rwth-i6/returnn

Issues

• TF end layer independent of batch causes error in beam search

  #1606 opened 4 months ago by albertz
  0

• RF masked computation / masking (like masked_select but without the packing)

  #1605 opened 4 months ago by albertz
  3

• Make `FileCache` able to detect updated remote files

  #1602 opened 4 months ago by NeoLegends
  1

• `rf.RelPosCausalSelfAttention` fails with `single_step_dim`

  #1585 opened 5 months ago by LucaG1
  9

• Torch print step info on crash

  #1597 opened 5 months ago by albertz
  0

• ConcatSeqsDataset with extended functionality

  #1573 opened 5 months ago by Stefanwuu
  3

• Make batch_size configurable for cross validation

  #1567 opened 5 months ago by michelwi
  1

• Torch: gradient_clip wrong when grad_scaler is used

  #1590 opened 5 months ago by michelwi
  0

• `rf.pack_padded` with PyTorch takes a lot of memory

  #1584 opened 5 months ago by albertz
  1

• Torch `report_profile` `check_events` based tests maybe unstable

  #1589 opened 5 months ago by albertz
  1

• Torch gradient_checkpoint_scope could trigger segmentation fault?

  #1581 opened 5 months ago by albertz
  16

• Tensor deepcopy does not copy raw_tensor

  #1541 opened 6 months ago by albertz
  1

• Torch multiple simultaneous gradient_checkpoint_scope

  #1583 opened 5 months ago by albertz
  0

• Torch gradient_checkpoint_scope potential memory leak

  #1582 opened 5 months ago by albertz
  0

• Torch gradient_checkpoint_scope _unregister_custom_saved_tensors_hooks error

  #1579 opened 5 months ago by albertz
  4

• RF parametrization breaks Conv

  #1580 opened 5 months ago by albertz
  0

• RF weight dropout and variational noise

  #1518 opened 5 months ago by albertz
  9

• RuntimeError: CUDA error: an illegal memory access was encountered

  #1577 opened 5 months ago by albertz
  1

• Gradient checkpointing for weight noise etc in PyTorch

  #1552 opened 5 months ago by albertz
  7

• Torch: print model at log verbosity 3

  #1575 opened 5 months ago by NeoLegends
  1

• PyTorch/RF (?): choosing on which epochs to save optimizer state

  #1565 opened 5 months ago by NeoLegends
  0

• DistributeFilesDataset: _distribute_evenly_by_size suboptimal for multi-gpu sharding

  #1570 opened 5 months ago by michelwi
  8

• multiprocessing: OSError: AF_UNIX path too long

  #1571 opened 5 months ago by michelwi
  11

• Ignore a single broken gradient

  #1568 opened 5 months ago by JackTemaki
  2

• Dataset ctx_left/ctx_right extension: ctx_clip_to_valid option

  #1564 opened 5 months ago by albertz
  5

• PyTorch Distributed Training: File descriptors opened and never closed

  #1560 opened 5 months ago by NeoLegends
  8

• Datasets: blocklist in addition to allowlist for segment list file

  #1566 opened 5 months ago by NeoLegends
  0

• Hang in training (often with multi GPU training)

  #1558 opened 6 months ago by albertz
  1

• DistributeFilesDataset Sharding with PT Dataloader breaks

  #1556 opened 6 months ago by michelwi
  3

• RF scaled_dot_product_attention

  #1555 opened 6 months ago by albertz
  0

• DistributeFilesDataset has issues with DataLoader and `num_workers > 0`

  #1554 opened 6 months ago by NeoLegends
  1

• SlowMo (BMUF) support for PyTorch distributed training

  #1553 opened 6 months ago by albertz
  0

• `DistributeFilesDataset`: copying files blocks `init_seq_order`

  #1549 opened 6 months ago by albertz
  2

• Ideas for generic `CachedFile` support across all datasets

  #1544 opened 6 months ago by NeoLegends
  18

• `FileCache`: Race condition when removing empty directories

  #1550 opened 6 months ago by NeoLegends
  5

• `FileCache`: avoid cache-wide dir lock

  #1548 opened 6 months ago by albertz
  0

• `FileCache`: better cleaning, free more than just the minimum

  #1547 opened 6 months ago by albertz
  0

• Possible race condition in `FileCache`?

  #1542 opened 6 months ago by NeoLegends
  5

• `DistributeFilesDataset` with sharding on file level

  #1531 opened 6 months ago by albertz
  6

• `DistributeFilesDataset`, allow kwargs in `get_sub_epoch_dataset`

  #1540 opened 6 months ago by Icemole
  10

• RF BatchNorm running var small diff between TF-layers, pure RF and direct PyTorch, biased vs unbiased

  #1539 opened 6 months ago by albertz
  0

• Support for larger scale datasets

  #1519 opened 6 months ago by albertz
  33

• `ConcatFilesDataset` needs a better name

  #1535 opened 6 months ago by NeoLegends
  10

• ConcatFilesDataset: Reshuffle files per subepoch after every full epoch

  #1534 opened 6 months ago by NeoLegends
  2

• RF torch `lstm` fails with torch amp option.

  #1529 opened 6 months ago by LucaG1
  6

• `ConcatFilesDataset` combines poorly with `MetaDataset`

  #1524 opened 6 months ago by NeoLegends
  6

• Compilation of custom operations failing on TF 2.15/CUDA 12

  #1523 opened 7 months ago by Icemole
  5

• PyTorch debug_add_check_numerics_ops

  #1522 opened 7 months ago by albertz
  0

• RuntimeError: CUDA error: unknown error

  #1520 opened 7 months ago by albertz
  0

• torch.onnx.export requires input_names and output_names to be in order

  #1517 opened 7 months ago by kuacakuaca
  12