Implementation of Deep Learning unit tests, as well as paths to the 2d data for windows users in the UNet jupyter notebook.

Unit tests have been implemented for classes and functions that relate to Deep Learning tasks such as defining Models, Hyperparameters, Datasets / Dataloaders, as well as the training process.

The process is documented in the following notion file:
https://www.notion.so/qim-dtu/Unit-tests-to-DL-funtions-e8ad1fc1aa134da2ab26474cc6480ca2

the following scripts have been made under qim3d/tests/:

• models/test_unet.py (unit tests for UNet() and Hyperparameters())
• utils/test_augmentations.py (unit tests for Augmentations() and ValueErrors)
• utils/test_data.py (unit tests for Dataset() class, resizing of images with crop, padding, prepare_datasets() and prepare_dataloaders())
• utils/test_models.py (unit tests for output of model_summary() and inference())

Problem:

• Some deep learning tests require considerably more time to run (around 15s to 20s) especially when a Dataloader is involved. This is the case for prepare_dataloaders() as well as the model_summary(). Should they still be included?
• I couldn't think of a way to test the train_model() function, since that requires training a DL model, and would also take too much time.

qim3d/utils/models.py

@@ -10,7 +10,7 @@ from tqdm.auto import tqdm
 from tqdm.contrib.logging import logging_redirect_tqdm
 
 
-def train_model(model, hyperparameters, train_loader, val_loader, eval_every = 1, print_every = 5, plot = True):
+def train_model(model, hyperparameters, train_loader, val_loader, eval_every = 1, print_every = 5, plot = True, return_loss = False):
     """ Function for training Neural Network models.
     
     Args:
@@ -21,6 +21,7 @@ def train_model(model, hyperparameters, train_loader, val_loader, eval_every = 1
         eval_every (int, optional): frequency of model evaluation. Defaults to every epoch.
         print_every (int, optional): frequency of log for model performance. Defaults to every 5 epochs.
         
+
     Returns:
         tuple:
             train_loss (dict): dictionary with average losses and batch losses for training loop.
@@ -65,10 +66,11 @@ def train_model(model, hyperparameters, train_loader, val_loader, eval_every = 1
             for data in train_loader:
                 inputs, targets = data
                 inputs = inputs.to(device)
-                targets = targets.to(device).type(torch.cuda.FloatTensor).unsqueeze(1)
+                targets = targets.to(device).unsqueeze(1)
     
                 optimizer.zero_grad()
                 outputs = model(inputs)
+                
                 loss = criterion(outputs, targets)
     
                 # Backpropagation
@@ -94,7 +96,7 @@ def train_model(model, hyperparameters, train_loader, val_loader, eval_every = 1
                 for data in val_loader:
                     inputs, targets = data
                     inputs = inputs.to(device)
-                    targets = targets.to(device).type(torch.cuda.FloatTensor).unsqueeze(1)
+                    targets = targets.to(device).unsqueeze(1)
                     
                     with torch.no_grad():
                         outputs = model(inputs)
@@ -122,6 +124,9 @@ def train_model(model, hyperparameters, train_loader, val_loader, eval_every = 1
         plot_metrics(val_loss,color = 'orange', label = 'Valid.')
         fig.show()
 
+    if return_loss:
+        return train_loss,val_loss
+
 
 def model_summary(dataloader,model):
     """Prints the summary of a PyTorch model.
@@ -196,7 +201,7 @@ def inference(data,model):
     else:
         raise ValueError("Input image must be (C,H,W) format")
 
-    
+    model.to(device)
     model.eval()
 
     # Make new list such that possible augmentations remain identical for all three rows