returnn
returnn copied to clipboard
rwth-i6
The RWTH extensible training framework for universal recurrent neural networks
#508 Thus effectively removing the default data key "data". I don't have a strong opinion on this, so I just put this here and leave it open for discussion. This...
I think this depends on #530. In case of `concat_sources=False`, the source names `data:0` etc are not nice, and I think we can have better ways (this is #530). (I'm...
#508 This is old Theano behavior.
Discussion related to #508 Many parameters that are part of the dataset can be set globally: ```python3 set_or_remove("window", config.int('window', 0) or None) set_or_remove("context_window", config.typed_value("context_window")) set_or_remove("chunking", config.opt_typed_value("chunking", None)) set_or_remove("seq_ordering", config.value("batching", None))...
Discussion related to #508 Right now there are both a `learning_rate` and a `learning_rates` parameter, which often causes confusion to the users. Here I am not sure what would be...
in #294 there was a discussion about the Dataset API but it was rather concetrated on every single method of the class. Would be nice to have a high level...
Added `get_all_tags`, `init_seq_order` and `get_current_seq_order` for **DummyDatasetMultipleSequenceLength**
Fix bug where custom_parameter_importer = "subset" would not operate on parameters inside rec layers.
For complicated `custom_construction_network` or `MultiStage` training such as in [returnn-experiments # 60](https://github.com/rwth-i6/returnn-experiments/pull/60) it would be nice to have some dummy dataset for fast testing. With testing, I mean checking if...
It seems to me that the `HDFDataset` does only support data streams with a single time-axis, and not data with no or multiple time axes. I know that I can...