🪣 Gradient Accumulation Visualizer
Every weight has a bucket underneath it. Backward passes fill it. Only zero_grad() empties it.
📋 How To Use This Lab
Next Step ▶
→
runs one backward pass — a gradient droplet falls into each weight's bucket.
⚡ optimizer.step()
→
updates every weight using its bucket's current total —
the bucket stays full!
🧹 zero_grad()
→
the only action that actually empties the buckets back to zero.
↺ Reset
→
clears everything back to batch 0 with fresh starting weights.
Buckett
◀ Previous Step
Next Step ▶ (backward)
Auto Run ▶▶
↺ Reset
⚡ optimizer.step()
🧹 zero_grad()
🤖 BUCKETT
Press Next Step to run a backward pass — watch a gradient droplet land in every bucket.
🧑🏫 PROF. TORCH
Gradients accumulate by default in PyTorch. Forgetting zero_grad() is one of the most common real training bugs.