语言: 英文
数据集:

• librispeech_asr
  标签:
• 自动语音识别
  许可证: apache-2.0

测试模型

运行以下代码来测试此模型：

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_robust")  

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)  

# 前向传播  
logits = model(input_values).logits  
pred_ids = torch.argmax(logits, dim=-1)  

# 模拟损失  
dummy_labels = pred_ids.clone()  
dummy_labels[dummy_labels == model.config.pad_token_id] = 1  # 标签中不能有CTC空白标记  
dummy_labels = dummy_labels[:, -(dummy_labels.shape[1] // 4):] # 确保标签更短以避免“无限”损失（但仍可能发生...）  
loss = model(input_values, labels=dummy_labels).loss