inference.py for simple inference
Browse files- inference.py +91 -0
inference.py
ADDED
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torchvision.transforms as transforms
|
3 |
+
from PIL import Image
|
4 |
+
import os
|
5 |
+
|
6 |
+
# Define the device
|
7 |
+
device = (
|
8 |
+
"cuda"
|
9 |
+
if torch.cuda.is_available()
|
10 |
+
else "mps"
|
11 |
+
if torch.backends.mps.is_available()
|
12 |
+
else "cpu"
|
13 |
+
)
|
14 |
+
|
15 |
+
class Params:
|
16 |
+
def __init__(self):
|
17 |
+
self.batch_size = 512
|
18 |
+
self.name = "resnet_50"
|
19 |
+
self.workers = 16
|
20 |
+
self.lr = 0.1
|
21 |
+
self.momentum = 0.9
|
22 |
+
self.weight_decay = 1e-4
|
23 |
+
self.lr_step_size = 30
|
24 |
+
self.lr_gamma = 0.1
|
25 |
+
|
26 |
+
def __repr__(self):
|
27 |
+
return str(self.__dict__)
|
28 |
+
|
29 |
+
def __eq__(self, other):
|
30 |
+
return self.__dict__ == other.__dict__
|
31 |
+
|
32 |
+
params = Params()
|
33 |
+
|
34 |
+
# Path to the saved model checkpoint
|
35 |
+
checkpoint_path = "checkpoints/resnet_50/checkpoint.pth"
|
36 |
+
|
37 |
+
# Load the model architecture
|
38 |
+
from model import ResNet50 # Assuming resnet.py contains your model definition
|
39 |
+
|
40 |
+
num_classes = 1000 # Adjust this to match your dataset
|
41 |
+
model = ResNet50(num_classes=num_classes).to(device)
|
42 |
+
|
43 |
+
# Load the trained model weights
|
44 |
+
checkpoint = torch.load(checkpoint_path)
|
45 |
+
model.load_state_dict(checkpoint["model"])
|
46 |
+
|
47 |
+
model.eval()
|
48 |
+
|
49 |
+
# Define transformations for inference
|
50 |
+
inference_transforms = transforms.Compose([
|
51 |
+
transforms.ToTensor(),
|
52 |
+
transforms.Resize(size=256),
|
53 |
+
transforms.CenterCrop(224),
|
54 |
+
transforms.Normalize(mean=[0.485, 0.485, 0.406], std=[0.229, 0.224, 0.225]),
|
55 |
+
])
|
56 |
+
|
57 |
+
# Load class names from the text file
|
58 |
+
def load_class_names(file_path):
|
59 |
+
with open(file_path, 'r') as f:
|
60 |
+
class_names = [line.strip() for line in f]
|
61 |
+
return class_names
|
62 |
+
|
63 |
+
# Function to make predictions on a single image
|
64 |
+
def predict(image_path, model, transforms, class_names=None):
|
65 |
+
# Load and transform the image
|
66 |
+
image = Image.open(image_path).convert("RGB")
|
67 |
+
image_tensor = transforms(image).unsqueeze(0).to(device)
|
68 |
+
|
69 |
+
# Forward pass
|
70 |
+
with torch.no_grad():
|
71 |
+
output = model(image_tensor)
|
72 |
+
probabilities = torch.nn.functional.softmax(output[0], dim=0)
|
73 |
+
top_prob, top_class = probabilities.topk(5, largest=True, sorted=True)
|
74 |
+
|
75 |
+
# Display the top predictions
|
76 |
+
print("Predictions:")
|
77 |
+
for i in range(top_prob.size(0)):
|
78 |
+
class_name = class_names[top_class[i]] if class_names else f"Class {top_class[i].item()}"
|
79 |
+
print(f"{class_name}: {top_prob[i].item() * 100:.2f}%")
|
80 |
+
|
81 |
+
return top_prob, top_class
|
82 |
+
|
83 |
+
# Path to the ImageNet classes text file
|
84 |
+
imagenet_classes_file = "imagenet-classes.txt" # Replace with the actual path to your text file
|
85 |
+
class_names = load_class_names(imagenet_classes_file)
|
86 |
+
|
87 |
+
# Path to the image for inference
|
88 |
+
image_path = "dog.png" # Replace with the actual path to your test image
|
89 |
+
|
90 |
+
# Make a prediction
|
91 |
+
predict(image_path, model, inference_transforms, class_names=class_names)
|