Question about the epoch-2-avg-6.pt checkpoint
Hello,
I hope this message finds you well! I have a question regarding the epoch-2-avg-6.pt checkpoint located in the icefall_asr_multi-hans_whisper_qwen2_1.5B directory. Is this checkpoint an early one that has not yet been fully trained?
Additionally, I would greatly appreciate it if you could provide the checkpoints corresponding to the results table found in the RESULTS.md at https://github.com/k2-fsa/icefall/blob/master/egs/speech_llm/ASR_LLM/RESULTS.md.
Thank you for your time and assistance!
Best regards,
Is this checkpoint an early one that has not yet been fully trained?
I don't understand it. Could you explain more please?
The epoch-2-avg-6.pt in the repo could give decoding results of second column in https://github.com/k2-fsa/icefall/blob/master/egs/speech_llm/ASR_LLM/RESULTS.md.
Thank you for your reply. I loaded the epoch-2-avg-6.pt checkpoint, but I obtained a very high CER on the two test sets of WenetSpeech, which is different from the RESULTS.md. Then I noticed that the CER in the log file /log-beam-search-beam1/log-decode-epoch-100-avg-1-2024-06-14-07-36-47 is similar to the RESULTS.md, but it used the epoch-100-avg-1 checkpoint. Therefore, I am unsure if my results are due to the checkpoint or my own code, so I am seeking your advice.
Thanks. I think the results are due to the inference code error. https://github.com/k2-fsa/icefall/blob/master/egs/speech_llm/ASR_LLM/RESULTS.md
Did you change the aishell example checkpoint with multi_hans_zh checkpoint?
Feel free to discuss more details through wechat (13679162366).
Thank you for your assistance. The root cause was identified as inconsistent preprocessing steps when switching to a different data loader I/O. After resolving these implementation discrepancies, I have successfully reproduced all results shown in the performance table.