Wav2vec2 Tiny Random

🚀 測試模型

本項目提供了一個測試模型的代碼示例，用於加載數據集、模型，並進行前向傳播和損失計算。

🚀 快速開始

要測試此模型，請運行以下代碼：

from datasets import load_dataset
from transformers import Wav2Vec2ForCTC
import torchaudio
import torch

ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")

model = Wav2Vec2ForCTC.from_pretrained("patrickvonplaten/wav2vec2_tiny_random")

def load_audio(batch):
    batch["samples"], _ = torchaudio.load(batch["file"])
    return batch
    
ds = ds.map(load_audio)

input_values = torch.nn.utils.rnn.pad_sequence([torch.tensor(x[0]) for x in ds["samples"][:10]], batch_first=True)

# forward
logits = model(input_values).logits
pred_ids = torch.argmax(logits, dim=-1)

# dummy loss
dummy_labels = pred_ids.clone()
dummy_labels[dummy_labels == model.config.pad_token_id] = 1  # can't have CTC blank token in label
dummy_labels = dummy_labels[:, -(dummy_labels.shape[1] // 4):] # make sure labels are shorter to avoid "inf" loss (can still happen though...)
loss = model(input_values, labels=dummy_labels).loss

💻 使用示例

基礎用法

from datasets import load_dataset
from transformers import Wav2Vec2ForCTC
import torchaudio
import torch

ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")

model = Wav2Vec2ForCTC.from_pretrained("patrickvonplaten/wav2vec2_tiny_random")

def load_audio(batch):
    batch["samples"], _ = torchaudio.load(batch["file"])
    return batch
    
ds = ds.map(load_audio)

input_values = torch.nn.utils.rnn.pad_sequence([torch.tensor(x[0]) for x in ds["samples"][:10]], batch_first=True)

# forward
logits = model(input_values).logits
pred_ids = torch.argmax(logits, dim=-1)

# dummy loss
dummy_labels = pred_ids.clone()
dummy_labels[dummy_labels == model.config.pad_token_id] = 1  # can't have CTC blank token in label
dummy_labels = dummy_labels[:, -(dummy_labels.shape[1] // 4):] # make sure labels are shorter to avoid "inf" loss (can still happen though...)
loss = model(input_values, labels=dummy_labels).loss

高級用法

此代碼示例已經涵蓋了基本的測試流程，目前暫無更高級的使用場景。若有其他需求，可根據實際情況對代碼進行修改和擴展。