Every step has the same logit output in the wav2vec2ForCTC inference phase after fine-tuning

🤗Transformers
1

My fine-tuning code is similar as code
In order to find out if it’s the reason for the data., input features and target ids of tarin_dataset are as follows:

{'input_values': tentor([[-0.0409, -0.0684, -0.0607, ...,0.000, 0.000, 0.000]
...]), 'labels': tensor([[5381, 1742, 3236, 1949, 2392, 127, 2463, 573, 244, 4344, 2956, 1446,4504, 7431, 6559, -100, -100, -100, -100, -100, -100, -100], ...])}

after fine-tuning, I put dev data to fine-tuned model and got logits as follows:

tensor([[[7.7506, -11.7912, -11.7508, ..., -7.8345, -116404, -6.9154],
[7.7506, -11.7912, -11.7508, ..., -7.8345, -116404, -6.9154],
[7.7506, -11.7912, -11.7508, ..., -7.8345, -116404, -6.9154],
...
[7.7506, -11.7912, -11.7508, ..., -7.8345, -116404, -6.9154],
[7.7506, -11.7912, -11.7508, ..., -7.8345, -116404, -6.9154]]])

It was so weird. Every step has the same logit output. I don’t know what happened. What could be the cause of this? How am I supposed to debug？