Tutorial 06: Convolution Pipeline

Goal: build a vision-style feature pipeline with convolution, normalization, and pooling.

Step 1: Create modules

import grilly.nn as nn

conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1)
bn1 = nn.BatchNorm2d(16)
pool = nn.MaxPool2d(kernel_size=2, stride=2)
act = nn.ReLU()

Step 2: Prepare image batch

import numpy as np

x = np.random.randn(8, 3, 64, 64).astype(np.float32)

Step 3: Run feature extraction block

y = conv1(x)
y = bn1(y)
y = act(y)
y = pool(y)

print("feature shape:", y.shape)

Step 4: Attach a classifier head

head = nn.Linear(16 * 32 * 32, 10)
logits = head(y.reshape(y.shape[0], -1))
print("logits shape:", logits.shape)

Step 5: Extend to training

You can now attach this block to the training loop pattern from tutorial 03:

run forward
compute output gradient
backward through modules
optimizer step