swiss-landmarks-cnn

Custom CNN for classifying 6 Swiss landmarks. Trained as part of an AI Applications course (HS24).

App: https://huggingface.co/spaces/eceleo/swiss-landmarks

Classes

Key	Label
matterhorn	Matterhorn
chapel_bridge	Chapel Bridge, Lucerne
zurich_oldtown	Zurich Old Town
rhine_falls	Rhine Falls
bern_oldtown	Bern Old Town
lake_geneva	Lake Geneva

Architecture

Custom LandmarkNet — MobileNet-style depthwise separable CNN, ~135k parameters, input 96×96 px.

Training

• Dataset: 360 images (1 real photo per class, augmented ×60)
• Optimizer: AdamW, lr=2e-3
• Epochs: 25
• Val accuracy: 100%

Usage

import torch, json, io
from torch import nn
from torchvision import transforms
from PIL import Image

classes = json.load(open("classes.json"))

class DWConv(nn.Module):
    def __init__(self, ci, co, s=1):
        super().__init__()
        self.net = nn.Sequential(
            nn.Conv2d(ci,ci,3,stride=s,padding=1,groups=ci,bias=False),
            nn.Conv2d(ci,co,1,bias=False), nn.BatchNorm2d(co), nn.ReLU6())
    def forward(self, x): return self.net(x)

class LandmarkNet(nn.Module):
    def __init__(self, n):
        super().__init__()
        self.net = nn.Sequential(
            nn.Conv2d(3,32,3,stride=2,padding=1,bias=False), nn.BatchNorm2d(32), nn.ReLU6(),
            DWConv(32,64), DWConv(64,128,2), DWConv(128,128),
            DWConv(128,256,2), DWConv(256,256),
            nn.AdaptiveAvgPool2d(1), nn.Flatten(),
            nn.Dropout(0.35), nn.Linear(256, n))
    def forward(self, x): return self.net(x)

model = LandmarkNet(len(classes))
model.load_state_dict(torch.load("model.pth", map_location="cpu", weights_only=False))
model.eval()

tf = transforms.Compose([
    transforms.Resize((96, 96)),
    transforms.ToTensor(),
    transforms.Normalize([0.5]*3, [0.5]*3),
])

img = Image.open("your_image.jpg").convert("RGB")
with torch.no_grad():
    probs = torch.softmax(model(tf(img).unsqueeze(0))[0], 0)
print(classes[probs.argmax()], f"{probs.max()*100:.1f}%")