harry
commited on
Commit
•
0dee387
1
Parent(s):
aaea685
feat: add prediction script and model file for MNIST digit classification
Browse files- models/mnist_model_lr0.001_bs64_ep20.pth +3 -0
- predict.py +58 -0
- test/image.jpg +0 -0
- torchvision.pyi +4 -1
models/mnist_model_lr0.001_bs64_ep20.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:aead1b3223333f05acf8494c6a73aec8bdaa9e32d3f0c239b16e5e12a3a07a8f
|
3 |
+
size 4803144
|
predict.py
ADDED
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
# mypy: ignore-errors
|
3 |
+
import torchvision.transforms as transforms
|
4 |
+
from PIL import Image
|
5 |
+
from mnist_classifier.model import MNISTModel
|
6 |
+
import torch.nn.functional as F
|
7 |
+
|
8 |
+
def load_model(model_path):
|
9 |
+
"""Load the trained model."""
|
10 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
11 |
+
model = MNISTModel().to(device)
|
12 |
+
model.load_state_dict(torch.load(model_path, weights_only=True))
|
13 |
+
model.eval()
|
14 |
+
return model, device
|
15 |
+
|
16 |
+
def preprocess_image(image_path):
|
17 |
+
"""Preprocess the input image."""
|
18 |
+
transform = transforms.Compose([
|
19 |
+
transforms.Grayscale(num_output_channels=1),
|
20 |
+
transforms.Resize((28, 28)),
|
21 |
+
transforms.ToTensor(), # This converts PIL Image to tensor
|
22 |
+
transforms.Normalize((0.5,), (0.5,))
|
23 |
+
])
|
24 |
+
|
25 |
+
image = Image.open(image_path)
|
26 |
+
image_tensor = transform(image) # Now image_tensor is already a tensor
|
27 |
+
return image_tensor.unsqueeze(0) # type: ignore # Add batch dimension using tensor method
|
28 |
+
|
29 |
+
def predict(model, image, device):
|
30 |
+
"""Make prediction on the input image."""
|
31 |
+
with torch.no_grad():
|
32 |
+
image = image.to(device)
|
33 |
+
output = model(image)
|
34 |
+
probabilities = F.softmax(output, dim=1)
|
35 |
+
pred = output.argmax(dim=1, keepdim=True)
|
36 |
+
return pred.item(), probabilities[0]
|
37 |
+
|
38 |
+
def main():
|
39 |
+
# Path to your trained model
|
40 |
+
model_path = "./models/mnist_model_lr0.001_bs64_ep10.pth"
|
41 |
+
|
42 |
+
# Load model
|
43 |
+
model, device = load_model(model_path)
|
44 |
+
|
45 |
+
# Path to input image
|
46 |
+
image_path = "./test/image.jpg"
|
47 |
+
|
48 |
+
# Preprocess image and get prediction
|
49 |
+
image = preprocess_image(image_path)
|
50 |
+
prediction, probabilities = predict(model, image, device)
|
51 |
+
|
52 |
+
print(f"Predicted digit: {prediction}")
|
53 |
+
print("\nProbabilities for each digit:")
|
54 |
+
for digit, prob in enumerate(probabilities):
|
55 |
+
print(f"{digit}: {prob.item():.4f}")
|
56 |
+
|
57 |
+
if __name__ == "__main__":
|
58 |
+
main()
|
test/image.jpg
ADDED
torchvision.pyi
CHANGED
@@ -6,4 +6,7 @@ class datasets:
|
|
6 |
class transforms:
|
7 |
Compose: Any
|
8 |
ToTensor: Any
|
9 |
-
Normalize: Any
|
|
|
|
|
|
|
|
6 |
class transforms:
|
7 |
Compose: Any
|
8 |
ToTensor: Any
|
9 |
+
Normalize: Any
|
10 |
+
Grayscale: Any
|
11 |
+
Resize: Any
|
12 |
+
|