Nay San fauxneticien

fauxneticien / signal2int.py

Created October 19, 2023 04:33

A solution for Colab exercise

	def signal2int(signal, sample_rate=16_000, seq_length=4):
	"""
	Decode a sine wave signal back into digits
	"""
	digits = []

	for start_index in range(0, sample_rate * seq_length, sample_rate):
	end_index = start_index + sample_rate - 1

	window_samples = signal[start_index:end_index]

fauxneticien / reduce-list-of-funcs.py

Created October 2, 2023 15:50

Apply a list of functions to some input

	from functools import reduce

	input = 1
	f1 = lambda i: i+1
	f2 = lambda i: i+2
	f3 = lambda i: i+3

	output = reduce(lambda i, f: f(i), [f1, f2, f3], input)

fauxneticien / benchmark-fastconformer.py

Created September 21, 2023 02:40

benchmark-fastconformer.py

	import torch
	import torchaudio

	fakewav_80mins = torch.rand(1, 16_000 * 60 * 80)
	torchaudio.save("tmp/80mins.wav", fakewav_80mins, 16_000)

	import nemo.collections.asr as nemo_asr

	asr_model = nemo_asr.models.EncDecCTCModelBPE.from_pretrained(model_name="nvidia/stt_en_fastconformer_ctc_large")
	asr_model.eval()

fauxneticien / librilight10h_tsv.py

Created September 17, 2023 02:09

Fetch LibriLightLimited 10h fine-tuning dataset as TSV file

	import torchaudio

	import pandas as pd

	from pathlib import Path
	from tqdm import tqdm

	llight10h = torchaudio.datasets.LibriLightLimited(root="tmp/", subset="10h", download=True)

	manifest_rows = []

fauxneticien / run-ddp.py

Last active September 8, 2023 19:57

Lhotse DDP test

	import torch
	import torch.nn.functional as F
	from torch.utils.data import Dataset, DataLoader

	import os
	import torchaudio
	from lhotse import CutSet, Fbank, FbankConfig
	from lhotse.dataset import IterableDatasetWrapper, SpeechSynthesisDataset, DynamicBucketingSampler, OnTheFlyFeatures, make_worker_init_fn
	from lhotse.recipes import download_librispeech, prepare_librispeech
	from tqdm import tqdm

fauxneticien / import-hf-to-torchaudio.py

Created July 30, 2023 20:12

Import HF model to torchaudio

	import torch
	from transformers import Wav2Vec2Model
	from torchaudio.models.wav2vec2.utils import import_huggingface_model

	hf_model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-xls-r-300m")
	ta_model = import_huggingface_model(hf_model)

	torch.save(ta_model.state_dict(), "tmp/wav2vec2-300m_xls-r.pt")

	## Usage

fauxneticien / lhotse-shar-outputs.txt

Created July 11, 2023 18:12

Tree output of Lhotse shar files

fauxneticien / hydra-demo.yaml

Last active June 21, 2023 22:26

Example class

	my_instantiated_object:
	_target_: my_classes.my_abstract_class.MyAbstractClass
	my_variable: used Hydra YAML!

fauxneticien / lhotse-token-collator-for-ctc.py

Last active June 12, 2023 20:24

Lhotse token collator for CTC

	# Modified version with diff (see history)

	class TokenCollater:
	"""Collate list of tokens

	Map sentences to integers. Sentences are padded to equal length.
	Beginning and end-of-sequence symbols can be added.
	Call .inverse(tokens_batch, tokens_lens) to reconstruct batch as string sentences.

	Example:

fauxneticien / backup.sh

Created April 9, 2023 15:36

Backup model checkpoints without cronjob

	#!/usr/bin/bash
	# Basic range in for loop
	for value in {1..8}
	do
	echo "Starting backup $value"
	rclone copy -P --exclude checkpoint_last.pt /home/nay/w2v2-fairseq-pretrain/outputs/2023-04-09/08-12-57/checkpoints/ gdrive_slasr:asr/checkpoints/xls-r_cpt_malay-140h/checkpoints/35-ga16/
	echo "Sleeping for 1 hour"
	sleep 1h
	done