foreverhell

How can I use Online Hard Example Mining (OHEM) in MMClassification

1

I have seen the supported OHEM, and I have found the OHEM in MMDetection `train_cfg=dict(rcnn=dict(sampler=dict(type='OHEMSampler'))))`, but when I use it in MMClassification, I encountered a problem 'FileNotFoundError: [Errno 2] No...

 How to gurantee the two the same? When I train a custom LLM in DPO, the loss cannot divergence. Is the reason for the two are different?

Create api_v4.py

4

支持API调用v4

流式生成测试结果

3

我使用streamer.py测试了流式生成的效果。 在默认参数 CHUNK_SIZE = 30 OVERLAP_SECONDS = 10 时，每个音频chunk的长度是20s，音频质量和非流式相近。 为了缩短首chunk时长，我尝试减小这两个值为 CHUNK_SIZE = 3 OVERLAP_SECONDS = 1 此时每个音频chunk的长度是2s，但是音量、音色等出现不稳定的状况。 或许适用于播客这种对话型任务的模型，对实时性要求不那么高？