Whisper Inference RuntimeError: The expanded size of the tensor (3000) must match the existing size (3392) at non-singleton dimension 1. Target sizes: [80, 3000]. Tensor sizes: [80, 3392]

🤗Transformers
1

I have trained a fine-tuned whisper model for an ultra low resource Malasar language.

While doing inference to evaluate the performance on a test set using the code, It performs decoding for few entries in data_test and then produces the error given in the title.

from tqdm import tqdm
from transformers.pipelines.pt_utils import KeyDataset

all_predictions = []

# run streamed inference
for prediction in tqdm(
    pipe(
        KeyDataset(data_test, "audio_path"),
        max_new_tokens=1024,
        generate_kwargs={"task": "transcribe"},
        batch_size=8,
    ),
    total=len(data_test),
):
    all_predictions.append(prediction["text"])

What does that error indicate, and what could be the possible solutions?

2

Did you find out what was the problem?