returnn
returnn copied to clipboard
rwth-i6
The RWTH extensible training framework for universal recurrent neural networks
There are various cases, e.g. whether we import for train, or for recog, or also just randomly initialize. Maybe it depends on the case whether it is ok to ignore...
Currently `rf.set_default_device` is called *after* the model was created. Specifically, we only use `rf.set_default_device_ctx(self._device)` around the run step, and not otherwise. The model creation happens on CPU, and then we...
This here is open for discussion on what we want in RETURNN. Packed tensors / packed sequences / ragged tensors / jagged tensors / flattened / flat tensors, however you...
Running some simple selected test takes a long time. Most of that time is spent in the collecting phase. That's very annoying for debugging. Also, I don't quite understand why...
For the latest RETURNN, when I use DistributeFilesDataset, I have this error. ``` File "/nas/models/asr/am/multilingual/16kHz/2024-11-08--jxu-best-rq-pretrain/work/i6_core/tools/git/CloneGitRepositoryJob.LD5f1wKK7LPo/output/returnn/returnn/datasets/basic.py", line 227, in Dataset._create_from_reduce line: ds = cls(**kwargs) locals: ds = cls = kwargs =...
We currently do this in Torch `Engine.forward_with_callback`: ```python ... elif dataset.supports_seq_order_sorting(): # We can sort it. Sort it in reverse to make sure that we have enough memory right at...
For the latest RETURNN, when using torch backend and SprintCacheDataset, I get this error. ``` File "/usr/local/lib/python3.10/dist-packages/torch/utils/data/datapipes/datapipe.py", line 179, in IterDataPipe.__reduce_ex__ line: return super().__reduce_ex__(*args, **kwargs) locals: super = __reduce_ex__ =...
In `SimpleHDFWriter.insert_batch`, when using `extra` to put some other data (despite for the main data stream "data"), the seq lens are currently not handled correctly. Currently the logic is: ```python...
Now that TF1 support was dropped (#1668), we can cleanup `returnn.tf.compat` (maybe also other things, but that is the only case I currently know about).
Current behavior in the torch engine when using a Checkpoint during training via "import_model_train_epoch1" is to reset the epoch to 0 but keeping the global train step count of the...
