48. Small CNN Image Project in PyTorch
48.1 Project Goal (Fictional)
Why. Theory sticks after a tiny end-to-end run — even on fake images.
What. Fictional project: classify 28×28 grayscale patches as “packet” vs “spill” on a Blinkit-style warehouse camera (synthetic tensors, not real CCTV).
How.
Write the one-pager
- 2 classes.
- Input shape (batch, 1, 28, 28).
- TinyCNN + CrossEntropyLoss.
- Success = loss drops and accuracy rises on a holdout batch.
print("Classes: 0=packet, 1=spill (fictional labels)")
Fictional data
We use random tensors with planted patterns — not real warehouse footage.