fibber.paraphrase_strategies.asrs_strategy module

class fibber.paraphrase_strategies.asrs_strategy.ASRSStrategy(arg_dict, dataset_name, strategy_gpu_id, output_dir, metric_bundle, field)[source]

Bases: fibber.paraphrase_strategies.strategy_base.StrategyBase

Initialize the paraphrase_strategies.

This function initialize the self._strategy_config, self._metric_bundle, self._device, self._output_dir, self._dataset_name.

You should not overwrite this function.

  • self._strategy_config (dict): a dictionary that stores the strategy name and all hyperparameter values. The dict is also saved to the results.

  • self._metric_bundle (MetricBundle): the metrics that will be used to evaluate paraphrases. Strategies can compute metrics during paraphrasing.

  • self._device (torch.Device): any computation that requires a GPU accelerator should use this device.

  • self._output_dir (str): the dir name where the strategy can save files.

  • self._dataset_name (str): the dataset name.

Parameters

  • arg_dict (dict) – all args load from command line.

  • dataset_name (str) – the name of the dataset.

  • strategy_gpu_id (int) – the gpu id to run the strategy.

  • output_dir (str) – a directory to save any models or temporary files.

  • metric_bundle (MetricBundle) – a MetricBundle object.

fit(trainset)[source]

Fit the paraphrase strategy on a training set.

Parameters

trainset (dict) – a fibber dataset.

paraphrase_example(data_record, n, early_stop=False)[source]

Paraphrase one data record.

This function should be overwritten by subclasses. When overwriting this class, you can use self._strategy_config, self._metric_bundle, self._device, self._output_dir, and self._dataset_name

Parameters

  • data_record (dict) – a dict storing one data of a dataset.

  • n (int) – number of paraphrases.

Returns

A list contain at most n strings.

Return type

([str,])

fibber.paraphrase_strategies.asrs_strategy.all_accept_criteria(candidate_ids, stats, **kwargs)[source]

Always accept proposed words.

Parameters

  • candidate_ids (torch.Tensor) – proposed word ids in this sampling step with size (batch_size, pos_ed-pos_st).

  • stats (dict) – a dict to keep track the accept rate.

Returns

(np.array, None)

np.array is the same as candidate_ids. None means this criteria does not have any state.

fibber.paraphrase_strategies.asrs_strategy.allow_list_constraint(allow_list, **kwargs)[source]
fibber.paraphrase_strategies.asrs_strategy.clf_criteria_score(origin, paraphrases, data_record, field, clf_metric, clf_weight)[source]
fibber.paraphrase_strategies.asrs_strategy.joint_weighted_criteria(tokenizer, data_record, field, origin, batch_tensor, pos_st, pos_ed, previous_ids, candidate_ids, sim_metric, sim_threshold, sim_weight, clf_metric, clf_weight, ppl_metric, ppl_weight, burnin_weight, stats, state, device, seq_len, log_prob_previous_ids, log_prob_candidate_ids, **kwargs)[source]

Accept or reject candidate word using the joint weighted criteria.

Parameters

  • tokenizer (transformers.BertTokenizer) – a bert tokenizer.

  • data_record (dict) – the data record dict.

  • field (str) – the field to rewritten.

  • origin (str) – original text. Same as data_record[field].

  • batch_tensor (torch.Tensor) – tensor of a batch of text with size (batch_size, L).

  • pos_st (int) – the start position of sampling (include).

  • pos_ed (int) – the end position of sampling (exclude).

  • previous_ids (torch.Tensor) – word ids before current step of sampling with size (batch_size, pos_ed-pos_st).

  • candidate_ids (torch.Tensor) – proposed word ids in this sampling step with size (batch_size, pos_ed-pos_st).

  • sim_metric (USESimilarityMetric) – a universal sentence encoder metric object.

  • sim_threshold (float) – the universal sentence encoder similarity threshold.

  • sim_weight (float) – the weight for USE criteria score.

  • clf_metric (BertClassifier) – a BertClassifier metric.

  • clf_weight (float) – the weight for BERT criteria score.

  • ppl_metric (GPT2PerplexityMetric) – a GPT2PerplexityMetric metric.

  • ppl_weight (float) – the weight for GPT2 criteria score.

  • burnin_weight (float) – the discount factor.

  • stats (dict) – a dict to keep track the accept rate.

  • state (np.array) – the state is criteria score from the previous iteration.

  • seq_len (np.array) – the valid length for each sentence in the batch.

  • device (torch.Device) – the device that batch_tensor is on.

Returns

(np.array, np.array)
a 2-D int array of size batch_size, pos_ed - pos_st. Each row i is

either previous_ids[i, :] if rejected, or candidate_ids[i, :] if accepted.

a 1-D float array of criteria score.

fibber.paraphrase_strategies.asrs_strategy.none_constraint(**kwargs)[source]
fibber.paraphrase_strategies.asrs_strategy.ppl_criteria_score(origin, paraphrases, ppl_metric, ppl_weight)[source]

Estimate the score of a sentence using USE.

Parameters

  • origin (str) – original sentence.

  • paraphrases ([str]) – a list of paraphrase_list.

  • ppl_metric (GPT2PerplexityMetric) – a GPT2PerplexityMetric metric object.

  • ppl_weight (float) – the weight parameter for the criteria.

Returns

a numpy array of size (batch_size,). All entries <=0.

Return type

(np.array)

fibber.paraphrase_strategies.asrs_strategy.process_text(text, patterns)[source]

Processing the text using regex patterns.

Parameters

  • text (str) – the str to be post processed.

  • patterns (list) – a list of substitution patterns.

fibber.paraphrase_strategies.asrs_strategy.sample_word_from_logits(logits, temperature=1.0, top_k=0)[source]

Sample a word from a distribution.

Parameters

  • logits (torch.Tensor) – tensor of logits with size (batch_size, vocab_size).

  • temperature (float) – the temperature of softmax. The PMF is softmax(logits/temperature).

  • top_k (int) – if k>0, only sample from the top k most probable words.

fibber.paraphrase_strategies.asrs_strategy.sim_criteria_score(origin, paraphrases, sim_metric, sim_threshold, sim_weight)[source]

Estimate the score of a sentence using USE.

Parameters

  • origin (str) – original sentence.

  • paraphrases ([str]) – a list of paraphrase_list.

  • sim_metric (MetricBase) – a similarity metric object.

  • sim_threshold (float) – the universal sentence encoder similarity threshold.

  • sim_weight (float) – the weight parameter for the criteria.

Returns

a numpy array of size (batch_size,). All entries <=0.

Return type

(np.array)

fibber.paraphrase_strategies.asrs_strategy.tostring(tokenizer, seq)[source]

Convert a sequence of word ids to a sentence. The post prossing is applied.

Parameters

  • tokenizer (transformers.BertTokenizer) – a BERT tokenizer.

  • seq (list) – a list-like sequence of word ids.

fibber.paraphrase_strategies.asrs_strategy.wpe_constraint(target_emb, word_embs, batch_tensor, pos_st, pos_ed, wpe_threshold, wpe_weight, attention_mask_paraphrase_text_only, **kwargs)[source]

previous

fibber.paraphrase_strategies package

next

fibber.paraphrase_strategies.asrs_utils_wpe module