SimCLRPL/
├─ conf/
│ ├─ config.yaml
│ ├─ StyleCLR_base_format.yaml
├─ data/
│ ├─ pretrained_model_adain/
│ │ ├─ decored.pth
│ │ ├─ vgg_normalised.pth
│ ├─ style_augmentation_images/
│ │ ├─ 1.jpg
│ │ ├─ ...
│ │ ├─ 1000.jpg
│ ├─ train_info_1000.csv
├─ data_aug/
│ ├─ adain.py
│ ├─ style_transforms.py
├─ exceptions/
│ ├─ exceptions.py
├─ models/
│ ├─ resnet_simclr.py
├─ dataset.py
├─ lightning_modules.py
├─ pl_main.py
├─ README.md
├─ requirements.txt
Set configuration in pl_main.py with
@hydra.main(config_path="conf", config_name="StyleCLR_base_format")Can create a new configuration file folowing the base format available in conf/.
For detailed information on how to modify configuration files through CLI, check Hydra Documentation.
model_name: StyleCLR_base # Model name for loggin (neptune).
root_path: /ROOT/ # Project main root.
seed: 9999 # Random seed in random, torch, pandas and cuda.
gpu_ids: # GPU ID's as list.
- 1
- 2
- 3
- 4
train: # PL trainer params. See PL trainer documentation.
strategy: ddp
precision: 32
max_epochs: 100
log_every_n_steps: 4
check_val_every_n_epoch: 0
amp_backend: native
enable_checkpointing: True
model:
base_model: resnet50 # Backbone model for SimCLR.
out_dim: 128 # SimCLR output dimension.
temperature: 0.07 # InfoNCE temperature parameters.
style_alpha: 0.8 # AdaIn alpha (higher -> stronger style).
style_vgg_path: /ROOT/data/pretrained_models_adain/vgg_normalised.pth # Style transfer VGG backbone network pretrained weights' path.
style_decoder_path: /ROOT/data/pretrained_models_adain/decoder.pth # Style transfer decoder network pretrained weights' path.
dataset:
style:
recalculate_vgg_embeddings: False # Whether to calculate pretrained embeddings for style images.
data_path: /ROOT/data/ # Style data path (metadata and images).
image_path: style_augmentation_images/ # Style images folder name.
metadata_filename: train_info_1000.csv # Style dataset metadata CSV filename.
pickle_filename: style_precalculated_features.pkl # Precalculated style images embeddings generated with frozen backbone VGG.
image_size: 96 # Style image warp size (square).
device: 0 # Device to calculate preprocessed style embeddings.
dataset_name: stl10 # Dataset to use (stl10/cifar10)
batch_size: 512 # Main batch size.
num_workers: 4 # Main dataloader number of workers.
n_views: 2 # Number of views to generate (currently supports 2)
data_path: /ROOT/datasets # Content images path.
len_train_loader: set_at_runtime # Holder for train dataloader size (used in scheduler).
optimizer:
lr: 0.0003 # Main Learning rate.
weight_decay: 0.0001 # Weight decay.
logger:
project_name: example/StyleCLR # Logger name for neptune.
tags:
- base
- example_dataset_name
- example_only_adain_augmentations # Tags for neptune logger as list.
api_key: NEPTUNE_API_TOKEN # Neptune API token.
callbacks:
monitor: val_loss # Torch Lightning callback parameters
monitor_mode: min # Example - Checkpointer, early stopping, etc.
save_top_k: 5
patience: 1000As an example, after setting the ROOT variable in the config file above, can run the model on 4 GPUs and 4 workers with
python pl_main.py gpu_ids=[1,2,3,4] dataset.num_workers=4