Wang Changxu

@gfjiangly @JerryLead The way I solve this problem is to shuffle and distribute TFRecord files to different GPUs evenly before each epoch. The problem is, how can to handle these...

@linkun-1998 Yes, IterableDataset does not has `__len__` method by default, so `len(dataset)` is unavailable for it. You must add `__len__` method by your self.

@linkun-1998 Due to company's compliance reason, I cannot upload the full code. This is the core part of `MultiTFRecordDataset`: ```python class MultiTFRecordDataset(torch.utils.data.IterableDataset): """Parse multiple (generic) TFRecords datasets into an `IterableDataset`...

Firstly, manually install `PyAudio`. Then, Just `pip install PyBaiduYuyin`. I'll put this on README

星级：5星 名称：Airmail 官网：http://airmailapp.com 推荐理由：轻巧、美观、好用的客户端，App Store的价格为68元RMB，性价比绝对比Postbox高

哎呀、抱歉发错位置了 王昌旭 > 在 2015年7月22日，21:28，Jak Wings [email protected] 写道： > > @DelightRun @Haixing-Hu 請把付費 App 加放到另一個主題介紹好嗎？ > > — > Reply to this email directly or view it on GitHub.

> UPDATE: > > This seems a memory-related issue. > When using a small model and setting `batch_size=1` and `num_workers=1`, training runs well. > But increasing `batch_size` or `num_workers` easily...

> I don't think you will benefit from gpu decoding if you use `num_worker>0` for data loading because it's using multiple cpu core to speed up the loading which will...

> @DelightRun Are you using gpu for training or inference? If so, data need to be copied to gpu again from batch input. @zhreshold After I read PyTorch's code, the...

After do some deep profiling, I think the problem is due to the `record_stream(torch.cuda.current_stream())`, which will block the default stream until copy complete. Here is what PyTorch's document says about...