PyTorch-Encoding

GluonCV-Torch

Load GluonCV Models in PyTorch. Simply import gluoncvth to getting better pretrained model than torchvision:

import gluoncvth as gcv
model = gcv.models.resnet50(pretrained=True)

Installation:

pip install gluoncv-torch

Available Models

ImageNet

ImageNet models single-crop error rates, comparing to the torchvision models:

More models available at GluonCV Image Classification ModelZoo

Semantic Segmentation

Results on Pascal VOC dataset:

Results on ADE20K dataset:

Quick Demo

import torch
import gluoncvth

# Get the model
model = gluoncvth.models.get_deeplab_resnet101_ade(pretrained=True)
model.eval()

# Prepare the image
url = 'https://github.com/zhanghang1989/image-data/blob/master/encoding/' + \
    'segmentation/ade20k/ADE_val_00001142.jpg?raw=true'
filename = 'example.jpg'
img = gluoncvth.utils.load_image(
    gluoncvth.utils.download(url, filename)).unsqueeze(0)

# Make prediction
output = model.evaluate(img)
predict = torch.max(output, 1)[1].cpu().numpy() + 1

# Get color pallete for visualization
mask = gluoncvth.utils.get_mask_pallete(predict, 'ade20k')
mask.save('output.png')

More models available at GluonCV Semantic Segmentation ModelZoo

API Reference

ResNet

• gluoncvth.models.resnet18(pretrained=True)
• gluoncvth.models.resnet34(pretrained=True)
• gluoncvth.models.resnet50(pretrained=True)
• gluoncvth.models.resnet101(pretrained=True)
• gluoncvth.models.resnet152(pretrained=True)

FCN

• gluoncvth.models.get_fcn_resnet101_voc(pretrained=True)
• gluoncvth.models.get_fcn_resnet101_ade(pretrained=True)

PSPNet

• gluoncvth.models.get_psp_resnet101_voc(pretrained=True)
• gluoncvth.models.get_psp_resnet101_ade(pretrained=True)

DeepLabV3

• gluoncvth.models.get_deeplab_resnet101_voc(pretrained=True)
• gluoncvth.models.get_deeplab_resnet101_ade(pretrained=True)

Why GluonCV?

1. State-of-the-art Implementations

2. Pretrained Models and Tutorials

3. Community Support

We expect this PyTorch inference API for GluonCV models will be beneficial to the entire computer vision comunity.