1.18.72

[1.18.72]

Changed

• Removed use of expand_dims in favor of reshape to save memory.

[1.18.71]

Fixed

• Fixed default setting of source factor combination to be 'concat' for backwards compatibility.

[1.18.70]

Added

• Sockeye now outputs fields found in a JSON input object, if they are not overwritten by Sockeye. This behavior can be enabled by selecting --json-input (to read input as a JSON object) and --output-type json (to write a JSON object to output).

[1.18.69]

Added

• Source factors can now be added to the embeddings instead of concatenated with --source-factors-combine sum (default: concat)

[1.18.68]

• Fixed training crashes with --learning-rate-decay-optimizer-states-reset initial option.

1.18.67

[1.18.67]

Added

• Added fertility as a further type of attention coverage.
• Added an option for training to keep the initializations of the model via --keep-initializations. When set, the trainer will avoid deleting the params file for the first checkpoint, no matter what --keep-last-params is set to.

[1.18.66]

Fixed

• Fix to argument names that are allowed to differ for resuming training.

[1.18.65]

Changed

• More informative error message about inconsistent --shared-vocab setting.

[1.18.64]

Added

• Adding translation sampling via --sample [N]. This causes the decoder to sample each next step from the target distribution probabilities at each
  timestep. An optional value of N causes the decoder to sample only from the top N vocabulary items for each hypothesis at each timestep (the
  default is 0, meaning to sample from the entire vocabulary).

[1.18.63]

Changed

• The checkpoint decoder and nvidia-smi subprocess are now launched from a forkserver, allowing for a better separation between processes.

[1.18.62]

Added

• Add option to make TranslatorInputs directly from a dict.

1.18.61

[1.18.61]

Changed

• Update to MXNet 1.3.1. Removed requirements/requirements.gpu-cu{75,91}.txt as CUDA 7.5 and 9.1 are deprecated.

[1.18.60]

Fixed

• Performance optimization to skip the softmax operation for single model greedy decoding is now only applied if no translation scores are required in the output.

[1.18.59]

Added

• Full training state is now returned from EarlyStoppingTrainer's fit().

Changed

• Training state cleanup will not be performed for training runs that did not converge yet.
• Switched to portalocker for locking files (Windows compatibility).

[1.18.58]

Added

• Added nbest translation, exposed as --nbest-size. Nbest translation means to not only output the most probable translation according to a model, but the top n most probable hypotheses. If --nbest-size > 1 and the option --output-type is not explicitly specified, the output type will be changed to one JSON list of nbest translations per line. --nbest-size can never be larger than --beam-size.

Changed

• Changed sockeye.rerank CLI to be compatible with nbest translation JSON output format.

1.18.57

[1.18.57]

Added

• Added sockeye.score CLI for quickly scoring existing translations (documentation).

Fixed

• Entry-point clean-up after the contrib/ rename

1.18.56

[1.18.56]

Changed

• Update to MXNet 1.3.0.post0

[1.18.55]

• Renamed contrib to less-generic sockeye_contrib

1.18.54

[1.18.54]

Added

• --source-factor-vocabs can be set to provide source factor vocabularies.

[1.18.53]

Added

• Always skipping softmax for greedy decoding by default, only for single models.
• Added option --skip-topk for greedy decoding.

[1.18.52]

Fixed

• Fixed bug in constrained decoding to make sure best hypothesis satifies all constraints.

[1.18.51]

Added

• Added a CLI for reranking of an nbest list of translations.

[1.18.50]

Fixed

• Check for equivalency of training and validation source factors was incorrectly indented.

[1.18.49]

Changed

• Removed dependence on the nvidia-smi tool. The number of GPUs is now determined programatically.

[1.18.48]

Changed

• Translator.max_input_length now reports correct maximum input length for TranslatorInput objects, independent of the internal representation, where an additional EOS gets added.

1.18.47

[1.18.47]

Changed

• translate CLI: no longer rely on external, user-given input id for sorting translations. Also allow string ids for sentences.

[1.18.46]

Fixed

• Fixed issue with --num-words 0:0 in image captioning and another issue related to loading all features to memory with variable length.

[1.18.45]

Added

• Added an 8 layer LSTM model similar (but not exactly identical) to the 'GNMT' architecture to autopilot.

[1.18.44]

Fixed

• Fixed an issue with --max-num-epochs causing training to stop before the update/batch that actually completes the epoch was made.

[1.18.43]

Added

• <s> now supported as the first token in a multi-word negative constraint
  (e.g., <s> I think to prevent a sentence from starting with I think)

Fixed

• Bugfix in resetting the state of a multiple-word negative constraint

[1.18.42]

Changed

• Simplified gluon blocks for length calculation

1.18.41

[1.18.41]

Changed

• Require numpy 1.14 or later to avoid MKL conflicts between numpy as mxnet-mkl.

[1.18.40]

Fixed

• Fixed bad check for existence of negative constraints.
• Resolved conflict for phrases that are both positive and negative constraints.
• Fixed softmax temperature at inference time.

[1.18.39]

Added

• Image Captioning now supports constrained decoding.
• Image Captioning: zero padding of features now allows input features of different shape for each image.

[1.18.38]

Fixed

• Fixed issue with the incorrect order of translations when empty inputs are present and translating in chunks.

[1.18.37]

Fixed

• Determining the max output length for each sentence in a batch by the bucket length rather than the actual in order to match the behavior of a single sentence translation.

[1.18.36]

Changed

• Updated to MXNet 1.2.1

1.18.35

[1.18.35]

Added

• ROUGE scores are now available in sockeye-evaluate.
• Enabled CHRF as an early-stopping metric.
• Added support for --beam-search-stop first for decoding jobs with --batch-size > 1.
• Now supports negative constraints, which are phrases that must not appear in the output.
  • Global constraints can be listed in a (pre-processed) file, one per line: --avoid-list FILE
  • Per-sentence constraints are passed using the avoid keyword in the JSON object, with a list of strings as its field value.
• Added option to pad vocabulary to a multiple of x: e.g. --pad-vocab-to-multiple-of 16.
• Pre-training the RNN decoder. Usage:
  1. Train with flag --decoder-only.
  2. Feed identical source/target training data.

Fixed

• Preserving max output length for each sentence to allow having identical translations for both with and without batching.

Changed

• No longer restrict the vocabulary to 50,000 words by default, but rather create the vocabulary from all words which occur at least --word-min-count times. Specifying --num-words explicitly will still lead to a restricted
  vocabulary.

1.18.28

[1.18.28]

Changed

• Temporarily fixing the pyyaml version to 3.12 as version 4.1 introduced some backwards incompatible changes.

[1.18.27]

Fixed

• Fix silent failing of NDArray splits during inference by using a version that always returns a list. This was causing incorrect behavior when using lexicon restriction and batch inference with a single source factor.

[1.18.26]

Added

• ROUGE score evaluation. It can be used as the stopping criterion for tasks such as summarization.

[1.18.25]

Changed

• Update requirements to use MKL versions of MXNet for fast CPU operation.

[1.18.24]

Added

• Dockerfiles and convenience scripts for running fast_align to generate lexical tables.
  These tables can be used to create top-K lexicons for faster decoding via vocabulary selection (documentation).

Changed

• Updated default top-K lexicon size from 20 to 200.