Skip to content

Latest commit

 

History

History
208 lines (151 loc) · 6.96 KB

README.md

File metadata and controls

208 lines (151 loc) · 6.96 KB
Raw

STKLIA

Trivia

The overall pipeline for training a speaker representation network has two main components, which are referred to in this repo as a generator and a classifier. The generator is the network which actually produces the embedding:

input (Acoustic features) -> generator (i.e. ResNet) -> embedding

Acting on this embedding is the classifier:

embedding -> classifier (i.e. NN projected to num_classes with Softmax) -> class prediction

In a classic scenario, this classifier is usually a feed forward network which projects to the number of classes, and trained using Cross Entropy Loss. This repo includes some alternate options such as angular penalty losses.

How to install

To install the Simple-tklia toolkit, do the following steps:

  1. We recomend using a conda venv : https://anaconda.org/anaconda/conda
  2. Install PyTorch (http://pytorch.org/).
  3. Clone the Simple-tklia repository:
git clone https://github.com/Chaanks/stklia
  1. Go into the project folder and Install the needed packages with:
pip install -r requirement.txt

Exemples

Exemples script can be found in exemples/. These exemples are ready to be run on the LIA's servers. We recommend trying them out before going further 😉.

Data preparation

To use this toolkit, please prepare your data with kaldi. When specifying a kaldi dataset folder to our toolkit, please ensure that this folder contains thes files feats.scp, utt2spk, spk2utt.
Tutorials on how to prepare some popular datasets can be found here.

How to run

The training and testing of a model is handled with the script run.py :

python run.py [-h] -m {train,test} --cfg CFG [--checkpoint CHECKPOINT]

Train a model

To train a model, simply specify the train mode and a configuration file to run.py.
Exemple :

python run.py --mode train --cfg config/example_speaker.cfg

In order to resume an experiment from an existing checkpoint interval, add the --checkpoint arguement.
Exemple:

python run.py --mode train --cfg config/example_speaker.cfg --checkpoint 1000

Test a model

To test a model, simply specify the test mode and a configuration file to run.py.
Exemple :

python run.py --mode test --cfg config/example_speaker.cfg

A checkpoint can be added to the argument to test. If no checkpoint is specified, the last iteration will be used. Exemple :

python run.py --mode test --cfg config/example_speaker.cfg --checkpoint 1250

Tensorboard

Training and validation are saved in a tensorboard in the folder stklia/runs/. To visualize the data, use the command :

tensorboard --logdir runs/

If launching tensorboard from a remote server, use the --port <your_port>and --bind_all options :

tensorboard --logdir runs/ --port <your_port> --bind_all

Note: make sure to be in the conda venv or to have tensorboard installed

Extract X-Vectors

To extract the x-vector of a dataset, use the extract.py script with the following command :

python extract.py [-h] -m MODELDIR [--checkpoint CHECKPOINT] -d DATA [-f {ark,txt,pt}]
  • --modeldir should be the path to the trained model you want to extract the xvectors with. This folder should at least contain de checkpoints folder and the experiment_settings.cfg file. The experiment_settings.cfg (or experiment_settings_resume.cfg if present) will be used to create the generator.
  • --checkpoint Is an optional parameter, it can be used to specify a checkpoint for the extraction. If not specifed the last checkpoint will be used.
  • --data can be used in 2 manners : You can specify a kaldi folder, and the folder's data will be extracted. Or, you can simply specify test/eval/train, and the corresponding dataset of experiment_settings.cfg will be extracted.
  • --format is used to specify the output format of the xvectors. It can be kaldi (ark), text (txt), or pytorch (pt). Default is kaldi.

Configuration files

An example .cfg file for speaker training is provided below and in configs/example_speaker.cfg:

Dataset specification

These are the locations of the datasets. train field is mandatory for train mode.
test and test_trial are mandatory for test mode.
eval and eval_trial are optional field. If they are not specified, no evaluation is done during training.
It is possible to specify multiple folders. If so, the folder will be merged into one dataset class containing all the data. Make sure to specify the number of features of your data with features_per_frame.

[Datasets]
train = path/to/kaldi/train/data/

eval = path/to/kaldi/eval/data/
eval_trial = path/to/trials/file1

test = path/to/kaldi/test/data/
    path/to/kaldi/enroll/data
test_trial = path/to/trials/file1
    path/to/trials/file2
    path/to/trials/file3
    
features_per_frame = 61

The format of trials is as follows:

1 <utterance_a> <utterance_b>
0 <utterance_a> <utterance_c>

Hyperparameters

Most of these configurable hyper-parameters are fairly self-explanatory.

[Hyperparams]
lr = 0.2
batch_size = 128
max_seq_len = 400
no_cuda = False
seed = 1234
num_iterations = 2000 # total num batches to train
momentum = 0.5
scheduler_steps = [1000, 1500, 1750]
scheduler_lambda = 0.5 # multiplies lr by this value at each step above
multi_gpu = False # dataparallel

Model

This section is used to specify the model size, the embeddings size, pooling mode. Pooling can be min, max, mean, std, statistical.

[Model]
emb_size = 256
layers = [3, 4, 6, 3]
num_filters = [32, 64, 128, 256]
zero_init_residual = True
pooling = statistical

Outputs

The model_dir is the folder in which models are stored. At every checkpoint_interval iterations, both the generator and classifier will be stored as a .pt model inside this folder. Each model has the form: g_<iterations>.pt, c_<iterations>.pt. This is relevant to the above section of how to resume from a previous checkpoint. For example, to resume from the 1000th iteration, both g_1000.pt, c_1000.pt must exist in checkpoints_dir.

[Outputs]
model_dir = exp/example_exp_speaker # place where models are stored
checkpoint_interval = 10 # Interval to save models and also evaluate
checkpoints_dir = checkpoints # checkpoints will be stored in <model_dir>/<checkpoints_dir>/
log_interval = 1

Slurm recommendation

With a batch size of 128, since batch size too big can lead to Cuda out of memory.

#!/bin/bash
#SBATCH --job-name="👾"
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=4
#SBATCH --partition=gpu
#SBATCH --gpus-per-node=2
#SBATCH --mem=16G
#SBATCH --time=120:00:00

source /etc/profile.d/conda.sh
source /etc/profile.d/cuda.sh

conda activate <venv name>

python run.py ...

References

https://github.com/cvqluu/dropclass_speaker
https://github.com/4uiiurz1/pytorch-adacos
https://github.com/ZhaoJ9014/face.evoLVe.PyTorch