2025 Contest Days

🐔 Chicken Counting

Count chickens via density maps — with a frozen encoder you're not allowed to touch.

CVModel surgeryDensity estimation★★★☆ pure model surgery

📜The task, in plain English

Count the chickens in farm images. The organizers hand you a frozen pretrained image encoder — you may not fine-tune it. Your job is to design and train the decoder that turns its features into a count (via a density map that sums to the number of chickens).

Input	Images → frozen encoder features
Output	Chicken count per image (via predicted density map)
Really tests	PyTorch model surgery: reading someone else's module, matching shapes, building your own head

🔧The baseline you're given

A trivial decoder (e.g., global pooling + linear regression on the count). Limitations: throwing away spatial information, no density-map supervision, weak on crowded images.

🚀Baseline vs. solution

⚠️ The baseline (what you are given)

Pool everything → regress one number
No spatial supervision
Crowded scenes fail

✅ The winning approach

Density-map decoder: a few upsampling conv layers producing a heat map whose SUM is the count — supervise with per-pixel targets built from dot annotations
Count = density.sum(); loss = MSE on maps (optionally + count loss)
Freeze means freeze: requires_grad=False and encoder in .eval() — verify with a param count
Precompute encoder features ONCE (encoder is frozen!) → decoder training becomes lightning fast

🧒Explain it like I'm brand new

Why density maps? Counting by detecting each chicken fails when they overlap. Instead, predict a "chicken heat" image — each bird contributes a small blob summing to 1 — and integrate the heat. 30 blobs → sum ≈ 30.

The precompute trick is the big transferable lesson: if a component is frozen, its outputs never change, so compute them once and train only the small part on top. Turns hours into minutes — and reappears everywhere (John Wilkins' precomputed table is the same idea in disguise).

💬The Gemma 4 playthrough (2000-token limit)

Chat 1 · Feature caching

YOU

PyTorch: frozen encoder `enc` maps (B,3,224,224) -> (B,512,14,14). Write ONLY code to iterate a DataLoader, run enc under torch.no_grad(), and save features+targets to features.pt. Then a TensorDataset loading them. Max 25 lines.

Chat 2 · Decoder design

YOU

Design ONLY an nn.Module decoder: input (B,512,14,14), output a (B,1,56,56) non-negative density map. Two ConvTranspose2d upsampling steps + final 1x1 conv + ReLU. Code only, max 25 lines.

Shape ritual: print(decoder(torch.randn(2,512,14,14)).shape) → [2,1,56,56]. Then overfit 10 images until train loss ≈ 0 — if you can't overfit 10 images, the wiring is broken (classic debugging trick).

🎯Takeaways & what Day 1 might do with this

Pattern family: frozen encoder + trainable head — the IOAI yearly ritual (2026 Night Watch = same skeleton).
Overfit-10-samples is the fastest wiring test in deep learning. Keep it in your ritual bank.