CNN model
- Machine Learning course
- Advanced
- Lesson 60/78
- 7 min
Define a CNN with Conv2d and MaxPool2d.
The CNN model
A small model of two convolution blocks and a linear classifier is enough. The 16×16 image is halved at every pooling: 16 → 8 → 4:
model = nn.Sequential(
nn.Conv2d(1, 8, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2), # (8, 8, 8)
nn.Conv2d(8, 16, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2), # (16, 4, 4)
nn.Flatten(), # 256
nn.Linear(16 * 4 * 4, 2))After training, the first layer's filters start to look like horizontal and vertical edge detectors.
Printing intermediate shapes
The modules of an nn.Sequential can be walked one by one; printing the shape after each layer is a fast way to verify the architecture:
x = torch.zeros(1, 1, 16, 16)
for katman in model:
x = katman(x)
print(type(katman).__name__, tuple(x.shape))The parameter count of the weighted layers is also computed to check the total model size.
Practice Build the model
For the given model, print the shape of a (1, 1, 16, 16) input after every MaxPool2d on separate lines. Then print the model's total parameter count.
Starter code
import torch
from torch import nn
torch.manual_seed(0)
def veri(n):
X = torch.rand(n, 1, 16, 16) * 0.3
y = torch.randint(0, 2, (n,))
for k in range(n):
if y[k] == 0:
X[k, 0, 7:9, :] += 1.0
else:
X[k, 0, :, 7:9] += 1.0
return X, y
X_train, y_train = veri(256)
X_test, y_test = veri(64)
model = nn.Sequential(
nn.Conv2d(1, 8, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),
nn.Conv2d(8, 16, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),
nn.Flatten(), nn.Linear(16 * 4 * 4, 2))
# buraya yazSample solution
import torch
from torch import nn
torch.manual_seed(0)
def veri(n):
X = torch.rand(n, 1, 16, 16) * 0.3
y = torch.randint(0, 2, (n,))
for k in range(n):
if y[k] == 0:
X[k, 0, 7:9, :] += 1.0
else:
X[k, 0, :, 7:9] += 1.0
return X, y
X_train, y_train = veri(256)
X_test, y_test = veri(64)
model = nn.Sequential(
nn.Conv2d(1, 8, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),
nn.Conv2d(8, 16, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),
nn.Flatten(), nn.Linear(16 * 4 * 4, 2))
x = torch.zeros(1, 1, 16, 16)
for katman in model:
x = katman(x)
if isinstance(katman, nn.MaxPool2d):
print(x.shape)
print(sum(p.numel() for p in model.parameters()))Expected output
torch.Size([1, 8, 8, 8]) torch.Size([1, 16, 4, 4]) 1762
Questions and answers
Review the lesson questions together with their answers. To test yourself, solve the lesson interactively.
1 How many features are there after Flatten in this model?
- 1024
- 64
- 256 Correct answer
- 16
Explanation: 16 channels × 4 × 4 = 256.
2 Print the layer's name.
print(type(katman).____, tuple(x.shape))print(type(katman).__name__, tuple(x.shape))Explanation: __name__ gives the class's name.
3 Match the shapes (input 1×16×16).
| Term | Meaning |
|---|---|
Conv2d(1, 8) + Pool |
(8, 8, 8) |
Conv2d(8, 16) + Pool |
(16, 4, 4) |
Flatten |
256 |
Linear |
2 |