# coding=utf-8 # Copyright 2020 The TensorFlow Datasets Authors. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # Lint as: python3 """SCAN tasks with various different splits.""" from __future__ import absolute_import from __future__ import division from __future__ import print_function import json import os import tensorflow.compat.v2 as tf import tensorflow_datasets.public_api as tfds _CITATION = """ @inproceedings{Lake2018GeneralizationWS, title={Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks}, author={Brenden M. Lake and Marco Baroni}, booktitle={ICML}, year={2018}, url={https://arxiv.org/pdf/1711.00350.pdf}, } @inproceedings{Keysers2020, title={Measuring Compositional Generalization: A Comprehensive Method on Realistic Data}, author={Daniel Keysers and Nathanael Sch\"{a}rli and Nathan Scales and Hylke Buisman and Daniel Furrer and Sergii Kashubin and Nikola Momchev and Danila Sinopalnikov and Lukasz Stafiniak and Tibor Tihon and Dmitry Tsarkov and Xiao Wang and Marc van Zee and Olivier Bousquet}, note={Additional citation for MCD splits}, booktitle={ICLR}, year={2020}, url={https://arxiv.org/abs/1912.09713.pdf}, } """ _DESCRIPTION = """SCAN tasks with various splits. SCAN is a set of simple language-driven navigation tasks for studying compositional learning and zero-shot generalization. Most splits are described at https://github.com/brendenlake/SCAN. For the MCD splits please see https://arxiv.org/abs/1912.09713.pdf. Basic usage: ``` data = tfds.load('scan/length') ``` More advanced example: ``` data = tfds.load( 'scan', builder_kwargs=dict( config=tfds.text.ScanConfig( name='simple_p8', directory='simple_split/size_variations'))) ``` """ _DATA_URL = 'https://github.com/brendenlake/SCAN/archive/master.zip' _MCD_SPLITS_URL = ( 'https://storage.googleapis.com/cfq_dataset/scan-splits.tar.gz') class ScanConfig(tfds.core.BuilderConfig): """BuilderConfig for SCAN. Splits can be read in two formats: 1) As a pair of train and test files where each file contains one example input and output per line. 2) With a 'splitfile' which contains for each split the indices into the full (unsplit) dataset. """ @tfds.core.disallow_positional_args def __init__(self, name, directory=None, splitfile=None, **kwargs): """BuilderConfig for SCAN. Args: name: Unique name of the split. directory: Which subdirectory to read the data files from. splitfile: If set the samples are read from the original archive (tasks.txt) but the splits are created using this index file. **kwargs: keyword arguments forwarded to super. """ # Version history: super(ScanConfig, self).__init__( name=name, version=tfds.core.Version('1.1.1'), description=_DESCRIPTION, **kwargs) self.splitfile = splitfile if 'mcd' in name: self.splitfile = name + '.json' if self.splitfile and directory is None: self.directory = '' elif directory is None: self.directory = name + '_split' else: self.directory = directory _COMMANDS = 'commands' _ACTIONS = 'actions' class Scan(tfds.core.GeneratorBasedBuilder): """SCAN task / splits as proposed by Brenden M. Lake and Marco Baroni.""" BUILDER_CONFIGS = [ ScanConfig(name='simple'), ScanConfig(name='addprim_jump', directory='add_prim_split'), ScanConfig(name='addprim_turn_left', directory='add_prim_split'), ScanConfig(name='filler_num0', directory='filler_split'), ScanConfig(name='filler_num1', directory='filler_split'), ScanConfig(name='filler_num2', directory='filler_split'), ScanConfig(name='filler_num3', directory='filler_split'), ScanConfig(name='length'), ScanConfig(name='template_around_right', directory='template_split'), ScanConfig(name='template_jump_around_right', directory='template_split'), ScanConfig(name='template_opposite_right', directory='template_split'), ScanConfig(name='template_right', directory='template_split'), ScanConfig(name='mcd1'), ScanConfig(name='mcd2'), ScanConfig(name='mcd3'), ] def _info(self): return tfds.core.DatasetInfo( builder=self, description=_DESCRIPTION, features=tfds.features.FeaturesDict({ _COMMANDS: tfds.features.Text(), _ACTIONS: tfds.features.Text(), }), supervised_keys=(_COMMANDS, _ACTIONS), homepage='https://github.com/brendenlake/SCAN', citation=_CITATION, ) def _split_generators(self, dl_manager): """Returns SplitGenerators.""" data_dir = dl_manager.download_and_extract( tfds.download.Resource( url=_DATA_URL, # Specify extract method manually as filename reported by github.com # misses the .zip extension so auto-detection doesn't work. extract_method=tfds.download.ExtractMethod.ZIP)) data_dir = os.path.join(data_dir, 'SCAN-master', self.builder_config.directory) split = self.builder_config.name splitfile = self.builder_config.splitfile if 'mcd' in split: split_dir = dl_manager.download_and_extract(_MCD_SPLITS_URL) split_dir = os.path.join(split_dir, 'scan-splits') splitfile = os.path.join(split_dir, splitfile) if splitfile: kwargs = { 'datapath': os.path.join(data_dir, 'tasks.txt'), 'splitpath': splitfile } train_kwargs = kwargs.copy() train_kwargs['splitname'] = 'train' test_kwargs = kwargs.copy() test_kwargs['splitname'] = 'test' else: train_kwargs = { 'datapath': os.path.join(data_dir, 'tasks_train_' + split + '.txt') } test_kwargs = { 'datapath': os.path.join(data_dir, 'tasks_test_' + split + '.txt') } return [ tfds.core.SplitGenerator( name=tfds.Split.TRAIN, gen_kwargs=train_kwargs), tfds.core.SplitGenerator(name=tfds.Split.TEST, gen_kwargs=test_kwargs) ] def _read_examples(self, datapath): with tf.io.gfile.GFile(datapath) as infile: for i, line in enumerate(infile): if not line.startswith('IN: '): continue # Chop the prefix and split string between input and output commands, actions = line[len('IN: '):].strip().split(' OUT: ', 1) yield i, {_COMMANDS: commands, _ACTIONS: actions} def _generate_examples(self, datapath, splitpath=None, splitname=None): """Yields examples.""" if splitpath: all_samples = list(self._read_examples(datapath)) with tf.io.gfile.GFile(splitpath) as infile: split = json.load(infile) for idx in split[splitname + 'Idxs']: yield all_samples[idx] else: for example in self._read_examples(datapath): yield example