Skip to content

Latest commit

 

History

History
151 lines (120 loc) · 7.4 KB

LOWBUDGET.md

File metadata and controls

151 lines (120 loc) · 7.4 KB
Raw

Description of the baseline

Similarly to the baselines for the ZeroSpeech 2021 challenge - audio-only track [1], the low-budget baseline is mainly composed of two main elements:

  • A visually-grounded (VG) model which replaces the CPC-based acoustic model of the audio-only baselines. This model is similar to the speech-image model described in [2-3].
  • A language model (LM) which is trained on activations extracted from the VG model and quantized through K-means clustering. We re-use here BERT-small which was introduced in the audio-only track.

Pretrained models

Pretrained models can be downloaded from the ZeroSpeech challenge website. Simply unzip the archive and copy the exps folder under the repository root directory.

Training procedure

If alternatively you want to retrain the baseline from scratch, you will need to go through following steps. We assume that you already have the datasets stored under ~/corpora (otherwise, you will first need to follow the instructions provided here.

Training the VG model.

To train the visually-grounded model, you will first need to preprocess the dataset to extract visual and audio features. This can conveniantly be done by running:

python -m platalea.utils.preprocessing spokencoco --spokencoco_root ~/corpora/spokencoco

The model can then be trained with:

mkdir -p exps/vg/vg-spokencoco
cd exps/vg/vg-spokencoco
python -m scripts.train_vg spokencoco --epochs 12 --spokencoco_root ~/corpora/spokencoco --image_features_fn resnet_features.memmap --audio_features_fn mfcc_features.memmap
python -m platalea.utils.copy_best
cd ../../..

After running these commands, the folder exps/vg/vg-spokencoco should contain, among other files:

  • result.json, which contains the performance of the model after each epoch.
  • net.best.pt, corresponding to the best checkpoint of the model (according to the R@10 metric).

Training K-means clustering

We train K-means clustering on LibriSpeech-100, using the activations of the first recurrent layer of the VG model. We use 50 clusters.

Extracting activations

To extract the relevant activations run (the GRU layers are named rnn0 to rnn3):

python -m scripts.extract_activations exps/vg/vg-spokencoco/net.best.pt \
    ~/corpora/LibriSpeech/train-clean-100 \
    data/activations/vg-spokencoco/librispeech/train-clean-100 \
    --batch_size 8 --layer rnn0 --output_file_extension '.pt' \
    --file_extension '.flac' --recursionLevel 2

Training the model

We can now train K-means clustering:

python -m scripts.clustering data/activations/vg-spokencoco/librispeech/train-clean-100 \
    exps/kmeans/vg-spokencoco-rnn0_kmeans-librispeech100-50 \
    --nClusters 50 --MAX_ITER 150 --batchSizeGPU 500 --recursionLevel 2 --save

Training the language model

We finally need to train the BERT-small language model on LibriSpeech 960. To do that, we need to extract activations for LibriSpeech and quantize them using K-means.

Extracting activations

python -m scripts.extract_activations exps/vg/vg-spokencoco/net.best.pt \
    ~/corpora/LibriSpeech/train-full-960 \
    data/activations/vg-spokencoco/librispeech/train-full-960 \
    --batch_size 8 --layer rnn0 --output_file_extension '.pt' \
    --file_extension '.flac' --recursionLevel 2
python -m scripts.extract_activations exps/vg/vg-spokencoco/net.best.pt \
    ~/corpora/LibriSpeech/dev-clean \
    data/activations/vg-spokencoco/librispeech/dev-clean \
    --batch_size 8 --layer rnn0 --output_file_extension '.pt' \
    --file_extension '.flac' --recursionLevel 2
python -m scripts.extract_activations exps/vg/vg-spokencoco/net.best.pt \
    ~/corpora/LibriSpeech/test-clean \
    data/activations/vg-spokencoco/librispeech/test-clean \
    --batch_size 8 --layer rnn0 --output_file_extension '.pt' \
    --file_extension '.flac' --recursionLevel 2

Quantizing the activations

python -m scripts.quantize_activations exps/kmeans/vg-spokencoco-rnn0_kmeans-librispeech100-50/checkpoint_last.pt \
    data/activations/vg-spokencoco/librispeech/train-full-960/rnn0 \
    data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/train-full-960 \
    --recursionLevel 2
python -m scripts.quantize_activations exps/kmeans/vg-spokencoco-rnn0_kmeans-librispeech100-50/checkpoint_last.pt \
    data/activations/vg-spokencoco/librispeech/dev-clean/rnn0 \
    data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/dev-clean \
    --recursionLevel 2
python -m scripts.quantize_activations exps/kmeans/vg-spokencoco-rnn0_kmeans-librispeech100-50/checkpoint_last.pt \
    data/activations/vg-spokencoco/librispeech/test-clean/rnn0 \
    data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/test-clean \
    --recursionLevel 2

Training the model

We can now train the model:

# Converting quantized output for fairseq
python -m scripts.convert_for_fairseq \
    data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/train-full-960/quantized_outputs.txt \
    data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/train-full-960/fairseq.txt
python -m scripts.convert_for_fairseq \
    data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/dev-clean/quantized_outputs.txt \
    data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/dev-clean/fairseq.txt
python -m scripts.convert_for_fairseq \
    data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/test-clean/quantized_outputs.txt \
    data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/test-clean/fairseq.txt

# Preprocessing of the data
fairseq-preprocess --only-source \
    --trainpref data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/train-full-960/fairseq.txt \
    --validpref data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/dev-clean/fairseq.txt \
    --testpref data/quantized/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/test-clean/fairseq.txt \
    --destdir data/fairseq-bin-data/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/train-full-960 \
    --workers 20

# Training
fairseq-train --fp16 \
    data/fairseq-bin-data/vg-spokencoco-rnn0_kmeans-librispeech100-50/librispeech/train-full-960 \
    --save-dir exps/lm/vg-spokencoco-rnn0_kmeans-librispeech100-50_lm-bert-small-librispeech960 \
    --task masked_lm \
    --keep-last-epochs 1 \
    --tensorboard-logdir tensorboard \
    --train-subset train \
    --num-workers 4 \
    --criterion masked_lm \
    --arch roberta_base \
    --sample-break-mode eos --tokens-per-sample 3072 --max-positions 6144 \
    --optimizer adam --adam-betas '(0.9, 0.98)' --adam-eps 1e-06 --clip-norm 0.0 \
    --lr-scheduler polynomial_decay --lr 0.0005 --total-num-update 250000 --warmup-updates 10000 \
    --dropout 0.1 --attention-dropout 0.1 --weight-decay 0.01 \
    --mask-multiple-length 5 --mask-prob 0.5 --mask-stdev 5 \
    --max-tokens 4096 --max-update 5000000 --encoder-embed-dim 512 --encoder-ffn-embed-dim 2048 --encoder-attention-heads 8 --encoder-layers 8 \
    --seed 5 --log-format simple --log-interval 10 --skip-invalid-size-inputs-valid-test

For examples of alternative language models (LSTM or BERT-big), please see the instructions for the high-budget baseline.