I am working on a classification problem using pre-trained VGG16 on imagenet. The training loss and Val loss are changing in each epoch, but the evaluation metrics like accuracy, F1-score, precision, recall remain exactly static throughout the epochs as can be seen in the image attached
. How can I make sure that evaluation metrics are changing according, as I have to choose the checkpoint based on the best F1-score for the test set. Below is the code for validation and metrics calculation:
def validate():
model.eval()
valid_losses = []
y_true, y_pred = [], []
with torch.no_grad():
for batch, target in valid_loader:
# move data to the device
batch = batch.to(device)
target = target.to(device)
# make predictions
predictions = model(batch)
# calculate loss
loss = criterion(predictions, target)
# track losses and predictions
valid_losses.append(float(loss.item()))
#print(valid_losses)
y_true.extend(target.cpu().numpy())
y_pred.extend(predictions.argmax(dim=1).cpu().numpy())
y_true = np.array(y_true)
y_pred = np.array(y_pred)
valid_losses = np.array(valid_losses)
# calculate the mean validation loss
valid_loss = valid_losses.mean()
# return valid_loss, valid_accuracy, macro_f1, macro_p, macro_r, micro_f1, cm
return valid_loss, y_true, y_pred
def evaluation_analysis(y_true, y_pred):
# calculate validation accuracy from y_true and y_pred
valid_accuracy = np.mean(y_pred == y_true)
# macro_average
macro_f1 = metrics.f1_score(y_true, y_pred, average='macro')
macro_p = metrics.precision_score(y_true, y_pred, average='macro')
macro_r = metrics.recall_score(y_true, y_pred, average='macro')
# micro average
micro_f1 = metrics.f1_score(y_true, y_pred, average='micro')
# confusion matrix
cm = metrics.confusion_matrix(y_true, y_pred)
return acc_score, macro_f1, macro_p, macro_r, micro_f1, cm```