v2. See requirements.txt for all prerequisites, and you can also install them using the following command.

pip install -r requirements.txt

Pretraining

To pretrain the model on CIFAR-10 with a single GPU, try the following command:

python run.py --train_mode=pretrain \
  --train_batch_size=512 --train_epochs=1000 \
  --learning_rate=1.0 --weight_decay=1e-4 --temperature=0.5 \
  --dataset=cifar10 --image_size=32 --eval_split=test --resnet_depth=18 \
  --use_blur=False --color_jitter_strength=0.5 \
  --model_dir=/tmp/simclr_test --use_tpu=False

To pretrain the model on ImageNet with Cloud TPUs, first check out the Google Cloud TPU tutorial for basic information on how to use Google Cloud TPUs.

Once you have created virtual machine with Cloud TPUs, and pre-downloaded the ImageNet data for tensorflow_datasets, please set the following enviroment variables:

TPU_NAME=<tpu-name>
STORAGE_BUCKET=gs://<storage-bucket>
DATA_DIR=$STORAGE_BUCKET/<path-to-tensorflow-dataset>
MODEL_DIR=$STORAGE_BUCKET/<path-to-store-checkpoints>

The following command can be used to pretrain a ResNet-50 on ImageNet (which reflects the default hyperparameters in our paper):

python run.py --train_mode=pretrain \
  --train_batch_size=4096 --train_epochs=100 --temperature=0.1 \
  --learning_rate=0.075 --learning_rate_scaling=sqrt --weight_decay=1e-4 \
  --dataset=imagenet2012 --image_size=224 --eval_split=validation \
  --data_dir=$DATA_DIR --model_dir=$MODEL_DIR \
  --use_tpu=True --tpu_name=$TPU_NAME --train_summary_steps=0

A batch size of 4096 requires at least 32 TPUs. 100 epochs takes around 6 hours with 32 TPU v3s. Note that learning rate of 0.3 with learning_rate_scaling=linear is equivalent to that of 0.075 with learning_rate_scaling=sqrt when the batch size is 4096. However, using sqrt scaling allows it to train better when smaller batch size is used.

Finetuning the linear head (linear eval)

To fine-tune a linear head (with a single GPU), try the following command:

python run.py --mode=train_then_eval --train_mode=finetune \
  --fine_tune_after_block=4 --zero_init_logits_layer=True \
  --variable_schema='(?!global_step|(?:.*/|^)Momentum|head)' \
  --global_bn=False --optimizer=momentum --learning_rate=0.1 --weight_decay=0.0 \
  --train_epochs=100 --train_batch_size=512 --warmup_epochs=0 \
  --dataset=cifar10 --image_size=32 --eval_split=test --resnet_depth=18 \
  --checkpoint=/tmp/simclr_test --model_dir=/tmp/simclr_test_ft --use_tpu=False

You can check the results using tensorboard, such as

python -m tensorboard.main --logdir=/tmp/simclr_test

As a reference, the above runs on CIFAR-10 should give you around 91% accuracy, though it can be further optimized.

For fine-tuning a linear head on ImageNet using Cloud TPUs, first set the CHKPT_DIR to pretrained model dir and set a new MODEL_DIR, then use the following command:

python run.py --mode=train_then_eval --train_mode=finetune \
  --fine_tune_after_block=4 --zero_init_logits_layer=True \
  --variable_schema='(?!global_step|(?:.*/|^)Momentum|head)' \
  --global_bn=False --optimizer=momentum --learning_rate=0.1 --weight_decay=1e-6 \
  --train_epochs=90 --train_batch_size=4096 --warmup_epochs=0 \
  --dataset=imagenet2012 --image_size=224 --eval_split=validation \
  --data_dir=$DATA_DIR --model_dir=$MODEL_DIR --checkpoint=$CHKPT_DIR \
  --use_tpu=True --tpu_name=$TPU_NAME --train_summary_steps=0

As a reference, the above runs on ImageNet should give you around 64.5% accuracy.

Semi-supervised learning and fine-tuning the whole network

You can access 1% and 10% ImageNet subsets used for semi-supervised learning via tensorflow datasets: simply set dataset=imagenet2012_subset/1pct and dataset=imagenet2012_subset/10pct in the command line for fine-tuning on these subsets.

You can also find image IDs of these subsets in imagenet_subsets/.

To fine-tune the whole network on ImageNet (1% of labels), refer to the following command:

python run.py --mode=train_then_eval --train_mode=finetune \
  --fine_tune_after_block=-1 --zero_init_logits_layer=True \
  --variable_schema='(?!global_step|(?:.*/|^)Momentum|head_supervised)' \
  --global_bn=True --optimizer=lars --learning_rate=0.005 \
  --learning_rate_scaling=sqrt --weight_decay=0 \
  --train_epochs=60 --train_batch_size=1024 --warmup_epochs=0 \
  --dataset=imagenet2012_subset/1pct --image_size=224 --eval_split=validation \
  --data_dir=$DATA_DIR --model_dir=$MODEL_DIR --checkpoint=$CHKPT_DIR \
  --use_tpu=True --tpu_name=$TPU_NAME --train_summary_steps=0 \
  --num_proj_layers=3 --ft_proj_selector=1

Set the checkpoint to those that are only pre-trained but not fine-tuned. Given that SimCLRv1 checkpoints do not contain projection head, it is recommended to run with SimCLRv2 checkpoints (you can still run with SimCLRv1 checkpoints, but variable_schema needs to exclude head). The num_proj_layers and ft_proj_selector need to be adjusted accordingly following SimCLRv2 paper to obtain best performances.

Other resources

Model convertion to Pytorch format

This repo provides a solution for converting the pretrained Tensorflow checkpoints into Pytorch ones.

Other non-offical implementations

(Feel free to share your implementation by creating an issue)

Implementations in PyTorch:

Implementations in Tensorflow 2 / Keras:

Cite

SimCLR paper:

@article{chen2020simple,
  title={A Simple Framework for Contrastive Learning of Visual Representations},
  author={Chen, Ting and Kornblith, Simon and Norouzi, Mohammad and Hinton, Geoffrey},
  journal={arXiv preprint arXiv:2002.05709},
  year={2020}
}

SimCLRv2 paper:

@article{chen2020big,
  title={Big Self-Supervised Models are Strong Semi-Supervised Learners},
  author={Chen, Ting and Kornblith, Simon and Swersky, Kevin and Norouzi, Mohammad and Hinton, Geoffrey},
  journal={arXiv preprint arXiv:2006.10029},
  year={2020}
}

Disclaimer

This is not an official Google product.