changed landmarks to environment variable, updated Dockerfile

README.md

@@ -56,7 +56,7 @@ docker run -it -v $DATA_PATH:/BFP/dataset -v $PWD/src:/BFP/src -e csv_path=/BFP/
 
 BCDR:
 ```bash
-DATA_PATH=/home/lidia/Datasets/BCDR/cropped/
+DATA_PATH=/home/lidia-garrucho/datasets/BCDR/cropped/
 CSV_FILENAME=None
 DATA_LOADER_TYPE=bcdr
 

docker/Dockerfile

@@ -20,12 +20,13 @@ RUN source fl_env/bin/activate
 RUN pwd
 
 COPY . /BFP/docker
-RUN pip3 install -r requirements.txt
-RUN pip3 install --timeout 1000 torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio===0.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
+RUN pip install -r requirements.txt
+RUN pip3 install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio===0.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
 RUN apt-get install ffmpeg libsm6 libxext6 -y
 
 RUN mkdir /BFP/dataset
 ENV PYTHONPATH "${PYTHONPATH}:/BFP"
+ENV landmarks "/BFP/src/preprocessing/optimam_train_hologic_landmarks.pth"
 
 WORKDIR "/BFP/src"
 #RUN python3 client.py

src/client.py

@@ -39,6 +39,7 @@ def import_class(name):
 # SERVER= os.getenv('server',"[::]:8080")
 SERVER= os.getenv('server',"161.116.4.137:8080") 
 DATA_LOADER_TYPE= os.getenv('data_loader_type',"optimam") #env variable data_loader if not given default to optimam type dataloading
+
 # Docker ip is: 172.17.0.3
 print(f'Here dataset path {DATASET_PATH}')
 print(f'Here csv path {CSV_PATH}')
@@ -50,7 +51,7 @@ def import_class(name):
 
 # args = parser.parse_args()
 
-DEVICE = torch.device("cuda") #if torch.cuda.is_available() else "cpu")
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 CRITERION = import_class(CONFIG['hyperparameters']['criterion'])
 
 def load_data():

src/config.yaml

@@ -6,7 +6,7 @@ name: 'sanity_check' # use 4digit name tag to overwrite experiment with same par
 hyperparameters:
     num_epochs : 20
     iteration_number: 8 # this will define the number of batches when doing federated (we want the same number of iterations)
-    batch_size : 4  # if 0, it will adapt to center
+    batch_size : 32  # if 0, it will adapt to center
     test_batch_size : 10 #use a higher one for speed up if CUDA memory allows  
     lr : 0.01
     early_stop_counter: 20 # number of iterations after performance stagnation before stopping
@@ -23,7 +23,6 @@ model:
             in_ch: 3
             out_ch: 1
             linear_ch: 512
-#             out_ch: 1 #
             early_layers_learning_rate: 1e-5 #10^-5, if set to 0 early layers will be frozen 
             seed: 42 # Particularly important for federated. Models will make no sense if we aggregate from different initializations
     pretrained: # if we have our own weights

src/data_loader.py

@@ -47,6 +47,8 @@
 
 image_ctr = 0
 
+LANDMARKS = os.environ['landmarks']# CONFIG['paths']['landmarks']
+
 # Handle DICOMs:
 # "/home/kaisar/Datasets/InBreast/AllDICOMs"
 
@@ -65,7 +67,7 @@ def preprocess_one_image_OPTIMAM(image): # Read as nifti without saving
     image = torch.from_numpy(rescaled_img).permute(2,0,1)
 
     # Histogram Matching 
-    landmarks_values = torch.load(HOME_PATH / CONFIG['paths']['landmarks'])
+    landmarks_values = torch.load(HOME_PATH / LANDMARKS)
     apply_hist_stand_landmarks(image, landmarks_values)
 
     paddedimg = torch.zeros(3,224,224)