/pr-dad

Pytorch implementation of paper: PR-DAD: Phase Retrieval Using Deep Auto-Decoders

Primary LanguagePython

PR-DAD

PyTorch implementation of the fellwoing paper: PR-DAD: Phase Retrieval Using Deep Auto-Decoders join with prof. Shai Dekel, School of mathematical sciences, Tel-Aviv University

Algorithm Pipeline

Package Intsallation

  • Python3 =>3.8
  • PyTorch =>1.9
  • Cuda =>11
  • Hardware requred: Ubuntu, NVIDIA Tesla V100 16Gib and 8x Intel Xeon E5-2686v4, recomeded AWS EC2 type: p3.2xlarge
  • Requred packages requirements.txt

Train Model

We agregate For each dataset per type of features we trained model and json config with hyperparameters in Table

  • Run trainer: Trainer uses ClearML logger.
python training/phase_retrieval_trainer.py --experiment_name my-experiment --config_path url_path/config-trainer.json **kwargs

my-experiment - ClearML experiment name

config_path - path(local/s3) to json with trainings hyperparameters

**kwargs (optinal) - change spefic parameters Example:

python training/phase_retrieval_trainer.py 
  --experiment_name my-experiment 
  --config_path s3://url_path/config-trainer.json 
  --path_pretrained s3://model_url/model.pt
  --batch_size 16
  --ae_type wavelet-net 
  --wavelet_type haar
  • Run evaluation:
 python training/phase_retrival_evaluator.py --model_type path_to_model/model.pt --config  url_path/config-trainer.json

Trained models and Configurations

Dataset Haar Features ConvNet Features
MNIST Model, Config-Trainer Model, Config-Trainer
EMNIST Model, Config-Trainer Model, Config-Trainer
KMNIST Model, Config-Trainer Model, Config-Trainer
Fashion-MNIST Model, Config-Trainer Model, Config-Trainer
CelebA Model, Config-Trainer Model, Config-Trainer