M09 / ARCHITECTURES · FLAGSHIP

Build a CNN. Layer by layer.

Drop layers onto the canvas, watch tensor shapes propagate forward in real time. Activations animate down the pipeline. Compare your stack against ResNet-18, VGG-11, MobileNet, your own arbitrary mess.

EXPERIMENT 04.01 · CNN.BUILDER STATUS · FORWARD-PASS LIVE LAYERS · 8 PARAMS · 0
PRESET ARCHITECTURES / 28×28×1 → 10 0 params
LAYER PALETTE DRAG TO CANVAS
Conv2D3×3
MaxPool2×2
AvgPool2×2
BatchNorm·
ReLUact
GELUact
Dropout0.2
Flatten·
DenseFC
Residual+
Attention⊕
INSPECTOR no selection
Click a layer in the network to edit its hyperparameters.
NETWORK GRAPH · INPUT 28×28×1 (MNIST-style)
ACTIVATIONS · live feature maps after each layer layer 1 / 8
SPATIAL SIZE · per layer
CHANNEL DEPTH · per layer
The rule of thumb → spatial size shrinks, channel depth grows. Each Conv extracts more features; each Pool halves the spatial dimension. By the end you have a fat-but-tiny tensor packed with abstract semantics, ready to flatten and classify.