#!/usr/bin/env python3 # Copyright 2017-present, Facebook, Inc. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """DrQA Document Reader model""" import torch import torch.optim as optim import torch.nn.functional as F import numpy as np import logging import copy from .config import override_model_args from .rnn_reader import RnnDocReader logger = logging.getLogger(__name__) class DocReader(object): """High level model that handles intializing the underlying network architecture, saving, updating examples, and predicting examples. """ # -------------------------------------------------------------------------- # Initialization # -------------------------------------------------------------------------- def __init__(self, args, word_dict, feature_dict, state_dict=None, normalize=True): # Book-keeping. self.args = args self.word_dict = word_dict self.args.vocab_size = len(word_dict) self.feature_dict = feature_dict self.args.num_features = len(feature_dict) self.updates = 0 self.use_cuda = False self.parallel = False # Building network. If normalize if false, scores are not normalized # 0-1 per paragraph (no softmax). if args.model_type == 'rnn': self.network = RnnDocReader(args, normalize) else: raise RuntimeError('Unsupported model: %s' % args.model_type) # Load saved state if state_dict: # Load buffer separately if 'fixed_embedding' in state_dict: fixed_embedding = state_dict.pop('fixed_embedding') self.network.load_state_dict(state_dict) self.network.register_buffer('fixed_embedding', fixed_embedding) else: self.network.load_state_dict(state_dict) def expand_dictionary(self, words): """Add words to the DocReader dictionary if they do not exist. The underlying embedding matrix is also expanded (with random embeddings). Args: words: iterable of tokens to add to the dictionary. Output: added: set of tokens that were added. """ to_add = {self.word_dict.normalize(w) for w in words if w not in self.word_dict} # Add words to dictionary and expand embedding layer if len(to_add) > 0: logger.info('Adding %d new words to dictionary...' % len(to_add)) for w in to_add: self.word_dict.add(w) self.args.vocab_size = len(self.word_dict) logger.info('New vocab size: %d' % len(self.word_dict)) old_embedding = self.network.embedding.weight.data self.network.embedding = torch.nn.Embedding(self.args.vocab_size, self.args.embedding_dim, padding_idx=0) new_embedding = self.network.embedding.weight.data new_embedding[:old_embedding.size(0)] = old_embedding # Return added words return to_add def load_embeddings(self, words, embedding_file): """Load pretrained embeddings for a given list of words, if they exist. Args: words: iterable of tokens. Only those that are indexed in the dictionary are kept. embedding_file: path to text file of embeddings, space separated. """ words = {w for w in words if w in self.word_dict} logger.info('Loading pre-trained embeddings for %d words from %s' % (len(words), embedding_file)) embedding = self.network.embedding.weight.data # When normalized, some words are duplicated. (Average the embeddings). vec_counts = {} with open(embedding_file) as f: # Skip first line if of form count/dim. line = f.readline().rstrip().split(' ') if len(line) != 2: f.seek(0) for line in f: parsed = line.rstrip().split(' ') assert(len(parsed) == embedding.size(1) + 1) w = self.word_dict.normalize(parsed[0]) if w in words: vec = torch.Tensor([float(i) for i in parsed[1:]]) if w not in vec_counts: vec_counts[w] = 1 embedding[self.word_dict[w]].copy_(vec) else: logging.warning( 'WARN: Duplicate embedding found for %s' % w ) vec_counts[w] = vec_counts[w] + 1 embedding[self.word_dict[w]].add_(vec) for w, c in vec_counts.items(): embedding[self.word_dict[w]].div_(c) logger.info('Loaded %d embeddings (%.2f%%)' % (len(vec_counts), 100 * len(vec_counts) / len(words))) def tune_embeddings(self, words): """Unfix the embeddings of a list of words. This is only relevant if only some of the embeddings are being tuned (tune_partial = N). Shuffles the N specified words to the front of the dictionary, and saves the original vectors of the other N + 1:vocab words in a fixed buffer. Args: words: iterable of tokens contained in dictionary. """ words = {w for w in words if w in self.word_dict} if len(words) == 0: logger.warning('Tried to tune embeddings, but no words given!') return if len(words) == len(self.word_dict): logger.warning('Tuning ALL embeddings in dictionary') return # Shuffle words and vectors embedding = self.network.embedding.weight.data for idx, swap_word in enumerate(words, self.word_dict.START): # Get current word + embedding for this index curr_word = self.word_dict[idx] curr_emb = embedding[idx].clone() old_idx = self.word_dict[swap_word] # Swap embeddings + dictionary indices embedding[idx].copy_(embedding[old_idx]) embedding[old_idx].copy_(curr_emb) self.word_dict[swap_word] = idx self.word_dict[idx] = swap_word self.word_dict[curr_word] = old_idx self.word_dict[old_idx] = curr_word # Save the original, fixed embeddings self.network.register_buffer( 'fixed_embedding', embedding[idx + 1:].clone() ) def init_optimizer(self, state_dict=None): """Initialize an optimizer for the free parameters of the network. Args: state_dict: network parameters """ if self.args.fix_embeddings: for p in self.network.embedding.parameters(): p.requires_grad = False parameters = [p for p in self.network.parameters() if p.requires_grad] if self.args.optimizer == 'sgd': self.optimizer = optim.SGD(parameters, self.args.learning_rate, momentum=self.args.momentum, weight_decay=self.args.weight_decay) elif self.args.optimizer == 'adamax': self.optimizer = optim.Adamax(parameters, weight_decay=self.args.weight_decay) else: raise RuntimeError('Unsupported optimizer: %s' % self.args.optimizer) # -------------------------------------------------------------------------- # Learning # -------------------------------------------------------------------------- def update(self, ex): """Forward a batch of examples; step the optimizer to update weights.""" if not self.optimizer: raise RuntimeError('No optimizer set.') # Train mode self.network.train() # Transfer to GPU if self.use_cuda: inputs = [e if e is None else e.cuda(non_blocking=True) for e in ex[:5]] target_s = ex[5].cuda(non_blocking=True) target_e = ex[6].cuda(non_blocking=True) else: inputs = [e if e is None else e for e in ex[:5]] target_s = ex[5] target_e = ex[6] # Run forward score_s, score_e = self.network(*inputs) # Compute loss and accuracies loss = F.nll_loss(score_s, target_s) + F.nll_loss(score_e, target_e) # Clear gradients and run backward self.optimizer.zero_grad() loss.backward() # Clip gradients torch.nn.utils.clip_grad_norm_(self.network.parameters(), self.args.grad_clipping) # Update parameters self.optimizer.step() self.updates += 1 # Reset any partially fixed parameters (e.g. rare words) self.reset_parameters() return loss.item(), ex[0].size(0) def reset_parameters(self): """Reset any partially fixed parameters to original states.""" # Reset fixed embeddings to original value if self.args.tune_partial > 0: if self.parallel: embedding = self.network.module.embedding.weight.data fixed_embedding = self.network.module.fixed_embedding else: embedding = self.network.embedding.weight.data fixed_embedding = self.network.fixed_embedding # Embeddings to fix are the last indices offset = embedding.size(0) - fixed_embedding.size(0) if offset >= 0: embedding[offset:] = fixed_embedding # -------------------------------------------------------------------------- # Prediction # -------------------------------------------------------------------------- def predict(self, ex, candidates=None, top_n=1, async_pool=None): """Forward a batch of examples only to get predictions. Args: ex: the batch candidates: batch * variable length list of string answer options. The model will only consider exact spans contained in this list. top_n: Number of predictions to return per batch element. async_pool: If provided, non-gpu post-processing will be offloaded to this CPU process pool. Output: pred_s: batch * top_n predicted start indices pred_e: batch * top_n predicted end indices pred_score: batch * top_n prediction scores If async_pool is given, these will be AsyncResult handles. """ # Eval mode self.network.eval() # Transfer to GPU if self.use_cuda: inputs = [e if e is None else e.cuda(non_blocking=True) for e in ex[:5]] else: inputs = [e for e in ex[:5]] # Run forward with torch.no_grad(): score_s, score_e = self.network(*inputs) # Decode predictions score_s = score_s.data.cpu() score_e = score_e.data.cpu() if candidates: args = (score_s, score_e, candidates, top_n, self.args.max_len) if async_pool: return async_pool.apply_async(self.decode_candidates, args) else: return self.decode_candidates(*args) else: args = (score_s, score_e, top_n, self.args.max_len) if async_pool: return async_pool.apply_async(self.decode, args) else: return self.decode(*args) @staticmethod def decode(score_s, score_e, top_n=1, max_len=None): """Take argmax of constrained score_s * score_e. Args: score_s: independent start predictions score_e: independent end predictions top_n: number of top scored pairs to take max_len: max span length to consider """ pred_s = [] pred_e = [] pred_score = [] max_len = max_len or score_s.size(1) for i in range(score_s.size(0)): # Outer product of scores to get full p_s * p_e matrix scores = torch.ger(score_s[i], score_e[i]) # Zero out negative length and over-length span scores scores.triu_().tril_(max_len - 1) # Take argmax or top n scores = scores.numpy() scores_flat = scores.flatten() if top_n == 1: idx_sort = [np.argmax(scores_flat)] elif len(scores_flat) < top_n: idx_sort = np.argsort(-scores_flat) else: idx = np.argpartition(-scores_flat, top_n)[0:top_n] idx_sort = idx[np.argsort(-scores_flat[idx])] s_idx, e_idx = np.unravel_index(idx_sort, scores.shape) pred_s.append(s_idx) pred_e.append(e_idx) pred_score.append(scores_flat[idx_sort]) return pred_s, pred_e, pred_score @staticmethod def decode_candidates(score_s, score_e, candidates, top_n=1, max_len=None): """Take argmax of constrained score_s * score_e. Except only consider spans that are in the candidates list. """ pred_s = [] pred_e = [] pred_score = [] for i in range(score_s.size(0)): # Extract original tokens stored with candidates tokens = candidates[i]['input'] cands = candidates[i]['cands'] if not cands: # try getting from globals? (multiprocessing in pipeline mode) from ..pipeline.drqa import PROCESS_CANDS cands = PROCESS_CANDS if not cands: raise RuntimeError('No candidates given.') # Score all valid candidates found in text. # Brute force get all ngrams and compare against the candidate list. max_len = max_len or len(tokens) scores, s_idx, e_idx = [], [], [] for s, e in tokens.ngrams(n=max_len, as_strings=False): span = tokens.slice(s, e).untokenize() if span in cands or span.lower() in cands: # Match! Record its score. scores.append(score_s[i][s] * score_e[i][e - 1]) s_idx.append(s) e_idx.append(e - 1) if len(scores) == 0: # No candidates present pred_s.append([]) pred_e.append([]) pred_score.append([]) else: # Rank found candidates scores = np.array(scores) s_idx = np.array(s_idx) e_idx = np.array(e_idx) idx_sort = np.argsort(-scores)[0:top_n] pred_s.append(s_idx[idx_sort]) pred_e.append(e_idx[idx_sort]) pred_score.append(scores[idx_sort]) return pred_s, pred_e, pred_score # -------------------------------------------------------------------------- # Saving and loading # -------------------------------------------------------------------------- def save(self, filename): if self.parallel: network = self.network.module else: network = self.network state_dict = copy.copy(network.state_dict()) if 'fixed_embedding' in state_dict: state_dict.pop('fixed_embedding') params = { 'state_dict': state_dict, 'word_dict': self.word_dict, 'feature_dict': self.feature_dict, 'args': self.args, } try: torch.save(params, filename) except BaseException: logger.warning('WARN: Saving failed... continuing anyway.') def checkpoint(self, filename, epoch): if self.parallel: network = self.network.module else: network = self.network params = { 'state_dict': network.state_dict(), 'word_dict': self.word_dict, 'feature_dict': self.feature_dict, 'args': self.args, 'epoch': epoch, 'optimizer': self.optimizer.state_dict(), } try: torch.save(params, filename) except BaseException: logger.warning('WARN: Saving failed... continuing anyway.') @staticmethod def load(filename, new_args=None, normalize=True): logger.info('Loading model %s' % filename) saved_params = torch.load( filename, map_location=lambda storage, loc: storage ) word_dict = saved_params['word_dict'] feature_dict = saved_params['feature_dict'] state_dict = saved_params['state_dict'] args = saved_params['args'] if new_args: args = override_model_args(args, new_args) return DocReader(args, word_dict, feature_dict, state_dict, normalize) @staticmethod def load_checkpoint(filename, normalize=True): logger.info('Loading model %s' % filename) saved_params = torch.load( filename, map_location=lambda storage, loc: storage ) word_dict = saved_params['word_dict'] feature_dict = saved_params['feature_dict'] state_dict = saved_params['state_dict'] epoch = saved_params['epoch'] optimizer = saved_params['optimizer'] args = saved_params['args'] model = DocReader(args, word_dict, feature_dict, state_dict, normalize) model.init_optimizer(optimizer) return model, epoch # -------------------------------------------------------------------------- # Runtime # -------------------------------------------------------------------------- def cuda(self): self.use_cuda = True self.network = self.network.cuda() def cpu(self): self.use_cuda = False self.network = self.network.cpu() def parallelize(self): """Use data parallel to copy the model across several gpus. This will take all gpus visible with CUDA_VISIBLE_DEVICES. """ self.parallel = True self.network = torch.nn.DataParallel(self.network)