phase 2 neural networks

[MahdiEsrafili]

• [x] add confusion matrix
• [x] add classification report (precision recall f-score)
• [x] more data augmentation and transforms
• [x] make intel classification network bigger
• [x] plot loss after training
• [ ] use tensorboard for monitoring
• [x] use batch normalization and drop out in network
• [ ] sample data for training

[MahdiEsrafili]

• [x] use .to(device) for putting tensors to cpu or gpu
• [x] for testing, use with torch.no_grad():
• [x] in training loop, put model.train inside each epoch
• [x] for counting number of correct samples in batches in testing, you can use this snippet:

images= images.to(device)
ground_truth.append(labels.tolist())
labels = labels.float().to(device)
out = model(clips)
out = out.argmax(1)
predicted_label.append(out.tolist())
n_corrects += (out == labels).sum()
n_samples += out.shape[0]

• [ ] does network overfit?

[MahdiEsrafili]

• [ ] use batch normalization
• [ ] use step learning rate scheduler instead of Exponential (step size=8 seems good)
• [ ] use another optimizer, like Adam
• [ ] train more steps (+20)

[MahdiEsrafili]

• [ ] check out tqdm documentation for set_description method, at the moment you use print inside training loop to show the loss and other info