eustlb

MLE @huggingface, Open Source Audio 🤗

eustlb / benchmark_transformers_whisper.py

Last active November 22, 2024 10:36

Benchmark WER and RTFx for transformers whisper.

	TRANSFORMERS_SRC_PATH = "/admin/home/eustache_lebihan/dev/benchmark-whisper/transformers-fix/src"

	import sys
	sys.path.insert(0, TRANSFORMERS_SRC_PATH)

	import wandb
	from tqdm import tqdm
	import evaluate
	import os
	import torch

eustlb / benchmark_openai_whisper.py

Last active November 21, 2024 22:40

Benchmark WER and RTFx for openai whisper.

	OPENAI_SRC_PATH = "/admin/home/eustache_lebihan/dev/benchmark-whisper/whisper"

	import sys
	sys.path.insert(0, OPENAI_SRC_PATH)

	import wandb
	from tqdm import tqdm
	import evaluate
	import os
	import torch

eustlb / benchmark_openai_whisper_myfork.py

Last active November 22, 2024 10:38

	OPENAI_SRC_PATH = "/admin/home/eustache_lebihan/dev/benchmark-whisper/whisper-myfork"

	import sys
	sys.path.insert(0, OPENAI_SRC_PATH)


	import wandb
	from tqdm import tqdm
	import evaluate
	import os

eustlb / benchmark_moonshine_tiny.py

Last active January 21, 2025 13:40

Benchmark moonshine/ whisper for varying batch sizes (FLEURS test set)

	import torch
	import evaluate
	from transformers.models.whisper.english_normalizer import EnglishTextNormalizer
	from transformers import MoonshineForConditionalGeneration, AutoProcessor, WhisperProcessor
	from datasets import load_dataset, Audio
	from tqdm import tqdm
	import json

	wer_metric = evaluate.load("wer")
	device = "cuda:0" if torch.cuda.is_available() else "cpu"

eustlb / benchmark_moonshine_tiny_artif.py

Last active January 21, 2025 16:13

Benchmark moonshine/ whisper for varying audio length and artificial number of generate tokens.

	import torch
	from transformers import MoonshineForConditionalGeneration, AutoProcessor
	from tqdm import tqdm
	import json

	device = "cuda:0" if torch.cuda.is_available() else "cpu"
	torch_dtype = torch.float32
	attn_implementation = "sdpa"

	model = MoonshineForConditionalGeneration.from_pretrained("UsefulSensors/moonshine-tiny", attn_implementation=attn_implementation).to(device, torch_dtype)

eustlb / .gitignore

Last active February 25, 2025 15:07

Benchmark seq vs batched - Style TTS2

traces-*/

eustlb / hub_to_s3.py

Created February 26, 2025 15:55

HF hub to s3 bucket

	from typing import IO

	from datatrove.io import get_datafolder
	from datatrove.executor import SlurmPipelineExecutor
	from datatrove.pipeline.readers import ParquetReader
	from datatrove.pipeline.writers import ParquetWriter
	from datatrove.utils.typeshelper import StatHints


	class ParquetReaderInMemory(ParquetReader):

eustlb / fine_tune_csm.py

Last active May 19, 2025 14:49

	from datasets import load_dataset, Audio
	from transformers import (
	CsmForConditionalGeneration,
	TrainingArguments,
	CsmProcessor,
	Trainer
	)

	processor = CsmProcessor.from_pretrained("eustlb/csm-1b")
	model = CsmForConditionalGeneration.from_pretrained("eustlb/csm-1b")

eustlb / infer_whisper.py

Last active March 31, 2025 12:44

infer whisper

	import torch
	from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
	from datasets import load_dataset

	device = "cuda:0" if torch.cuda.is_available() else "cpu"
	torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

	model_id = "openai/whisper-large-v3-turbo"

	model = AutoModelForSpeechSeq2Seq.from_pretrained(

eustlb / csm_test_1b_model_integration_generate.py

Last active April 23, 2025 08:42

Reproducer for CSM Transformers integration

	# TEST GREEDY FLOAT 32
	# make sure to clone [email protected]:eustlb/csm.git and checkout compare-trfms
	import sys
	sys.path.insert(0, "./csm")

	from generator import load_csm_1b, Segment

	from datasets import load_dataset, Audio
	from huggingface_hub import hf_hub_download
	import torch