bjascob/amrlib

📋 Changes

• Fix issue in models/parse_xfm/inference.py and trainer.py that for transformers v5, the

📋 Changes

• Add generate_xfm and removed generate_t5/generate_t5wtense
• Fix parse_gsii annotator multiprocessing error under Windows and Mac
• Add "quiet" option to the parse_gsii model

📋 Changes

• Fix missing resources directory for the parse_spring model in the pypi package
• No code changes except setup scripts.

📋 Changes

• Add parse_xfm model, configs and training scripts (this replaces parse_t5)
• Add WikiAdderBlink and update parse_xmf / parse_spring train code to use it.
• Added tensorboard smatch logging for parse_xfm training
• Added smatch log redirection to utils/logging.py
• Fix interrogative', 'imperative', 'expressive' as possible nodes

📋 Changes

• Added model_parse_spring
• Updated code for model_parse_t5 (faster training and inference)
• Fix dynamic load of model_stog when running amr_view (for transformers 4.4.2)
• Retest with torch 1.10.0 and transformers 4.12.3
• Add additional tests for parse models

📋 Changes

• Fix faa_aligner fail on empty strings
• Fix potential crash from line-feed in parse_gsii inference
• Fix potential crash from deserialization in parse_t5 inference
• Add isi aligner scripts and documentation
• Update faa_aligner no binaries error to be more explicit
• Retest with spaCy 3.x and torch 1.8

📋 Changes

• Added Fast Align Algorithm (FAA) Aligner
• Fix potential crash from None in penman triple in rbw_aligner
• Add alignment rule for "me" in rbw_aligner
• Added rule not to repeat ARGx for nodes in model_parse_gsii
• Changed test for rbw aligner to run against ISI hand alignments

📋 Changes

• Add model_generate_t5wtense
• Remove eos token from input in T5 modules, trainer/inference for parse_t5 and generate_t5
• Fix clip detection in generate_t5/parse_t5 inference.py
• Add remove_surface_alignments to RBWAligner

📋 Changes

• Added model_parse_t5 (stog) model based on the HuggingFace T5 transformer
• Added dynamic loading infrastructure
• Fix AMRPlot single attribute box for attribs of the same name

📋 Changes

• Fix parse model for inference compatibility with transformers 4.0.0
• Filter warning message in generate model inference from transformers 4.0.0
• Fixed generate model training issue with transformers 4.0.0
• Fixed parse model training with torch 1.7

📋 Changes in Release 0.2.0 ####

• Added Rule Based Word Aligner
• Changed AMRPlot() to get render_fn programatically for Windows compatibility
• Changed AMRPlot() to take a "format" param and changed default to pdf
• Added config options in amr_view to allow setting plotted graph format (ie.. pdf)
• Changed model/parse_gsii/post_processor.py to graph_builder.py with rules and filtering
• Documentation updates

📋 Changes in Release 0.1.1 ####

• Specify np.int64 in parse_gsii.data_loader for Windows Compatibility
• Linux numpy arrays default to 64-bit where Windows defaults to 32
• Update generate_t5.inference.generate() to filter out meta-data to avoid generating garbage
• Update model parse_gsii to deserialize to RAM so GPU is not required to run
• Removed all symlinks (in scripts for setup_run_dir.py file) and replace with the actual file
• Changed spacy model in defaults.py to be en_core_web_sm
• Updated requirements.txt / setup.py
• Documentation updates