New PyTorch Library Releases in PyTorch 1.9, including TorchVision, TorchAudio, and more

Today, we are announcing updates to a number of PyTorch libraries, alongside the PyTorch 1.9 release. The updates include new releases for the domain libraries including TorchVision, TorchText and TorchAudio. These releases, along with the PyTorch 1.9 release, include a number of new features and improvements that will provide a broad set of updates for the PyTorch community.

Some highlights include:

  • TorchVision – Added new SSD and SSDLite models, quantized kernels for object detection, GPU Jpeg decoding, and iOS support. See release notes here.
  • TorchAudio – Added wav2vec 2.0 model deployable in non-Python environments (including C++, Android, and iOS). Many performance improvements in lfilter, spectral operations, resampling. Added options for quality control in sampling (i.e. Kaiser window support). Initiated the migration of complex tensors operations. Improved autograd support. See release notes here.
  • TorchText – Added a new high-performance Vocab module that provides common functional APIs for NLP workflows. See release notes here.

We’d like to thank the community for their support and work on this latest release.

Features in PyTorch releases are classified as Stable, Beta, and Prototype. You can learn more about the definitions in this blog post.

TorchVision 0.10

(Stable) Quantized kernels for object detection

The forward pass of the nms and roi_align operators now support tensors with a quantized dtype, which can help lower the memory footprint of object detection models, particularly on mobile environments. For more details, refer to the documentation.

(Stable) Speed optimizations for Tensor transforms

The resize and flip transforms have been optimized and its runtime improved by up to 5x on the CPU.

(Stable) Documentation improvements

Significant improvements were made to the documentation. In particular, a new gallery of examples is available. These examples visually illustrate how each transform acts on an image, and also properly documents and illustrates the output of the segmentation models.

The example gallery will be extended in the future to provide more comprehensive examples and serve as a reference for common torchvision tasks. For more details, refer to the documentation.

(Beta) New models for detection

SSD and SSDlite are two popular object detection architectures that are efficient in terms of speed and provide good results for low resolution pictures. In this release, we provide implementations for the original SSD model with VGG16 backbone and for its mobile-friendly variant SSDlite with MobileNetV3-Large backbone.

The models were pre-trained on COCO train2017 and can be used as follows:

import torch
import torchvision

# Original SSD variant
x = [torch.rand(3, 300, 300), torch.rand(3, 500, 400)]
m_detector = torchvision.models.detection.ssd300_vgg16(pretrained=True)
m_detector.eval()
predictions = m_detector(x)

# Mobile-friendly SSDlite variant
x = [torch.rand(3, 320, 320), torch.rand(3, 500, 400)]
m_detector = torchvision.models.detection.ssdlite320_mobilenet_v3_large(pretrained=True)
m_detector.eval()
predictions = m_detector(x)

The following accuracies can be obtained on COCO val2017 (full results available in #3403 and #3757):

Model mAP mAP@50 mAP@75
SSD300 VGG16 25.1 41.5 26.2
SSDlite320 MobileNetV3-Large 21.3 34.3 22.1

For more details, refer to the documentation.

(Beta) JPEG decoding on the GPU

Decoding jpegs is now possible on GPUs with the use of nvjpeg, which should be readily available in your CUDA setup. The decoding time of a single image should be about 2 to 3 times faster than with libjpeg on CPU. While the resulting tensor will be stored on the GPU device, the input raw tensor still needs to reside on the host (CPU), because the first stages of the decoding process take place on the host:
from torchvision.io.image import read_file, decode_jpeg

data = read_file('path_to_image.jpg')  # raw data is on CPU
img = decode_jpeg(data, device='cuda')  # decoded image in on GPU

For more details, see the documentation.

(Beta) iOS support

TorchVision 0.10 now provides pre-compiled iOS binaries for its C++ operators, which means you can run Faster R-CNN and Mask R-CNN on iOS. An example app on how to build a program leveraging those ops can be found here.

TorchAudio 0.9.0

(Stable) Complex Tensor Migration

TorchAudio has functions that handle complex-valued tensors. These functions follow a convention to use an extra dimension to represent real and imaginary parts. In PyTorch 1.6, the native complex type was introduced. As its API is getting stable, torchaudio has started to migrate to the native complex type.

In this release, we added support for native complex tensors, and you can opt-in to use them. Using the native complex types, we have verified that affected functions continue to support autograd and TorchScript, moreover, switching to native complex types improves their performance. For more details, refer to pytorch/audio#1337.

(Stable) Filtering Improvement

In release 0.8, we added the C++ implementation of the core part of lfilter for CPU, which improved the performance. In this release, we optimized some internal operations of the CPU implementation for further performance improvement. We also added autograd support to both CPU and GPU. Now lfilter and all the biquad filters (biquad, band_biquad, bass_biquad, treble_biquad, allpass_biquad, lowpass_biquad, highpass_biquad, bandpass_biquad, equalizer_biquad and bandrefect_biquad) benefit from the performance improvement and support autograd. We also moved the implementation of overdrive to C++ for performance improvement. For more details, refer to the documentation.

(Stable) Improved Autograd Support

Along with the work of Complex Tensor Migration and Filtering Improvement, we also added autograd tests to transforms. lfilter, biquad and its variants, and most transforms are now guaranteed to support autograd. For more details, refer to the release note.

(Stable) Improved Windows Support

Torchaudio implements some operations in C++ for reasons such as performance and integration with third-party libraries. These C++ components were only available on Linux and macOS. In this release, we have added support to Windows. With this, the efficient filtering implementation mentioned above is also available on Windows.

However, please note that not all the C++ components are available for Windows. “sox_io” backend and torchaudio.functional.compute_kaldi_pitch are not supported.

(Stable) I/O Functions Migration

Since the 0.6 release, we have continuously improved I/O functionality. Specifically, in 0.8 we changed the default backend from “sox” to “sox_io” and applied the same switch to API of the “soundfile” backend. The 0.9 release concludes this migration by removing the deprecated backends. For more details, please refer to #903.

(Beta) Wav2Vec2.0 Model

We have added the model architectures from Wav2Vec2.0. You can import fine-tuned models parameters published on fairseq and Hugging Face Hub. Our model definition supports TorchScript, and it is possible to deploy the model to non-Python environments, such as C++, Android and iOS.

The following code snippet illustrates such a use case. Please check out our c++ example directory for the complete example. Currently, it is designed for running inference. If you would like more support for training, please file a feature request.

# Import fine-tuned model from Hugging Face Hub
import transformers
from torchaudio.models.wav2vec2.utils import import_huggingface_model

original = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")
imported = import_huggingface_model(original)
# Import fine-tuned model from fairseq
import fairseq
from torchaudio.models.wav2vec2.utils import import_fairseq_model

original, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
    ["wav2vec_small_960h.pt"], arg_overrides={'data': "<data_dir>"})
imported = import_fairseq_model(original[0].w2v_encoder)
# Build uninitialized model and load state dict
from torchaudio.models import wav2vec2_base

model = wav2vec2_base(num_out=32)
model.load_state_dict(imported.state_dict())

# Quantize / script / optimize for mobile
quantized_model = torch.quantization.quantize_dynamic(
    model, qconfig_spec={torch.nn.Linear}, dtype=torch.qint8)
scripted_model = torch.jit.script(quantized_model)
optimized_model = optimize_for_mobile(scripted_model)
optimized_model.save("model_for_deployment.pt")

For more details, see the documentation.

(Beta) Resampling Improvement

In release 0.8, we vectorized the operation in torchaudio.compliance.kaldi.resample_waveform, which improved the performance of resample_waveform and torchaudio.transforms.Resample. In this release, we have further revised the way the resampling algorithm is implemented.

We have:

  • Added Kaiser Window support for a wider range of resampling quality.
  • Added rolloff parameter for anti-aliasing control.
  • Added the mechanism to precompute the kernel and cache it in torchaudio.transforms.Resample for even faster operation.
  • Moved the implementation from torchaudio.compliance.kaldi.resample_waveform to torchaudio.functional.resample and deprecated torchaudio.compliance.kaldi.resample_waveform.

For more details, see the documentation.

(Prototype) RNN Transducer Loss

The RNN transducer loss is used in training RNN transducer models, which is a popular architecture for speech recognition tasks. The prototype loss in torchaudio currently supports autograd, torchscript, float16 and float32, and can also be run on both CPU and CUDA. For more details, please refer to the documentation.

TorchText 0.10.0

(Beta) New Vocab Module

In this release, we introduce a new Vocab module that replaces the current Vocab class. The new Vocab provides common functional APIs for NLP workflows. This module is backed by an efficient C++ implementation that reduces batch look-up time by up-to ~85% (refer to summary of #1248 and #1290 for further information on benchmarks), and provides support for TorchScript. We provide accompanying factory functions that can be used to build the Vocab object either through a python ordered dictionary or an Iterator that yields lists of tokens.

#creating Vocab from text file
import io
from torchtext.vocab import build_vocab_from_iterator
#generator that yield list of tokens
def yield_tokens(file_path):
   with io.open(file_path, encoding = 'utf-8') as f:
      for line in f:
          yield line.strip().split()
#get Vocab object
vocab_obj = build_vocab_from_iterator(yield_tokens(file_path), specials=["<unk>"])

#creating Vocab through ordered dict
from torchtext.vocab import vocab
from collections import Counter, OrderedDict
counter = Counter(["a", "a", "b", "b", "b"])
sorted_by_freq_tuples = sorted(counter.items(), key=lambda x: x[1], reverse=True)
ordered_dict = OrderedDict(sorted_by_freq_tuples)
vocab_obj = vocab(ordered_dict)

#common API usage

#look-up index
vocab_obj["a"]

#batch look-up indices
vocab_obj.looup_indices(["a","b"])
#support forward API of PyTorch nn Modules
vocab_obj(["a","b"])

#batch look-up tokens
vocab_obj.lookup_tokens([0,1])

#set default index to return when token not found
vocab_obj.set_default_index(0)
vocab_obj["out_of_vocabulary"] #prints 0

For more details, refer to the documentation.

Thanks for reading. If you’re interested in these updates and want to join the PyTorch community, we encourage you to join the discussion forums and open GitHub issues. To get the latest news from PyTorch, follow us on Facebook, Twitter, Medium, YouTube or LinkedIn.

Cheers!

-Team PyTorch

Read More