Tensors are the language of the model
At the beginning, tensors can look like just another data structure. But in practice they are the core representation through which models store inputs, outputs, weights, and intermediate calculations.
Understanding shapes, dimensions, and operations gives you the confidence to reason about what the network is doing.
Loss gives learning a direction
The model does not improve just because it runs. It improves because the loss function measures how wrong the output is. That measurement becomes the signal used to adjust the model.
This is why loss should not be treated like a number to print and ignore. It represents the gap between current behavior and desired behavior.
Intuition grows when the flow becomes clear
Forward pass produces a prediction. Loss compares it against the truth. Backpropagation computes gradients. The optimizer uses those gradients to update parameters. Repeating this process gradually improves performance.
Once that flow becomes intuitive, PyTorch starts to feel much less intimidating.
