Merge pull request #3 from aidotse/master

Small correction

.gitignore

@@ -5,12 +5,12 @@ __pycache__/
 *jpg
 *pth 
 *npz
-*txt*
 *even*
 *tar*
 *jpeg
 *json*
 *tsv
 .ipynb_checkpoints/
 *batch*
-wandb/
+wandb/
+data/

Dockerfile

@@ -0,0 +1,24 @@
+FROM nvcr.io/nvidia/pytorch:21.11-py3
+
+ARG USER_ID
+ARG GROUP_ID
+
+ENV PYTHONDONTWRITEBYTECODE 1
+ENV PYTHONUNBUFFERED 1
+
+WORKDIR /workspace
+
+COPY requirements.txt .
+
+RUN pip install -r requirements.txt
+
+# Unset TORCH_CUDA_ARCH_LIST and exec.  This makes pytorch run-time
+# extension builds significantly faster as we only compile for the
+# currently active GPU configuration.
+#RUN (printf '#!/bin/bash\nunset TORCH_CUDA_ARCH_LIST\nexec \"$@\"\n' >> /entry.sh) && chmod a+x /entry.sh
+#ENTRYPOINT ["/entry.sh"]
+
+RUN addgroup --gid $GROUP_ID user
+RUN adduser --disabled-password --gecos '' --uid $USER_ID --gid $GROUP_ID user
+USER user
+

README.md

@@ -1,5 +1,8 @@
 # Federated Learning in Healthcare with Flower
 
+[![Watch the video](https://img.youtube.com/vi/kifLAY_5JA0/maxresdefault.jpg)](https://youtu.be/kifLAY_5JA0)
+Result from the project was presented at [Flower Summit, Cambridge 2022](https://flower.dev/conf/flower-summit-2022/).
+
 Harnessing the potential of AI in healthcare requires access to huge amounts of data to build robust models. One solution to overcome the problems of sharing healthcare data to develop better models is federated learning. In this scenario, different models are trained on each hospital's local data and share their knowledge (parameters) with a central server that performs the aggregation in order to achieve a more robust and fair model.
 
 This repository contains the code to reproduce the experiments performed in the framework of the Decentralized AI in Healthcare project at Sahlgrenska University Hospital and AI Sweden. In this example repository we are working on publicly available data (ISIC Archive) and simulating the decentralised setup internally. We have two different tasks on which we are actively working :
@@ -9,7 +12,8 @@ This repository contains the code to reproduce the experiments performed in the
 * Image classification in FL/SL setup. 
 
 Our main use case is connected with Melanoma Diagnosis using ISIC Dataset:
-    * **ISIC 2020**: Download the [ISIC 2020 dataset](https://www.kaggle.com/nroman/melanoma-external-malignant-256) 
+
+* **ISIC 2020**: Download the [ISIC 2020 dataset](https://www.kaggle.com/nroman/melanoma-external-malignant-256) 
 
 ## Flower framework
 Flower is a user-friendly framework designed for implementing the Federated Learning approach. 
@@ -18,7 +22,7 @@ Flower is a user-friendly framework designed for implementing the Federated Lear
 
 Installing the Flower framework requires Python 3.6 or higher version.
 
-To install its stabte version found on PyPI:
+To install its stable version found on PyPI:
 
  ```pip install flwr ```
 
@@ -30,6 +34,22 @@ To install its latest version from GitHub
 
  ```pip install git+https://github.com/adap/flower.git ```
 
+
+### Requirements
+
+Use the provided Dockerfile to build and image with the required library dependencies, provided in requirements.txt.
+
+```docker build -t decentralized_ai_dermatology --build-arg USER_ID=$(id -u) --build-arg GROUP_ID=$(id -g) -f Dockerfile .```
+
+Start the container
+
+```docker run -d --rm -it --volume $(pwd):/workspace --shm-size 8G --name decentralized_ai_dermatology decentralized_ai_dermatology```
+
+Execute container
+
+```docker exec -it decentralized_ai_dermatology bash```
+
+
 ### Federated learning pipeline
 
 A federated learning system needs two parts
@@ -434,7 +454,7 @@ With both client and server ready, we can now run everything and see federated l
 Once the server is running we can start the clients in different terminals. Open a new terminal per client and start the client:
 
 [`client_isic.py`](https://github.com/aidotse/decentralizedAI_dermatology/blob/master/client_isic.py) per terminal:
-```python client_isic.py –path_data <path> –num_partitions <>  –partition  <> –gpu  <gpu ID>```
+```python client_isic.py –-path <path> –-num_partitions <>  –-partition  <> –-gpu  <gpu ID>```
 
 Note: Use `--nowandb` flag if you want to disable wandb logging.
 
@@ -443,7 +463,7 @@ Note: Use `--nowandb` flag if you want to disable wandb logging.
 To train the model in a centralized way in case you want to make a comparison, you can run: 
 
 [`train_local.py`](https://github.com/aidotse/decentralizedAI_dermatology/blob/master/train_local.py)  
-```python train_local.py –path_data <path> –num_partitions <>  –partition  <> –gpu  <gpu ID>```
+```python train_local.py –-path_data <path> –-num_partitions <>  –-partition  <> –-gpu  <gpu ID>```
 
 Note: Use `--nowandb` flag if you want to disable wandb logging.
 
@@ -465,7 +485,7 @@ For StylGAN2-ADA implementation we used:
   The model is evaluated in a decentralized manner. For some GANs parameters see the script.
 
   2. Launch one [`client_isic_gan.py`](https://github.com/aidotse/decentralizedAI_dermatology/blob/master/client_isic_gan.py) per terminal:
-```python client_isic_gan.py –data <path> –num_partitions <>  –partition  <> –gpu  <gpu ID>```
+```python client_isic_gan.py –-data <path> –-num_partitions <>  –-partition  <> –gpu  <gpu ID>```
 
 Note: Use `--wandb` flag if you want to enable wandb logging.
 

client_isic.py

@@ -111,18 +111,20 @@ def evaluate(
 
 if __name__ == "__main__":
     parser = ArgumentParser() 
-    parser.add_argument("--model", type=str, default='efficientnet-b2') 
+    parser.add_argument("--model", type=str, default="efficientnet-b2") 
     parser.add_argument("--log_interval", type=int, default=100)  
     parser.add_argument("--epochs", type=int, default=2)  
+    parser.add_argument("--batch_train", type=int, default=32) 
     parser.add_argument("--num_partitions", type=int, default=20) 
     parser.add_argument("--partition", type=int, default=0)   
     parser.add_argument("--gpu", type=int, default=0)   
-    parser.add_argument("--tags", type=str, default='Exp 5. FedBN') 
+    parser.add_argument("--tags", type=str, default="Exp 5. FedBN") 
     parser.add_argument("--nowandb", action="store_true") 
-    parser.add_argument("--path", type=str, default='/workspace/melanoma_isic_dataset') 
+    parser.add_argument("--path", type=str, default="/workspace/melanoma_isic_dataset")
+    parser.add_argument("--host", type=str, default="0.0.0.0")
+    parser.add_argument("--port", type=str, default="8080")
     args = parser.parse_args()
 
-
     # Setting up GPU for processing or CPU if GPU isn't available
     os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
     device = torch.device( "cuda" if torch.cuda.is_available() else "cpu")
@@ -132,13 +134,13 @@ def evaluate(
     model = utils.load_model(args.model, device)
 
     if not args.nowandb:
-        wandb.init(project="dai-healthcare" , entity='eyeforai', group='FL', tags=[args.tags], config={"model": args.model})
+        wandb.init(project="dai-healthcare" , entity="eyeforai", group="FL", tags=[args.tags], config={"model": args.model})
         wandb.config.update(args) 
         # wandb.watch(model, log="all")
 
     # Load data
     # Normal partition
-    trainset, valset, num_examples = utils.load_isic_data()
+    trainset, valset, num_examples = utils.load_isic_data(args.path)
     trainset, valset, num_examples = utils.load_partition(trainset, valset, num_examples, idx=args.partition, num_partitions=args.num_partitions)
     # Exp 1
     # trainset, testset, num_examples = utils.load_exp1_partition(trainset, testset, num_examples, idx=args.partition)
@@ -150,12 +152,12 @@ def evaluate(
 
     print(f"Train dataset: {len(trainset)}, Val dataset: {len(valset)}, Test dataset: {len(testset)}")
 
-    train_loader = DataLoader(trainset, batch_size=32, num_workers=4, worker_init_fn=utils.seed_worker, shuffle=True) 
+    train_loader = DataLoader(trainset, batch_size=args.batch_train, num_workers=4, worker_init_fn=utils.seed_worker, shuffle=True) 
     val_loader = DataLoader(valset, batch_size=16, num_workers=4, worker_init_fn=utils.seed_worker, shuffle = False)   
     test_loader = DataLoader(testset, batch_size=16, num_workers=4, worker_init_fn=utils.seed_worker, shuffle = False)   
 
     # Start client 
     client = Client(model, train_loader, val_loader, test_loader, num_examples)
-    fl.client.start_numpy_client("0.0.0.0:8080", client)
+    fl.client.start_numpy_client(args.host + ":" + args.port, client)
 
 

requirements.txt

@@ -0,0 +1,5 @@
+imageio-ffmpeg==0.4.3 
+pyspng==0.1.0
+flwr==0.18.0
+efficientnet_pytorch==0.7.1
+wandb==0.12.17

server_advanced.py

@@ -31,7 +31,7 @@ def get_eval_fn(model, path):
     # Load data and model here to avoid the overhead of doing it in `evaluate` itself
 
     # Exp 1
-    trainset, testset, num_examples = utils.load_isic_data()
+    trainset, testset, num_examples = utils.load_isic_data(path)
     trainset, testset, num_examples = utils.load_partition(trainset, testset, num_examples, idx=3, num_partitions=10)  # Use validation set partition 3 for evaluation of the whole model
 
     # Exp 2

train_local.py

@@ -48,13 +48,13 @@
 
     # Load data
     # trainset, testset, num_examples = utils.load_exp1_partition(trainset, testset, num_examples, idx=args.partition)
-    train_df, validation_df, num_examples = utils.load_isic_by_patient(args.partition) 
+    train_df, validation_df, num_examples = utils.load_isic_by_patient(args.partition, args.path_data) 
 
     trainset = utils.CustomDataset(df = train_df, train = True, transforms = training_transforms) 
     valset = utils.CustomDataset(df = validation_df, train = True, transforms = testing_transforms ) 
     train_loader = DataLoader(trainset, batch_size=32, num_workers=8, worker_init_fn=utils.seed_worker ,shuffle=True) 
     val_loader = DataLoader(valset, batch_size=16, num_workers=4, worker_init_fn=utils.seed_worker, shuffle = False)  
-    testset = utils.load_isic_by_patient(-1) 
+    testset = utils.load_isic_by_patient(-1, args.path_data) 
     test_loader = DataLoader(testset, batch_size=16, num_workers=4, worker_init_fn=utils.seed_worker, shuffle = False)  
     print(f"Train dataset: {len(trainset)}, Val dataset: {len(valset)}, Test dataset: {len(testset)}")