rwth-i6/returnn issues and pull requests

#1633 - Dim declare_same_as, fix when existing same_as

Pull Request - State: closed - Opened by albertz 5 days ago

#1632 - `LaplaceOrdering`: avoid spiky CPU utilization

Pull Request - State: closed - Opened by NeoLegends 6 days ago - 1 comment
Labels: bug

#1631 - `LaplaceOrdering` interacts badly w/ MultiProcDataset

Issue - State: closed - Opened by NeoLegends 6 days ago
Labels: bug

#1630 - Datasets: implement support for within-dataset sharding

Pull Request - State: open - Opened by NeoLegends 8 days ago - 3 comments

#1629 - DeepCopyError in gradient checkpointing when using `param_variational_noise`

Issue - State: closed - Opened by mmz33 11 days ago - 5 comments

#1628 - PT: regularly sync progress during eval, fix tensor assignment

Pull Request - State: closed - Opened by NeoLegends 12 days ago - 1 comment

#1627 - PT: regularly sync progress during eval

Pull Request - State: closed - Opened by NeoLegends 12 days ago
Labels: bug

#1621 - PT distributed training: RuntimeError: Socket Timeout in eval

Issue - State: closed - Opened by albertz 23 days ago - 10 comments

#1620 - TF prepare_gradient_checkpointing, fix for newer TF

Pull Request - State: closed - Opened by albertz 25 days ago

#1619 - TF prepare_gradient_checkpointing, avoid deep recursion

Pull Request - State: closed - Opened by albertz 25 days ago

#1618 - some type fixes (TF code)

Pull Request - State: closed - Opened by albertz 25 days ago

#1617 - TF fix combined variational noise and weight dropout

Pull Request - State: closed - Opened by albertz 25 days ago

#1616 - `param_variational_noise` causes recursion limit error in TF backend

Issue - State: closed - Opened by mmz33 26 days ago - 10 comments

#1615 - PostprocessingDataset: `get_data_keys` fails if called before `init_seq_order`

Issue - State: closed - Opened by NeoLegends 26 days ago
Labels: bug

#1614 - DistributeFilesDataset does not interact well w/ datasets that don't have a known num_seqs

Issue - State: closed - Opened by NeoLegends 26 days ago - 12 comments

#1613 - Use util function to generate forwards compat kwargs

Pull Request - State: closed - Opened by NeoLegends 27 days ago - 1 comment

#1612 - torch distributed: add support for user-specified parameter synchronization

Pull Request - State: open - Opened by NeoLegends 27 days ago - 5 comments

#1611 - Dim math, fix potential leak

Pull Request - State: closed - Opened by albertz 28 days ago

#1610 - PT preload_from_files ext: random init, external part

Pull Request - State: closed - Opened by albertz about 1 month ago - 2 comments

#1609 - `PostprocessingDataset`: add composition function

Pull Request - State: closed - Opened by NeoLegends about 1 month ago - 2 comments

#1608 - `PostprocessingDataset`: add laplace ordering `map_seq_stream` iterator

Pull Request - State: closed - Opened by NeoLegends about 1 month ago
Labels: enhancement

#1607 - Simplify assertion in bliss-to-ogg-zip tool

Pull Request - State: closed - Opened by NeoLegends about 1 month ago

#1606 - TF end layer independent of batch causes error in beam search

Issue - State: open - Opened by albertz about 1 month ago

#1605 - RF masked computation / masking (like masked_select but without the packing)

Issue - State: closed - Opened by albertz about 1 month ago - 3 comments

#1604 - `PostprocessingDataset`: implement laplace sequence ordering

Pull Request - State: closed - Opened by NeoLegends about 2 months ago - 5 comments

#1603 - `FileCache`: detect modified source files

Pull Request - State: closed - Opened by NeoLegends 2 months ago - 5 comments
Labels: enhancement

#1602 - Make `FileCache` able to detect updated remote files

Issue - State: closed - Opened by NeoLegends 2 months ago - 1 comment

#1601 - Remove some usages of `num_output`

Pull Request - State: open - Opened by NeoLegends 2 months ago

#1600 - Bump tensorflow from 2.11.1 to 2.12.1 in /docs

Pull Request - State: open - Opened by dependabot[bot] 2 months ago
Labels: dependencies

#1599 - OggZipDataset various fixes

Pull Request - State: closed - Opened by NeoLegends 2 months ago
Labels: bug

#1598 - Add nose to requirements-dev

Pull Request - State: closed - Opened by NeoLegends 2 months ago - 2 comments

#1597 - Torch print step info on crash

Issue - State: open - Opened by albertz 2 months ago

#1596 - PostprocessingDataset

Pull Request - State: closed - Opened by NeoLegends 2 months ago - 6 comments

#1595 - Torch: unscale gradients before noise and clipping

Pull Request - State: closed - Opened by michelwi 2 months ago

#1594 - Torch: allow setting custom batch_size for each dataset

Pull Request - State: closed - Opened by michelwi 2 months ago

#1593 - Torch masked_select, no custom nonzero for now

Pull Request - State: closed - Opened by albertz 2 months ago - 2 comments

#1592 - RF num_elements_of_shape fix for multiple dyn dims

Pull Request - State: closed - Opened by albertz 2 months ago

#1591 - Move and extend Torch test `report_profile`

Pull Request - State: closed - Opened by albertz 2 months ago

#1590 - Torch: gradient_clip wrong when grad_scaler is used

Issue - State: closed - Opened by michelwi 2 months ago

#1589 - Torch `report_profile` `check_events` based tests maybe unstable

Issue - State: closed - Opened by albertz 3 months ago - 1 comment

#1588 - Torch tests also new PyTorch

Pull Request - State: closed - Opened by albertz 3 months ago

#1587 - RF sequence_mask for more dyn dims

Pull Request - State: closed - Opened by albertz 3 months ago

#1586 - Optimize RF PT pack_padded

Pull Request - State: closed - Opened by albertz 3 months ago - 9 comments

#1585 - `rf.RelPosCausalSelfAttention` fails with `single_step_dim`

Issue - State: open - Opened by LucaG1 3 months ago - 9 comments
Labels: returnn-frontend

#1584 - `rf.pack_padded` with PyTorch takes a lot of memory

Issue - State: closed - Opened by albertz 3 months ago - 1 comment

#1583 - Torch multiple simultaneous gradient_checkpoint_scope

Issue - State: open - Opened by albertz 3 months ago

#1582 - Torch gradient_checkpoint_scope potential memory leak

Issue - State: open - Opened by albertz 3 months ago

#1581 - Torch gradient_checkpoint_scope could trigger segmentation fault?

Issue - State: open - Opened by albertz 3 months ago - 16 comments

#1580 - RF parametrization breaks Conv

Issue - State: closed - Opened by albertz 3 months ago

#1579 - Torch gradient_checkpoint_scope _unregister_custom_saved_tensors_hooks error

Issue - State: closed - Opened by albertz 3 months ago - 4 comments

#1578 - RF: weight parametrization

Pull Request - State: closed - Opened by albertz 3 months ago - 3 comments

#1577 - RuntimeError: CUDA error: an illegal memory access was encountered

Issue - State: open - Opened by albertz 3 months ago - 1 comment

#1576 - Torch: print model at log verbosity 3

Pull Request - State: open - Opened by NeoLegends 3 months ago

#1575 - Torch: print model at log verbosity 3

Issue - State: open - Opened by NeoLegends 3 months ago - 1 comment

#1574 - ConcatSeqsDataset pad_narrow_data_to_multiple_of_target_len

Pull Request - State: closed - Opened by Stefanwuu 3 months ago - 5 comments

#1573 - ConcatSeqsDataset with extended functionality

Issue - State: closed - Opened by Stefanwuu 3 months ago - 3 comments

#1572 - DistributeFilesDataset: Distribute files more evenly

Pull Request - State: closed - Opened by NeoLegends 3 months ago

#1571 - multiprocessing: OSError: AF_UNIX path too long

Issue - State: open - Opened by michelwi 3 months ago - 11 comments

#1570 - DistributeFilesDataset: _distribute_evenly_by_size suboptimal for multi-gpu sharding

Issue - State: closed - Opened by michelwi 3 months ago - 8 comments

#1569 - Torch: optionally apply `cleanup_old_models` logic to optimizer states

Pull Request - State: closed - Opened by NeoLegends 3 months ago - 7 comments

#1568 - Ignore a single broken gradient

Issue - State: open - Opened by JackTemaki 3 months ago - 2 comments

#1567 - Make batch_size configurable for cross validation

Issue - State: closed - Opened by michelwi 3 months ago - 1 comment

#1566 - Datasets: blocklist in addition to allowlist for segment list file

Issue - State: open - Opened by NeoLegends 3 months ago

#1565 - PyTorch/RF (?): choosing on which epochs to save optimizer state

Issue - State: closed - Opened by NeoLegends 3 months ago
Labels: PyTorch

#1564 - Dataset ctx_left/ctx_right extension: ctx_clip_to_valid option

Issue - State: closed - Opened by albertz 3 months ago - 5 comments

#1563 - Default torch DataLoader num_workers to 1

Pull Request - State: closed - Opened by NeoLegends 3 months ago - 7 comments

#1562 - CI PyCharm update Torch

Pull Request - State: closed - Opened by albertz 3 months ago

#1561 - Add warning for not using `num_workers > 0` in torch

Pull Request - State: closed - Opened by NeoLegends 3 months ago - 2 comments

#1560 - PyTorch Distributed Training: File descriptors opened and never closed

Issue - State: closed - Opened by NeoLegends 3 months ago - 8 comments
Labels: PyTorch, MultiGPU

#1559 - Torch `gradient_checkpoint_scope`

Pull Request - State: closed - Opened by albertz 3 months ago - 5 comments

#1558 - Hang in training (often with multi GPU training)

Issue - State: open - Opened by albertz 3 months ago - 1 comment

#1557 - DistributeFilesDataset: pass worker group info from parent process to child via pickle

Pull Request - State: closed - Opened by NeoLegends 3 months ago - 1 comment
Labels: bug

#1556 - DistributeFilesDataset Sharding with PT Dataloader breaks

Issue - State: closed - Opened by michelwi 3 months ago - 3 comments

#1555 - RF scaled_dot_product_attention

Issue - State: open - Opened by albertz 3 months ago

#1554 - DistributeFilesDataset has issues with DataLoader and `num_workers > 0`

Issue - State: closed - Opened by NeoLegends 3 months ago - 1 comment
Labels: bug

#1553 - SlowMo (BMUF) support for PyTorch distributed training

Issue - State: open - Opened by albertz 3 months ago
Labels: PyTorch, MultiGPU

#1552 - Gradient checkpointing for weight noise etc in PyTorch

Issue - State: closed - Opened by albertz 3 months ago - 7 comments
Labels: PyTorch

#1551 - FileCache reliability improvements

Pull Request - State: closed - Opened by NeoLegends 3 months ago - 1 comment

#1550 - `FileCache`: Race condition when removing empty directories

Issue - State: closed - Opened by NeoLegends 3 months ago - 5 comments

#1549 - `DistributeFilesDataset`: copying files blocks `init_seq_order`

Issue - State: closed - Opened by albertz 3 months ago - 2 comments

#1548 - `FileCache`: avoid cache-wide dir lock

Issue - State: closed - Opened by albertz 3 months ago

#1547 - `FileCache`: better cleaning, free more than just the minimum

Issue - State: closed - Opened by albertz 3 months ago

#1546 - `FileCache`: add generic support for all datasets

Pull Request - State: closed - Opened by NeoLegends 3 months ago - 5 comments

#1545 - HDFDataset: add support for CachedFile

Pull Request - State: closed - Opened by michelwi 3 months ago - 2 comments

#1544 - Ideas for generic `CachedFile` support across all datasets

Issue - State: closed - Opened by NeoLegends 3 months ago - 18 comments

#1543 - `FileCache`: retry on ENOSPC, prealloc space, lock on whole cache dir

Pull Request - State: closed - Opened by NeoLegends 3 months ago - 8 comments

#1542 - Possible race condition in `FileCache`?

Issue - State: closed - Opened by NeoLegends 3 months ago - 5 comments

#1541 - Tensor deepcopy does not copy raw_tensor

Issue - State: open - Opened by albertz 4 months ago - 1 comment

#1540 - `DistributeFilesDataset`, allow kwargs in `get_sub_epoch_dataset`

Issue - State: closed - Opened by Icemole 4 months ago - 10 comments

#1539 - RF BatchNorm running var small diff between TF-layers, pure RF and direct PyTorch, biased vs unbiased

Issue - State: open - Opened by albertz 4 months ago

#1538 - `DistributeFilesDataset`: allow sharding files across GPU workers

Pull Request - State: closed - Opened by NeoLegends 4 months ago - 12 comments

#1537 - Rename `ConcatFilesDataset` to `DistributeFilesDataset`

Pull Request - State: closed - Opened by NeoLegends 4 months ago - 4 comments

#1536 - Gradient checkpointing experiments

Pull Request - State: closed - Opened by NeoLegends 4 months ago - 6 comments

#1535 - `ConcatFilesDataset` needs a better name

Issue - State: closed - Opened by NeoLegends 4 months ago - 10 comments

#1534 - ConcatFilesDataset: Reshuffle files per subepoch after every full epoch

Issue - State: closed - Opened by NeoLegends 4 months ago - 2 comments

#1533 - Pass `random_seed_offset` via env var

Pull Request - State: closed - Opened by NeoLegends 4 months ago - 2 comments

#1532 - HDFDataset, use bisect in _get_file_index()

Pull Request - State: closed - Opened by Icemole 4 months ago - 1 comment

#1531 - `DistributeFilesDataset` with sharding on file level

Issue - State: closed - Opened by albertz 4 months ago - 6 comments

#1530 - ConcatFilesDataset (DistributeFilesDataset): restart dataset worker proc once on error

Pull Request - State: closed - Opened by NeoLegends 4 months ago - 1 comment

#1529 - RF torch `lstm` fails with torch amp option.

Issue - State: closed - Opened by LucaG1 4 months ago - 6 comments

GitHub / rwth-i6/returnn issues and pull requests