batch_size
32
Images processed by GPU at once. Larger = faster, but needs more VRAM. Out of Memory? Reduce this first.
num_workers
4
CPU threads reading images from disk in parallel. Too few = GPU waits idle. Think of workers as librarians fetching books simultaneously.
pinned_memory
True
Reserves fast CPU RAM for GPU transfers — like a dedicated highway lane instead of shared road. Makes PCIe transfer significantly faster.
non_blocking
True
CPU doesn't wait for GPU transfer to finish — immediately loads the next batch. GPU and CPU run simultaneously = maximum efficiency.
Forward Pass
→ ŷ
Model sees the image, runs through all neural network layers and makes a prediction: "Is this a leopard? Where is it?"
Loss Function
ℒ
Measures how wrong the prediction was. Lower loss = smarter model. Goal of training is to reduce this number every epoch.
Backpropagation
∇
Maths that traces the error backwards through every layer, calculating how much each weight contributed to the mistake (gradient).
Weight Update
w←
Optimizer (Adam/SGD) nudges each weight slightly to reduce loss. After thousands of iterations, the model learns to detect leopards accurately.
training.log · ai-server-01
[SYSTEM] AI Training Server online · NVIDIA GPU ready · CUDA initialised · Press "Run Training Epoch" to start.