def init_meters(self, args): self.meters = OrderedDict() self.meters['train_loss'] = AverageMeter() self.meters['train_nll_loss'] = AverageMeter() self.meters['valid_loss'] = AverageMeter() self.meters['valid_nll_loss'] = AverageMeter() self.meters['wps'] = TimeMeter() # words per second self.meters['ups'] = TimeMeter() # updates per second self.meters['wpb'] = AverageMeter() # words per batch self.meters['bsz'] = AverageMeter() # sentences per batch self.meters['gnorm'] = AverageMeter() # gradient norm self.meters['clip'] = AverageMeter() # % of updates clipped self.meters['oom'] = AverageMeter() # out of memory if args.fp16: self.meters['loss_scale'] = AverageMeter() # dynamic loss scale self.meters['wall'] = TimeMeter() # wall time in seconds self.meters['train_wall'] = StopwatchMeter( ) # train wall time in seconds
def init_meters(self, args): self.meters = OrderedDict() self.meters["train_loss"] = AverageMeter() self.meters["train_nll_loss"] = AverageMeter() self.meters["valid_loss"] = AverageMeter() self.meters["valid_nll_loss"] = AverageMeter() self.meters["wps"] = TimeMeter() # words per second self.meters["ups"] = TimeMeter() # updates per second self.meters["wpb"] = AverageMeter() # words per batch self.meters["bsz"] = AverageMeter() # sentences per batch self.meters["gnorm"] = AverageMeter() # gradient norm self.meters["clip"] = AverageMeter() # % of updates clipped self.meters["oom"] = AverageMeter() # out of memory if args.fp16: self.meters["loss_scale"] = AverageMeter() # dynamic loss scale self.meters["wall"] = TimeMeter() # wall time in seconds self.meters["train_wall"] = StopwatchMeter( ) # train wall time in seconds
def __init__(self, args, task, model, criterion, dummy_batch): if not torch.cuda.is_available(): raise NotImplementedError('Training on CPU is not supported') self.args = args self.task = task # copy model and criterion to current device self.criterion = criterion.cuda() if args.fp16: self._model = model.half().cuda() else: self._model = model.cuda() # initialize meters self.meters = OrderedDict() self.meters['train_loss'] = AverageMeter() self.meters['train_nll_loss'] = AverageMeter() self.meters['valid_loss'] = AverageMeter() self.meters['valid_nll_loss'] = AverageMeter() self.meters['train_acc'] = AverageMeter() self.meters['valid_acc'] = AverageMeter() self.meters['wps'] = TimeMeter() # words per second self.meters['ups'] = TimeMeter() # updates per second self.meters['wpb'] = AverageMeter() # words per batch self.meters['bsz'] = AverageMeter() # sentences per batch self.meters['gnorm'] = AverageMeter() # gradient norm self.meters['clip'] = AverageMeter() # % of updates clipped self.meters['oom'] = AverageMeter() # out of memory if args.fp16: self.meters['loss_scale'] = AverageMeter() # dynamic loss scale self.meters['wall'] = TimeMeter() # wall time in seconds self.meters['train_wall'] = StopwatchMeter( ) # train wall time in seconds self._dummy_batch = dummy_batch self._num_updates = 0 self._optim_history = None self._optimizer = None self._wrapped_model = None
def init_meters(self, args): self.meters = OrderedDict() self.meters['train_loss'] = AverageMeter() self.meters['train_nll_loss'] = AverageMeter() self.meters['valid_loss'] = AverageMeter() self.meters['valid_nll_loss'] = AverageMeter() self.meters['wps'] = TimeMeter() # words per second self.meters['ups'] = TimeMeter() # updates per second self.meters['wpb'] = AverageMeter() # words per batch self.meters['bsz'] = AverageMeter() # sentences per batch self.meters['gnorm'] = AverageMeter() # gradient norm self.meters['clip'] = AverageMeter() # % of updates clipped self.meters['oom'] = AverageMeter() # out of memory if args.fp16: self.meters['loss_scale'] = AverageMeter() # dynamic loss scale self.meters['wall'] = TimeMeter() # wall time in seconds self.meters['train_wall'] = StopwatchMeter( ) # train wall time in seconds # add_norm for i in range(args.encoder_layers): self.meters['enc{}_attn_input_gnorm'.format( i)] = AverageMeter() # gradient norm self.meters['enc{}_ffn_fc1_gnorm'.format(i)] = AverageMeter()
def main(args, init_distributed=False): utils.import_user_module(args) assert args.max_tokens is not None or args.max_sentences is not None, \ 'Must specify batch size either with --max-tokens or --max-sentences' # Initialize CUDA and distributed training if torch.cuda.is_available() and not args.cpu: torch.cuda.set_device(args.device_id) np.random.seed(args.seed) torch.manual_seed(args.seed) if init_distributed: args.distributed_rank = distributed_utils.distributed_init(args) if distributed_utils.is_master(args): checkpoint_utils.verify_checkpoint_directory(args.save_dir) # Print args print(args) # Setup task, e.g., translation, language modeling, etc. task = tasks.setup_task(args) # Load valid dataset (we load training data below, based on the latest checkpoint) for valid_sub_split in args.valid_subset.split(','): task.load_dataset(valid_sub_split, combine=False, epoch=0) # Build model and criterion model = task.build_model(args) criterion = task.build_criterion(args) print(model) print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__)) print('| num. model params: {} (num. trained: {})'.format( sum(p.numel() for p in model.parameters()), sum(p.numel() for p in model.parameters() if p.requires_grad), )) # Build trainer trainer = Trainer(args, task, model, criterion) print('| training on {} GPUs'.format(args.distributed_world_size)) print('| max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) # Load the latest checkpoint if one is available and restore the # corresponding train iterator extra_state, epoch_itr = checkpoint_utils.load_checkpoint(args, trainer) # Train until the learning rate gets too small max_epoch = args.max_epoch or math.inf max_update = args.max_update or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() valid_subsets = args.valid_subset.split(',') while lr > args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates( ) < max_update: # train for one epoch train(args, trainer, task, epoch_itr) if not args.disable_validation and epoch_itr.epoch % args.validate_interval == 0: valid_losses = validate(args, trainer, task, epoch_itr, valid_subsets) else: valid_losses = [None] # only use first validation loss to update the learning rate lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0]) # save checkpoint if epoch_itr.epoch % args.save_interval == 0: checkpoint_utils.save_checkpoint(args, trainer, epoch_itr, valid_losses[0]) reload_dataset = ':' in getattr(args, 'data', '') # sharded data: get train iterator for next epoch epoch_itr = trainer.get_train_iterator(epoch_itr.epoch, load_dataset=reload_dataset) train_meter.stop() print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args): assert args.path is not None, '--path required for generation!' assert not args.sampling or args.nbest == args.beam, \ '--sampling requires --nbest to be equal to --beam' utils.import_user_module(args) if args.max_tokens is None and args.max_sentences is None: args.max_tokens = 12000 print(args) use_cuda = torch.cuda.is_available() and not args.cpu # Load dataset split task = tasks.setup_task(args) task.load_dataset(args.gen_subset) # Load ensemble print('| loading model(s) from {}'.format(args.path)) model_paths = args.path.split(':') models, _model_args = checkpoint_utils.load_model_ensemble( model_paths, arg_overrides=eval(args.model_overrides), task=task, ) # Optimize ensemble for generation for model in models: # model.make_generation_fast_() if args.fp16: model.half() if use_cuda: model.cuda() # Load dataset (possibly sharded) itr = task.get_batch_iterator( dataset=task.dataset(args.gen_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=utils.resolve_max_positions( task.max_positions(), *[model.max_positions() for model in models]), ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test, required_batch_size_multiple=args.required_batch_size_multiple, num_shards=args.num_shards, shard_id=args.shard_id, num_workers=args.num_workers, ).next_epoch_itr(shuffle=False) # Initialize generator gen_timer = StopwatchMeter() generator = task.build_generator(args) # Generate and compute score coco = task.dataset(args.gen_subset).coco iou_types = ['bbox'] scorer = CocoEvaluator(coco, iou_types) num_images = 0 with progress_bar.build_progress_bar( args, itr, prefix='inference on \'{}\' subset'.format(args.gen_subset), no_progress_bar='simple', ) as progress: wps_meter = TimeMeter() for sample in progress: sample = utils.move_to_cuda(sample) if use_cuda else sample gen_timer.start() hypos = task.inference_step(generator, models, sample) num_generated_boxes = sum(len(h['scores']) for h in hypos) gen_timer.stop(num_generated_boxes) result = {} for i, sample_id in enumerate(sample['id'].tolist()): result[sample_id] = hypos[i] scorer.update(result) wps_meter.update(num_generated_boxes) progress.log({'wps': round(wps_meter.avg)}) num_images += sample['nsentences'] print( '| Detected {} images ({} tokens) in {:.1f}s ({:.2f} images/s, {:.2f} tokens/s)' .format(num_images, gen_timer.n, gen_timer.sum, num_images / gen_timer.sum, 1. / gen_timer.avg)) # gather the stats from all processes scorer.synchronize_between_processes() # accumulate predictions from all images scorer.accumulate() scorer.summarize() return scorer
def _generate_score(models, args, task, dataset, optimize=True): use_cuda = torch.cuda.is_available() and not args.cpu # Load ensemble if not args.quiet: print("| loading model(s) from {}".format(", ".join(args.path.split(":")))) # Optimize ensemble for generation if optimize: for model in models: model.make_generation_fast_( beamable_mm_beam_size=None if args.no_beamable_mm else args.beam, need_attn=True, ) translator = build_sequence_generator(args, task, models) # Load alignment dictionary for unknown word replacement # (None if no unknown word replacement, empty if no path to align dictionary) align_dict = utils.load_align_dict(args.replace_unk) # Keep track of translations # Initialize with empty translations # and zero probs scores translated_sentences = [""] * len(dataset) translated_scores = [0.0] * len(dataset) # Generate and compute BLEU score dst_dict = task.target_dictionary scorer = bleu.Scorer(dst_dict.pad(), dst_dict.eos(), dst_dict.unk()) itr = get_eval_itr(args, models, task, dataset) num_sentences = 0 translation_samples = [] with progress_bar.build_progress_bar(args, itr) as t: wps_meter = TimeMeter() gen_timer = StopwatchMeter() translations = translator.generate_batched_itr( t, maxlen_a=args.max_len_a, maxlen_b=args.max_len_b, cuda=use_cuda, timer=gen_timer, prefix_size=1 if pytorch_translate_data.is_multilingual(args) else 0, ) if pytorch_translate_data.is_multilingual(args): first_best_translations = _iter_first_best_multilingual else: first_best_translations = _iter_first_best_bilingual for trans_info in first_best_translations( args, task, dataset, translations, align_dict ): scorer.add(trans_info.target_tokens, trans_info.hypo_tokens) translated_sentences[trans_info.sample_id] = trans_info.hypo_str translated_scores[trans_info.sample_id] = trans_info.hypo_score translation_samples.append( collections.OrderedDict( { "sample_id": trans_info.sample_id.item(), "src_str": trans_info.src_str, "target_str": trans_info.target_str, "hypo_str": trans_info.hypo_str, } ) ) wps_meter.update(trans_info.src_tokens.size(0)) t.log({"wps": round(wps_meter.avg)}) num_sentences += 1 # If applicable, save the translations to the output file # For eg. external evaluation if getattr(args, "translation_output_file", False): with open(args.translation_output_file, "w") as out_file: for hypo_str in translated_sentences: print(hypo_str, file=out_file) if getattr(args, "translation_probs_file", False): with open(args.translation_probs_file, "w") as out_file: for hypo_score in translated_scores: print(np.exp(hypo_score), file=out_file) return scorer, num_sentences, gen_timer, translation_samples
def main(args): check_args(args) import_user_module(args) if args.max_tokens is None and args.max_sentences is None: args.max_tokens = 30000 logger.info(args) use_cuda = torch.cuda.is_available() and not args.cpu # Load dataset splits task = tasks.setup_task(args) task.load_dataset(args.gen_subset) logger.info("| {} {} {} examples".format( args.data, args.gen_subset, len(task.dataset(args.gen_subset)))) # Set dictionary tgt_dict = task.target_dictionary logger.info("| decoding with criterion {}".format(args.criterion)) # Load ensemble logger.info("| loading model(s) from {}".format(args.path)) models, criterions, _model_args = load_models_and_criterions( args.path.split(":"), arg_overrides=eval(args.model_overrides), # noqa task=task, ) optimize_models(args, use_cuda, models) # hack to pass transitions to W2lDecoder if args.criterion == "asg_loss": trans = criterions[0].asg.trans.data args.asg_transitions = torch.flatten(trans).tolist() # Load dataset (possibly sharded) itr = get_dataset_itr(args, task) # Initialize generator gen_timer = StopwatchMeter() generator = task.build_generator(args) num_sentences = 0 if not os.path.exists(args.results_path): os.makedirs(args.results_path) sp = spm.SentencePieceProcessor() sp.Load(os.path.join(args.data, "spm.model")) res_files = prepare_result_files(args) with progress_bar.build_progress_bar(args, itr) as t: wps_meter = TimeMeter() for sample in t: sample = utils.move_to_cuda(sample) if use_cuda else sample if "net_input" not in sample: continue prefix_tokens = None if args.prefix_size > 0: prefix_tokens = sample["target"][:, :args.prefix_size] gen_timer.start() hypos = task.inference_step(generator, models, sample, prefix_tokens) num_generated_tokens = sum(len(h[0]["tokens"]) for h in hypos) gen_timer.stop(num_generated_tokens) for i, sample_id in enumerate(sample["id"].tolist()): speaker = task.dataset( args.gen_subset).speakers[int(sample_id)] id = task.dataset(args.gen_subset).ids[int(sample_id)] target_tokens = (utils.strip_pad(sample["target"][i, :], tgt_dict.pad()).int().cpu()) # Process top predictions process_predictions(args, hypos[i], sp, tgt_dict, target_tokens, res_files, speaker, id) wps_meter.update(num_generated_tokens) t.log({"wps": round(wps_meter.avg)}) num_sentences += sample["nsentences"] logger.info("| Processed {} sentences ({} tokens) in {:.1f}s ({:.2f}" "sentences/s, {:.2f} tokens/s)".format( num_sentences, gen_timer.n, gen_timer.sum, num_sentences / gen_timer.sum, 1.0 / gen_timer.avg, )) logger.info("| Generate {} with beam={}".format(args.gen_subset, args.beam))
def main(args): import_user_module(args) if args.max_tokens is None: args.max_tokens = 6000 print(args) if torch.cuda.is_available() and not args.cpu: torch.cuda.set_device(args.device_id) torch.manual_seed(args.seed) # Setup task, e.g., translation, language modeling, etc. task = tasks.setup_task(args) # Load dataset splits load_dataset_splits(task, ['train', 'valid']) # Build model and criterion model = task.build_model(args) criterion = task.build_criterion(args) print(model) print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__)) print('| num. model params: {} (num. trained: {})'.format( sum(p.numel() for p in model.parameters()), sum(p.numel() for p in model.parameters() if p.requires_grad), )) # Make a dummy batch to (i) warm the caching allocator and (ii) as a # placeholder DistributedDataParallel when there's an uneven number of # batches per worker. max_positions = utils.resolve_max_positions( task.max_positions(), model.max_positions(), ) dummy_batch = task.dataset('train').get_dummy_batch( args.max_tokens, max_positions) oom_batch = task.dataset('train').get_dummy_batch(1, max_positions) # Build trainer trainer = Trainer(args, task, model, criterion, dummy_batch, oom_batch) print('| training on {} GPUs'.format(args.distributed_world_size)) print('| max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) # Initialize dataloader epoch_itr = task.get_batch_iterator( dataset=task.dataset(args.train_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=max_positions, ignore_invalid_inputs=True, required_batch_size_multiple=8, seed=args.seed, num_shards=args.distributed_world_size, shard_id=args.distributed_rank, num_workers=args.num_workers, ) # Load the latest checkpoint if one is available if not load_checkpoint(args, trainer, epoch_itr): trainer.dummy_train_step([dummy_batch]) # Train until the learning rate gets too small max_epoch = args.max_epoch or math.inf max_update = args.max_update or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() valid_losses = [None] valid_subsets = args.valid_subset.split(',') while lr > args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates( ) < max_update: # train for one epoch train(args, trainer, task, epoch_itr) if epoch_itr.epoch % args.validate_interval == 0: valid_losses = validate(args, trainer, task, epoch_itr, valid_subsets) # only use first validation loss to update the learning rate lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0]) # save checkpoint if epoch_itr.epoch % args.save_interval == 0: save_checkpoint(args, trainer, epoch_itr, valid_losses[0]) train_meter.stop() print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(parsed_args): assert parsed_args.path is not None, '--path required for evaluation!' print(parsed_args) use_cuda = torch.cuda.is_available() and not parsed_args.cpu task = tasks.setup_task(parsed_args) # Load ensemble print('| loading model(s) from {}'.format(parsed_args.path)) models, args = utils.load_ensemble_for_inference( parsed_args.path.split(':'), task, model_arg_overrides=eval(parsed_args.model_overrides)) for arg in vars(parsed_args).keys(): if arg not in { 'self_target', 'future_target', 'past_target', 'tokens_per_sample', 'output_size_dictionary' }: setattr(args, arg, getattr(parsed_args, arg)) task = tasks.setup_task(args) # Load dataset splits task.load_dataset(args.gen_subset) print('| {} {} {} examples'.format(args.data, args.gen_subset, len(task.dataset(args.gen_subset)))) # Optimize ensemble for generation and set the source and dest dicts on the model (required by scorer) for model in models: model.make_generation_fast_() if args.fp16: model.half() assert len(models) > 0 print('num. model params: {}'.format( sum(p.numel() for p in models[0].parameters()))) itr = task.get_batch_iterator( dataset=task.dataset(args.gen_subset), max_tokens=args.max_tokens or 36000, max_sentences=args.max_sentences, max_positions=utils.resolve_max_positions( *[model.max_positions() for model in models]), num_shards=args.num_shards, shard_id=args.shard_id, ignore_invalid_inputs=True, ).next_epoch_itr(shuffle=False) gen_timer = StopwatchMeter() scorer = SequenceScorer(models, task.target_dictionary) if use_cuda: scorer.cuda() score_sum = 0. count = 0 if args.remove_bpe is not None: bpe_cont = args.remove_bpe.rstrip() bpe_toks = set(i for i in range(len(task.dictionary)) if task.dictionary[i].endswith(bpe_cont)) bpe_len = len(bpe_cont) else: bpe_toks = None bpe_len = 0 word_stats = dict() with progress_bar.build_progress_bar(args, itr) as t: results = scorer.score_batched_itr(t, cuda=use_cuda, timer=gen_timer) wps_meter = TimeMeter() for _, src_tokens, __, hypos in results: for hypo in hypos: pos_scores = hypo['positional_scores'] skipped_toks = 0 if bpe_toks is not None: for i in range(len(hypo['tokens']) - 1): if hypo['tokens'][i].item() in bpe_toks: skipped_toks += 1 pos_scores[i + 1] += pos_scores[i] pos_scores[i] = 0 inf_scores = pos_scores.eq(float('inf')) | pos_scores.eq( float('-inf')) if inf_scores.any(): print( '| Skipping tokens with inf scores:', task.target_dictionary.string( hypo['tokens'][inf_scores.nonzero()])) pos_scores = pos_scores[(~inf_scores).nonzero()] score_sum += pos_scores.sum().cpu() count += pos_scores.numel() - skipped_toks if args.output_word_probs or args.output_word_stats: w = '' word_prob = [] is_bpe = False for i in range(len(hypo['tokens'])): w_ind = hypo['tokens'][i].item() w += task.dictionary[w_ind] if bpe_toks is not None and w_ind in bpe_toks: w = w[:-bpe_len] is_bpe = True else: word_prob.append((w, pos_scores[i].item())) next_prob = None ind = i + 1 while ind < len(hypo['tokens']): if pos_scores[ind].item() != 0: next_prob = pos_scores[ind] break ind += 1 word_stats.setdefault(w, WordStat(w, is_bpe)).add( pos_scores[i].item(), next_prob) is_bpe = False w = '' if args.output_word_probs: print('\t'.join('{} [{:2f}]'.format(x[0], x[1]) for x in word_prob)) wps_meter.update(src_tokens.size(0)) t.log({'wps': round(wps_meter.avg)}) avg_nll_loss = -score_sum / count print('| Evaluated {} tokens in {:.1f}s ({:.2f} tokens/s)'.format( gen_timer.n, gen_timer.sum, 1. / gen_timer.avg)) print('| Loss: {:.4f}, Perplexity: {:.2f}'.format(avg_nll_loss, np.exp(avg_nll_loss))) if args.output_word_stats: for ws in sorted(word_stats.values(), key=lambda x: x.count, reverse=True): print(ws)
def main(args): if args.max_tokens is None: args.max_tokens = 6000 print(args) if not torch.cuda.is_available(): raise NotImplementedError('Training on CPU is not supported') torch.cuda.set_device(args.device_id) torch.manual_seed(args.seed) # Setup task, e.g., translation, language modeling, etc. task = tasks.setup_task(args) # Load dataset splits load_dataset_splits(task, ['train', 'valid']) # Build model and criterion model = task.build_model(args) criterion = task.build_criterion(args) print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__)) print('| num. model params: {}'.format( sum(p.numel() for p in model.parameters()))) # Make a dummy batch to (i) warm the caching allocator and (ii) as a # placeholder DistributedDataParallel when there's an uneven number of # batches per worker. max_positions = utils.resolve_max_positions( task.max_positions(), model.max_positions(), ) dummy_batch = task.dataset('train').get_dummy_batch( args.max_tokens, max_positions) # Build trainer trainer = Trainer(args, task, model, criterion, dummy_batch) print('| training on {} GPUs'.format(args.distributed_world_size)) print('| max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) ## "Prune" heads (actually mask but shh...) #if len(args.transformer_mask_heads) > 0: # # Determine which head to prune # to_prune = parse_head_pruning_descriptors( # args.transformer_mask_heads, # reverse_descriptors=args.transformer_mask_all_but_one_head, # n_heads=model.encoder.layers[0].self_attn.num_heads # ) # print(to_prune) # # Apply pruning # mask_heads(model, to_prune, args.transformer_mask_rescale) # Save initial model initial = os.path.join(args.save_dir, "checkpoint_initial.pt") trainer.save_checkpoint(initial, {}) # Initialize dataloader epoch_itr = task.get_batch_iterator( dataset=task.dataset(args.train_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=max_positions, ignore_invalid_inputs=True, required_batch_size_multiple=8, seed=args.seed, num_shards=args.distributed_world_size, shard_id=args.distributed_rank, ) # Load the latest checkpoint if one is available if not load_checkpoint(args, trainer, epoch_itr): trainer.dummy_train_step([dummy_batch]) # Train until the learning rate gets too small max_epoch = args.max_epoch or math.inf max_update = args.max_update or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() valid_losses = [None] valid_subsets = args.valid_subset.split(',') while lr > args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates( ) < max_update: # train for one epoch train(args, trainer, task, epoch_itr) if epoch_itr.epoch % args.validate_interval == 0: valid_losses = validate(args, trainer, task, epoch_itr, valid_subsets) # only use first validation loss to update the learning rate lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0]) #if epoch_itr.epoch % args.save_interval == 0: # save_checkpoint(args, trainer, epoch_itr, valid_losses[0]) # ***********************************Below Changed****************************** # save checkpoint #if epoch_itr.epoch % args.save_interval == 0: save_interval = 5 # prune and save checkpoint for every five epoch if epoch_itr.epoch % save_interval == 0: #****** changed # do prunning before saving prune2(args, task, model, trainer, epoch_itr) #****** changed2 # save checkpoint save_checkpoint(args, trainer, epoch_itr, valid_losses[0]) prune2(args, task, model, trainer, epoch_itr ) #****** changed2 do last prunning on the last chekcpoint saved # ***********************************Above Changed****************************** train_meter.stop() print('| done training in {:.1f} seconds'.format(train_meter.sum))
def prune(args): if args.max_tokens is None: args.max_tokens = 6000 print(args) if not torch.cuda.is_available(): raise NotImplementedError('Training on CPU is not supported') torch.cuda.set_device(args.device_id) torch.manual_seed(args.seed) # Setup task, e.g., translation, language modeling, etc. task = tasks.setup_task(args) # Load dataset splits load_dataset_splits(task, ['train', "valid"]) # Build model and criterion model = task.build_model(args) criterion = task.build_criterion(args) print('| model {}, criterion {},'.format(args.arch, criterion.__class__.__name__)) print('| num. model params: {}'.format( sum(p.numel() for p in model.parameters()))) # Make a dummy batch to (i) warm the caching allocator and (ii) as a # placeholder DistributedDataParallel when there's an uneven number of # batches per worker. max_positions = utils.resolve_max_positions( task.max_positions(), model.max_positions(), ) dummy_batch = task.dataset('train').get_dummy_batch( args.max_tokens, max_positions) # Build trainer trainer = Trainer(args, task, model, criterion, dummy_batch) print('| training on {} GPUs'.format(args.distributed_world_size)) print('| max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) print('| Optimizer {}'.format(trainer.optimizer.__class__.__name__)) # Initialize dataloader epoch_itr = task.get_batch_iterator( dataset=task.dataset(args.train_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=max_positions, ignore_invalid_inputs=True, required_batch_size_multiple=8, seed=args.seed, num_shards=args.distributed_world_size, shard_id=args.distributed_rank, ) # Load the latest checkpoint if one is available if not load_checkpoint(args, trainer, epoch_itr): trainer.dummy_train_step([dummy_batch]) # Train until the learning rate gets too small prune_meter = StopwatchMeter() prune_meter.start() # Estimate head importance scores head_importance, head_stats = estimate_head_importance( args, trainer, task, epoch_itr) prune_meter.stop() print('| done estimating head importance in {:.1f} seconds'.format( prune_meter.sum)) torch.save(head_stats, f"{os.path.dirname(args.restore_file)}/heads_stats.bin") # Print print("Head importances") print("Encoder self attention") for layer in range(head_importance["encoder_self"].size(0)): print("\t".join(f"{x:.5f}" for x in head_importance["encoder_self"][layer])) print("Encoder decoder attention") for layer in range(head_importance["encoder_decoder"].size(0)): print("\t".join(f"{x:.5f}" for x in head_importance["encoder_decoder"][layer])) print("Decoder self attention") for layer in range(head_importance["decoder_self"].size(0)): print("\t".join(f"{x:.5f}" for x in head_importance["decoder_self"][layer])) # Print sorted pruning profile encoder_self_profile = get_profile(head_importance["encoder_self"], prefix="E") encoder_decoder_profile = get_profile(head_importance["encoder_decoder"], prefix="A") decoder_self_profile = get_profile(head_importance["decoder_self"], prefix="D") # Join all all_profiles = {} if not (args.decoder_self_only or args.encoder_decoder_only): all_profiles.update(encoder_self_profile) if not (args.encoder_self_only or args.decoder_self_only): all_profiles.update(encoder_decoder_profile) if not (args.encoder_self_only or args.encoder_decoder_only): all_profiles.update(decoder_self_profile) sorted_profiles = sorted(all_profiles.items(), key=lambda x: x[1], reverse=args.one_minus) print("Heads sorted by importance:") print(" ".join(p for p, _ in sorted_profiles)) print("Sorted head importance scores:") print(" ".join(f"{v.data:.5f}" for _, v in sorted_profiles)) if args.only_importance: return tot_n_heads = len(sorted_profiles) # Eval pruning if args.one_head: kept_layers = set() to_prune_profile = [] for p, _ in reversed(sorted_profiles): layer_name = ":".join(p.split(":")[:-1]) if layer_name not in kept_layers: kept_layers.add(layer_name) continue else: to_prune_profile.insert(0, p) to_prune = parse_head_pruning_descriptors(to_prune_profile, reverse_descriptors=False) print(f"Evaluating following profile: \t{' '.join(to_prune_profile)}") # Apply pruning mask_heads(model, to_prune, args.transformer_mask_rescale) bleu = eval_bleu_score( model, task, task.dataset(args.valid_subset), beam=args.beam, replace_unk=args.replace_unk, lenpen=args.lenpen, buffer_size=100, use_cuda=torch.cuda.is_available() and not args.cpu, remove_bpe=args.remove_bpe, max_sentences=args.max_sentences, max_tokens=args.max_tokens, stop_early=not args.no_early_stop, normalize_scores=not args.unnormalized, min_len=args.min_len, ) print(f"BLEU score: \t{bleu.score:.2f}") sys.stdout.flush() return for i in range(0, 10): n_to_prune = int(ceil(tot_n_heads * i / 10)) to_prune_profile = [p for p, _ in sorted_profiles[:n_to_prune]] to_prune = parse_head_pruning_descriptors(to_prune_profile, reverse_descriptors=False) print(f"Evaluating following profile: \t{' '.join(to_prune_profile)}") # Apply pruning mask_heads(model, to_prune, args.transformer_mask_rescale) bleu = eval_bleu_score( model, task, task.dataset(args.valid_subset), beam=args.beam, replace_unk=args.replace_unk, lenpen=args.lenpen, buffer_size=100, use_cuda=torch.cuda.is_available() and not args.cpu, remove_bpe=args.remove_bpe, max_sentences=args.max_sentences, max_tokens=args.max_tokens, stop_early=not args.no_early_stop, normalize_scores=not args.unnormalized, min_len=args.min_len, ) print(f"BLEU score: \t{bleu.score:.2f}") sys.stdout.flush()
def prune2(args, task, model, trainer, epoch_itr): # changed2 if args.max_tokens is None: args.max_tokens = 6000 print(args) # avoid aliasing task = copy.deepcopy(task) model = copy.deepcopy(model) trainer = copy.deepcopy(trainer) epoch_itr = copy.deepcopy(trainer) if not torch.cuda.is_available(): raise NotImplementedError('Training on CPU is not supported') torch.cuda.set_device(args.device_id) torch.manual_seed(args.seed) # Make a dummy batch to (i) warm the caching allocator and (ii) as a # placeholder DistributedDataParallel when there's an uneven number of # batches per worker. max_positions = utils.resolve_max_positions( task.max_positions(), model.max_positions(), ) dummy_batch = task.dataset('train').get_dummy_batch( args.max_tokens, max_positions) print('| training on {} GPUs'.format(args.distributed_world_size)) print('| max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) print('| Optimizer {}'.format(trainer.optimizer.__class__.__name__)) # Train until the learning rate gets too small prune_meter = StopwatchMeter() prune_meter.start() # Estimate head importance scores head_importance, head_stats = estimate_head_importance( args, trainer, task, epoch_itr) prune_meter.stop() print('| done estimating head importance in {:.1f} seconds'.format( prune_meter.sum)) torch.save(head_stats, f"{os.path.dirname(args.restore_file)}/heads_stats.bin") # Print print("Head importances") print("Encoder self attention") for layer in range(head_importance["encoder_self"].size(0)): print("\t".join(f"{x:.5f}" for x in head_importance["encoder_self"][layer])) print("Encoder decoder attention") for layer in range(head_importance["encoder_decoder"].size(0)): print("\t".join(f"{x:.5f}" for x in head_importance["encoder_decoder"][layer])) print("Decoder self attention") for layer in range(head_importance["decoder_self"].size(0)): print("\t".join(f"{x:.5f}" for x in head_importance["decoder_self"][layer])) # Print sorted pruning profile encoder_self_profile = get_profile(head_importance["encoder_self"], prefix="E") encoder_decoder_profile = get_profile(head_importance["encoder_decoder"], prefix="A") decoder_self_profile = get_profile(head_importance["decoder_self"], prefix="D") # Join all all_profiles = {} if not (args.decoder_self_only or args.encoder_decoder_only): all_profiles.update(encoder_self_profile) if not (args.encoder_self_only or args.decoder_self_only): all_profiles.update(encoder_decoder_profile) if not (args.encoder_self_only or args.encoder_decoder_only): all_profiles.update(decoder_self_profile) sorted_profiles = sorted(all_profiles.items(), key=lambda x: x[1], reverse=args.one_minus) print("Heads sorted by importance:") print(" ".join(p for p, _ in sorted_profiles)) print("Sorted head importance scores:") print(" ".join(f"{v.data:.5f}" for _, v in sorted_profiles)) tot_n_heads = len(sorted_profiles) for i in range(0, 10): n_to_prune = int(ceil(tot_n_heads * i / 10)) to_prune_profile = [p for p, _ in sorted_profiles[:n_to_prune]] to_prune = parse_head_pruning_descriptors(to_prune_profile, reverse_descriptors=False) print(f"Evaluating following profile: \t{' '.join(to_prune_profile)}") # Apply pruning mask_heads(model, to_prune, args.transformer_mask_rescale) bleu = eval_bleu_score( model, task, task.dataset(args.valid_subset), beam=args.beam, replace_unk=args.replace_unk, lenpen=args.lenpen, buffer_size=100, use_cuda=torch.cuda.is_available() and not args.cpu, remove_bpe=args.remove_bpe, max_sentences=args.max_sentences, max_tokens=args.max_tokens, stop_early=not args.no_early_stop, normalize_scores=not args.unnormalized, min_len=args.min_len, ) print(f"BLEU score: \t{bleu.score:.2f}") sys.stdout.flush()
def main(args, init_distributed=False): utils.import_user_module(args) assert args.max_tokens is not None or args.max_sentences is not None, \ 'Must specify batch size either with --max-tokens or --max-sentences' # Initialize CUDA and distributed training if torch.cuda.is_available() and not args.cpu: torch.cuda.set_device(args.device_id) np.random.seed(args.seed) torch.manual_seed(args.seed) if init_distributed: args.distributed_rank = distributed_utils.distributed_init(args) if distributed_utils.is_master(args): checkpoint_utils.verify_checkpoint_directory(args.save_dir) # Print args logger.info(args) # Setup task, e.g., translation, language modeling, etc. task = tasks.setup_task(args) # Load valid dataset (we load training data below, based on the latest checkpoint) for valid_sub_split in args.valid_subset.split(','): task.load_dataset(valid_sub_split, combine=False, epoch=0) # Build model and criterion model = task.build_model(args) criterion = task.build_criterion(args) logger.info(model) logger.info('model {}, criterion {}'.format(args.arch, criterion.__class__.__name__)) logger.info('num. model params: {} (num. trained: {})'.format( sum(p.numel() for p in model.parameters()), sum(p.numel() for p in model.parameters() if p.requires_grad), )) # Build trainer trainer = Trainer(args, task, model, criterion) logger.info('training on {} GPUs'.format(args.distributed_world_size)) logger.info( 'max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) # Load the latest checkpoint if one is available and restore the # corresponding train iterator extra_state, epoch_itr = checkpoint_utils.load_checkpoint(args, trainer) # Train until the learning rate gets too small train_meter = StopwatchMeter() train_meter.start() valid_subsets = args.valid_subset.split(',') tokenize = sacrebleu.DEFAULT_TOKENIZER if not args.eval_tokenized_bleu else 'none' hyps, refs = validate(args, trainer, task, epoch_itr, valid_subsets) for h, r, split in zip(hyps, refs, args.valid_subset.split(',')): assert len(h) == len(r) sacrebleu_score, _, _ = sacrebleu.corpus_bleu( h, [r], tokenize=tokenize), hyps, refs bleu = compute_cvpr_bleu(h, r) rouge_score = rouge.rouge(h, r) print('{} set has {} samples,\n' 'sacrebleu: {},\n' 'CVPR BLEU scripts: {}\n' 'CVPR ROUGE: {}'.format(split, len(h), sacrebleu_score, bleu, rouge_score)) print('performance: {:.2f} {}'.format( rouge_score['rouge_l/f_score'] * 100, ' '.join([str(b) for b in bleu])))
def _generate_score(models, args, dataset, dataset_split): use_cuda = torch.cuda.is_available() and not args.cpu # Load ensemble if not args.quiet: print("| loading model(s) from {}".format(", ".join(args.path))) # Optimize ensemble for generation for model in models: model.make_generation_fast_( beamable_mm_beam_size=None if args.no_beamable_mm else args.beam) # Initialize generator model_weights = None if args.model_weights: model_weights = [ float(w.strip()) for w in args.model_weights.split(",") ] translator = beam_decode.SequenceGenerator( models, beam_size=args.beam, stop_early=(not args.no_early_stop), normalize_scores=(not args.unnormalized), len_penalty=args.lenpen, unk_penalty=args.unkpen, word_reward=args.word_reward, model_weights=model_weights, ) if use_cuda: translator.cuda() # Load alignment dictionary for unknown word replacement # (None if no unknown word replacement, empty if no path to align dictionary) align_dict = utils.load_align_dict(args.replace_unk) # Generate and compute BLEU score scorer = bleu.Scorer(dataset.dst_dict.pad(), dataset.dst_dict.eos(), dataset.dst_dict.unk()) max_positions = min(model.max_encoder_positions() for model in models) itr = dataset.eval_dataloader( dataset_split, max_sentences=args.max_sentences, max_positions=max_positions, skip_invalid_size_inputs_valid_test=( args.skip_invalid_size_inputs_valid_test), ) if args.num_shards > 1: if args.shard_id < 0 or args.shard_id >= args.num_shards: raise ValueError("--shard-id must be between 0 and num_shards") itr = data.sharded_iterator(itr, args.num_shards, args.shard_id) num_sentences = 0 with progress_bar.build_progress_bar(args, itr) as t: wps_meter = TimeMeter() gen_timer = StopwatchMeter() translations = translator.generate_batched_itr( t, maxlen_a=args.max_len_a, maxlen_b=args.max_len_b, cuda=use_cuda, timer=gen_timer, ) for sample_id, src_tokens, target_tokens, hypos in translations: # Process input and ground truth target_tokens = target_tokens.int().cpu() # Either retrieve the original sentences or regenerate them from tokens. if align_dict is not None: src_str = dataset.splits[dataset_split].src.get_original_text( sample_id) target_str = dataset.splits[ dataset_split].dst.get_original_text(sample_id) else: src_str = dataset.src_dict.string(src_tokens, args.remove_bpe) target_str = dataset.dst_dict.string(target_tokens, args.remove_bpe, escape_unk=True) if not args.quiet: print(f"S-{sample_id}\t{src_str}") print(f"T-{sample_id}\t{target_str}") # Process top predictions for i, hypo in enumerate(hypos[:min(len(hypos), args.nbest)]): hypo_tokens, hypo_str, alignment = utils.post_process_prediction( hypo_tokens=hypo["tokens"].int().cpu(), src_str=src_str, alignment=hypo["alignment"].int().cpu(), align_dict=align_dict, dst_dict=dataset.dst_dict, remove_bpe=args.remove_bpe, ) if not args.quiet: print(f"H-{sample_id}\t{hypo['score']}\t{hypo_str}") print("A-{}\t{}".format( sample_id, " ".join(map(lambda x: str(utils.item(x)), alignment)), )) # Score only the top hypothesis if i == 0: if align_dict is not None or args.remove_bpe is not None: # Convert back to tokens for evaluation with unk replacement # and/or without BPE target_tokens = tokenizer.Tokenizer.tokenize( target_str, dataset.dst_dict, add_if_not_exist=True) scorer.add(target_tokens, hypo_tokens) wps_meter.update(src_tokens.size(0)) t.log({"wps": round(wps_meter.avg)}) num_sentences += 1 return scorer, num_sentences, gen_timer
def main(args, init_distributed=False): utils.import_user_module(args) assert args.max_tokens is not None or args.max_sentences is not None, \ 'Must specify batch size either with --max-tokens or --max-sentences' # Initialize CUDA and distributed training if torch.cuda.is_available() and not args.cpu: torch.cuda.set_device(args.device_id) np.random.seed(args.seed) torch.manual_seed(args.seed) if init_distributed: args.distributed_rank = distributed_utils.distributed_init(args) if distributed_utils.is_master(args): checkpoint_utils.verify_checkpoint_directory(args.save_dir) # Print args logger.info(args) # Setup task, e.g., translation, language modeling, etc. task = tasks.setup_task(args) # Load valid dataset (we load training data below, based on the latest checkpoint) for valid_sub_split in args.valid_subset.split(','): task.load_dataset(valid_sub_split, combine=False, epoch=0) # Build model and criterion model = task.build_model(args) criterion = task.build_criterion(args) logger.info(model) logger.info('model {}, criterion {}'.format(args.arch, criterion.__class__.__name__)) logger.info('num. model params: {} (num. trained: {})'.format( sum(p.numel() for p in model.parameters()), sum(p.numel() for p in model.parameters() if p.requires_grad), )) # Build trainer trainer = Trainer(args, task, model, criterion) logger.info('training on {} GPUs'.format(args.distributed_world_size)) logger.info('max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) # Load the latest checkpoint if one is available and restore the # corresponding train iterator extra_state, epoch_itr = checkpoint_utils.load_checkpoint(args, trainer) # Train until the learning rate gets too small max_epoch = args.max_epoch or math.inf max_update = args.max_update or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() valid_subsets = args.valid_subset.split(',') print(args.multi_views) while ( lr > args.min_lr and ( epoch_itr.epoch < max_epoch # allow resuming training from the final checkpoint or epoch_itr._next_epoch_itr is not None ) and trainer.get_num_updates() < max_update ): # train for one epoch train(args, trainer, task, epoch_itr) if not args.disable_validation and epoch_itr.epoch % args.validate_interval == 0: valid_losses = validate(args, trainer, task, epoch_itr, valid_subsets) else: valid_losses = [None] # only use first validation loss to update the learning rate lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0]) bart = BARTHubInterface(args, task, trainer.model).cuda() #print(bart.device) bart.eval() count = 1 bsz = 8 print("Test on val set: ") with open('../data/val_sent_trans_cons_label.source') as source, open('../data/val_sent_c99_label.source') as source2, open('./val_best_multi_attn_'+str(args.lr_weight)+'_.hypo', 'wt', encoding='utf-8') as fout: s1 = source.readlines() s2 = source2.readlines() slines = [s1[0].strip()] slines2 = [s2[0].strip()] for i in tqdm(range(1, len(s1))): if count % bsz == 0: with torch.no_grad(): if args.multi_views: hypotheses_batch = bart.sample(slines, sentences2 = slines2, balance = True, beam=4, lenpen=2.0, max_len_b=100, min_len=5, no_repeat_ngram_size=3) else: hypotheses_batch = bart.sample(slines, beam=4, lenpen=2.0, max_len_b=100, min_len=5, no_repeat_ngram_size=3) for hypothesis in hypotheses_batch: fout.write(hypothesis + '\n') fout.flush() slines = [] slines2 = [] slines.append(s1[i].strip()) slines2.append(s2[i].strip()) count += 1 if slines != []: if args.multi_views: hypotheses_batch = bart.sample(slines, sentences2 = slines2, balance = True, beam=4, lenpen=2.0, max_len_b=100, min_len=5, no_repeat_ngram_size=3) else: hypotheses_batch = bart.sample(slines, beam=4, lenpen=2.0, max_len_b=100, min_len=5, no_repeat_ngram_size=3) #hypotheses_batch = bart.sample(slines, sentences2 = slines2, balance = True, beam=4, lenpen=2.0, max_len_b=100, min_len=5, no_repeat_ngram_size=3) for hypothesis in hypotheses_batch: fout.write(hypothesis + '\n') fout.flush() hyp_path = './val_best_multi_attn_'+str(args.lr_weight)+'_.hypo' ref_path = '../data/val_sent_trans_cons_label.target' hypothesis = [] with open(hyp_path, 'r') as f: lines = f.readlines() for l in lines: hypothesis.append(l[:-1]) reference = [] with open(ref_path, 'r') as f: lines = f.readlines() for l in lines: reference.append(l[:-1]) rouge = Rouge() print("Val", rouge.get_scores(hypothesis, reference, avg = True)) # save checkpoint if epoch_itr.epoch % args.save_interval == 0: checkpoint_utils.save_checkpoint(args, trainer, epoch_itr, valid_losses[0]) print("Test on testing set: ") count = 1 bsz = 8 with open('../data/test_sent_trans_cons_label.source') as source, open('../data/test_sent_c99_label.source') as source2, open('./test_best_multi_attn_'+str(args.lr_weight)+'_.hypo', 'wt', encoding='utf-8') as fout: s1 = source.readlines() s2 = source2.readlines() slines = [s1[0].strip()] slines2 = [s2[0].strip()] for i in tqdm(range(1, len(s1))): if count % bsz == 0: with torch.no_grad(): if args.multi_views: hypotheses_batch = bart.sample(slines, sentences2 = slines2, balance = True, beam=4, lenpen=2.0, max_len_b=100, min_len=5, no_repeat_ngram_size=3) else: hypotheses_batch = bart.sample(slines, beam=4, lenpen=2.0, max_len_b=100, min_len=5, no_repeat_ngram_size=3) for hypothesis in hypotheses_batch: fout.write(hypothesis + '\n') fout.flush() slines = [] slines2 = [] slines.append(s1[i].strip()) slines2.append(s2[i].strip()) count += 1 if slines != []: if args.multi_views: hypotheses_batch = bart.sample(slines, sentences2 = slines2, balance = True, beam=4, lenpen=2.0, max_len_b=100, min_len=5, no_repeat_ngram_size=3) else: hypotheses_batch = bart.sample(slines, beam=4, lenpen=2.0, max_len_b=100, min_len=5, no_repeat_ngram_size=3) for hypothesis in hypotheses_batch: fout.write(hypothesis + '\n') fout.flush() hyp_path = './test_best_multi_attn_'+str(args.lr_weight)+'_.hypo' ref_path = '../data/test_sent_trans_cons_label.target' hypothesis = [] with open(hyp_path, 'r') as f: lines = f.readlines() for l in lines: hypothesis.append(l[:-1]) reference = [] with open(ref_path, 'r') as f: lines = f.readlines() for l in lines: reference.append(l[:-1]) rouge = Rouge() print('Test', rouge.get_scores(hypothesis, reference, avg = True)) # early stop if should_stop_early(args, valid_losses[0]): logger.info('early stop since valid performance hasn\'t improved for last {} runs'.format(args.patience)) break epoch_itr = trainer.get_train_iterator( epoch_itr.epoch, # sharded data: get train iterator for next epoch load_dataset=(os.pathsep in getattr(args, 'data', '')), ) train_meter.stop() logger.info('done training in {:.1f} seconds'.format(train_meter.sum))
def _generate_score(models, args, task, dataset, modify_target_dict): use_cuda = torch.cuda.is_available() and not args.cpu # Load ensemble if not args.quiet: print("| loading model(s) from {}".format(", ".join(args.path.split(":")))) # Optimize ensemble for generation for model in models: model.make_generation_fast_( beamable_mm_beam_size=None if args.no_beamable_mm else args.beam, need_attn=True, ) translator = build_sequence_generator(args, task, models) # Load alignment dictionary for unknown word replacement # (None if no unknown word replacement, empty if no path to align dictionary) align_dict = utils.load_align_dict(args.replace_unk) print("seed number is" + str(args.max_examples_to_evaluate_seed)) if args.max_examples_to_evaluate > 0: pytorch_translate_data.subsample_pair_dataset( dataset, args.max_examples_to_evaluate, args.max_examples_to_evaluate_seed ) # Keep track of translations # Initialize with empty translations # and zero probs scores translated_sentences = [""] * len(dataset) translated_scores = [0.0] * len(dataset) hypos_list = [] collect_output_hypos = getattr(args, "output_hypos_binary_path", False) if collect_output_hypos: output_hypos_token_arrays = [None] * len(dataset) # Generate and compute BLEU score dst_dict = task.target_dictionary if args.sacrebleu: scorer = bleu.SacrebleuScorer() else: scorer = bleu.Scorer(dst_dict.pad(), dst_dict.eos(), dst_dict.unk()) itr = task.get_batch_iterator( dataset=dataset, max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=utils.resolve_max_positions( task.max_positions(), *[model.max_positions() for model in models] ), ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test, required_batch_size_multiple=8, num_shards=args.num_shards, shard_id=args.shard_id, num_workers=args.num_workers, ).next_epoch_itr(shuffle=False) oracle_scorer = None if args.report_oracle_bleu: oracle_scorer = bleu.Scorer(dst_dict.pad(), dst_dict.eos(), dst_dict.unk()) rescorer = None num_sentences = 0 translation_samples = [] translation_info_list = [] with progress_bar.build_progress_bar(args, itr) as t: wps_meter = TimeMeter() gen_timer = StopwatchMeter() translations = translator.generate_batched_itr( t, maxlen_a=args.max_len_a, maxlen_b=args.max_len_b, cuda=use_cuda, timer=gen_timer, prefix_size=1 if pytorch_translate_data.is_multilingual_many_to_one(args) else 0, ) for trans_info in _iter_translations( args, task, dataset, translations, align_dict, rescorer, modify_target_dict ): if hasattr(scorer, "add_string"): scorer.add_string(trans_info.target_str, trans_info.hypo_str) else: scorer.add(trans_info.target_tokens, trans_info.hypo_tokens) if oracle_scorer is not None: oracle_scorer.add(trans_info.target_tokens, trans_info.best_hypo_tokens) if getattr(args, "translation_output_file", False): translated_sentences[trans_info.sample_id] = trans_info.hypo_str if getattr(args, "translation_probs_file", False): translated_scores[trans_info.sample_id] = trans_info.hypo_score if getattr(args, "hypotheses_export_path", False): hypos_list.append(trans_info.hypos) if collect_output_hypos: output_hypos_token_arrays[ trans_info.sample_id ] = trans_info.best_hypo_tokens if args.translation_info_export_path is not None: # Strip expensive data from hypotheses before saving hypos = [ {k: v for k, v in hypo.items() if k in ["tokens", "score"]} for hypo in trans_info.hypos ] # Make sure everything is on cpu before exporting hypos = [ {"score": hypo["score"], "tokens": hypo["tokens"].cpu()} for hypo in hypos ] translation_info_list.append( { "src_tokens": trans_info.src_tokens.cpu(), "target_tokens": trans_info.target_tokens, "hypos": hypos, } ) translation_samples.append( collections.OrderedDict( { "sample_id": trans_info.sample_id.item(), "src_str": trans_info.src_str, "target_str": trans_info.target_str, "hypo_str": trans_info.hypo_str, } ) ) wps_meter.update(trans_info.src_tokens.size(0)) t.log({"wps": round(wps_meter.avg)}) num_sentences += 1 # If applicable, save collected hypothesis tokens to binary output file if collect_output_hypos: output_dataset = pytorch_translate_data.InMemoryIndexedDataset() output_dataset.load_from_sequences(output_hypos_token_arrays) output_dataset.save(args.output_hypos_binary_path) if args.output_source_binary_path: dataset.src.save(args.output_source_binary_path) if args.translation_info_export_path is not None: f = open(args.translation_info_export_path, "wb") pickle.dump(translation_info_list, f) f.close() # If applicable, save the translations and scores to the output files # These two ouputs are used in dual learning for weighted backtranslation if getattr(args, "translation_output_file", False) and getattr( args, "translation_probs_file", False ): with open(args.translation_output_file, "w") as translation_file, open( args.translation_probs_file, "w" ) as score_file: for hypo_str, hypo_score in zip(translated_sentences, translated_scores): if len(hypo_str.strip()) > 0: print(hypo_str, file=translation_file) print(np.exp(hypo_score), file=score_file) # For eg. external evaluation if getattr(args, "hypotheses_export_path", False): with open(args.hypotheses_export_path, "w") as out_file: for hypos in hypos_list: for hypo in hypos: print( task.tgt_dict.string( hypo["tokens"], bpe_symbol=args.remove_bpe ), file=out_file, ) if oracle_scorer is not None: print(f"| Oracle BLEU (best hypo in beam): {oracle_scorer.result_string()}") return scorer, num_sentences, gen_timer, translation_samples
def main(args, init_distributed=False): import_user_module(args) if args.max_tokens is None: args.max_tokens = 6000 print(args) #args.distributed_world_size = 1 if torch.cuda.is_available() and not args.cpu: torch.cuda.set_device(args.device_id) torch.manual_seed(args.seed) # Setup task, e.g., translation, language modeling, etc. task = tasks.setup_task(args) # Load dataset splits load_dataset_splits(args, task) # Build model and criterion model = task.build_model(args) criterion = task.build_criterion(args) print(model) print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__)) print('| num. model params: {} (num. trained: {})'.format( sum(p.numel() for p in model.parameters()), sum(p.numel() for p in model.parameters() if p.requires_grad), )) # Make a dummy batch to (i) warm the caching allocator and (ii) as a # placeholder DistributedDataParallel when there's an uneven number of # batches per worker. max_positions = utils.resolve_max_positions( task.max_positions(), model.max_positions(), ) dummy_batch = task.dataset(args.train_subset).get_dummy_batch( args.max_tokens, max_positions) oom_batch = task.dataset(args.train_subset).get_dummy_batch( 1, max_positions) # Build trainer print("Building trainer...") trainer = Trainer(args, task, model, criterion, dummy_batch, oom_batch) print('| training on {} GPUs'.format(args.distributed_world_size)) print('| max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) # Initialize dataloader print("Initialize dataloader...") epoch_itr = task.get_batch_iterator( dataset=task.dataset(args.train_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=max_positions, ignore_invalid_inputs=True, required_batch_size_multiple=args.required_batch_size_multiple, seed=args.seed, num_shards=args.distributed_world_size, shard_id=args.distributed_rank, num_workers=args.num_workers, ) # Initialize distributed training (after data loading) print("Initialize distributed training (after data loading)...") if init_distributed: import socket args.distributed_rank = distributed_utils.distributed_init(args) print('| initialized host {} as rank {}'.format( socket.gethostname(), args.distributed_rank)) # Load the latest checkpoint if one is available print("Load the latest checkpoint if one is available...") if not load_checkpoint(args, trainer, epoch_itr): trainer.dummy_train_step([dummy_batch]) if args.reset_target_embedding: trainer.init_meters(args) print("reset trainer.meters") # Train until the learning rate gets too small max_epoch = args.max_epoch or math.inf max_update = args.max_update or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() valid_losses = [None] valid_subsets = args.valid_subset.split(',') if args.distributed_rank == 0: if os.path.basename(args.save_dir) != "": log_file = os.path.join( args.save_dir, "({0})-params.log".format(os.path.basename(args.save_dir))) else: log_file = os.path.join( args.save_dir, "({0})-params.log".format(args.save_dir.split('/')[-2])) # create log file args.log_file = log_file if os.path.exists(log_file): w = open(log_file, "a+", encoding="utf-8") else: w = open(log_file, "w", encoding="utf-8") w.write(str(args).replace(", ", ",\n") + "\n") w.write(str(model) + "\n") w.flush() w.close() print("saving params file into{}...".format(log_file)) while lr > args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates( ) < max_update: # train for one epoch train(args, trainer, task, epoch_itr) if epoch_itr.epoch % args.validate_interval == 0: valid_losses = validate(args, trainer, task, epoch_itr, valid_subsets) # only use first validation loss to update the learning rate lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0]) # save checkpoint if epoch_itr.epoch % args.save_interval == 0: save_checkpoint(args, trainer, epoch_itr, valid_losses[0]) train_meter.stop() print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args): assert args.path is not None, '--path required for generation!' assert not args.sampling or args.nbest == args.beam, \ '--sampling requires --nbest to be equal to --beam' assert args.replace_unk is None or args.raw_text, \ '--replace-unk requires a raw text dataset (--raw-text)' import_user_module(args) if args.max_tokens is None and args.max_sentences is None: args.max_tokens = 12000 print(args) use_cuda = torch.cuda.is_available() and not args.cpu # ========== for bartsv task, rebuild dictionary after model args are loaded ========== # assert not hasattr(args, 'node_freq_min'), 'node_freq_min should be read from model args' # args.node_freq_min = 5 # temporarily set before model loading, as this is needed in tasks.setup_task(args) # ===================================================================================== # Load dataset splits task = tasks.setup_task(args) # Note: states are not needed since they will be provided by the state # machine task.load_dataset(args.gen_subset, state_machine=False) # Set dictionaries try: src_dict = getattr(task, 'source_dictionary', None) except NotImplementedError: src_dict = None tgt_dict = task.target_dictionary # Load ensemble print('| loading model(s) from {}'.format(args.path)) try: models, _model_args = checkpoint_utils.load_model_ensemble( args.path.split(':'), arg_overrides=eval(args.model_overrides), task=task, ) except: # NOTE this is for "bartsv" models when default "args.node_freq_min" (5) is not equal to the model # when loading model with the above task there will be an error when building the model with the task's # target vocabulary, which would be of different size # TODO better handle these cases (without sacrificing compatibility with other model archs) models, _model_args = checkpoint_utils.load_model_ensemble( args.path.split(':'), arg_overrides=eval(args.model_overrides), task=None, ) # ========== for bartsv task, rebuild the dictionary based on model args ========== if 'bartsv' in _model_args.arch and args.node_freq_min != _model_args.node_freq_min: args.node_freq_min = _model_args.node_freq_min # Load dataset splits task = tasks.setup_task(args) # Note: states are not needed since they will be provided by the state machine task.load_dataset(args.gen_subset, state_machine=False) # Set dictionaries try: src_dict = getattr(task, 'source_dictionary', None) except NotImplementedError: src_dict = None tgt_dict = task.target_dictionary # ================================================================================== # import pdb; pdb.set_trace() # print(_model_args) # ========== for previous model trained when new arguments were not there ========== if not hasattr(_model_args, 'shift_pointer_value'): _model_args.shift_pointer_value = 1 # ================================================================================== # Optimize ensemble for generation for model in models: model.make_generation_fast_( beamable_mm_beam_size=None if args.no_beamable_mm else args.beam, need_attn=args.print_alignment, ) if args.fp16: model.half() if use_cuda: model.cuda() # Load alignment dictionary for unknown word replacement # (None if no unknown word replacement, empty if no path to align # dictionary) align_dict = utils.load_align_dict(args.replace_unk) # Load dataset (possibly sharded) itr = task.get_batch_iterator( dataset=task.dataset(args.gen_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=None, ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test, required_batch_size_multiple=args.required_batch_size_multiple, num_shards=args.num_shards, shard_id=args.shard_id, num_workers=args.num_workers, # large_sent_first=False # not in fairseq ).next_epoch_itr(shuffle=False) # Initialize generator gen_timer = StopwatchMeter() generator = task.build_generator(args, _model_args) # Generate and compute BLEU score if args.sacrebleu: scorer = bleu.SacrebleuScorer() else: scorer = bleu.Scorer(tgt_dict.pad(), tgt_dict.eos(), tgt_dict.unk()) num_sentences = 0 has_target = True examples = Examples(args.path, args.results_path, args.gen_subset, args.nbest) error_stats = {'num_sub_start': 0} with progress_bar.build_progress_bar(args, itr) as t: wps_meter = TimeMeter() for sample in t: sample = utils.move_to_cuda(sample) if use_cuda else sample if 'net_input' not in sample: raise Exception("Did not expect empty sample") continue prefix_tokens = None if args.prefix_size > 0: prefix_tokens = sample['target'][:, :args.prefix_size] # breakpoint() gen_timer.start() hypos = task.inference_step(generator, models, sample, args, prefix_tokens) num_generated_tokens = sum(len(h[0]['tokens']) for h in hypos) gen_timer.stop(num_generated_tokens) # breakpoint() for i, sample_id in enumerate(sample['id'].tolist()): has_target = sample['target'] is not None # Remove padding src_tokens = utils.strip_pad( sample['net_input']['src_tokens'][i, :], tgt_dict.pad()) target_tokens = None if has_target: target_tokens = utils.strip_pad( sample['target'][i, :], tgt_dict.pad()).int().cpu() # Either retrieve the original sentences or regenerate them from tokens. if align_dict is not None: src_str = task.dataset( args.gen_subset).src.get_original_text(sample_id) target_str = task.dataset( args.gen_subset).tgt.get_original_text(sample_id) else: if src_dict is not None: src_str = src_dict.string(src_tokens, args.remove_bpe) else: src_str = "" if has_target: target_str = tgt_dict.string(target_tokens, args.remove_bpe, escape_unk=True) # debug: '<<unk>>' is added to the dictionary # if 'unk' in target_str: # breakpoint() # ==========> NOTE we do not really have the ground truth target (with the same alignments) # target_str might have <unk> as the target dictionary is only built on training data # but it doesn't matter. It should not affect the target dictionary! if not args.quiet: if src_dict is not None: print('S-{}\t{}'.format(sample_id, src_str)) if has_target: print('T-{}\t{}'.format(sample_id, target_str)) # Process top predictions for j, hypo in enumerate(hypos[i][:args.nbest]): # hypo_tokens, hypo_str, alignment = utils.post_process_prediction( # hypo_tokens=hypo['tokens'].int().cpu(), # src_str=src_str, # alignment=hypo['alignment'].int().cpu() if hypo['alignment'] is not None else None, # align_dict=align_dict, # tgt_dict=tgt_dict, # remove_bpe=args.remove_bpe, # # FIXME: AMR specific # split_token="\t", # line_tokenizer=task.tokenize, # ) if 'bartsv' in _model_args.arch: if not tgt_dict[hypo['tokens'][0]].startswith( tgt_dict.bpe.INIT): error_stats['num_sub_start'] += 1 try: actions_nopos, actions_pos, actions = post_process_action_pointer_prediction_bartsv( hypo, tgt_dict) except: breakpoint() else: actions_nopos, actions_pos, actions = post_process_action_pointer_prediction( hypo, tgt_dict) # breakpoint() if args.clean_arcs: actions_nopos, actions_pos, actions, invalid_idx = clean_pointer_arcs( actions_nopos, actions_pos, actions) # TODO these are just dummy for the reference below to run hypo_tokens = hypo['tokens'].int().cpu() hypo_str = '/t'.join(actions) alignment = None # update the list of examples examples.append({ 'actions_nopos': actions_nopos, 'actions_pos': actions_pos, 'actions': actions, 'reference': target_str, 'src_str': src_str, 'sample_id': sample_id }) if not args.quiet: print('H-{}\t{}\t{}'.format(sample_id, hypo_str, hypo['score'])) print('P-{}\t{}'.format( sample_id, ' '.join( map( lambda x: '{:.4f}'.format(x), hypo['positional_scores'].tolist(), )))) if args.print_alignment: print('A-{}\t{}'.format( sample_id, ' '.join( map(lambda x: str(utils.item(x)), alignment)))) # Score only the top hypothesis if has_target and j == 0: if align_dict is not None or args.remove_bpe is not None: # Convert back to tokens for evaluation with unk replacement and/or without BPE target_tokens = tgt_dict.encode_line( target_str, add_if_not_exist=False) # NOTE do not modify the tgt dictionary with 'add_if_not_exist=True'! if hasattr(scorer, 'add_string'): scorer.add_string(target_str, hypo_str) else: scorer.add(target_tokens, hypo_tokens) wps_meter.update(num_generated_tokens) t.log({'wps': round(wps_meter.avg)}) num_sentences += sample['nsentences'] # Save examples to files examples.save() print('| Error case (handled by manual fix) statistics:') print(error_stats) print( '| Translated {} sentences ({} tokens) in {:.1f}s ({:.2f} sentences/s, {:.2f} tokens/s)' .format(num_sentences, gen_timer.n, gen_timer.sum, num_sentences / gen_timer.sum, 1. / gen_timer.avg)) if has_target: print('| Generate {} with beam={}: {}'.format(args.gen_subset, args.beam, scorer.result_string())) return scorer
def train(args, extra_state, trainer, dataset): # offset for current epoch (may be different from checkpoint offset) starting_offset = extra_state["batch_offset"] # Train until the learning rate gets too small max_epoch = args.max_epoch or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() while lr > args.min_lr and extra_state["epoch"] <= max_epoch: """Train the model for one epoch.""" itr, progress, extra_meters = setup_epoch( args=args, epoch=extra_state["epoch"], batch_offset=starting_offset, trainer=trainer, dataset=dataset, ) for i, sample in enumerate(itr, start=starting_offset): log_output = trainer.train_step(sample) train_stats = log_mid_epoch_stats( trainer=trainer, progress=progress, extra_meters=extra_meters, log_output=log_output, ) if (args.continuous_averaging_after_epochs >= 0 and extra_state["epoch"] > args.continuous_averaging_after_epochs): model_param_dict = trainer.model.state_dict() if "param_totals" not in extra_state: extra_state["param_totals"] = {} for name, value in model_param_dict.items(): extra_state["param_totals"][name] = value.clone() extra_state["param_accum_count"] = 1 else: for name, value in model_param_dict.items(): extra_state["param_totals"][name] += value extra_state["param_accum_count"] += 1 if i == starting_offset: # ignore the first mini-batch in words-per-second calculation trainer.get_meter("wps").reset() num_updates = trainer.get_num_updates() do_validate = (args.subepoch_validate_interval > 0 and num_updates % args.subepoch_validate_interval == 0) do_save = (not args.no_save and args.save_interval > 0 and num_updates % args.save_interval == 0) do_eval_bleu = ( # We can only do BLEU eval when we have a new checkpoint to load. do_save and args.generate_bleu_eval_interval > 0 and num_updates - extra_state["last_bleu_eval"] >= args.generate_bleu_eval_interval) if do_eval_bleu: extra_state["last_bleu_eval"] = num_updates extra_state["batch_offset"] = i + 1 ( _, val_ppl, val_bleu, stop_training_mid_epoch, ) = validate_save_and_evaluate_bleu( args=args, trainer=trainer, dataset=dataset, extra_state=extra_state, do_validate=do_validate, do_save=do_save, do_eval_bleu=do_eval_bleu, ) yield ( trainer.get_num_updates(), { "train_ppl": train_stats["ppl"], "tune_ppl": val_ppl, "tune_bleu": val_bleu, }, ) if stop_training_mid_epoch: break # log end-of-epoch stats train_stats = log_end_epoch_stats(trainer=trainer, progress=progress, extra_meters=extra_meters) if stop_training_mid_epoch: break # batch_offset being None denotes the end of an epoch. extra_state["batch_offset"] = None ( val_loss, val_ppl, val_bleu, stop_training_end_of_epoch, ) = validate_save_and_evaluate_bleu( args=args, trainer=trainer, dataset=dataset, extra_state=extra_state, do_validate=True, do_save=not args.no_save and not args.no_end_of_epoch_checkpoints, do_eval_bleu=args.generate_bleu_eval_per_epoch, ) extra_state["val_loss"] = val_loss yield ( trainer.get_num_updates(), { "train_ppl": train_stats["ppl"], "tune_ppl": val_ppl, "tune_bleu": val_bleu, }, ) if stop_training_end_of_epoch: break lr = trainer.lr_step(extra_state["epoch"], val_loss) extra_state["epoch"] += 1 extra_state["batch_offset"] = 0 starting_offset = 0 if is_training_over_time_limit(extra_state["start_time"], args.stop_time_hr): break train_meter.stop() print(f"| done training in {train_meter.sum:.1f} seconds") if "evaluate_bleu" in extra_state: print( f"| Best BLEU score of {extra_state['evaluate_bleu']['best']} was from " f"epoch {extra_state['evaluate_bleu']['best_epoch']}")
def main(args): print(args) if not torch.cuda.is_available(): raise NotImplementedError('Training on CPU is not supported') torch.cuda.set_device(args.device_id) torch.manual_seed(args.seed) # Load dataset splits = ['train', 'valid'] if data.has_binary_files(args.data, splits): dataset = data.load_dataset( args.data, splits, args.source_lang, args.target_lang) else: dataset = data.load_raw_text_dataset( args.data, splits, args.source_lang, args.target_lang) if args.source_lang is None or args.target_lang is None: # record inferred languages in args, so that it's saved in checkpoints args.source_lang, args.target_lang = dataset.src, dataset.dst print('| [{}] dictionary: {} types'.format(dataset.src, len(dataset.src_dict))) print('| [{}] dictionary: {} types'.format(dataset.dst, len(dataset.dst_dict))) for split in splits: print('| {} {} {} examples'.format(args.data, split, len(dataset.splits[split]))) # Build model and criterion model = models.build_model(args, dataset.src_dict, dataset.dst_dict) criterion = criterions.build_criterion(args, dataset.src_dict, dataset.dst_dict) print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__)) print('| num. model params: {}'.format(sum(p.data.numel() for p in model.parameters()))) # Build trainer trainer = Trainer(args, model, criterion) print('| training on {} GPUs'.format(args.distributed_world_size)) print('| max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) # Load the latest checkpoint if one is available os.makedirs(args.save_dir, exist_ok=True) checkpoint_path = os.path.join(args.save_dir, args.restore_file) extra_state = trainer.load_checkpoint(checkpoint_path) if extra_state is not None: epoch = extra_state['epoch'] batch_offset = extra_state['batch_offset'] print('| loaded checkpoint {} (epoch {})'.format(checkpoint_path, epoch)) if batch_offset == 0: trainer.lr_step(epoch) epoch += 1 else: epoch, batch_offset = 1, 0 # Train until the learning rate gets too small max_epoch = args.max_epoch or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() while lr > args.min_lr and epoch <= max_epoch: # train for one epoch train(args, trainer, dataset, epoch, batch_offset) # evaluate on validate set for k, subset in enumerate(args.valid_subset.split(',')): val_loss = validate(args, trainer, dataset, subset, epoch) if k == 0: # only use first validation loss to update the learning schedule lr = trainer.lr_step(epoch, val_loss) # save checkpoint if not args.no_save: save_checkpoint(trainer, args, epoch, 0, val_loss) epoch += 1 batch_offset = 0 train_meter.stop() print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args, init_distributed=False): utils.import_user_module(args) assert args.max_tokens is not None or args.max_sentences is not None, \ 'Must specify batch size either with --max-tokens or --max-sentences' # Initialize CUDA and distributed training if torch.cuda.is_available() and not args.cpu: torch.cuda.set_device(args.device_id) torch.manual_seed(args.seed) if init_distributed: args.distributed_rank = distributed_utils.distributed_init(args) # Print args print(args) # Setup task, e.g., translation, language modeling, etc. task = tasks.setup_task(args) # Load dataset splits task.load_dataset(args.train_subset, combine=True, epoch=0) for valid_sub_split in args.valid_subset.split(','): task.load_dataset(valid_sub_split, combine=True, epoch=0) # Build model and criterion model = task.build_model(args) criterion = task.build_criterion(args) print(model) print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__)) print('| num. model params: {} (num. trained: {})'.format( sum(p.numel() for p in model.parameters()), sum(p.numel() for p in model.parameters() if p.requires_grad), )) # Build trainer trainer = Trainer(args, task, model, criterion) print('| training on {} GPUs'.format(args.distributed_world_size)) print('| max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) max_positions = utils.resolve_max_positions( task.max_positions(), model.max_positions(), ) # Initialize dataloader epoch_itr = task.get_batch_iterator( dataset=task.dataset(args.train_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=max_positions, ignore_invalid_inputs=True, required_batch_size_multiple=args.required_batch_size_multiple, seed=args.seed, num_shards=args.distributed_world_size, shard_id=args.distributed_rank, num_workers=args.num_workers, ) # Load the latest checkpoint if one is available load_checkpoint(args, trainer, epoch_itr, max_positions, task) # Train until the learning rate gets too small max_epoch = args.max_epoch or math.inf max_update = args.max_update or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() valid_losses = [None] valid_subsets = args.valid_subset.split(',') while lr > args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates( ) < max_update: # train for one epoch train(args, trainer, task, epoch_itr) if epoch_itr.epoch % args.validate_interval == 0: valid_losses = validate(args, trainer, task, epoch_itr, valid_subsets) # only use first validation loss to update the learning rate lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0]) # save checkpoint if epoch_itr.epoch % args.save_interval == 0: save_checkpoint(args, trainer, epoch_itr, valid_losses[0]) epoch_itr = reload_train(args, epoch_itr, max_positions, task) train_meter.stop() print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args): assert args.path is not None, '--path required for generation!' assert not args.sampling or args.nbest == args.beam, \ '--sampling requires --nbest to be equal to --beam' assert args.replace_unk is None or args.raw_text, \ '--replace-unk requires a raw text dataset (--raw-text)' import_user_module(args) if args.max_tokens is None and args.max_sentences is None: args.max_tokens = 12000 print(args) tgt_file = None hypo_file = None if args.output_dir: os.makedirs(args.output_dir, exist_ok=True) tgt_fn = os.path.join(args.output_dir, 'gold') hypo_fn = os.path.join(args.output_dir, 'candidate') tgt_file = open(tgt_fn, 'w', encoding='utf-8') hypo_file = open(hypo_fn, 'w', encoding='utf-8') use_cuda = torch.cuda.is_available() and not args.cpu # Load dataset splits task = tasks.setup_task(args) task.load_dataset(args.gen_subset) print('| {} {} {} examples'.format(args.data, args.gen_subset, len(task.dataset(args.gen_subset)))) # Set dictionaries try: src_dict = getattr(task, 'source_dictionary', None) except NotImplementedError: src_dict = None tgt_dict = task.target_dictionary # Load ensemble print('| loading model(s) from {}'.format(args.path)) models, _model_args = utils.load_ensemble_for_inference( args.path.split(':'), task, model_arg_overrides=eval(args.model_overrides), ) # Optimize ensemble for generation for model in models: model.make_generation_fast_( beamable_mm_beam_size=None if args.no_beamable_mm else args.beam, need_attn=args.print_alignment, ) if args.fp16: model.half() if use_cuda: model.cuda() # Load alignment dictionary for unknown word replacement # (None if no unknown word replacement, empty if no path to align dictionary) align_dict = utils.load_align_dict(args.replace_unk) # Load dataset (possibly sharded) itr = task.get_batch_iterator( dataset=task.dataset(args.gen_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=utils.resolve_max_positions( task.max_positions(), *[model.max_positions() for model in models]), ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test, required_batch_size_multiple=args.required_batch_size_multiple, num_shards=args.num_shards, shard_id=args.shard_id, num_workers=args.num_workers, ).next_epoch_itr(shuffle=False) # Initialize generator gen_timer = StopwatchMeter() generator = task.build_generator(args) # Generate and compute BLEU score if args.sacrebleu: scorer = bleu.SacrebleuScorer() else: scorer = bleu.Scorer(tgt_dict.pad(), tgt_dict.eos(), tgt_dict.unk()) num_sentences = 0 has_target = True with progress_bar.build_progress_bar(args, itr) as t: wps_meter = TimeMeter() for sample in t: sample = utils.move_to_cuda(sample) if use_cuda else sample if 'net_input' not in sample: continue prefix_tokens = None if args.prefix_size > 0: prefix_tokens = sample['target'][:, :args.prefix_size] gen_timer.start() hypos = task.inference_step(generator, models, sample, prefix_tokens) num_generated_tokens = sum(len(h[0]['tokens']) for h in hypos) gen_timer.stop(num_generated_tokens) for i, sample_id in enumerate(sample['id'].tolist()): has_target = sample['target'] is not None # Remove padding src_tokens = utils.strip_pad( sample['net_input']['src_tokens'][i, :], tgt_dict.pad()) target_tokens = None if has_target: target_tokens = utils.strip_pad( sample['target'][i, :], tgt_dict.pad()).int().cpu() # Either retrieve the original sentences or regenerate them from tokens. if align_dict is not None: src_str = task.dataset( args.gen_subset).src.get_original_text(sample_id) target_str = task.dataset( args.gen_subset).tgt.get_original_text(sample_id) else: if src_dict is not None: src_str = src_dict.string(src_tokens, args.remove_bpe) else: src_str = "" if has_target: target_str = tgt_dict.string(target_tokens, args.remove_bpe, escape_unk=True) if not args.quiet: if src_dict is not None: print('S-{}\t{}'.format( sample_id, src_str.encode(encoding='utf-8'))) if has_target: print('T-{}\t{}'.format( sample_id, target_str.encode(encoding='utf-8'))) # Process top predictions for j, hypo in enumerate( hypos[i][:min(len(hypos[i]), args.nbest)]): hypo_tokens, hypo_str, alignment = utils.post_process_prediction( hypo_tokens=hypo['tokens'].int().cpu(), src_str=src_str, alignment=hypo['alignment'].int().cpu() if hypo['alignment'] is not None else None, align_dict=align_dict, tgt_dict=tgt_dict, remove_bpe=args.remove_bpe, ) if not args.quiet: print('H-{}\t{}\t{}'.format( sample_id, hypo['score'], hypo_str.encode(encoding='utf-8'))) print('P-{}\t{}'.format( sample_id, ' '.join( map( lambda x: '{:.4f}'.format(x), hypo['positional_scores'].tolist(), )))) if args.print_alignment: print('A-{}\t{}'.format( sample_id, ' '.join( map(lambda x: str(utils.item(x)), alignment)))) # Score only the top hypothesis if has_target and j == 0: if align_dict is not None or args.remove_bpe is not None: # Convert back to tokens for evaluation with unk replacement and/or without BPE target_tokens = tgt_dict.encode_line( target_str, add_if_not_exist=True) if hasattr(scorer, 'add_string'): scorer.add_string(target_str, hypo_str) else: scorer.add(target_tokens, hypo_tokens) if args.output_dir: tgt_file.writelines(target_str + '\n') hypo_file.writelines(hypo_str + '\n') wps_meter.update(num_generated_tokens) t.log({'wps': round(wps_meter.avg)}) num_sentences += sample['nsentences'] tgt_file.close() hypo_file.close() print( '| Translated {} sentences ({} tokens) in {:.1f}s ({:.2f} sentences/s, {:.2f} tokens/s)' .format(num_sentences, gen_timer.n, gen_timer.sum, num_sentences / gen_timer.sum, 1. / gen_timer.avg)) if has_target: print('| Generate {} with beam={}: {}'.format(args.gen_subset, args.beam, scorer.result_string())) return scorer
def save_checkpoint(args, trainer, epoch_itr, val_loss): if args.no_save or not distributed_utils.is_master(args): return write_timer = StopwatchMeter() write_timer.start() epoch = epoch_itr.epoch end_of_epoch = epoch_itr.end_of_epoch() updates = trainer.get_num_updates() checkpoint_conds = collections.OrderedDict() checkpoint_conds['checkpoint{}.pt'.format(epoch)] = ( end_of_epoch and not args.no_epoch_checkpoints and epoch % args.save_interval == 0) checkpoint_conds['checkpoint_{}_{}.pt'.format( epoch, updates)] = (not end_of_epoch and args.save_interval_updates > 0 and updates % args.save_interval_updates == 0) checkpoint_conds['checkpoint_best.pt'] = ( val_loss is not None and (not hasattr(save_checkpoint, 'best') or val_loss < save_checkpoint.best)) checkpoint_conds[ 'checkpoint_last.pt'] = True # keep this last so that it's a symlink prev_best = getattr(save_checkpoint, 'best', val_loss) if val_loss is not None: save_checkpoint.best = min(val_loss, prev_best) extra_state = { 'train_iterator': epoch_itr.state_dict(), 'val_loss': val_loss, } if hasattr(save_checkpoint, 'best'): extra_state.update({'best': save_checkpoint.best}) checkpoints = [ os.path.join(args.save_dir, fn) for fn, cond in checkpoint_conds.items() if cond ] if len(checkpoints) > 0: trainer.save_checkpoint(checkpoints[0], extra_state) for cp in checkpoints[1:]: shutil.copyfile(checkpoints[0], cp) write_timer.stop() print( '| saved checkpoint {} (epoch {} @ {} updates) (writing took {} seconds)' .format(checkpoints[0], epoch, updates, write_timer.sum)) if not end_of_epoch and args.keep_interval_updates > 0: # remove old checkpoints; checkpoints are sorted in descending order checkpoints = checkpoint_utils.checkpoint_paths( args.save_dir, pattern=r'checkpoint_\d+_(\d+)\.pt', ) for old_chk in checkpoints[args.keep_interval_updates:]: if os.path.lexists(old_chk): os.remove(old_chk) if args.keep_last_epochs > 0: # remove old epoch checkpoints; checkpoints are sorted in descending order checkpoints = checkpoint_utils.checkpoint_paths( args.save_dir, pattern=r'checkpoint(\d+)\.pt', ) for old_chk in checkpoints[args.keep_last_epochs:]: if os.path.lexists(old_chk): os.remove(old_chk)
def main(args): if args.max_tokens is None: args.max_tokens = 6000 print(args) if not torch.cuda.is_available(): raise NotImplementedError('Training on CPU is not supported') torch.cuda.set_device(args.device_id) torch.manual_seed(args.seed) # Setup task, e.g., translation, language modeling, etc. task = tasks.setup_task(args) # Load dataset splits load_dataset_splits(args, task, ['train', 'valid']) # Build model and criterion model = task.build_model(args) criterion = task.build_criterion(args) print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__)) print('| num. model params: {}'.format( sum(p.numel() for p in model.parameters()))) # Build trainer if args.fp16: trainer = FP16Trainer(args, task, model, criterion) else: if torch.cuda.get_device_capability(0)[0] >= 7: print( '| NOTICE: your device may support faster training with --fp16' ) trainer = Trainer(args, task, model, criterion) print('| training on {} GPUs'.format(args.distributed_world_size)) print('| max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) # Initialize dataloader max_positions = trainer.get_model().max_positions() epoch_itr = data.EpochBatchIterator( dataset=task.dataset(args.train_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences_valid, max_positions=max_positions, ignore_invalid_inputs=True, required_batch_size_multiple=8, seed=args.seed, num_shards=args.distributed_world_size, shard_id=args.distributed_rank, ) # Load the latest checkpoint if one is available load_checkpoint(args, trainer, epoch_itr) # Send a dummy batch to warm the caching allocator dummy_batch = task.dataset('train').get_dummy_batch( args.max_tokens, max_positions) trainer.dummy_train_step(dummy_batch) # Train until the learning rate gets too small max_epoch = args.max_epoch or math.inf max_update = args.max_update or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() valid_losses = [None] valid_subsets = args.valid_subset.split(',') while lr > args.min_lr and epoch_itr.epoch <= max_epoch and trainer.get_num_updates( ) < max_update: # train for one epoch train(args, trainer, task, epoch_itr) if epoch_itr.epoch % args.validate_interval == 0: valid_losses = validate(args, trainer, task, epoch_itr, valid_subsets) # only use first validation loss to update the learning rate lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0]) # save checkpoint if epoch_itr.epoch % args.save_interval == 0: save_checkpoint(args, trainer, epoch_itr, valid_losses[0]) train_meter.stop() print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args): assert args.path is not None, '--path required for generation!' assert not args.sampling or args.nbest == args.beam, \ '--sampling requires --nbest to be equal to --beam' assert args.replace_unk is None or args.raw_text, \ '--replace-unk requires a raw text dataset (--raw-text)' if args.max_tokens is None and args.max_sentences is None: args.max_tokens = 12000 print(args) use_cuda = torch.cuda.is_available() and not args.cpu # Load dataset splits task = tasks.setup_task(args) task.load_dataset(args.gen_subset) print('| {} {} {} examples'.format(args.data, args.gen_subset, len(task.dataset(args.gen_subset)))) # Set dictionaries src_dict = task.source_dictionary tgt_dict = task.target_dictionary # Load ensemble print('| loading model(s) from {}'.format(args.path)) models, _model_args = utils.load_ensemble_for_inference( args.path.split(':'), task, model_arg_overrides=eval(args.model_overrides), ) # Optimize ensemble for generation for model in models: model.make_generation_fast_( beamable_mm_beam_size=None if args.no_beamable_mm else args.beam, need_attn=args.print_alignment, ) if args.fp16: model.half() # Load alignment dictionary for unknown word replacement # (None if no unknown word replacement, empty if no path to align dictionary) align_dict = utils.load_align_dict(args.replace_unk) # Load dataset (possibly sharded) itr = task.get_batch_iterator( dataset=task.dataset(args.gen_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=utils.resolve_max_positions( task.max_positions(), *[model.max_positions() for model in models] ), ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test, required_batch_size_multiple=8, num_shards=args.num_shards, shard_id=args.shard_id, num_workers=args.num_workers, ).next_epoch_itr(shuffle=False) # Initialize generator gen_timer = StopwatchMeter() if args.score_reference: translator = SequenceScorer(models, task.target_dictionary) else: translator = SequenceGenerator( models, task.target_dictionary, beam_size=args.beam, minlen=args.min_len, stop_early=(not args.no_early_stop), normalize_scores=(not args.unnormalized), len_penalty=args.lenpen, unk_penalty=args.unkpen, sampling=args.sampling, sampling_topk=args.sampling_topk, sampling_temperature=args.sampling_temperature, diverse_beam_groups=args.diverse_beam_groups, diverse_beam_strength=args.diverse_beam_strength, match_source_len=args.match_source_len, no_repeat_ngram_size=args.no_repeat_ngram_size, ) if use_cuda: translator.cuda() # Generate and compute BLEU score scorer = bleu.Scorer(tgt_dict.pad(), tgt_dict.eos(), tgt_dict.unk()) num_sentences = 0 has_target = True # output the result result=['']*21678 with progress_bar.build_progress_bar(args, itr) as t: if args.score_reference: translations = translator.score_batched_itr(t, cuda=use_cuda, timer=gen_timer) else: translations = translator.generate_batched_itr( t, maxlen_a=args.max_len_a, maxlen_b=args.max_len_b, cuda=use_cuda, timer=gen_timer, prefix_size=args.prefix_size, ) wps_meter = TimeMeter() for sample_id, src_tokens, target_tokens, hypos in translations: # Process input and ground truth has_target = target_tokens is not None target_tokens = target_tokens.int().cpu() if has_target else None # Either retrieve the original sentences or regenerate them from tokens. if align_dict is not None: src_str = task.dataset(args.gen_subset).src.get_original_text(sample_id) target_str = task.dataset(args.gen_subset).tgt.get_original_text(sample_id) else: src_str = src_dict.string(src_tokens, args.remove_bpe) if has_target: target_str = tgt_dict.string(target_tokens, args.remove_bpe, escape_unk=True) if not args.quiet: print('S-{}\t{}'.format(sample_id, src_str)) if has_target: print('T-{}\t{}'.format(sample_id, target_str)) # Process top predictions for i, hypo in enumerate(hypos[:min(len(hypos), args.nbest)]): hypo_tokens, hypo_str, alignment = utils.post_process_prediction( hypo_tokens=hypo['tokens'].int().cpu(), src_str=src_str, alignment=hypo['alignment'].int().cpu() if hypo['alignment'] is not None else None, align_dict=align_dict, tgt_dict=tgt_dict, remove_bpe=args.remove_bpe, ) result[sample_id]=hypo_str if not args.quiet: print('H-{}\t{}\t{}'.format(sample_id, hypo['score'], hypo_str)) print('P-{}\t{}'.format( sample_id, ' '.join(map( lambda x: '{:.4f}'.format(x), hypo['positional_scores'].tolist(), )) )) if args.print_alignment: print('A-{}\t{}'.format( sample_id, ' '.join(map(lambda x: str(utils.item(x)), alignment)) )) # Score only the top hypothesis if has_target and i == 0: if align_dict is not None or args.remove_bpe is not None: # Convert back to tokens for evaluation with unk replacement and/or without BPE target_tokens = tokenizer.Tokenizer.tokenize( target_str, tgt_dict, add_if_not_exist=True) scorer.add(target_tokens, hypo_tokens) wps_meter.update(src_tokens.size(0)) t.log({'wps': round(wps_meter.avg)}) num_sentences += 1 print('| Translated {} sentences ({} tokens) in {:.1f}s ({:.2f} sentences/s, {:.2f} tokens/s)'.format( num_sentences, gen_timer.n, gen_timer.sum, num_sentences / gen_timer.sum, 1. / gen_timer.avg)) if has_target: print('| Generate {} with beam={}: {}'.format(args.gen_subset, args.beam, scorer.result_string())) #output the result return result
def main(args): # we should not do this! ''' if args.max_tokens is None: args.max_tokens = 6000 ''' utils.xpprint(args) if not torch.cuda.is_available(): raise NotImplementedError('Training on CPU is not supported') torch.cuda.set_device(args.device_id) torch.manual_seed(args.seed) # Setup task, e.g., translation, language modeling, etc. task = tasks.setup_task(args) utils.xprintln('setup task done!') # Load dataset splits load_dataset_splits(args, task, ['train']) valid_dataset = args.valid_subset.split(',') load_dataset_splits(args, task, valid_dataset, shuffle=False) utils.xprintln('load dataset done!') if args.task.startswith('extractive_summarization'): if distributed_utils.is_master(args): from sum_eval import MultiProcSumEval sum_eval_pool = MultiProcSumEval(args.ncpu_eval) sum_valid_pool_params = dict( article_file=args.raw_valid + '.article', summary_file=args.raw_valid + '.summary', entity_map_file=None, length=-1, eval_type='predict', topk=args.topk_sent_eval, rerank=False, with_m=False, cmd='-a -c 95 -m -n 4 -w 1.2', trigram_block=args.trigram_block, ) sum_test_pool_params = dict( article_file=args.raw_test + '.article', summary_file=args.raw_test + '.summary', entity_map_file=None, length=-1, eval_type='predict', topk=args.topk_sent_eval, rerank=False, with_m=False, cmd='-a -c 95 -m -n 4 -w 1.2', trigram_block=args.trigram_block, ) sum_pool_params = dict(valid=sum_valid_pool_params, test=sum_test_pool_params) def make_params(default_dict, result_file, out_rouge_file, rerank=False, with_m=False): para_dict = dict(default_dict) para_dict['result_file'] = result_file para_dict['out_rouge_file'] = out_rouge_file para_dict['rerank'] = rerank para_dict['with_m'] = with_m return para_dict # Build model and criterion model = task.build_model(args) criterion = task.build_criterion(args) print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__)) print('| num. model params: {}'.format( sum(p.numel() for p in model.parameters()))) # print(model) import sys sys.stdout.flush() # if summarization try to load pretrained model # if args.task.startswith('extractive_summarization') or args.task == 'pretrain_document_modeling': # # assume this is a single GPU program if args.init_from_pretrained_doc_model: task.load_pretrained_model(model, args.pretrained_doc_model_path) sys.stdout.flush() # Build trainer trainer = Trainer(args, task, model, criterion) print('| training on {} GPUs'.format(args.distributed_world_size)) print('| max tokens per GPU = {} and max sentences per GPU = {}'.format( args.max_tokens, args.max_sentences, )) # Initialize dataloader max_positions = trainer.get_model().max_positions() epoch_itr = trainer.get_train_iterator(epoch=0, load_dataset=False) # Load the latest checkpoint if one is available # load_checkpoint(args, trainer, epoch_itr) # make sure training from a different checkpoint will use different random seed cur_dataset = task.dataset('train') if hasattr(cur_dataset, 'rng'): print('epoch ', epoch_itr.epoch) cur_dataset.rng = numpy.random.RandomState(args.seed + epoch_itr.epoch) # Train until the learning rate gets too small max_epoch = args.max_epoch or math.inf max_update = args.max_update or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() valid_losses = [None] valid_subsets = args.valid_subset.split(',') for alpha in range(10, 9, -1): # train for one epoch # train(args, trainer, task, epoch_itr) epoch_itr.next_epoch_itr() if epoch_itr.epoch % args.validate_interval == 0: if args.task.startswith('extractive_summarization'): if distributed_utils.is_master(args): validate_metric(args, trainer, task, epoch_itr, valid_subsets)
def infer_onebyone(args, models, task, input): assert args.path is not None, '--path required for generation!' assert not args.sampling or args.nbest == args.beam, \ '--sampling requires --nbest to be equal to --beam' assert args.replace_unk is None or args.raw_text, \ '--replace-unk requires a raw text dataset (--raw-text)' utils.import_user_module(args) if args.max_tokens is None and args.max_sentences is None: args.max_tokens = 12000 print(args) use_cuda = torch.cuda.is_available() and not args.cpu # Set dictionaries try: src_dict = getattr(task, 'source_dictionary', None) except NotImplementedError: src_dict = None tgt_dict = task.target_dictionary input = ' '.join([i for i in input]) src_tokens = tgt_dict.encode_line(input).type(torch.LongTensor) input_sample = { 'id': torch.Tensor([0]), 'nsentences': 1, 'ntokens': len(src_tokens), 'net_input': { 'src_tokens': src_tokens.unsqueeze(0), 'src_lengths': torch.tensor([len(src_tokens)]), 'prev_output_tokens': torch.tensor([[tgt_dict.eos()]]) }, 'target': src_tokens.unsqueeze(0), } # Load ensemble print('| loading model(s) from {}'.format(args.path)) # Optimize ensemble for generation # Load alignment dictionary for unknown word replacement # (None if no unknown word replacement, empty if no path to align dictionary) align_dict = utils.load_align_dict(args.replace_unk) # Load dataset (possibly sharded) # Initialize generator gen_timer = StopwatchMeter() generator = task.build_generator(args) # Generate and compute BLEU score if args.sacrebleu: scorer = bleu.SacrebleuScorer() else: scorer = bleu.Scorer(tgt_dict.pad(), tgt_dict.eos(), tgt_dict.unk()) num_sentences = 0 has_target = True wps_meter = TimeMeter() sample = input_sample sample = utils.move_to_cuda(sample) if use_cuda else sample prefix_tokens = None if args.prefix_size > 0: prefix_tokens = sample['target'][:, :args.prefix_size] gen_timer.start() #pdb.set_trace() hypos = task.inference_step(generator, models, sample, prefix_tokens) num_generated_tokens = sum(len(h[0]['tokens']) for h in hypos) gen_timer.stop(num_generated_tokens) for i, sample_id in enumerate(sample['id'].tolist()): has_target = sample['target'] is not None # Remove padding src_tokens = utils.strip_pad(sample['net_input']['src_tokens'][i, :], tgt_dict.pad()) target_tokens = None if has_target: target_tokens = utils.strip_pad(sample['target'][i, :], tgt_dict.pad()).int().cpu() # Either retrieve the original sentences or regenerate them from tokens. if align_dict is not None: src_str = task.dataset( args.gen_subset).src.get_original_text(sample_id) target_str = task.dataset( args.gen_subset).tgt.get_original_text(sample_id) else: if src_dict is not None: src_str = src_dict.string(src_tokens, args.remove_bpe) else: src_str = "" if has_target: target_str = tgt_dict.string(target_tokens, args.remove_bpe, escape_unk=True) if not args.quiet: if src_dict is not None: print('S-{}\t{}'.format(sample_id, src_str)) if has_target: print('T-{}\t{}'.format(sample_id, target_str)) # Process top predictions for j, hypo in enumerate(hypos[i][:args.nbest]): hypo_tokens, hypo_str, alignment = utils.post_process_prediction( hypo_tokens=hypo['tokens'].int().cpu(), src_str=src_str, alignment=hypo['alignment'], align_dict=align_dict, tgt_dict=tgt_dict, remove_bpe=args.remove_bpe, ) if not args.quiet: output = ''.join(hypo_str.split(' ')) print('H-{}\t{}\t{}'.format(sample_id, hypo['score'], hypo_str)) print('P-{}\t{}'.format( sample_id, ' '.join( map( lambda x: '{:.4f}'.format(x), hypo['positional_scores'].tolist(), )))) if args.print_alignment: print('A-{}\t{}'.format( sample_id, ' '.join([ '{}-{}'.format(src_idx, tgt_idx) for src_idx, tgt_idx in alignment ]))) if args.print_step: print('I-{}\t{}'.format(sample_id, hypo['steps'])) # Score only the top hypothesis if has_target and j == 0: if align_dict is not None or args.remove_bpe is not None: # Convert back to tokens for evaluation with unk replacement and/or without BPE target_tokens = tgt_dict.encode_line(target_str, add_if_not_exist=True) if hasattr(scorer, 'add_string'): scorer.add_string(target_str, hypo_str) else: scorer.add(target_tokens, hypo_tokens) wps_meter.update(num_generated_tokens) num_sentences += sample['nsentences'] print( '| Translated {} sentences ({} tokens) in {:.1f}s ({:.2f} sentences/s, {:.2f} tokens/s)' .format(num_sentences, gen_timer.n, gen_timer.sum, num_sentences / gen_timer.sum, 1. / gen_timer.avg)) if has_target: print('| Generate {} with beam={}: {}'.format(args.gen_subset, args.beam, scorer.result_string())) return scorer, output
def main(args): assert args.path is not None, '--path required for recognition!' assert not args.sampling or args.nbest == args.beam, \ '--sampling requires --nbest to be equal to --beam' utils.import_user_module(args) if args.max_tokens is None and args.max_sentences is None: args.max_tokens = 12000 print(args) use_cuda = torch.cuda.is_available() and not args.cpu # Load dataset split task = tasks.setup_task(args) task.load_dataset(args.gen_subset) # Set dictionary dict = task.target_dictionary # Load ensemble print('| loading model(s) from {}'.format(args.path)) models, _model_args = checkpoint_utils.load_model_ensemble( args.path.split(':'), arg_overrides=eval(args.model_overrides), task=task, ) for i, m in enumerate(models): if hasattr(m, 'is_wordlm') and m.is_wordlm: # assume subword LM comes before word LM if isinstance(models[i - 1], FairseqLanguageModel): models[i - 1] = MultiLevelLanguageModel( m, models[i - 1], subwordlm_weight=args.subwordlm_weight, oov_penalty=args.oov_penalty, open_vocab=not args.disable_open_vocab, ) del models[i] print('| LM fusion with Multi-level LM') else: models[i] = TensorizedLookaheadLanguageModel( m, dict, oov_penalty=args.oov_penalty, open_vocab=not args.disable_open_vocab, ) print('| LM fusion with Look-ahead Word LM') # assume subword LM comes after E2E models elif i == len(models) - 1 and isinstance(m, FairseqLanguageModel): print('| LM fusion with Subword LM') if args.lm_weight != 0.0: print('| using LM fusion with lm-weight={:.2f}'.format(args.lm_weight)) # Optimize ensemble for generation for model in models: model.make_generation_fast_( beamable_mm_beam_size=None if args.no_beamable_mm else args.beam, need_attn=args.print_alignment, ) if args.fp16: model.half() if use_cuda: model.cuda() # Load dataset (possibly sharded) itr = task.get_batch_iterator( dataset=task.dataset(args.gen_subset), max_tokens=args.max_tokens, max_sentences=args.max_sentences, max_positions=utils.resolve_max_positions( task.max_positions(), *[ model.max_positions() if hasattr(model, 'encoder') else (None, model.max_positions()) for model in models ]), ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test, required_batch_size_multiple=8, num_shards=args.num_shards, shard_id=args.shard_id, num_workers=args.num_workers, ).next_epoch_itr(shuffle=False) # Initialize generator if args.match_source_len: print('| The option match_source_len is not applicable to ' 'speech recognition. Ignoring it.') gen_timer = StopwatchMeter() generator = task.build_generator(args) # Generate and compute WER scorer = wer.Scorer(dict, wer_output_filter=args.wer_output_filter) num_sentences = 0 has_target = True with progress_bar.build_progress_bar(args, itr) as t: wps_meter = TimeMeter() for sample in t: sample = utils.move_to_cuda(sample) if use_cuda else sample if 'net_input' not in sample: continue prefix_tokens = None if args.prefix_size > 0: prefix_tokens = sample['target'][:, :args.prefix_size] gen_timer.start() hypos = task.inference_step( generator, models, sample, prefix_tokens, lm_weight=args.lm_weight, ) num_generated_tokens = sum(len(h[0]['tokens']) for h in hypos) gen_timer.stop(num_generated_tokens) # obtain nonpad mask of encoder output to plot attentions if args.print_alignment: net_input = sample['net_input'] src_tokens = net_input['src_tokens'] output_lengths = models[0].encoder.output_lengths( net_input['src_lengths']) nonpad_idxs = sequence_mask( output_lengths, models[0].encoder.output_lengths(src_tokens.size(1))) for i, sample_id in enumerate(sample['id'].tolist()): has_target = sample['target'] is not None utt_id = sample['utt_id'][i] # Retrieve the original sentences if has_target: target_str = task.dataset( args.gen_subset).tgt.get_original_tokens(sample_id) if not args.quiet: target_sent = dict.tokens_to_sentence( target_str, use_unk_sym=False, bpe_symbol=args.remove_bpe, ) print('T-{}\t{}'.format(utt_id, target_sent)) # Process top predictions for j, hypo in enumerate(hypos[i][:args.nbest]): hypo_str = dict.string(hypo['tokens'].int().cpu() ) # not removing bpe at this point if not args.quiet or i == 0: hypo_sent = dict.tokens_to_sentence( hypo_str, bpe_symbol=args.remove_bpe) if not args.quiet: print('H-{}\t{}\t{}'.format(utt_id, hypo_sent, hypo['score'])) # Score and obtain attention only the top hypothesis if j == 0: # src_len x tgt_len attention = hypo['attention'][nonpad_idxs[i]].float().cpu() \ if args.print_alignment and hypo['attention'] is not None else None if args.print_alignment and attention is not None: save_dir = os.path.join(args.results_path, 'attn_plots') os.makedirs(save_dir, exist_ok=True) plot_attention(attention, hypo_sent, utt_id, save_dir) scorer.add_prediction(utt_id, hypo_str, bpe_symbol=args.remove_bpe) if has_target: scorer.add_evaluation(utt_id, target_str, hypo_str, bpe_symbol=args.remove_bpe) wps_meter.update(num_generated_tokens) t.log({'wps': round(wps_meter.avg)}) num_sentences += sample['nsentences'] print( '| Recognized {} utterances ({} tokens) in {:.1f}s ({:.2f} sentences/s, {:.2f} tokens/s)' .format(num_sentences, gen_timer.n, gen_timer.sum, num_sentences / gen_timer.sum, 1. / gen_timer.avg)) if args.print_alignment: print('| Saved attention plots in ' + save_dir) if has_target: assert args.test_text_files is not None scorer.add_ordered_utt_list(*args.test_text_files) os.makedirs(args.results_path, exist_ok=True) fn = 'decoded_char_results.txt' with open(os.path.join(args.results_path, fn), 'w', encoding='utf-8') as f: f.write(scorer.print_char_results()) print('| Decoded char results saved as ' + f.name) fn = 'decoded_results.txt' with open(os.path.join(args.results_path, fn), 'w', encoding='utf-8') as f: f.write(scorer.print_results()) print('| Decoded results saved as ' + f.name) if has_target: header = ' Recognize {} with beam={}: '.format(args.gen_subset, args.beam) fn = 'wer' with open(os.path.join(args.results_path, fn), 'w', encoding='utf-8') as f: res = 'WER={:.2f}%, Sub={:.2f}%, Ins={:.2f}%, Del={:.2f}%'.format( *(scorer.wer())) print('|' + header + res) f.write(res + '\n') print('| WER saved in ' + f.name) fn = 'cer' with open(os.path.join(args.results_path, fn), 'w', encoding='utf-8') as f: res = 'CER={:.2f}%, Sub={:.2f}%, Ins={:.2f}%, Del={:.2f}%'.format( *(scorer.cer())) print('|' + ' ' * len(header) + res) f.write(res + '\n') print('| CER saved in ' + f.name) fn = 'aligned_results.txt' with open(os.path.join(args.results_path, fn), 'w', encoding='utf-8') as f: f.write(scorer.print_aligned_results()) print('| Aligned results saved as ' + f.name) return scorer
def main(): parser = options.get_parser('Trainer') dataset_args = options.add_dataset_args(parser) dataset_args.add_argument('--max-tokens', default=0, type=int, metavar='N', help='maximum number of tokens in a batch') dataset_args.add_argument('--batch-size', default=32, type=int, metavar='N', help='batch size') dataset_args.add_argument('--test-batch-size', default=32, type=int, metavar='N', help='batch size for test set') dataset_args.add_argument('--valid-batch-size', default=32, type=int, metavar='N', help='batch size for validation set') dataset_args.add_argument( '--train-subset', default='train', metavar='SPLIT', choices=['train', 'valid', 'test'], help='data subset to use for training (train, valid, test)') dataset_args.add_argument( '--valid-subset', default='valid', metavar='SPLIT', help='comma separated list ofdata subsets ' ' to use for validation (train, valid, valid1,test, test1)') dataset_args.add_argument('--test-subset', default='test', metavar='SPLIT', help='comma separated list ofdata subset ' 'to use for testing (train, valid, test)') dataset_args.add_argument( '--valid-script', nargs='+', metavar='PATH', help='path to external validation script (optional).') options.add_optimization_args(parser) options.add_checkpoint_args(parser) options.add_model_args(parser) args = utils.parse_args_and_arch(parser) print(args) if args.no_progress_bar: progress_bar.enabled = False progress_bar.print_interval = args.log_interval if not os.path.exists(args.save_dir): os.makedirs(args.save_dir) torch.manual_seed(args.seed) # Setting args.max_tokens to infinity(same as setting to None) if args.max_tokens == 0: args.max_tokens = None # Load dataset dataset = data.load_with_check(args.data, args.source_lang, args.target_lang) if args.source_lang is None or args.target_lang is None: # record inferred languages in args, so that it's saved in checkpoints args.source_lang, args.target_lang = dataset.src, dataset.dst print('| [{}] dictionary: {} types'.format(dataset.src, len(dataset.src_dict))) print('| [{}] dictionary: {} types'.format(dataset.dst, len(dataset.dst_dict))) for split in dataset.splits: print('| {} {} {} examples'.format(args.data, split, len(dataset.splits[split]))) if not torch.cuda.is_available(): raise NotImplementedError('Training on CPU is not supported') num_gpus = torch.cuda.device_count() print('| using {} GPUs (with max tokens per GPU = {})'.format( num_gpus, args.max_tokens)) # Build model print('| model {}'.format(args.arch)) model = utils.build_model(args, dataset) criterion = utils.build_criterion(args, dataset) # Start multiprocessing trainer = MultiprocessingTrainer(args, model) # Load the latest checkpoint if one is available epoch, batch_offset = trainer.load_checkpoint( os.path.join(args.save_dir, args.restore_file)) # Train until the learning rate gets too small val_loss = None max_epoch = args.max_epoch or math.inf lr = trainer.get_lr() train_meter = StopwatchMeter() train_meter.start() while lr > args.min_lr and epoch <= max_epoch: # train for one epoch train(args, epoch, batch_offset, trainer, criterion, dataset, num_gpus) # evaluate on validate set for k, subset in enumerate(args.valid_subset.split(',')): val_loss = validate(args, epoch, trainer, criterion, dataset, subset, num_gpus) if k == 0: if not args.no_save: # save checkpoint trainer.save_checkpoint( args, epoch, 0, val_loss, validation_script=args.valid_script) # only use first validation loss to update the learning schedule lr = trainer.lr_step(val_loss, epoch) epoch += 1 batch_offset = 0 train_meter.stop() print('| done training in {:.1f} seconds'.format(train_meter.sum)) # Generate on test set and compute BLEU score for beam in [1, 5, 10, 20]: for subset in args.test_subset.split(','): scorer = score_test(args, trainer.get_model(), dataset, subset, beam, cuda_device=(0 if num_gpus > 0 else None)) print('| Test on {} with beam={}: {}'.format( subset, beam, scorer.result_string())) # Stop multiprocessing trainer.stop()