hopwise.model.logits_processor

Common logits processor in recommender system

Classes

LogitsProcessor	Abstract base class for all logit processors that can be applied during generation.
LogitsProcessorList	This class can be used to create a list of [LogitsProcessor] to subsequently process a scores input tensor.
ConstrainedLogitsProcessorWordLevel	Force the last token to be one of the force_tokens if the total length is reached, in the path generation stage
PrefixConstrainedLogitsProcessorWordLevel	Force the last token to be one of the force_tokens if the total length is reached, in the path generation stage
PLMLogitsProcessorWordLevel	https://dl.acm.org/doi/pdf/10.1145/3485447.3511937

Module Contents

class hopwise.model.logits_processor.LogitsProcessor

Abstract base class for all logit processors that can be applied during generation. Copy of HuggingFace’s LogitsProcessor.

abstractmethod __call__(input_ids: torch.LongTensor, scores: torch.FloatTensor) → torch.FloatTensor

class hopwise.model.logits_processor.LogitsProcessorList

Bases: list

This class can be used to create a list of [LogitsProcessor] to subsequently process a scores input tensor. This class inherits from list and adds a specific __call__ method to apply each [LogitsProcessor] to the inputs. Copy of HuggingFace’s LogitsProcessorList.

__call__(input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) → torch.FloatTensor

Parameters:

• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) – Indices of input sequence tokens in the vocabulary. [What are input IDs?](../glossary#input-ids)

• scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) – Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

• kwargs (Dict[str, Any], optional) – Additional kwargs that are specific to a logits processor.

Returns:

The processed prediction scores.

Return type:

torch.FloatTensor of shape (batch_size, config.vocab_size)

class hopwise.model.logits_processor.ConstrainedLogitsProcessorWordLevel(tokenized_ckg, tokenized_used_ids, max_sequence_length, tokenizer, mask_cache_size=3 * 10**4, pos_candidates_cache_size=1 * 10**5, task=KnowledgeEvaluationType.REC, **kwargs)

Bases: LogitsProcessor

Force the last token to be one of the force_tokens if the total length is reached, in the path generation stage this means to limit the hop size. This is a word-level constraint, does not work with piece tokenizers. If task is link prediction (LP) logit processor forces last token to reachable ones

tokenized_ckg

tokenized_used_ids

max_sequence_length

tokenizer

bos_token_id

pos_candidates_cache

mask_cache

task

is_bos_token_in_input(input_ids)

Check if the input contains a BOS token. Checking the first sequence is enough.

__call__(input_ids, scores)

process_scores_rec(input_ids, idx)

Process each score based on input length and update mask list.

process_scores_lp(input_ids, idx)

Process each score based on input length or skip.

is_next_token_entity(input_ids)

get_current_key(input_ids, idx)

get_candidates_rec(key1, key2=None)

Parameters:

• key1

• key2 – if key2 is not None, it returns entity candidates, otherwise relation candidates

get_candidates_lp(key)

get_banned_mask(key, candidate_tokens)

Retrieve or cache the banned token mask for a specific key.

class hopwise.model.logits_processor.PrefixConstrainedLogitsProcessorWordLevel(tokenized_ckg, tokenized_used_ids, max_sequence_length, tokenizer, **kwargs)

Bases: ConstrainedLogitsProcessorWordLevel

Force the last token to be one of the force_tokens if the total length is reached, in the path generation stage this means to limit the hop size. This is a word-level constraint, does not work with piece tokenizers. If task is link prediction (LP) logit processor forces last token to reachable ones

mask_cache = None

__call__(input_ids, scores)

class hopwise.model.logits_processor.PLMLogitsProcessorWordLevel(tokenized_ckg, tokenized_used_ids, max_sequence_length, tokenizer, pos_candidates_cache_size=1 * 10**5, task=KnowledgeEvaluationType.REC, **kwargs)

Bases: LogitsProcessor

https://dl.acm.org/doi/pdf/10.1145/3485447.3511937 Constraint decoding strategy for PLM, it forces the model to generate alternatively entities and relations

tokenized_ckg

tokenized_used_ids

max_sequence_length

tokenizer

bos_token_id

pos_candidates_cache

task

entity_token_ids

relation_token_ids

__call__(input_ids, scores)

is_bos_token_in_input(input_ids)

Check if the input contains a BOS token. Checking the first sequence is enough.

is_next_token_entity(input_ids)

process_scores(input_ids, idx)

Process each score based on input length and update mask to allow only entities or only relations.