This release introduces improved support for mobile, with new mobile-friendly models, pre-compiled binaries for Android available in maven and an android demo app. It also improves image IO and provides new data augmentations including AutoAugment.

Highlights

Better mobile support

torchvision 0.9 adds support for the MobileNetV3 architecture with pre-trained weights for Classification, Object Detection and Segmentation tasks.
It also improves C++ operators so that they can be compiled and run on Android, and we are providing pre-compiled torchvision artifacts published to jcenter. An example application on how to use the torchvision ops on an Android app can be found in here.

Classification

We provide MobileNetV3 variants (including a quantized version) pre-trained on ImageNet 2012.

import torch
import torchvision

# Classification
x = torch.rand(1, 3, 224, 224)
m_classifier = torchvision.models.mobilenet_v3_large(pretrained=True)
# m_classifier = torchvision.models.mobilenet_v3_small(pretrained=True)
m_classifier.eval()
predictions = m_classifier(x)

# Quantized Classification
x = torch.rand(1, 3, 224, 224)
m_classifier = torchvision.models.quantization.mobilenet_v3_large(pretrained=True)
m_classifier.eval()
predictions = m_classifier(x)

The pre-trained models have the following accuracies on ImageNet 2012 val:

Model	Top-1 Acc	Top-5 Acc
MobileNetV3 Large	74.042	91.340
MobileNetV3 Large (Quantized)	73.004	90.858
MobileNetV3 Small	67.620	87.404

Object Detection

We provide two variants of Faster R-CNN with MobileNetV3 backbone pre-trained on COCO train2017. They can be obtained as follows

import torch
import torchvision

# Fast Low Resolution Model
x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
m_detector = torchvision.models.detection.fasterrcnn_mobilenet_v3_large_320_fpn(pretrained=True)
m_detector.eval()
predictions = m_detector(x)

# Highly Accurate High Resolution Model
x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
m_detector = torchvision.models.detection.fasterrcnn_mobilenet_v3_large_fpn(pretrained=True)
m_detector.eval()
predictions = m_detector(x)

And yield the following accuracies on COCO val 2017 (full results available in #3265):

Model	mAP	mAP@50	mAP@75
Faster R-CNN MobileNetV3-Large 320 FPN	22.8	38.0	23.2
Faster R-CNN MobileNetV3-Large FPN	32.8	52.5	34.3

Semantic Segmentation

We also provide pre-trained models for semantic segmentation. The models have been trained on a subset of COCO train2017, which contains the same 20 categories as those from Pascal VOC.

import torch
import torchvision

# Fast Mobile Model
x = torch.rand(1, 3, 520, 520)
m_segmenter = torchvision.models.segmentation.lraspp_mobilenet_v3_large(pretrained=True)
m_segmenter.eval()
predictions = m_segmenter(x)

# Highly Accurate Mobile Model
x = torch.rand(1, 3, 520, 520)
m_segmenter = torchvision.models.segmentation.deeplabv3_mobilenet_v3_large(pretrained=True)
m_segmenter.eval()
predictions = m_segmenter(x)

The pre-trained models give the following results on the subset of COCO val2017 which contain the same 20 categories as those present in Pascal VOC (full results in #3276):

Model	mean IoU	global pixelwise accuracy
Lite R-ASPP with Dilated MobileNetV3 Large Backbone	57.9	91.2
DeepLabV3 with Dilated MobileNetV3 Large Backbone	60.3	91.2

Addition of the AutoAugment method

AutoAugment is a common Data Augmentation technique that can improve the accuracy of Scene Classification models. Though the data augmentation policies are directly linked to their trained dataset, empirical studies show that ImageNet policies provide significant improvements when applied to other datasets.

In TorchVision we implemented 3 policies learned on the following datasets: ImageNet, CIFA10 and SVHN. The new transform can be used standalone or mixed-and-matched with existing transforms:

from torchvision import transforms

t = transforms.AutoAugment()
transformed = t(image)

transform=transforms.Compose([
    transforms.Resize(256),
    transforms.AutoAugment(),
    transforms.ToTensor()])

Improved Image IO and on-the-fly image type conversions

All the read and decode methods of the io.image package have been updated to:

Add support for Palette, Grayscale Alpha and RBG Alpha image types during PNG decoding.
Allow the on-the-fly conversion of image from one type to the other during read.

from torchvision.io.image import read_image, ImageReadMode

# keeps original type, channels unchanged
x1 = read_image("image.png")

# converts to grayscale, channels = 1
x2 = read_image("image.png", mode=ImageReadMode.GRAY)

# converts to grayscale with alpha transparency, channels = 2
x3 = read_image("image.png", mode=ImageReadMode.GRAY_ALPHA)

# coverts to RGB, channels = 3
x4 = read_image("image.png", mode=ImageReadMode.RGB)

# converts to RGB with alpha transparency, channels = 4
x5 = read_image("image.png", mode=ImageReadMode.RGB_ALPHA)

Python 3.9 and CUDA 11.1

This release adds official support for Python 3.9 and CUDA 11.1 (#3341, #3418)

Backwards Incompatible Changes

[Ops] Change default eps value of FrozenBN to better align with nn.BatchNorm (#2933)
[Ops] Remove deprecated _new_empty_tensor. (#3156)
[Transforms] ColorJitter gets its random params by calling get_params() (#3001)
[Transforms] Change rounding of transforms on integer tensors (#2964)
[Utils] Remove normalize from save_image (#3324)

New Features

[Datasets] Add WiderFace dataset (#2883)
[Models] Add MobileNetV3 architecture:
- Classification Models: (#3354, #3252, #3182, #3242, #3177)
- Object Detection Models: (#3265, #3253, #3223, #3243, #3244, #3248)
- Segmentation Models: (#3276)
- Quantized Models: (#3366, #3323)
[Models] Improve speed/accuracy of FasterRCNN by introducing a score threshold on RPN (#3205)
[Mobile] Add Android gradle project with demo test app (#2897)
[Transforms] Implemented AutoAugment, along with required new transforms + Policies (#3123)
[Ops] Added support of Autocast in all Operators: #2938, #2926, #2922, #2928, #2905, #2906, #2907, #2898
[Ops] Add modulation input for DeformConv2D (#2791)
[IO] Improved io.image with on-the-fly image type conversions: (#3193, #3069, #3024, #2988, #2984)
[IO] Add option to write audio to video file (#2304)
[Utils] Added a utility to draw bounding boxes (#2785, #3296, #3075)

Improvements

Datasets

Concatenate small tensors in video datasets to reduce the use of shared file descriptor (#1795)
Improve testing for datasets (#3336, #3337, #3402, #3412, #3413, #3415, #3416, #3345, #3376, #3346, #3338)
Check if dataset file is located on Google Drive before downloading it (#3245)
Improve Coco implementation (#3417)
Make download_url follow redirects (#3236)
make_dataset as staticmethod of DatasetFolder (#3215)
Add a warning if any clip can't be obtained from a video in VideoClips. (#2513)

Models

Improve error message in AnchorGenerator (#2960)
Disable pretrained backbone downloading if pretrained is True in segmentation models (#3325)
Support for image with no annotations in RetinaNet (#3032)
Change RoIHeads reshape to support empty batches. (#3031)
Fixed typing exception throwing issues with JIT (#3029)
Replace deprecated functional.sigmoid with torch.sigmoid in RetinaNet (#3307)
Assert that inputs are floating point in Faster R-CNN normalize method (#3266)
Speedup RetinaNet's postprocessing (#2828)

Ops

Added eps in the __repr__ of FrozenBN (#2852)
Added __repr__ to MultiScaleRoIAlign (#2840)
Exposing LevelMapper params in MultiScaleRoIAlign (#3151)
Enable autocast for all operators and let them use the dispatcher (#2926, #2922, #2928, #2898)

Transforms

adjust_hue now accepts tensors with one channel (#3222)
Add fill color support for tensor affine transforms (#2904)
Remove torchscript workaround for center_crop (#3118)
Improved error message for RandomCrop (#2816)

IO

Enabling to import read_file and the other methods from torchvision.io (#2918)
accept python bytes in _read_video_from_memory() (#3347)
Enable rtmp timeout in decoder (#3076)
Specify tls cert file to decoder through config (#3289, #3374)
Add UUID in LOG() in decoder (#3080)

References

Add weight averaging and storing methods in references utils (#3352)
Adding Preset Transforms in reference scripts (#3317)
Load variables when --resume /path/to/checkpoint --test-only (#3285)
Updated video classification ref example with new transforms (#2935)

Misc

Various documentation improvements (#3039, #3271, #2820, #2808, #3131, #3062, #3061, #3000, #3299, #3400, #2899, #2901, #2908, #2851, #2909, #3005, #2821, #2957, #3360, #3019, #3124, #3217, #2879, #3234, #3180, #3425, #2979, #2935, #3298, #3268, #3203, #3290, #3295, #3200, #2663, #3153, #3147, #3232)
The documentation infrastructure was improved, in particular the docs are now built on every PR and uploaded to CircleCI (#3259, #3378, #3408, #3373, #3290)
Avoid some deprecation warnings from PyTorch (#3348)
Ensure operators are added in C++ (#2798, #3091, #3391)
Fixed compilation warnings on C++ codebase (#3390)
CI Improvements (#3401, #3329, #2990, #2978, #3189, #3230, #3254, #2844, #2872, #2825, #3144, #3137, #2827, #2848, #2914, #3419, #2895, #2837)
Installation improvements (#3302, #2969, #3113, #3202)
CMake improvements (#2801, #2805, #3212, #3381)

Mobile

Add Torch Selective macros in all C++ Ops for better support on mobile (#3218)

Code Quality, testing

[BC-breaking] Modernized C++ codebase & made it mobile-friendly (25% faster to compile): #2885, #2891, #2892, #2893, #2905, #2906, #2907, #2938, #2944, #2945, #3011, #3020, #3097, #3105, #3134, #3135, #3143, #3146, #3154, #3156, #3163, #3218, #3308, #3311, #3312, #3326, #3350, #3390
Cleaned up Python codebase & made it more Pythonic: #3263, #3239, #3059, #3055, #3045, #3382, #3159, #3171
Improve type annotations (#3288, #3045, #2862, #2858, #2857, #2863, #2865, #2856, #2860, #2864, #2875, #2859, #2854, #2861, #3174, #3059)
Code refactoring and static analysis improvements (#3379, #3335, #3229, #3204, #3095)
Miscellaneous test improvements (#2966, #2965, #3018, #3035, #2961, #2806, #2812, #2815, #2834, #2874, #3099, #3092, #3160, #3103, #2971, #3023, #2803, #3136, #3319, #3310, #3287, #3033, #2983, #3386, #3369, #3116, #2985, #3320)

Bug Fixes

[DATASETS] Fixes EMNIST split and label issues (#2673)
[DATASETS] Fix overflow in STL10 fold reading (#3353)
[MODELS] Fix incorrectly frozen BN on ResNet FPN backbone (#3396)
[MODELS] Fix scriptability support in Inception V3 (#2976)
[MODELS] Changed default value of eps in FrozenBatchNorm to match BatchNorm: #2940 #2933
[MODELS] Fixed warning in models.detection.transforms.resize_image_and_masks. (#3237)
[MODELS] Fix trainable_layers on RetinaNet (#3234)
[MODELS] Fix ShuffleNetV2 ONNX model export issue. (#3158)
[UTILS] Fixes no grad and range bugs in utils. (#3269)
[UTILS] make_grid uses a more correct normalization (#2967)
[OPS] fix GET_THREADS() for ROCm with DeformConv (#2997)
[OPS] Fix NMS and IoU overflows for fp16 (#3383, #3382)
[OPS] Fix ops registration on windows (#3380)
[OPS] Fix initialisation bug on FeaturePyramidNetwork (#2954)
[IO] Replace hardcoded error code with ENODATA (#3277)
[REFERENCES] Fix repeated UserWarning and add more flexibility to reference code for segmentation tasks (#2886)
[TRANSFORMS] Fix default fill value in RandomRotation (#3303)
[TRANSFORMS] Correct aspect ratio sampling in transforms.RandomErasing (#3344)
[TRANSFORMS] Fix CenterCrop for Tensor size is greater than imgsize (#3333)
[TRANSFORMS] Functional to_tensor returns float tensor of default dtype (#3398)
[TRANSFORMS] Add explicit check for number of channels (#3013)
[TRANSFORMS] pil_to_tensor with accimage backend now return uint8 (#3109)
[TRANSFORMS] Fix potential overflow in convert_image_dtype (#3107)
[TRANSFORMS] Check num of channels on adjust*_ transformations (#3069)

Deprecations

[TRANSFORMS] Introduced InterpolationModes and deprecated arguments: resample and fillcolor (#2952, #3055)

torchvision 0.9.0 Mobile support, AutoAugment, improved IO and more on Python PyPI