Pytorch-Training-Toolkit

• Training Loop
• HistoricalTraining
  • Display Tracked Value While Training Loop
  • Diagnostic Training
  • Plot Training/Validation Curves

---

Training Loop

• Train and Evaluate Model with Train and Test Dataloader
• Track and Analyze Data with HistoricalTraining instance (history)
• for an example of utilisation got to -> training.ipynb

history = train(
    model=model,
    train_dataloader=train_dataloader,
    test_dataloader=test_dataloader,
    optimizer=optimizer,
    loss_function=loss_function,
    metric_function=metric_function,
    device=device,
    epochs=5
)

---

HistoricalTraining

• Display Tracked Value While Training Loop
• Diagnostic Training Loop
• Plot Training/Validation Curves

---

---

• Initialize Instance Before Loop :

from training_history.core import HistoricalTraining

# create a HistoricalTraining instance
history = HistoricalTraining(max_epochs=NBR_EPOCHS)

# Initialize List for each tracked Value
history["Train BCE"] = list()
history["Validation BCE"] = list()
history["Train Accuracy"] = list()
history["Validation Accuracy"] = list()

Use Instance method in Loop :

• display_info() display with reccurance all Tracked Value
• diagnostic() Compute and Display Diagnostic of Model (OverFitting and UnderFitting), Give at method the Target metric to Diagnostic metric_name
• plot_curves Visualize Tracked Value in Pretty Subplot, Give at method the Target Loss and Metric to Plot loss_name and metric_name

for epoch in range(NBR_EPOCHS):

    # Activate Training Mode
    model.train()
    
	# forward, convert logits...
	
    # Get Value
    train_loss = loss_function(y_logits, y_train)
    train_score = metric_function(y_preds, y_train)
    
    # Compute Gradients, Optimize, ect...
    
    # Activate Evaluation Mode
    model.eval()

    # Make Accelerate Prediction in Inference Mode
    with torch.inference_mode():

        # forward, convert logits...
        
        # Get Value
        val_loss = loss_function(y_logits, y_test)
        val_score = metric_function(y_preds, y_test)

    # Tracked all New Value in Instance
    # Convert all Value to Numpy (detach from gpu)
    history["Train BCE"].append(to_numpy(val_loss))
    history["Train Accuracy"].append(to_numpy(train_score))
    history["Validation BCE"].append(to_numpy(train_loss))
    history["Validation Accuracy"].append(to_numpy(val_score))

	# print information with current epoch and reccurance
    history.display_info(current_epoch=epoch,
                          reccurance=100)

# At the end of The loop Diagnostic and Visualize Training Process
history.diagnostic(average=False)
history.diagnostic(average=True)
history.plot_curves(loss_name="BCE")

---

Display Tracked Information in history While Training

• Initialize Instance Before Loop :

from training_history.core import HistoricalTraining

# create a HistoricalTraining instance
history = HistoricalTraining(max_epochs=NBR_EPOCHS)

# Initialize List for each tracked Value
history["Train BCE"] = list()
history["Validation BCE"] = list()
history["Train Accuracy"] = list()
history["Validation Accuracy"] = list()

Use Instance method in Loop :

• display_info() display with reccurance all Tracked Value
  • give the current_epoch
  • give the reccurance of display

for epoch in range(NBR_EPOCHS):

    # Activate Training Mode
    model.train()
    
	# forward, convert logits...
	
    # Get Value
    train_loss = loss_function(y_logits, y_train)
    train_score = metric_function(y_preds, y_train)
    
    # Compute Gradients, Optimize, ect...
    
    # Activate Evaluation Mode
    model.eval()

    # Make Accelerate Prediction in Inference Mode
    with torch.inference_mode():

        # forward, convert logits...
        
        # Get Value
        val_loss = loss_function(y_logits, y_test)
        val_score = metric_function(y_preds, y_test)

    # Tracked all New Value in Instance
    # Convert all Value to Numpy (detach from gpu)
    history["Train BCE"].append(to_numpy(val_loss))
    history["Train Accuracy"].append(to_numpy(train_score))
    history["Validation BCE"].append(to_numpy(train_loss))
    history["Validation Accuracy"].append(to_numpy(val_score))

	# print information with current epoch and reccurance
    history.display_info(current_epoch=epoch,
                          reccurance=100)

---

Diagnostic Training

Use at the End or While Pytorch Training Loop

• Initialize Instance Before Loop :

from training_history.core import HistoricalTraining

# create a HistoricalTraining instance
history = HistoricalTraining(max_epochs=NBR_EPOCHS)

# Initialize List for each tracked Value
history["Train BCE"] = list()
history["Validation BCE"] = list()
history["Train F1Score"] = list()
history["Validation F1Score"] = list()

Use Instance method in Loop :

• diagnostic() Compute Diagnostic of Model (OverFitting and UnderFitting) and store in instance.
• Give at method the Target metric to Diagnostic metric_name

for epoch in range(NBR_EPOCHS):

    # Activate Training Mode
    model.train()
    
	# forward, convert logits...
	
    # Get Value
    train_loss = loss_function(y_logits, y_train)
    train_score = metric_function(y_preds, y_train)
    
    # Compute Gradients, Optimize, ect...
    
    # Activate Evaluation Mode
    model.eval()

    # Make Accelerate Prediction in Inference Mode
    with torch.inference_mode():

        # forward, convert logits...
        
        # Get Value
        val_loss = loss_function(y_logits, y_test)
        val_score = metric_function(y_preds, y_test)

    # Tracked all New Value in Instance
    # Convert all Value to Numpy (detach from gpu)
    history["Train BCE"].append(to_numpy(val_loss))
    history["Train F1Score"].append(to_numpy(train_score))
    history["Validation BCE"].append(to_numpy(train_loss))
    history["Validation F1Score"].append(to_numpy(val_score))

# At the end of The loop Diagnostic and Visualize Training Process
history.diagnostic(average=False, metric_name="F1")

>>> history

{'Epochs': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
 'Train Loss': [0.53915702799956,
                ...,
                0.18948346860706805],
 'Train Accuracy': [0.7875,
                    ...,
                    0.9270833333333334],
 'Val Loss': [0.28438338860869405,
              ...,
              0.3194670816883445],
 'Val Accuracy': [0.8863636374473571,
                  ...,
                  0.9125],
 'Training Time': 50.151948901999276,
 'Bias and UnderFitting': ['Medium Bias'],
 'Variance and OverFitting': ['Low Variance', 'Good Generalization !']}

---

Plot Training & Validation Curves

Use at the End of Pytorch Training Loop

• Initialize Instance Before Loop :

from training_history.core import HistoricalTraining

# create a HistoricalTraining instance
history = HistoricalTraining(max_epochs=NBR_EPOCHS)

# Initialize List for each tracked Value
history["Train BCE"] = list()
history["Validation BCE"] = list()
history["Train F1Score"] = list()
history["Validation F1Score"] = list()

Use Instance method in Loop :

• plot_curves Visualize Tracked Value in Pretty Subplot,
• Give at method the Target Loss and Metric to Plot loss_name and metric_name

for epoch in range(NBR_EPOCHS):

    # Activate Training Mode
    model.train()
    
	# forward, convert logits...
	
    # Get Value
    train_loss = loss_function(y_logits, y_train)
    train_score = metric_function(y_preds, y_train)
    
    # Compute Gradients, Optimize, ect...
    
    # Activate Evaluation Mode
    model.eval()

    # Make Accelerate Prediction in Inference Mode
    with torch.inference_mode():

        # forward, convert logits...
        
        # Get Value
        val_loss = loss_function(y_logits, y_test)
        val_score = metric_function(y_preds, y_test)

    # Tracked all New Value in Instance
    # Convert all Value to Numpy (detach from gpu)
    history["Train BCE"].append(to_numpy(val_loss))
    history["Train F1Score"].append(to_numpy(train_score))
    history["Validation BCE"].append(to_numpy(train_loss))
    history["Validation F1Score"].append(to_numpy(val_score))

# At the end of The loop Visualize Training/ Validation Process
history.plot_curves(loss_name="BCE", metric_name="F1")

Recover figure

fig = history["Curve Figure"]
fig.savefig("your_path")