Changelog

All notable changes to the project are documented in this file.

Version numbers are of the form 1.0.0. Any version bump in the last digit is backwards-compatible, in that a model trained with the previous version can still be used for translation with the new version. Any bump in the second digit indicates a backwards-incompatible change, e.g. due to changing the architecture or simply modifying model parameter names. Note that Sockeye has checks in place to not translate with an old model that was trained with an incompatible version.

Each version section may have have subsections for: Added, Changed, Removed, Deprecated, and Fixed.

[1.17.0]

Added

  • Source factors, as described in

    Linguistic Input Features Improve Neural Machine Translation (Sennrich & Haddow, WMT 2016) PDF bibtex

    Additional source factors are enabled by passing --source-factors file1 [file2 ...] (-sf), where file1, etc. are token-parallel to the source (-s). An analogous parameter, --validation-source-factors, is used to pass factors for validation data. The flag --source-factors-num-embed D1 [D2 ...] denotes the embedding dimensions and is required if source factor files are given. Factor embeddings are concatenated to the source embeddings dimension (--num-embed).

    At test time, the input sentence and its factors can be passed in via STDIN or command-line arguments.

    • For STDIN, the input and factors should be in a token-based factored format, e.g., word1|factor1|factor2|... w2|f1|f2|... ...1.
    • You can also use file arguments, which mirrors training: --input takes the path to a file containing the source, and --input-factors a list of files containing token-parallel factors. At test time, an exception is raised if the number of expected factors does not match the factors passed along with the input.
  • Removed bias parameters from multi-head attention layers of the transformer.

[1.16.6]

Changed

  • Loading/Saving auxiliary parameters of the models. Before aux parameters were not saved or used for initialization. Therefore the parameters of certain layers were ignored (e.g., BatchNorm) and randomly initialized. This change enables to properly load, save and initialize the layers which use auxiliary parameters.

[1.16.5]

Changed

  • Device locking: Only one process will be acquiring GPUs at a time. This will lead to consecutive device ids whenever possible.

[1.16.4]

Changed

  • Internal change: Standardized all data to be batch-major both at training and at inference time.

[1.16.3]

Changed

  • When a device lock file exists and the process has no write permissions for the lock file we assume that the device is locked. Previously this lead to an permission denied exception. Please note that in this scenario we an not detect if the original Sockeye process did not shut down gracefully. This is not an issue when the sockeye process has write permissions on existing lock files as in that case locking is based on file system locks, which cease to exist when a process exits.

[1.16.2]

Changed

  • Changed to a custom speedometer that tracks samples/sec AND words/sec. The original MXNet speedometer did not take variable batch sizes due to word-based batching into account.

[1.16.1]

Fixed

  • Fixed entry points in setup.py.

[1.16.0]

Changed

  • Update to MXNet 1.0.0 which adds more advanced indexing features, benefitting the beam search implementation.
  • --kvstore now accepts 'nccl' value. Only works if MXNet was compiled with USE_NCCL=1.

Added

[1.15.8]

Fixed

  • Taking the BOS and EOS tag into account when calculating the maximum input length at inference.

[1.15.7]

Fixed

  • fixed a problem with --num-samples-per-shard flag not being parsed as int.

[1.15.6]

Added

  • New CLI sockeye.prepare_data for preprocessing the training data only once before training, potentially splitting large datasets into shards. At training time only one shard is loaded into memory at a time, limiting the maximum memory usage.

Changed

  • Instead of using the --source and --target arguments sockeye.train now accepts a --prepared-data argument pointing to the folder containing the preprocessed and sharded data. Using the raw training data is still possible and now consumes less memory.

[1.15.5]

Added

  • Optionally apply query, key and value projections to the source and target hidden vectors in the CNN model before applying the attention mechanism. CLI parameter: --cnn-project-qkv.

[1.15.4]

Added

  • A warning will be printed if the checkpoint decoder slows down training.

[1.15.3]

Added

  • Exposing the xavier random number generator through --weight-init-xavier-rand-type.

[1.15.2]

Added

  • Exposing MXNet's Nesterov Accelerated Gradient, Adadelta and Adadelta optimizers.

[1.15.1]

Added

  • A tool that initializes embedding weights with pretrained word representations, sockeye.init_embedding.

[1.15.0]

Added

[1.14.3]

Changed

  • Fast decoding for transformer models. Caches keys and values of self-attention before softmax. Changed decoding flag --bucket-width to apply only to source length.

[1.14.2]

Added

  • Gradient norm clipping (--gradient-clipping-type) and monitoring.

Changed

  • Changed --clip-gradient to --gradient-clipping-threshold for consistency.

[1.14.1]

Changed

  • Sorting sentences during decoding before splitting them into batches.
  • Default chunk size: The default chunk size when batching is enabled is now batch_size * 500 during decoding to avoid users accidentally forgetting to increase the chunk size.

[1.14.0]

Changed

  • Downscaled fixed positional embeddings for CNN models.
  • Renamed --monitor-bleu flag to --decode-and-evaluate to illustrate that it computes other metrics in addition to BLEU.

Added

  • --decode-and-evaluate-use-cpu flag to use CPU for decoding validation data.
  • --decode-and-evaluate-device-id flag to use a separate GPU device for validation decoding. If not specified, the existing and still default behavior is to use the last acquired GPU for training.

[1.13.2]

Added

  • A tool that extracts specified parameters from params.x into a .npz file for downstream applications or analysis.

[1.13.1]

Added

[1.13.0]

Fixed

  • Transformer models do not ignore --num-embed anymore as they did silently before. As a result there is an error thrown if --num-embed != --transformer-model-size.
  • Fixed the attention in upper layers (--rnn-attention-in-upper-layers), which was previously not passed correctly to the decoder.

Removed

  • Removed RNN parameter (un-)packing and support for FusedRNNCells (removed --use-fused-rnns flag). These were not used, not correctly initialized, and performed worse than regular RNN cells. Moreover, they made the code much more complex. RNN models trained with previous versions are no longer compatible.
  • Removed the lexical biasing functionality (Arthur ETAL'16) (removed arguments --lexical-bias and --learn-lexical-bias).

[1.12.2]

Changed

  • Updated to MXNet 0.12.1, which includes an important bug fix for CPU decoding.

[1.12.1]

Changed

  • Removed dependency on sacrebleu pip package. Now imports directly from contrib/.

[1.12.0]

Changed

  • Transformers now always use the linear output transformation after combining attention heads, even if input & output depth do not differ.

[1.11.2]

Fixed

  • Fixed a bug where vocabulary slice padding was defaulting to CPU context. This was affecting decoding on GPUs with very small vocabularies.

[1.11.1]

Fixed

  • Fixed an issue with the use of ignore in CrossEntropyMetric::cross_entropy_smoothed. This was affecting runs with Eve optimizer and label smoothing. Thanks @kobenaxie for reporting.

[1.11.0]

Added

  • Lexicon-based target vocabulary restriction for faster decoding. New CLI for top-k lexicon creation, sockeye.lexicon. New translate CLI argument --restrict-lexicon.

Changed

  • Bleu computation based on Sacrebleu.

[1.10.5]

Fixed

  • Fixed yet another bug with the data iterator.

[1.10.4]

Fixed

  • Fixed a bug with the revised data iterator not correctly appending EOS symbols for variable-length batches. This reverts part of the commit added in 1.10.1 but is now correct again.

[1.10.3]

Changed

  • Fixed a bug with max_observed_{source,target}_len being computed on the complete data set, not only on the sentences actually added to the buckets based on --max_seq_len.

[1.10.2]

Added

  • --max-num-epochs flag to train for a maximum number of passes through the training data.

[1.10.1]

Changed

  • Reduced memory footprint when creating data iterators: integer sequences are streamed from disk when being assigned to buckets.

[1.10.0]

Changed

  • Updated MXNet dependency to 0.12 (w/ MKL support by default).
  • Changed --smoothed-cross-entropy-alpha to --label-smoothing. Label smoothing should now require significantly less memory due to its addition to MXNet's SoftmaxOutput operator.
  • --weight-normalization now applies not only to convolutional weight matrices, but to output layers of all decoders. It is also independent of weight tying.
  • Transformers now use --embed-dropout. Before they were using --transformer-dropout-prepost for this.
  • Transformers now scale their embedding vectors before adding fixed positional embeddings. This turns out to be crucial for effective learning.
  • .param files now use 5 digit identifiers to reduce risk of overflowing with many checkpoints.

Added

  • Added CUDA 9.0 requirements file.
  • --loss-normalization-type. Added a new flag to control loss normalization. New default is to normalize by the number of valid, non-PAD tokens instead of the batch size.
  • --weight-init-xavier-factor-type. Added new flag to control Xavier factor type when --weight-init=xavier.
  • --embed-weight-init. Added new flag for initialization of embeddings matrices.

Removed

  • --smoothed-cross-entropy-alpha argument. See above.
  • --normalize-loss argument. See above.

[1.9.0]

Added

  • Batch decoding. New options for the translate CLI: --batch-size and --chunk-size. Translator.translate() now accepts and returns lists of inputs and outputs.

[1.8.4]

Added

  • Exposing the MXNet KVStore through the --kvstore argument, potentially enabling distributed training.

[1.8.3]

Added

  • Optional smart rollback of parameters and optimizer states after updating the learning rate if not improved for x checkpoints. New flags: --learning-rate-decay-param-reset, --learning-rate-decay-optimizer-states-reset

[1.8.2]

Fixed

  • The RNN variational dropout mask is now independent of the input (previously any zero initial state led to the first state being canceled).
  • Correctly pass self.dropout_inputs float to mx.sym.Dropout in VariationalDropoutCell.

[1.8.1]

Changed

  • Instead of truncating sentences exceeding the maximum input length they are now translated in chunks.

[1.8.0]

Added

  • Convolutional decoder.
  • Weight normalization (for CNN only so far).
  • Learned positional embeddings for the transformer.

Changed

  • --attention-* CLI params renamed to --rnn-attention-*.
  • --transformer-no-positional-encodings generalized to --transformer-positional-embedding-type.