The most commonly used architecture for classification problems.

ResNet is the most popular architecture for classifiers with over 20,000 citations. The reason: the authors were able to build a very deep, powerful network without running into the problem of vanishing gradients.

It is also often used as backbone network for detection and segmentation models.

ResNet is often noted together with a number (e.g., ResNet18, ResNet50, ...). The number depicts the depth of the network, i.e., how many layers it contains.


One would assume that stacking additional layers to a deep neural network would improve its performance because it could learn more features.

However, the authors of the original ResNet paper showed that this is not the case, but that just adding more layers to a network will actually lead to a performance decrease at a certain amount of layers because they simply pass over the same features again and again.

Classification error with Cifar-10 data. Graphic taken from the orignal paper.

Residual Blocks

As a solution, the authors proposed residual blocks, an โ€œidentity shortcut connectionโ€ that skips one or more layers, as shown in the following figure:

Graphic from the original paper

Due to this skip connection, the output of the layer is not the same now. Without using this skip connection, the input โ€˜xโ€™ gets multiplied by the layer's weights, followed by adding a bias term. Next, this term goes through the activation function, f() and we get our output as H(x).

H(x)=f(wx+b)H(x)=f(wx + b)

Now with the introduction of the skip connection, the output is changed to

H(x)=f(wx+b)+xH(x)=f(wx + b)+x

The skip connections in ResNet solve the problem of vanishing gradient in deep neural networks by allowing this alternate shortcut path for the gradient to flow through.

Graphic from the original paper

Using ResNet has significantly enhanced the performance of neural networks with more layers, as shown above.

Code implementation

# Deep residual networks pre-trained on ImageNet
import torch
model = torch.hub.load('pytorch/vision:v0.9.0', 'resnet18', pretrained=True)
# or any of these variants
# model = torch.hub.load('pytorch/vision:v0.9.0', 'resnet34', pretrained=True)
# model = torch.hub.load('pytorch/vision:v0.9.0', 'resnet50', pretrained=True)
# model = torch.hub.load('pytorch/vision:v0.9.0', 'resnet101', pretrained=True)
# model = torch.hub.load('pytorch/vision:v0.9.0', 'resnet152', pretrained=True)
# Sample Execution
import urllib
url, filename = ("", "dog.jpg")
try: urllib.URLopener().retrieve(url, filename)
except: urllib.request.urlretrieve(url, filename)
rom PIL import Image
from torchvision import transforms
input_image =
preprocess = transforms.Compose([
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
# move the input and model to GPU for speed if available
if torch.cuda.is_available():
input_batch ='cuda')'cuda')
with torch.no_grad():
output = model(input_batch)
# Tensor of shape 1000, with confidence scores over Imagenet's 1000 classes
# The output has unnormalized scores. To get probabilities, you can run a softmax on it.
probabilities = torch.nn.functional.softmax(output[0], dim=0)
# Download ImageNet labels
# Read the categories
with open("imagenet_classes.txt", "r") as f:
categories = [s.strip() for s in f.readlines()]
# Show top categories per image
top5_prob, top5_catid = torch.topk(probabilities, 5)
for i in range(top5_prob.size(0)):
print(categories[top5_catid[i]], top5_prob[i].item())
#Model structure Top-1 error Top-5 error
#resnet18 30.24 10.92
#resnet34 26.70 8.58
#resnet50 23.85 7.13
#resnet101 22.63 6.44
#resnet152 21.69 5.94

Further Resources