Skip to main content

Restore the official code 100% and improve it to make it easier to research and facilitate production.

Project description

DenseNet-PyTorch

Note: Now supports the more efficient DenseNet-BC (DenseNet-Bottleneck-Compressed) networks. Using the DenseNet-BC-190-40 model, it obtaines state of the art performance on CIFAR-10 and CIFAR-100.

Update (Feb 18, 2020)

The update is for ease of use and deployment.

It is also now incredibly simple to load a pretrained model with a new number of classes for transfer learning:

from densenet_pytorch import DenseNet 
model = DenseNet.from_pretrained('densenet121', num_classes=10)

Update (January 15, 2020)

This update allows you to use NVIDIA's Apex tool for accelerated training. By default choice hybrid training precision + dynamic loss amplified version, if you need to learn more and details about apex tools, please visit https://github.com/NVIDIA/apex.

Update (January 6, 2020)

This update adds a modular neural network, making it more flexible in use. It can be deployed to many common dataset classification tasks. Of course, it can also be used in your products.

Overview

This repository contains an op-for-op PyTorch reimplementation of Densely Connected Convolutional Networks.

The goal of this implementation is to be simple, highly extensible, and easy to integrate into your own projects. This implementation is a work in progress -- new features are currently being implemented.

At the moment, you can easily:

  • Load pretrained DenseNet models
  • Use DenseNet models for classification or feature extraction

Upcoming features: In the next few days, you will be able to:

  • Quickly finetune an DenseNet on your own dataset
  • Export DenseNet models for production

Table of contents

  1. About DenseNet
  2. Installation
  3. Usage
  4. Contributing

About DenseNet

If you're new to DenseNets, here is an explanation straight from the official PyTorch implementation:

Dense Convolutional Network (DenseNet), connects each layer to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections - one between each layer and its subsequent layer - our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters.

Installation

Install from pypi:

pip install densenet_pytorch

Install from source:

git clone https://github.com/Lornatang/DenseNet-PyTorch
cd DenseNet-PyTorch
pip install -e .

Usage

Loading pretrained models

Load an densenet121 network:

from densenet_pytorch import DenseNet
model = DenseNet.from_name("densenet121")

Load a pretrained densenet11:

from densenet_pytorch import DenseNet
model = DenseNet.from_pretrained("densenet121")

Their 1-crop error rates on imagenet dataset with pretrained models are listed below.

Model structure Top-1 error Top-5 error
densenet121 25.35 7.83
densenet169 24.00 7.00
densenet201 22.80 6.43
densenet161 22.35 6.20

Example: Classification

We assume that in your current directory, there is a img.jpg file and a labels_map.txt file (ImageNet class names). These are both included in examples/simple.

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].

Here's a sample execution.

import json

import torch
import torchvision.transforms as transforms
from PIL import Image

from densenet_pytorch import DenseNet 

# Open image
input_image = Image.open("img.jpg")

# Preprocess image
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

# Load class names
labels_map = json.load(open("labels_map.txt"))
labels_map = [labels_map[str(i)] for i in range(1000)]

# Classify with DenseNet121
model = DenseNet.from_pretrained("densenet121")
model.eval()

# move the input and model to GPU for speed if available
if torch.cuda.is_available():
    input_batch = input_batch.to("cuda")
    model.to("cuda")

with torch.no_grad():
    logits = model(input_batch)
preds = torch.topk(logits, k=5).indices.squeeze(0).tolist()

print("-----")
for idx in preds:
    label = labels_map[idx]
    prob = torch.softmax(logits, dim=1)[0, idx].item()
    print(f"{label:<75} ({prob * 100:.2f}%)")

Example: Feature Extraction

You can easily extract features with model.extract_features:

import torch
from densenet_pytorch import DenseNet 
model = DenseNet.from_pretrained('densenet121')

# ... image preprocessing as in the classification example ...
inputs = torch.randn(1, 3, 224, 224)
print(inputs.shape) # torch.Size([1, 3, 224, 224])

features = model.extract_features(inputs)
print(features.shape) # torch.Size([1, 1024, 7, 7])

Example: Export to ONNX

Exporting to ONNX for deploying to production is now simple:

import torch 
from densenet_pytorch import DenseNet 

model = DenseNet.from_pretrained('densenet121')
dummy_input = torch.randn(16, 3, 224, 224)

torch.onnx.export(model, dummy_input, "demo.onnx", verbose=True)

Example: Visual

cd $REPO$/framework
sh start.sh

Then open the browser and type in the browser address http://127.0.0.1:10003/.

Enjoy it.

ImageNet

See examples/imagenet for details about evaluating on ImageNet.

For more datasets result. Please see research/README.md.

Contributing

If you find a bug, create a GitHub issue, or even better, submit a pull request. Similarly, if you have questions, simply post them as GitHub issues.

I look forward to seeing what the community does with these models!

Credit

Densely Connected Convolutional Networks

Gao Huang, Zhuang Liu, Laurens van der Maaten, Kilian Q. Weinberger

Abstract

Recent work has shown that convolutional networks can be substantially deeper, more accurate, and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper, we embrace this observation and introduce the Dense Convolutional Network (DenseNet), which connects each layer to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections

  • one between each layer and its subsequent layer - our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR-10, CIFAR-100, SVHN, and ImageNet). DenseNets obtain significant improvements over the state-of-the-art on most of them, whilst requiring less computation to achieve high performance. Code and pre-trained models are available at this https URL .

paper code

@article{DenseNet,
title:{Densely Connected Convolutional Networks},
author:{Gao Huang, Zhuang Liu, Laurens van der Maaten, Kilian Q. Weinberger},
journal={cvpr},
year={2016}
}

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

densenet_pytorch-0.2.0.tar.gz (11.5 kB view hashes)

Uploaded Source

Built Distribution

densenet_pytorch-0.2.0-py2.py3-none-any.whl (14.6 kB view hashes)

Uploaded Python 2 Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page