yujiepan / whisper-v3-tiny-random

This model is for debugging. It is randomly initialized with the config from openai/whisper-large-v3 but is of smaller size.

Codes:

import os

import torch

from huggingface_hub import create_repo, upload_folder
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    GenerationConfig,
    AutoConfig,
    pipeline,
    set_seed,
)
import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline, AutoConfig
from datasets import load_dataset

model_id = "openai/whisper-large-v3"
repo_id = "yujiepan/whisper-v3-tiny-random"
save_path = f"/tmp/{repo_id}"
os.system(f'rm -rf {save_path}')
os.makedirs(save_path, exist_ok=True)

device = "cuda"
torch_dtype = torch.float16
model_id = "openai/whisper-large-v3"

config = AutoConfig.from_pretrained(model_id)
config.num_hidden_layers = 2
config.d_model = 8
config.decoder_attention_heads = 2
config.decoder_ffn_dim = 16
config.decoder_layers = 2
config.encoder_ffn_dim = 16
config.encoder_attention_heads = 2
config.encoder_layers = 2

model = AutoModelForSpeechSeq2Seq.from_config(config)
model.to(device).to(torch_dtype)
model.generation_config = GenerationConfig.from_pretrained(model_id)
processor = AutoProcessor.from_pretrained(model_id)

set_seed(42)
num_params = 0
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        print(name, p.shape)
        torch.nn.init.uniform_(p, -0.5, 0.5)
        num_params += p.numel()
print("Total number of parameters:", num_params)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)

sample = load_dataset(
    "distil-whisper/librispeech_long", "clean",
    split="validation",
)[0]["audio"]
result = pipe(sample, return_timestamps=True)
print(result["text"])

create_repo(repo_id, exist_ok=True)
upload_folder(repo_id=repo_id, folder_path=save_path, repo_type='model')