Deep Residual Network (ResNet)

In this section, we are going to use the pre-trained resnet18 on ImageNet from PyTorch to classify the input image.

Import torch and models from TorchVision.

import torch
import torchvision.models as models

Load the pre-trained resnet18 model from TorchVision.

model = models.resnet18(pretrained=True)
# other resnet model variants from torchvision 
# resnet18 means 18 layers
# models.resnet34(pretrained=True)
# models.resnet50(pretrained=True)
# models.resnet101(pretrained=True)
# models.resnet152(pretrained=True)
print("Model loaded!")

Evaluate the pre-trained model.

model.eval()

All pre-trained models expect input images normalized in the same way, i.e. mini-batches of 3-channel RGB images of shape (3 x H x W) where H and W are expected to be at least 224. The images have to be loaded in to a range of [0, 1] and then normalized using mean =[0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225].

Optional: For this tutorial, we can use an example image from Pytorch or our own image.

To download the image from PyTorch, run the below section of code.

import urllib
url, filename = ("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "data/dog.jpg")
try: urllib.URLopener().retrieve(url, filename)
except: urllib.request.urlretrieve(url, filename)

Load the dog.jpg or your image from data directory and preprocess the data.

filename = "dog.jpg" # To try with other image, change this filename

# sample execution (requires torchvision)
from PIL import Image
from torchvision import transforms
input_image = Image.open(filename) # stored dog.jpg into input_image
# preprocess the image and transform it to tensor
preprocess = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
input_tensor = preprocess(input_image)
input_batch = input_tensor.unsqueeze(0) # create a mini-batch as expected by the model

If your machine is CUDA capable, it will transfer the input and model to your GPU.

if torch.cuda.is_available():
    input_batch = input_batch.to('cuda')
    model.to('cuda')

Pass input_batch into our model.

with torch.no_grad():
    output = model(input_batch)

Print the output.

# Tensor of shape 1000, with confidence scores over Imagenet's 1000 classes
print(output[0])
# The output has unnormalized scores. To get probabilities, you can run a softmax on it.
probabilities = torch.nn.functional.softmax(output[0], dim=0)
print(probabilities)

There are 1000 image classes in ImageNet. The probabilities of those classes are represented by each element of the output tensor (probabilities). We can see how accurate the output is by downloading ImageNet's label and comparing it to our tensor output.

Optional: If you don’t have the ImageNet label on your machine, you can download it from this link.

wget is a command for retrieving content and files from various web servers.

wget https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt

Read the categories from ImageNet label file and stores it as list.

# Read the categories
with open("imagenet_classes.txt", "r") as f:
    categories = [s.strip() for s in f.readlines()]

Show the top categories of our input image.

# Show top categories per image
top5_prob, top5_catid = torch.topk(probabilities, 5)
for i in range(top5_prob.size(0)):
        # the output is in 'class' 'probabilities' format 
        # 'probabilites' is floating point number between 0 and 1. 
    print(categories[top5_catid[i]], top5_prob[i].item())

The output

Samoyed 0.8847386837005615
Arctic fox 0.045728735625743866
white wolf 0.04426601156592369
Pomeranian 0.005614196881651878
Great Pyrenees 0.00464773690328002

This mean our dog.jpg is likely to be Samoyed!

Acknowledgement: The content of this document has been adapted from the original PyTorch ResNet.

Deep Residual Network (ResNet)

The output​

The output