WarpRNNT loss ported in Numba CPU/CUDA for Pytorch

Last update: Oct 22, 2022

Related tags

Overview

RNNT loss in Pytorch - Numba JIT compiled (warprnnt_numba)

Warp RNN Transducer Loss for ASR in Pytorch, ported from HawkAaron/warp-transducer and a replica of the stable version in NVIDIA Neural Module repository (NVIDIA NeMo).

NOTE: The code here will have experimental extensions and may be potentially unstable, use the version in NeMo for long term supported loss version of RNNT for PyTorch.

Supported Features

Currently supports :

WarpRNNT loss in pytorch for CPU / CUDA (jit compiled)
FastEmit
Gradient Clipping (from Torch Audio)

Installation

You will need PyTorch (usually the latest version should be used), plus installation of Numba in a Conda environment (pip only environment is untested but may work).

# Follow installation instructions to install pytorch from website (with cuda if required)
conda install -c conda-force numba or conda update -c conda-forge numba (to get latest version)

# Then install this library
pip install --upgrade git+https://github.com/titu1994/warprnnt_numba.git

Usage

Import warprnnt_numba and use RNNTLossNumba. If attempting to use CUDA version of loss, it is advisable to test that your installed CUDA version is compatible with numba version using numba_utils.

There is also included a very slow numpy/pytorch explicit-loop based loss implementation for verification of exact correct results.

import torch
import numpy as np
import warprnnt_numba

# Define the loss function
fastemit_lambda = 0.001  # any float >= 0.0
loss_pt = warprnnt_numba.RNNTLossNumba(blank=4, reduction='sum', fastemit_lambda=fastemit_lambda)

# --------------
# Example usage

device = "cuda"
torch.random.manual_seed(0)

# Assume Batchsize=2, Acoustic Timesteps = 8, Label Timesteps = 5 (including BLANK=BOS token),
# and Vocabulary size of 5 tokens (including RNNT BLANK)
acts = torch.randn(2, 8, 5, 5, device=device, requires_grad=True)
sequence_length = torch.tensor([5, 8], dtype=torch.int32,
                               device=device)  # acoustic sequence length. One element must be == acts.shape[1].

# Let 0 be MASK/PAD value, 1-3 be token ids, and 4 represent RNNT BLANK token
# The BLANK token is overloaded for BOS token as well here, but can be different token.
# Let first sample be padded with 0 (actual length = 3). Loss is computed according to supplied `label_lengths`.
# and gradients for the 4th index onwards (0 based indexing).
labels = torch.tensor([[4, 1, 1, 3, 0], [4, 2, 2, 3, 1]], dtype=torch.int32, device=device)
label_lengths = torch.tensor([3, 4], dtype=torch.int32,
                             device=device)  # Lengths here must be WITHOUT the BOS token.

# If on CUDA, log_softmax is computed internally efficiently (preserving memory and speed)
# Compute it explicitly for CPU, this is done automatically for you inside forward() of the loss.
# -1-th vocab index is RNNT blank token here.
loss_func = warprnnt_numba.RNNTLossNumba(blank=4, reduction='none',
                                         fastemit_lambda=0.0, clamp=0.0)
loss = loss_func(acts, labels, sequence_length, label_lengths)
print("Loss :", loss)
loss.sum().backward()

# When parsing the gradients, look at grads[0] -
# Since it was padded in T (sequence_length=5 < T=8), there are gradients only for grads[0, :5, :, :].
# Since it was padded in U (label_lengths=3+1 < U=5), there are gradeints only for grads[0, :5, :3+1, :].
grads = acts.grad
print("Gradients of activations :")
print(grads)

Tests

Tests will perform CPU only checks if there are no GPUs. If GPUs are present, will run all tests once for cuda:0 as well.

pytest tests/

Requirements

pytorch >= 1.10. Older versions might work, not tested.
numba - Minimum required version is 0.53.0, preferred is 0.54+.

Comments

GPU under utilization due to low occupancy.

Thank you for the warprnnt_numba, I got the warnning (show blow) when I use this loss in my code. Is this known issue? How can it be debugged and solved?

Thank you!

opened by jiay7 2
Fix runtime speed
Improve runtime speed of numba loss

Fix issue with data movement of costs tensor from llForward to pytorch data view in numba

This alone costs a linear loop (scaling with batch size) that is roughly 10x the kernel costs themselves.

Fix by writing a small kernel to copy the data and update the costs.
opened by titu1994 0

Releases(v0.4.0)

v0.4.0(Jan 30, 2022)
Supports

Simple RNNT loss with Atomic Locks implementation

Improvements

Improve runtime speed of numba loss

Fix issue with data movement of costs tensor from llForward to pytorch data view in numba

This alone costs a linear loop (scaling with batch size) that is roughly 10x the kernel costs themselves.

Fix by writing a small kernel to copy the data and update the costs.

Source code(tar.gz)
Source code(zip)
v0.2.2(Jan 24, 2022)
Initial release of Warp RNNT loss with Numba JIT compile (CPU/CUDA)

Supports:

Pytorch RNNT loss (CPU and JIT compiled CUDA)

FastEmit

Gradient clipping

Source code(tar.gz)
Source code(zip)

WarpRNNT loss ported in Numba CPU/CUDA for Pytorch

Related tags

Overview

RNNT loss in Pytorch - Numba JIT compiled (warprnnt_numba)

Supported Features

Installation

Usage

Tests

Requirements

You might also like...

This Repo is the official CUDA implementation of ICCV 2019 Oral paper for CARAFE: Content-Aware ReAssembly of FEatures

Example repository for custom C++/CUDA operators for TorchScript

Convert Python 3 code to CUDA code.

This demo showcase the use of onnxruntime-rs with a GPU on CUDA 11 to run Bert in a data pipeline with Rust.

LightSeq is a high performance training and inference library for sequence processing and generation implemented in CUDA

CUDA Python Low-level Bindings

A dead simple python wrapper for darknet that works with OpenCV 4.1, CUDA 10.1

Prevent `CUDA error: out of memory` in just 1 line of code.

An addernet CUDA version

Comments

GPU under utilization due to low occupancy.

Fix runtime speed

Improve runtime speed of numba loss

Releases(v0.4.0)

v0.4.0(Jan 30, 2022)

Supports

Improvements

v0.2.2(Jan 24, 2022)

Owner

Somshubra Majumdar

Source code for our Paper "Learning in High-Dimensional Feature Spaces Using ANOVA-Based Matrix-Vector Multiplication"

PyTorch implementation DRO: Deep Recurrent Optimizer for Structure-from-Motion

This project is used for the paper Differentiable Programming of Isometric Tensor Network

Gym environments used in the paper: "Developmental Reinforcement Learning of Control Policy of a Quadcopter UAV with Thrust Vectoring Rotors"

Code for the ICML 2021 paper: "ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision"

NeuroGen: activation optimized image synthesis for discovery neuroscience

Implementation of a protein autoregressive language model, but with autoregressive infilling objective (editing subsequences capability)

This is a collection of our NAS and Vision Transformer work.

RLHive: a framework designed to facilitate research in reinforcement learning.

A self-supervised 3D representation learning framework named viewpoint bottleneck.

Robotics environments

Towards Interpretable Deep Metric Learning with Structural Matching

Tensorflow implementation of Character-Aware Neural Language Models.

Memory efficient transducer loss computation

Direct LiDAR Odometry: Fast Localization with Dense Point Clouds

RAFT-Stereo: Multilevel Recurrent Field Transforms for Stereo Matching

Pytorch implementation of paper: "NeurMiPs: Neural Mixture of Planar Experts for View Synthesis"

Weakly-Supervised Semantic Segmentation Network with Deep Seeded Region Growing (CVPR 2018).

TSP: Temporally-Sensitive Pretraining of Video Encoders for Localization Tasks

GPOEO is a micro-intrusive GPU online energy optimization framework for iterative applications