def __init__(self, model_path, model_name): print("................................................") print("LOADING Biaffine Model") alphabet_path = os.path.join(model_path, 'alphabets/') model_name = os.path.join(model_path, model_name) self.word_alpha, self.char_alpha, self.tag_alpha, self.type_alpha = conllx_data.create_alphabets(alphabet_path, None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None) self.id2word = {v: k for k, v in self.word_alpha.instance2index.iteritems()} num_words = self.word_alpha.size() num_chars = self.char_alpha.size() num_pos = self.tag_alpha.size() num_types = self.type_alpha.size() print("Word Alphabet Size: %d" % num_words) print("Character Alphabet Size: %d" % num_chars) print("POS Alphabet Size: %d" % num_pos) print("Type Alphabet Size: %d" % num_types) def load_model_arguments_from_json(): arguments = json.load(open(arg_path, 'r')) return arguments['args'], arguments['kwargs'] arg_path = model_name + '.arg.json' args, kwargs = load_model_arguments_from_json() self.network = BiRecurrentConvBiAffine(*args, **kwargs) self.network.load_state_dict(torch.load(model_name)) self.network.id2word = self.id2word self.network.cuda() self.network.eval()
def main(): args_parser = argparse.ArgumentParser(description='Tuning with graph-based parsing') args_parser.add_argument('--test_phase', action='store_true', help='Load trained model and run testing phase.') args_parser.add_argument('--mode', choices=['RNN', 'LSTM', 'GRU', 'FastLSTM'], help='architecture of rnn', required=True) args_parser.add_argument('--cuda', action='store_true', help='using GPU') args_parser.add_argument('--num_epochs', type=int, default=200, help='Number of training epochs') args_parser.add_argument('--batch_size', type=int, default=64, help='Number of sentences in each batch') args_parser.add_argument('--hidden_size', type=int, default=256, help='Number of hidden units in RNN') args_parser.add_argument('--arc_space', type=int, default=128, help='Dimension of tag space') args_parser.add_argument('--type_space', type=int, default=128, help='Dimension of tag space') args_parser.add_argument('--num_layers', type=int, default=1, help='Number of layers of RNN') args_parser.add_argument('--num_filters', type=int, default=50, help='Number of filters in CNN') args_parser.add_argument('--pos', action='store_true', help='use part-of-speech embedding.') args_parser.add_argument('--char', action='store_true', help='use character embedding and CNN.') args_parser.add_argument('--pos_dim', type=int, default=50, help='Dimension of POS embeddings') args_parser.add_argument('--char_dim', type=int, default=50, help='Dimension of Character embeddings') args_parser.add_argument('--opt', choices=['adam', 'sgd', 'adamax'], help='optimization algorithm') args_parser.add_argument('--objective', choices=['cross_entropy', 'crf'], default='cross_entropy', help='objective function of training procedure.') args_parser.add_argument('--decode', choices=['mst', 'greedy'], help='decoding algorithm', required=True) args_parser.add_argument('--learning_rate', type=float, default=0.01, help='Learning rate') args_parser.add_argument('--decay_rate', type=float, default=0.05, help='Decay rate of learning rate') args_parser.add_argument('--clip', type=float, default=5.0, help='gradient clipping') args_parser.add_argument('--gamma', type=float, default=0.0, help='weight for regularization') args_parser.add_argument('--epsilon', type=float, default=1e-8, help='epsilon for adam or adamax') args_parser.add_argument('--p_rnn', nargs=2, type=float, required=True, help='dropout rate for RNN') args_parser.add_argument('--p_in', type=float, default=0.33, help='dropout rate for input embeddings') args_parser.add_argument('--p_out', type=float, default=0.33, help='dropout rate for output layer') args_parser.add_argument('--schedule', type=int, help='schedule for learning rate decay') args_parser.add_argument('--unk_replace', type=float, default=0., help='The rate to replace a singleton word with UNK') args_parser.add_argument('--punctuation', nargs='+', type=str, help='List of punctuations') args_parser.add_argument('--word_embedding', choices=['glove', 'senna', 'sskip', 'polyglot'], help='Embedding for words', required=True) args_parser.add_argument('--word_path', help='path for word embedding dict') args_parser.add_argument('--freeze', action='store_true', help='frozen the word embedding (disable fine-tuning).') args_parser.add_argument('--char_embedding', choices=['random', 'polyglot'], help='Embedding for characters', required=True) args_parser.add_argument('--char_path', help='path for character embedding dict') args_parser.add_argument('--train') # "data/POS-penn/wsj/split1/wsj1.train.original" args_parser.add_argument('--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original" args_parser.add_argument('--test') # "data/POS-penn/wsj/split1/wsj1.test.original" args_parser.add_argument('--model_path', help='path for saving model file.', required=True) args_parser.add_argument('--model_name', help='name for saving model file.', required=True) args = args_parser.parse_args() logger = get_logger("GraphParser") mode = args.mode obj = args.objective decoding = args.decode train_path = args.train dev_path = args.dev test_path = args.test model_path = args.model_path model_name = args.model_name num_epochs = args.num_epochs batch_size = args.batch_size hidden_size = args.hidden_size arc_space = args.arc_space type_space = args.type_space num_layers = args.num_layers num_filters = args.num_filters learning_rate = args.learning_rate opt = args.opt momentum = 0.9 betas = (0.9, 0.9) eps = args.epsilon decay_rate = args.decay_rate clip = args.clip gamma = args.gamma schedule = args.schedule p_rnn = tuple(args.p_rnn) p_in = args.p_in p_out = args.p_out unk_replace = args.unk_replace punctuation = args.punctuation freeze = args.freeze word_embedding = args.word_embedding word_path = args.word_path use_char = args.char char_embedding = args.char_embedding char_path = args.char_path use_pos = args.pos pos_dim = args.pos_dim word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path) char_dict = None char_dim = args.char_dim if char_embedding != 'random': char_dict, char_dim = utils.load_embedding_dict(char_embedding, char_path) logger.info("Creating Alphabets") alphabet_path = os.path.join(model_path, 'alphabets/') model_name = os.path.join(model_path, model_name) word_alphabet, char_alphabet, pos_alphabet, type_alphabet = conllx_data.create_alphabets(alphabet_path, train_path, data_paths=[dev_path, test_path], max_vocabulary_size=100000, embedd_dict=word_dict) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) logger.info("Reading Data") device = torch.device('cuda') if args.cuda else torch.device('cpu') data_train = conllx_data.read_data_to_tensor(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, symbolic_root=True, device=device) # data_train = conllx_data.read_data(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet) # num_data = sum([len(bucket) for bucket in data_train]) num_data = sum(data_train[1]) data_dev = conllx_data.read_data_to_tensor(dev_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, symbolic_root=True, device=device) data_test = conllx_data.read_data_to_tensor(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, symbolic_root=True, device=device) punct_set = None if punctuation is not None: punct_set = set(punctuation) logger.info("punctuations(%d): %s" % (len(punct_set), ' '.join(punct_set))) def construct_word_embedding_table(): scale = np.sqrt(3.0 / word_dim) table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.zeros([1, word_dim]).astype(np.float32) if freeze else np.random.uniform(-scale, scale, [1, word_dim]).astype(np.float32) oov = 0 for word, index in word_alphabet.items(): if word in word_dict: embedding = word_dict[word] elif word.lower() in word_dict: embedding = word_dict[word.lower()] else: embedding = np.zeros([1, word_dim]).astype(np.float32) if freeze else np.random.uniform(-scale, scale, [1, word_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('word OOV: %d' % oov) return torch.from_numpy(table) def construct_char_embedding_table(): if char_dict is None: return None scale = np.sqrt(3.0 / char_dim) table = np.empty([num_chars, char_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.random.uniform(-scale, scale, [1, char_dim]).astype(np.float32) oov = 0 for char, index, in char_alphabet.items(): if char in char_dict: embedding = char_dict[char] else: embedding = np.random.uniform(-scale, scale, [1, char_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('character OOV: %d' % oov) return torch.from_numpy(table) word_table = construct_word_embedding_table() char_table = construct_char_embedding_table() window = 3 if obj == 'cross_entropy': network = BiRecurrentConvBiAffine(word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, num_filters, window, mode, hidden_size, num_layers, num_types, arc_space, type_space, embedd_word=word_table, embedd_char=char_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, biaffine=True, pos=use_pos, char=use_char) elif obj == 'crf': raise NotImplementedError else: raise RuntimeError('Unknown objective: %s' % obj) def save_args(): arg_path = model_name + '.arg.json' arguments = [word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, num_filters, window, mode, hidden_size, num_layers, num_types, arc_space, type_space] kwargs = {'p_in': p_in, 'p_out': p_out, 'p_rnn': p_rnn, 'biaffine': True, 'pos': use_pos, 'char': use_char} json.dump({'args': arguments, 'kwargs': kwargs}, open(arg_path, 'w'), indent=4) if freeze: freeze_embedding(network.word_embedd) network = network.to(device) save_args() pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) def generate_optimizer(opt, lr, params): params = filter(lambda param: param.requires_grad, params) if opt == 'adam': return Adam(params, lr=lr, betas=betas, weight_decay=gamma, eps=eps) elif opt == 'sgd': return SGD(params, lr=lr, momentum=momentum, weight_decay=gamma, nesterov=True) elif opt == 'adamax': return Adamax(params, lr=lr, betas=betas, weight_decay=gamma, eps=eps) else: raise ValueError('Unknown optimization algorithm: %s' % opt) lr = learning_rate optim = generate_optimizer(opt, lr, network.parameters()) opt_info = 'opt: %s, ' % opt if opt == 'adam': opt_info += 'betas=%s, eps=%.1e' % (betas, eps) elif opt == 'sgd': opt_info += 'momentum=%.2f' % momentum elif opt == 'adamax': opt_info += 'betas=%s, eps=%.1e' % (betas, eps) word_status = 'frozen' if freeze else 'fine tune' char_status = 'enabled' if use_char else 'disabled' pos_status = 'enabled' if use_pos else 'disabled' logger.info("Embedding dim: word=%d (%s), char=%d (%s), pos=%d (%s)" % (word_dim, word_status, char_dim, char_status, pos_dim, pos_status)) logger.info("CNN: filter=%d, kernel=%d" % (num_filters, window)) logger.info("RNN: %s, num_layer=%d, hidden=%d, arc_space=%d, type_space=%d" % (mode, num_layers, hidden_size, arc_space, type_space)) logger.info("train: obj: %s, l2: %f, (#data: %d, batch: %d, clip: %.2f, unk replace: %.2f)" % (obj, gamma, num_data, batch_size, clip, unk_replace)) logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" % (p_in, p_out, p_rnn)) logger.info("decoding algorithm: %s" % decoding) logger.info(opt_info) num_batches = num_data / batch_size + 1 dev_ucorrect = 0.0 dev_lcorrect = 0.0 dev_ucomlpete_match = 0.0 dev_lcomplete_match = 0.0 dev_ucorrect_nopunc = 0.0 dev_lcorrect_nopunc = 0.0 dev_ucomlpete_match_nopunc = 0.0 dev_lcomplete_match_nopunc = 0.0 dev_root_correct = 0.0 best_epoch = 0 test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete_match = 0.0 test_lcomplete_match = 0.0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_match_nopunc = 0.0 test_lcomplete_match_nopunc = 0.0 test_root_correct = 0.0 test_total = 0 test_total_nopunc = 0 test_total_inst = 0 test_total_root = 0 if decoding == 'greedy': decode = network.decode elif decoding == 'mst': decode = network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) patient = 0 decay = 0 max_decay = 9 double_schedule_decay = 5 for epoch in range(1, num_epochs + 1): print('Epoch %d (%s, optim: %s, learning rate=%.6f, eps=%.1e, decay rate=%.2f (schedule=%d, patient=%d, decay=%d)): ' % (epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay)) train_err = 0. train_err_arc = 0. train_err_type = 0. train_total = 0. start_time = time.time() num_back = 0 network.train() for batch in range(1, num_batches + 1): word, char, pos, heads, types, masks, lengths = conllx_data.get_batch_tensor(data_train, batch_size, unk_replace=unk_replace) optim.zero_grad() loss_arc, loss_type = network.loss(word, char, pos, heads, types, mask=masks, length=lengths) loss = loss_arc + loss_type loss.backward() clip_grad_norm_(network.parameters(), clip) optim.step() with torch.no_grad(): num_inst = word.size(0) if obj == 'crf' else masks.sum() - word.size(0) train_err += loss * num_inst train_err_arc += loss_arc * num_inst train_err_type += loss_type * num_inst train_total += num_inst time_ave = (time.time() - start_time) / batch time_left = (num_batches - batch) * time_ave # update log if batch % 10 == 0: sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) log_info = 'train: %d/%d loss: %.4f, arc: %.4f, type: %.4f, time left: %.2fs' % (batch, num_batches, train_err / train_total, train_err_arc / train_total, train_err_type / train_total, time_left) sys.stdout.write(log_info) sys.stdout.flush() num_back = len(log_info) sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) print('train: %d loss: %.4f, arc: %.4f, type: %.4f, time: %.2fs' % (num_batches, train_err / train_total, train_err_arc / train_total, train_err_type / train_total, time.time() - start_time)) # evaluate performance on dev data with torch.no_grad(): network.eval() pred_filename = 'tmp/%spred_dev%d' % (str(uid), epoch) pred_writer.start(pred_filename) gold_filename = 'tmp/%sgold_dev%d' % (str(uid), epoch) gold_writer.start(gold_filename) dev_ucorr = 0.0 dev_lcorr = 0.0 dev_total = 0 dev_ucomlpete = 0.0 dev_lcomplete = 0.0 dev_ucorr_nopunc = 0.0 dev_lcorr_nopunc = 0.0 dev_total_nopunc = 0 dev_ucomlpete_nopunc = 0.0 dev_lcomplete_nopunc = 0.0 dev_root_corr = 0.0 dev_total_root = 0.0 dev_total_inst = 0.0 for batch in conllx_data.iterate_batch_tensor(data_dev, batch_size): word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.cpu().numpy() pos = pos.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.cpu().numpy() types = types.cpu().numpy() pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True) gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True) stats, stats_nopunc, stats_root, num_inst = parser.eval(word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root dev_ucorr += ucorr dev_lcorr += lcorr dev_total += total dev_ucomlpete += ucm dev_lcomplete += lcm dev_ucorr_nopunc += ucorr_nopunc dev_lcorr_nopunc += lcorr_nopunc dev_total_nopunc += total_nopunc dev_ucomlpete_nopunc += ucm_nopunc dev_lcomplete_nopunc += lcm_nopunc dev_root_corr += corr_root dev_total_root += total_root dev_total_inst += num_inst pred_writer.close() gold_writer.close() print('W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % ( dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 / dev_total, dev_lcorr * 100 / dev_total, dev_ucomlpete * 100 / dev_total_inst, dev_lcomplete * 100 / dev_total_inst)) print('Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % ( dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc, dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc * 100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 / dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst)) print('Root: corr: %d, total: %d, acc: %.2f%%' %(dev_root_corr, dev_total_root, dev_root_corr * 100 / dev_total_root)) if dev_lcorrect_nopunc < dev_lcorr_nopunc or (dev_lcorrect_nopunc == dev_lcorr_nopunc and dev_ucorrect_nopunc < dev_ucorr_nopunc): dev_ucorrect_nopunc = dev_ucorr_nopunc dev_lcorrect_nopunc = dev_lcorr_nopunc dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc dev_lcomplete_match_nopunc = dev_lcomplete_nopunc dev_ucorrect = dev_ucorr dev_lcorrect = dev_lcorr dev_ucomlpete_match = dev_ucomlpete dev_lcomplete_match = dev_lcomplete dev_root_correct = dev_root_corr best_epoch = epoch patient = 0 # torch.save(network, model_name) torch.save(network.state_dict(), model_name) pred_filename = 'tmp/%spred_test%d' % (str(uid), epoch) pred_writer.start(pred_filename) gold_filename = 'tmp/%sgold_test%d' % (str(uid), epoch) gold_writer.start(gold_filename) test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete_match = 0.0 test_lcomplete_match = 0.0 test_total = 0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_match_nopunc = 0.0 test_lcomplete_match_nopunc = 0.0 test_total_nopunc = 0 test_total_inst = 0 test_root_correct = 0.0 test_total_root = 0 for batch in conllx_data.iterate_batch_tensor(data_test, batch_size): word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.cpu().numpy() pos = pos.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.cpu().numpy() types = types.cpu().numpy() pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True) gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True) stats, stats_nopunc, stats_root, num_inst = parser.eval(word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root test_ucorrect += ucorr test_lcorrect += lcorr test_total += total test_ucomlpete_match += ucm test_lcomplete_match += lcm test_ucorrect_nopunc += ucorr_nopunc test_lcorrect_nopunc += lcorr_nopunc test_total_nopunc += total_nopunc test_ucomlpete_match_nopunc += ucm_nopunc test_lcomplete_match_nopunc += lcm_nopunc test_root_correct += corr_root test_total_root += total_root test_total_inst += num_inst pred_writer.close() gold_writer.close() else: if dev_ucorr_nopunc * 100 / dev_total_nopunc < dev_ucorrect_nopunc * 100 / dev_total_nopunc - 5 or patient >= schedule: # network = torch.load(model_name) network.load_state_dict(torch.load(model_name)) lr = lr * decay_rate optim = generate_optimizer(opt, lr, network.parameters()) if decoding == 'greedy': decode = network.decode elif decoding == 'mst': decode = network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) patient = 0 decay += 1 if decay % double_schedule_decay == 0: schedule *= 2 else: patient += 1 print('----------------------------------------------------------------------------------------------------------------------------') print('best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( dev_ucorrect, dev_lcorrect, dev_total, dev_ucorrect * 100 / dev_total, dev_lcorrect * 100 / dev_total, dev_ucomlpete_match * 100 / dev_total_inst, dev_lcomplete_match * 100 / dev_total_inst, best_epoch)) print('best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc, dev_ucorrect_nopunc * 100 / dev_total_nopunc, dev_lcorrect_nopunc * 100 / dev_total_nopunc, dev_ucomlpete_match_nopunc * 100 / dev_total_inst, dev_lcomplete_match_nopunc * 100 / dev_total_inst, best_epoch)) print('best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % ( dev_root_correct, dev_total_root, dev_root_correct * 100 / dev_total_root, best_epoch)) print('----------------------------------------------------------------------------------------------------------------------------') print('best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 / test_total, test_lcorrect * 100 / test_total, test_ucomlpete_match * 100 / test_total_inst, test_lcomplete_match * 100 / test_total_inst, best_epoch)) print('best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc, test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc, test_ucomlpete_match_nopunc * 100 / test_total_inst, test_lcomplete_match_nopunc * 100 / test_total_inst, best_epoch)) print('best test Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % ( test_root_correct, test_total_root, test_root_correct * 100 / test_total_root, best_epoch)) print('============================================================================================================================') if decay == max_decay: break
def parse(args): logger = get_logger("Parsing") args.cuda = torch.cuda.is_available() device = torch.device('cuda', 0) if args.cuda else torch.device('cpu') test_path = args.test model_path = args.model_path model_name = os.path.join(model_path, 'model.pt') punctuation = args.punctuation print(args) logger.info("Creating Alphabets") alphabet_path = os.path.join(model_path, 'alphabets') assert os.path.exists(alphabet_path) word_alphabet, char_alphabet, pos_alphabet, type_alphabet = conllx_data.create_alphabets( alphabet_path, None) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) result_path = os.path.join(model_path, 'tmp') if not os.path.exists(result_path): os.makedirs(result_path) punct_set = None if punctuation is not None: punct_set = set(punctuation) logger.info("punctuations(%d): %s" % (len(punct_set), ' '.join(punct_set))) logger.info("loading network...") hyps = json.load(open(os.path.join(model_path, 'config.json'), 'r')) model_type = hyps['model'] assert model_type in ['DeepBiAffine', 'NeuroMST', 'StackPtr'] word_dim = hyps['word_dim'] char_dim = hyps['char_dim'] use_pos = hyps['pos'] pos_dim = hyps['pos_dim'] mode = hyps['rnn_mode'] hidden_size = hyps['hidden_size'] arc_space = hyps['arc_space'] type_space = hyps['type_space'] p_in = hyps['p_in'] p_out = hyps['p_out'] p_rnn = hyps['p_rnn'] activation = hyps['activation'] prior_order = None alg = 'transition' if model_type == 'StackPtr' else 'graph' if model_type == 'DeepBiAffine': num_layers = hyps['num_layers'] network = DeepBiAffine(word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, mode, hidden_size, num_layers, num_types, arc_space, type_space, p_in=p_in, p_out=p_out, p_rnn=p_rnn, pos=use_pos, activation=activation) elif model_type == 'NeuroMST': num_layers = hyps['num_layers'] network = NeuroMST(word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, mode, hidden_size, num_layers, num_types, arc_space, type_space, p_in=p_in, p_out=p_out, p_rnn=p_rnn, pos=use_pos, activation=activation) elif model_type == 'StackPtr': encoder_layers = hyps['encoder_layers'] decoder_layers = hyps['decoder_layers'] num_layers = (encoder_layers, decoder_layers) prior_order = hyps['prior_order'] grandPar = hyps['grandPar'] sibling = hyps['sibling'] network = StackPtrNet(word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, mode, hidden_size, encoder_layers, decoder_layers, num_types, arc_space, type_space, prior_order=prior_order, activation=activation, p_in=p_in, p_out=p_out, p_rnn=p_rnn, pos=use_pos, grandPar=grandPar, sibling=sibling) else: raise RuntimeError('Unknown model type: %s' % model_type) network = network.to(device) network.load_state_dict(torch.load(model_name, map_location=device)) model = "{}-{}".format(model_type, mode) logger.info("Network: %s, num_layer=%s, hidden=%d, act=%s" % (model, num_layers, hidden_size, activation)) logger.info("Reading Data") if alg == 'graph': data_test = conllx_data.read_data(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, symbolic_root=True) else: data_test = conllx_stacked_data.read_data(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, prior_order=prior_order) beam = args.beam pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) pred_filename = os.path.join(result_path, 'pred.txt') pred_writer.start(pred_filename) gold_filename = os.path.join(result_path, 'gold.txt') gold_writer.start(gold_filename) with torch.no_grad(): print('Parsing...') start_time = time.time() eval(alg, data_test, network, pred_writer, gold_writer, punct_set, word_alphabet, pos_alphabet, device, beam, batch_size=args.test_batch_size) print('Time: %.2fs' % (time.time() - start_time)) pred_writer.close() gold_writer.close()
def train(args): logger = get_logger("Parsing") args.cuda = torch.cuda.is_available() device = torch.device('cuda', 0) if args.cuda else torch.device('cpu') train_path = args.train dev_path = args.dev test_path = args.test num_epochs = args.num_epochs batch_size = args.batch_size optim = args.optim learning_rate = args.learning_rate lr_decay = args.lr_decay amsgrad = args.amsgrad eps = args.eps betas = (args.beta1, args.beta2) warmup_steps = args.warmup_steps weight_decay = args.weight_decay grad_clip = args.grad_clip loss_ty_token = args.loss_type == 'token' unk_replace = args.unk_replace freeze = args.freeze model_path = args.model_path model_name = os.path.join(model_path, 'model.pt') punctuation = args.punctuation word_embedding = args.word_embedding word_path = args.word_path char_embedding = args.char_embedding char_path = args.char_path print(args) word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path) char_dict = None if char_embedding != 'random': char_dict, char_dim = utils.load_embedding_dict( char_embedding, char_path) else: char_dict = None char_dim = None logger.info("Creating Alphabets") alphabet_path = os.path.join(model_path, 'alphabets') word_alphabet, char_alphabet, pos_alphabet, type_alphabet = conllx_data.create_alphabets( alphabet_path, train_path, data_paths=[dev_path, test_path], embedd_dict=word_dict, max_vocabulary_size=200000) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) result_path = os.path.join(model_path, 'tmp') if not os.path.exists(result_path): os.makedirs(result_path) punct_set = None if punctuation is not None: punct_set = set(punctuation) logger.info("punctuations(%d): %s" % (len(punct_set), ' '.join(punct_set))) def construct_word_embedding_table(): scale = np.sqrt(3.0 / word_dim) table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.zeros([1, word_dim]).astype( np.float32) if freeze else np.random.uniform( -scale, scale, [1, word_dim]).astype(np.float32) oov = 0 for word, index in word_alphabet.items(): if word in word_dict: embedding = word_dict[word] elif word.lower() in word_dict: embedding = word_dict[word.lower()] else: embedding = np.zeros([1, word_dim]).astype( np.float32) if freeze else np.random.uniform( -scale, scale, [1, word_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('word OOV: %d' % oov) return torch.from_numpy(table) def construct_char_embedding_table(): if char_dict is None: return None scale = np.sqrt(3.0 / char_dim) table = np.empty([num_chars, char_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.random.uniform( -scale, scale, [1, char_dim]).astype(np.float32) oov = 0 for char, index, in char_alphabet.items(): if char in char_dict: embedding = char_dict[char] else: embedding = np.random.uniform(-scale, scale, [1, char_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('character OOV: %d' % oov) return torch.from_numpy(table) word_table = construct_word_embedding_table() char_table = construct_char_embedding_table() logger.info("constructing network...") hyps = json.load(open(args.config, 'r')) json.dump(hyps, open(os.path.join(model_path, 'config.json'), 'w'), indent=2) model_type = hyps['model'] assert model_type in ['DeepBiAffine', 'NeuroMST', 'StackPtr'] assert word_dim == hyps['word_dim'] if char_dim is not None: assert char_dim == hyps['char_dim'] else: char_dim = hyps['char_dim'] use_pos = hyps['pos'] pos_dim = hyps['pos_dim'] mode = hyps['rnn_mode'] hidden_size = hyps['hidden_size'] arc_space = hyps['arc_space'] type_space = hyps['type_space'] p_in = hyps['p_in'] p_out = hyps['p_out'] p_rnn = hyps['p_rnn'] activation = hyps['activation'] prior_order = None alg = 'transition' if model_type == 'StackPtr' else 'graph' if model_type == 'DeepBiAffine': num_layers = hyps['num_layers'] network = DeepBiAffine(word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, mode, hidden_size, num_layers, num_types, arc_space, type_space, embedd_word=word_table, embedd_char=char_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, pos=use_pos, activation=activation) elif model_type == 'NeuroMST': num_layers = hyps['num_layers'] network = NeuroMST(word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, mode, hidden_size, num_layers, num_types, arc_space, type_space, embedd_word=word_table, embedd_char=char_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, pos=use_pos, activation=activation) elif model_type == 'StackPtr': encoder_layers = hyps['encoder_layers'] decoder_layers = hyps['decoder_layers'] num_layers = (encoder_layers, decoder_layers) prior_order = hyps['prior_order'] grandPar = hyps['grandPar'] sibling = hyps['sibling'] network = StackPtrNet(word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, mode, hidden_size, encoder_layers, decoder_layers, num_types, arc_space, type_space, embedd_word=word_table, embedd_char=char_table, prior_order=prior_order, activation=activation, p_in=p_in, p_out=p_out, p_rnn=p_rnn, pos=use_pos, grandPar=grandPar, sibling=sibling) else: raise RuntimeError('Unknown model type: %s' % model_type) if freeze: freeze_embedding(network.word_embed) network = network.to(device) model = "{}-{}".format(model_type, mode) logger.info("Network: %s, num_layer=%s, hidden=%d, act=%s" % (model, num_layers, hidden_size, activation)) logger.info("dropout(in, out, rnn): %s(%.2f, %.2f, %s)" % ('variational', p_in, p_out, p_rnn)) logger.info('# of Parameters: %d' % (sum([param.numel() for param in network.parameters()]))) logger.info("Reading Data") if alg == 'graph': data_train = conllx_data.read_bucketed_data(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, symbolic_root=True) data_dev = conllx_data.read_data(dev_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, symbolic_root=True) data_test = conllx_data.read_data(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, symbolic_root=True) else: data_train = conllx_stacked_data.read_bucketed_data( train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, prior_order=prior_order) data_dev = conllx_stacked_data.read_data(dev_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, prior_order=prior_order) data_test = conllx_stacked_data.read_data(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, prior_order=prior_order) num_data = sum(data_train[1]) logger.info("training: #training data: %d, batch: %d, unk replace: %.2f" % (num_data, batch_size, unk_replace)) pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) optimizer, scheduler = get_optimizer(network.parameters(), optim, learning_rate, lr_decay, betas, eps, amsgrad, weight_decay, warmup_steps) best_ucorrect = 0.0 best_lcorrect = 0.0 best_ucomlpete = 0.0 best_lcomplete = 0.0 best_ucorrect_nopunc = 0.0 best_lcorrect_nopunc = 0.0 best_ucomlpete_nopunc = 0.0 best_lcomplete_nopunc = 0.0 best_root_correct = 0.0 best_total = 0 best_total_nopunc = 0 best_total_inst = 0 best_total_root = 0 best_epoch = 0 test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete = 0.0 test_lcomplete = 0.0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_nopunc = 0.0 test_lcomplete_nopunc = 0.0 test_root_correct = 0.0 test_total = 0 test_total_nopunc = 0 test_total_inst = 0 test_total_root = 0 patient = 0 beam = args.beam reset = args.reset num_batches = num_data // batch_size + 1 if optim == 'adam': opt_info = 'adam, betas=(%.1f, %.3f), eps=%.1e, amsgrad=%s' % ( betas[0], betas[1], eps, amsgrad) else: opt_info = 'sgd, momentum=0.9, nesterov=True' for epoch in range(1, num_epochs + 1): start_time = time.time() train_loss = 0. train_arc_loss = 0. train_type_loss = 0. num_insts = 0 num_words = 0 num_back = 0 num_nans = 0 network.train() lr = scheduler.get_lr()[0] print( 'Epoch %d (%s, lr=%.6f, lr decay=%.6f, grad clip=%.1f, l2=%.1e): ' % (epoch, opt_info, lr, lr_decay, grad_clip, weight_decay)) if args.cuda: torch.cuda.empty_cache() gc.collect() with torch.autograd.set_detect_anomaly(True): for step, data in enumerate( iterate_data(data_train, batch_size, bucketed=True, unk_replace=unk_replace, shuffle=True)): optimizer.zero_grad() bert_words = data["BERT_WORD"].to(device) sub_word_idx = data["SUB_IDX"].to(device) words = data['WORD'].to(device) chars = data['CHAR'].to(device) postags = data['POS'].to(device) heads = data['HEAD'].to(device) nbatch = words.size(0) if alg == 'graph': types = data['TYPE'].to(device) masks = data['MASK'].to(device) nwords = masks.sum() - nbatch BERT = True if BERT: loss_arc, loss_type = network.loss(bert_words, sub_word_idx, words, chars, postags, heads, types, mask=masks) else: loss_arc, loss_type = network.loss(words, chars, postags, heads, types, mask=masks) else: masks_enc = data['MASK_ENC'].to(device) masks_dec = data['MASK_DEC'].to(device) stacked_heads = data['STACK_HEAD'].to(device) children = data['CHILD'].to(device) siblings = data['SIBLING'].to(device) stacked_types = data['STACK_TYPE'].to(device) nwords = masks_enc.sum() - nbatch loss_arc, loss_type = network.loss(words, chars, postags, heads, stacked_heads, children, siblings, stacked_types, mask_e=masks_enc, mask_d=masks_dec) loss_arc = loss_arc.sum() loss_type = loss_type.sum() loss_total = loss_arc + loss_type # print("loss", loss_arc, loss_type, loss_total) if loss_ty_token: loss = loss_total.div(nwords) else: loss = loss_total.div(nbatch) loss.backward() if grad_clip > 0: grad_norm = clip_grad_norm_(network.parameters(), grad_clip) else: grad_norm = total_grad_norm(network.parameters()) if math.isnan(grad_norm): num_nans += 1 else: optimizer.step() scheduler.step() with torch.no_grad(): num_insts += nbatch num_words += nwords train_loss += loss_total.item() train_arc_loss += loss_arc.item() train_type_loss += loss_type.item() # update log if step % 100 == 0: torch.cuda.empty_cache() sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) curr_lr = scheduler.get_lr()[0] num_insts = max(num_insts, 1) num_words = max(num_words, 1) log_info = '[%d/%d (%.0f%%) lr=%.6f (%d)] loss: %.4f (%.4f), arc: %.4f (%.4f), type: %.4f (%.4f)' % ( step, num_batches, 100. * step / num_batches, curr_lr, num_nans, train_loss / num_insts, train_loss / num_words, train_arc_loss / num_insts, train_arc_loss / num_words, train_type_loss / num_insts, train_type_loss / num_words) sys.stdout.write(log_info) sys.stdout.flush() num_back = len(log_info) sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) print( 'total: %d (%d), loss: %.4f (%.4f), arc: %.4f (%.4f), type: %.4f (%.4f), time: %.2fs' % (num_insts, num_words, train_loss / num_insts, train_loss / num_words, train_arc_loss / num_insts, train_arc_loss / num_words, train_type_loss / num_insts, train_type_loss / num_words, time.time() - start_time)) print('-' * 125) # evaluate performance on dev data with torch.no_grad(): pred_filename = os.path.join(result_path, 'pred_dev%d' % epoch) pred_writer.start(pred_filename) gold_filename = os.path.join(result_path, 'gold_dev%d' % epoch) gold_writer.start(gold_filename) print('Evaluating dev:') dev_stats, dev_stats_nopunct, dev_stats_root = eval( alg, data_dev, network, pred_writer, gold_writer, punct_set, word_alphabet, pos_alphabet, device, beam=beam) pred_writer.close() gold_writer.close() dev_ucorr, dev_lcorr, dev_ucomlpete, dev_lcomplete, dev_total = dev_stats dev_ucorr_nopunc, dev_lcorr_nopunc, dev_ucomlpete_nopunc, dev_lcomplete_nopunc, dev_total_nopunc = dev_stats_nopunct dev_root_corr, dev_total_root, dev_total_inst = dev_stats_root if best_ucorrect_nopunc + best_lcorrect_nopunc < dev_ucorr_nopunc + dev_lcorr_nopunc: best_ucorrect_nopunc = dev_ucorr_nopunc best_lcorrect_nopunc = dev_lcorr_nopunc best_ucomlpete_nopunc = dev_ucomlpete_nopunc best_lcomplete_nopunc = dev_lcomplete_nopunc best_ucorrect = dev_ucorr best_lcorrect = dev_lcorr best_ucomlpete = dev_ucomlpete best_lcomplete = dev_lcomplete best_root_correct = dev_root_corr best_total = dev_total best_total_nopunc = dev_total_nopunc best_total_root = dev_total_root best_total_inst = dev_total_inst best_epoch = epoch patient = 0 torch.save(network.state_dict(), model_name) pred_filename = os.path.join(result_path, 'pred_test%d' % epoch) pred_writer.start(pred_filename) gold_filename = os.path.join(result_path, 'gold_test%d' % epoch) gold_writer.start(gold_filename) print('Evaluating test:') test_stats, test_stats_nopunct, test_stats_root = eval( alg, data_test, network, pred_writer, gold_writer, punct_set, word_alphabet, pos_alphabet, device, beam=beam) test_ucorrect, test_lcorrect, test_ucomlpete, test_lcomplete, test_total = test_stats test_ucorrect_nopunc, test_lcorrect_nopunc, test_ucomlpete_nopunc, test_lcomplete_nopunc, test_total_nopunc = test_stats_nopunct test_root_correct, test_total_root, test_total_inst = test_stats_root pred_writer.close() gold_writer.close() else: patient += 1 print('-' * 125) print( 'best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (best_ucorrect, best_lcorrect, best_total, best_ucorrect * 100 / best_total, best_lcorrect * 100 / best_total, best_ucomlpete * 100 / dev_total_inst, best_lcomplete * 100 / dev_total_inst, best_epoch)) print( 'best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (best_ucorrect_nopunc, best_lcorrect_nopunc, best_total_nopunc, best_ucorrect_nopunc * 100 / best_total_nopunc, best_lcorrect_nopunc * 100 / best_total_nopunc, best_ucomlpete_nopunc * 100 / best_total_inst, best_lcomplete_nopunc * 100 / best_total_inst, best_epoch)) print( 'best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % (best_root_correct, best_total_root, best_root_correct * 100 / best_total_root, best_epoch)) print('-' * 125) print( 'best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 / test_total, test_lcorrect * 100 / test_total, test_ucomlpete * 100 / test_total_inst, test_lcomplete * 100 / test_total_inst, best_epoch)) print( 'best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc, test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc, test_ucomlpete_nopunc * 100 / test_total_inst, test_lcomplete_nopunc * 100 / test_total_inst, best_epoch)) print( 'best test Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % (test_root_correct, test_total_root, test_root_correct * 100 / test_total_root, best_epoch)) print('=' * 125) if patient >= reset: logger.info('reset optimizer momentums') network.load_state_dict( torch.load(model_name, map_location=device)) scheduler.reset_state() patient = 0
def main(): parser = argparse.ArgumentParser( description='Tuning with bi-directional RNN-CNN-CRF') parser.add_argument('--mode', choices=['RNN', 'LSTM', 'GRU'], help='architecture of rnn', required=True) parser.add_argument('--num_epochs', type=int, default=1000, help='Number of training epochs') parser.add_argument('--batch_size', type=int, default=16, help='Number of sentences in each batch') parser.add_argument('--hidden_size', type=int, default=128, help='Number of hidden units in RNN') parser.add_argument('--num_filters', type=int, default=30, help='Number of filters in CNN') parser.add_argument('--char_dim', type=int, default=30, help='Dimension of Character embeddings') parser.add_argument('--learning_rate', type=float, default=0.01, help='Learning rate') parser.add_argument('--decay_rate', type=float, default=0.1, help='Decay rate of learning rate') parser.add_argument('--gamma', type=float, default=0.0, help='weight for regularization') parser.add_argument('--dropout', choices=['std', 'variational'], help='type of dropout', required=True) parser.add_argument('--p', type=float, default=0.5, help='dropout rate') parser.add_argument('--bigram', action='store_true', help='bi-gram parameter for CRF') parser.add_argument('--schedule', type=int, help='schedule for learning rate decay') parser.add_argument('--unk_replace', type=float, default=0., help='The rate to replace a singleton word with UNK') parser.add_argument( '--train') # "data/POS-penn/wsj/split1/wsj1.train.original" parser.add_argument( '--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original" parser.add_argument( '--test') # "data/POS-penn/wsj/split1/wsj1.test.original" args = parser.parse_args() logger = get_logger("POSCRFTagger") mode = args.mode train_path = args.train dev_path = args.dev test_path = args.test num_epochs = args.num_epochs batch_size = args.batch_size hidden_size = args.hidden_size num_filters = args.num_filters learning_rate = args.learning_rate momentum = 0.9 decay_rate = args.decay_rate gamma = args.gamma schedule = args.schedule p = args.p unk_replace = args.unk_replace bigram = args.bigram embedd_dict, embedd_dim = utils.load_embedding_dict( 'glove', "data/glove/glove.6B/glove.6B.100d.gz") logger.info("Creating Alphabets") word_alphabet, char_alphabet, pos_alphabet, \ type_alphabet = conllx_data.create_alphabets("data/alphabets/pos_crf/", train_path, data_paths=[dev_path, test_path], max_vocabulary_size=50000, embedd_dict=embedd_dict) logger.info("Word Alphabet Size: %d" % word_alphabet.size()) logger.info("Character Alphabet Size: %d" % char_alphabet.size()) logger.info("POS Alphabet Size: %d" % pos_alphabet.size()) logger.info("Reading Data") use_gpu = torch.cuda.is_available() data_train = conllx_data.read_data_to_variable(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu) # data_train = conllx_data.read_data(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet) # num_data = sum([len(bucket) for bucket in data_train]) num_data = sum(data_train[1]) num_labels = pos_alphabet.size() data_dev = conllx_data.read_data_to_variable(dev_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=True) data_test = conllx_data.read_data_to_variable(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=True) def construct_word_embedding_table(): scale = np.sqrt(3.0 / embedd_dim) table = np.empty([word_alphabet.size(), embedd_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.random.uniform( -scale, scale, [1, embedd_dim]).astype(np.float32) oov = 0 for word, index in word_alphabet.items(): if word in embedd_dict: embedding = embedd_dict[word] elif word.lower() in embedd_dict: embedding = embedd_dict[word.lower()] else: embedding = np.random.uniform( -scale, scale, [1, embedd_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('oov: %d' % oov) return torch.from_numpy(table) word_table = construct_word_embedding_table() logger.info("constructing network...") char_dim = args.char_dim window = 3 num_layers = 1 if args.dropout == 'std': network = BiRecurrentConvCRF(embedd_dim, word_alphabet.size(), char_dim, char_alphabet.size(), num_filters, window, mode, hidden_size, num_layers, num_labels, embedd_word=word_table, p_rnn=p, bigram=bigram) else: raise NotImplementedError if use_gpu: network.cuda() lr = learning_rate optim = SGD(network.parameters(), lr=lr, momentum=momentum, weight_decay=gamma) logger.info("Network: %s, num_layer=%d, hidden=%d, filter=%d, crf=%s" % (mode, num_layers, hidden_size, num_filters, 'bigram' if bigram else 'unigram')) logger.info( "training: l2: %f, (#training data: %d, batch: %d, dropout: %.2f, unk replace: %.2f)" % (gamma, num_data, batch_size, p, unk_replace)) num_batches = num_data / batch_size + 1 dev_correct = 0.0 best_epoch = 0 test_correct = 0.0 test_total = 0 for epoch in range(1, num_epochs + 1): print( 'Epoch %d (%s(%s), learning rate=%.4f, decay rate=%.4f (schedule=%d)): ' % (epoch, mode, args.dropout, lr, decay_rate, schedule)) train_err = 0. train_total = 0. start_time = time.time() num_back = 0 network.train() for batch in range(1, num_batches + 1): word, char, labels, _, _, masks, lengths = conllx_data.get_batch_variable( data_train, batch_size, unk_replace=unk_replace) optim.zero_grad() loss = network.loss(word, char, labels, mask=masks) loss.backward() optim.step() num_inst = word.size(0) train_err += loss.data[0] * num_inst train_total += num_inst time_ave = (time.time() - start_time) / batch time_left = (num_batches - batch) * time_ave # update log if batch % 100 == 0: sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) log_info = 'train: %d/%d loss: %.4f, time left (estimated): %.2fs' % ( batch, num_batches, train_err / train_total, time_left) sys.stdout.write(log_info) sys.stdout.flush() num_back = len(log_info) sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) print('train: %d loss: %.4f, time: %.2fs' % (num_batches, train_err / train_total, time.time() - start_time)) # evaluate performance on dev data network.eval() dev_corr = 0.0 dev_total = 0 for batch in conllx_data.iterate_batch_variable(data_dev, batch_size): word, char, labels, _, _, masks, lengths = batch preds, corr = network.decode( word, char, target=labels, mask=masks, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) num_tokens = masks.data.sum() dev_corr += corr dev_total += num_tokens print('dev corr: %d, total: %d, acc: %.2f%%' % (dev_corr, dev_total, dev_corr * 100 / dev_total)) if dev_correct < dev_corr: dev_correct = dev_corr best_epoch = epoch # evaluate on test data when better performance detected test_corr = 0.0 test_total = 0 for batch in conllx_data.iterate_batch_variable( data_test, batch_size): word, char, labels, _, _, masks, lengths = batch preds, corr = network.decode( word, char, target=labels, mask=masks, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) num_tokens = masks.data.sum() test_corr += corr test_total += num_tokens test_correct = test_corr print("best dev corr: %d, total: %d, acc: %.2f%% (epoch: %d)" % (dev_correct, dev_total, dev_correct * 100 / dev_total, best_epoch)) print("best test corr: %d, total: %d, acc: %.2f%% (epoch: %d)" % (test_correct, test_total, test_correct * 100 / test_total, best_epoch)) if epoch % schedule == 0: lr = learning_rate / (1.0 + epoch * decay_rate) optim = SGD(network.parameters(), lr=lr, momentum=momentum, weight_decay=gamma, nesterov=True)
custom_args.dataset_name) train_path, dev_path, test_path, _, _, _ = datasets.DATASET_FILES[ custom_args.dataset_name] if custom_args.bio_embeddings != 'none': if 'bio' in custom_args.bio_embeddings: embedding_vocab_dict = datasets.load_bio_word_embedding_vocab( custom_args.bio_embeddings) elif 'glove' in custom_args.bio_embeddings: embedding_vocab_dict = datasets.load_glove_word_embedding_vocab( custom_args.bio_embeddings) else: embedding_vocab_dict = None graph_word_alphabet, graph_char_alphabet, _, _, graph_ner_alphabet = conllx_data.create_alphabets( alphabet_path, train_path, data_paths=[dev_path, test_path], max_vocabulary_size=100000, embedd_dict=embedding_vocab_dict, normalize_digits=False, suffix='graph_') parser_data, graph_data, mentions, labels = datasets.load_data( (word_alphabet, graph_word_alphabet), (char_alphabet, graph_char_alphabet), pos_alphabet, type_alphabet, graph_ner_alphabet, custom_args) data_train, data_dev, data_test = parser_data graph_data_train, graph_data_dev, graph_data_test = graph_data mentions_train, mentions_dev, mentions_test = mentions labels_train, labels_dev, labels_test = labels all_labels = list( set(x.lower() for x in set(labels_train + labels_dev + labels_test))) all_labels.sort()
def main(): args_parser = argparse.ArgumentParser( description='Tuning with graph-based parsing') args_parser.register('type', 'bool', str2bool) args_parser.add_argument('--seed', type=int, default=1234, help='random seed for reproducibility') args_parser.add_argument('--mode', choices=['RNN', 'LSTM', 'GRU', 'FastLSTM'], help='architecture of rnn', required=True) args_parser.add_argument('--num_epochs', type=int, default=1000, help='Number of training epochs') args_parser.add_argument('--batch_size', type=int, default=64, help='Number of sentences in each batch') args_parser.add_argument('--hidden_size', type=int, default=256, help='Number of hidden units in RNN') args_parser.add_argument('--arc_space', type=int, default=128, help='Dimension of tag space') args_parser.add_argument('--type_space', type=int, default=128, help='Dimension of tag space') args_parser.add_argument('--num_layers', type=int, default=1, help='Number of layers of encoder.') args_parser.add_argument('--num_filters', type=int, default=50, help='Number of filters in CNN') args_parser.add_argument('--pos', action='store_true', help='use part-of-speech embedding.') args_parser.add_argument('--char', action='store_true', help='use character embedding and CNN.') args_parser.add_argument('--pos_dim', type=int, default=50, help='Dimension of POS embeddings') args_parser.add_argument('--char_dim', type=int, default=50, help='Dimension of Character embeddings') args_parser.add_argument('--opt', choices=['adam', 'sgd', 'adamax'], help='optimization algorithm') args_parser.add_argument('--objective', choices=['cross_entropy', 'crf'], default='cross_entropy', help='objective function of training procedure.') args_parser.add_argument('--decode', choices=['mst', 'greedy'], default='mst', help='decoding algorithm') args_parser.add_argument('--learning_rate', type=float, default=0.01, help='Learning rate') # args_parser.add_argument('--decay_rate', type=float, default=0.05, help='Decay rate of learning rate') args_parser.add_argument('--clip', type=float, default=5.0, help='gradient clipping') args_parser.add_argument('--gamma', type=float, default=0.0, help='weight for regularization') args_parser.add_argument('--epsilon', type=float, default=1e-8, help='epsilon for adam or adamax') args_parser.add_argument('--p_rnn', nargs='+', type=float, required=True, help='dropout rate for RNN') args_parser.add_argument('--p_in', type=float, default=0.33, help='dropout rate for input embeddings') args_parser.add_argument('--p_out', type=float, default=0.33, help='dropout rate for output layer') # args_parser.add_argument('--schedule', type=int, help='schedule for learning rate decay') args_parser.add_argument( '--unk_replace', type=float, default=0., help='The rate to replace a singleton word with UNK') args_parser.add_argument('--punctuation', nargs='+', type=str, help='List of punctuations') args_parser.add_argument( '--word_embedding', choices=['word2vec', 'glove', 'senna', 'sskip', 'polyglot'], help='Embedding for words', required=True) args_parser.add_argument('--word_path', help='path for word embedding dict') args_parser.add_argument( '--freeze', action='store_true', help='frozen the word embedding (disable fine-tuning).') args_parser.add_argument('--char_embedding', choices=['random', 'polyglot'], help='Embedding for characters', required=True) args_parser.add_argument('--char_path', help='path for character embedding dict') args_parser.add_argument('--data_dir', help='Data directory path') args_parser.add_argument( '--src_lang', required=True, help='Src language to train dependency parsing model') args_parser.add_argument('--aux_lang', nargs='+', help='Language names for adversarial training') args_parser.add_argument('--vocab_path', help='path for prebuilt alphabets.', default=None) args_parser.add_argument('--model_path', help='path for saving model file.', required=True) args_parser.add_argument('--model_name', help='name for saving model file.', required=True) # args_parser.add_argument('--attn_on_rnn', action='store_true', help='use self-attention on top of context RNN.') args_parser.add_argument('--no_word', type='bool', default=False, help='do not use word embedding.') args_parser.add_argument('--use_bert', type='bool', default=False, help='use multilingual BERT.') # # lrate schedule with warmup in the first iter. args_parser.add_argument('--use_warmup_schedule', type='bool', default=False, help="Use warmup lrate schedule.") args_parser.add_argument('--decay_rate', type=float, default=0.75, help='Decay rate of learning rate') args_parser.add_argument('--max_decay', type=int, default=9, help='Number of decays before stop') args_parser.add_argument('--schedule', type=int, help='schedule for learning rate decay') args_parser.add_argument('--double_schedule_decay', type=int, default=5, help='Number of decays to double schedule') args_parser.add_argument( '--check_dev', type=int, default=5, help='Check development performance in every n\'th iteration') # encoder selection args_parser.add_argument('--encoder_type', choices=['Transformer', 'RNN', 'SelfAttn'], default='RNN', help='do not use context RNN.') args_parser.add_argument( '--pool_type', default='mean', choices=['max', 'mean', 'weight'], help='pool type to form fixed length vector from word embeddings') # Tansformer encoder args_parser.add_argument( '--trans_hid_size', type=int, default=1024, help='#hidden units in point-wise feed-forward in transformer') args_parser.add_argument( '--d_k', type=int, default=64, help='d_k for multi-head-attention in transformer encoder') args_parser.add_argument( '--d_v', type=int, default=64, help='d_v for multi-head-attention in transformer encoder') args_parser.add_argument('--num_head', type=int, default=8, help='Value of h in multi-head attention') args_parser.add_argument( '--use_all_encoder_layers', type='bool', default=False, help='Use a weighted representations of all encoder layers') # - positional args_parser.add_argument( '--enc_use_neg_dist', action='store_true', help="Use negative distance for enc's relational-distance embedding.") args_parser.add_argument( '--enc_clip_dist', type=int, default=0, help="The clipping distance for relative position features.") args_parser.add_argument('--position_dim', type=int, default=50, help='Dimension of Position embeddings.') args_parser.add_argument( '--position_embed_num', type=int, default=200, help= 'Minimum value of position embedding num, which usually is max-sent-length.' ) args_parser.add_argument('--train_position', action='store_true', help='train positional encoding for transformer.') args_parser.add_argument('--input_concat_embeds', action='store_true', help="Concat input embeddings, otherwise add.") args_parser.add_argument('--input_concat_position', action='store_true', help="Concat position embeddings, otherwise add.") args_parser.add_argument( '--partitioned', type='bool', default=False, help= "Partition the content and positional attention for multi-head attention." ) args_parser.add_argument( '--partition_type', choices=['content-position', 'lexical-delexical'], default='content-position', help="How to apply partition in the self-attention.") # args_parser.add_argument( '--train_len_thresh', type=int, default=100, help='In training, discard sentences longer than this.') # # regarding adversarial training args_parser.add_argument('--pre_model_path', type=str, default=None, help='Path of the pretrained model.') args_parser.add_argument('--pre_model_name', type=str, default=None, help='Name of the pretrained model.') args_parser.add_argument('--adv_training', type='bool', default=False, help='Use adversarial training.') args_parser.add_argument( '--lambdaG', type=float, default=0.001, help='Scaling parameter to control generator loss.') args_parser.add_argument('--discriminator', choices=['weak', 'not-so-weak', 'strong'], default='weak', help='architecture of the discriminator') args_parser.add_argument( '--delay', type=int, default=0, help='Number of epochs to be run first for the source task') args_parser.add_argument( '--n_critic', type=int, default=5, help='Number of training steps for discriminator per iter') args_parser.add_argument( '--clip_disc', type=float, default=5.0, help='Lower and upper clip value for disc. weights') args_parser.add_argument('--debug', type='bool', default=False, help='Use debug portion of the training data') args_parser.add_argument('--train_level', type=str, default='word', choices=['word', 'sent'], help='Use X-level adversarial training') args_parser.add_argument('--train_type', type=str, default='GAN', choices=['GR', 'GAN', 'WGAN'], help='Type of adversarial training') # # regarding motivational training args_parser.add_argument( '--motivate', type='bool', default=False, help='This is opposite of the adversarial training') # args = args_parser.parse_args() # fix data-prepare seed random.seed(1234) np.random.seed(1234) # model's seed torch.manual_seed(args.seed) # if output directory doesn't exist, create it if not os.path.exists(args.model_path): os.makedirs(args.model_path) logger = get_logger("GraphParser") logger.info('\ncommand-line params : {0}\n'.format(sys.argv[1:])) logger.info('{0}\n'.format(args)) logger.info("Visible GPUs: %s", str(os.environ["CUDA_VISIBLE_DEVICES"])) args.parallel = False if torch.cuda.device_count() > 1: args.parallel = True mode = args.mode obj = args.objective decoding = args.decode train_path = args.data_dir + args.src_lang + "_train.debug.1_10.conllu" \ if args.debug else args.data_dir + args.src_lang + '_train.conllu' dev_path = args.data_dir + args.src_lang + "_dev.conllu" test_path = args.data_dir + args.src_lang + "_test.conllu" # vocab_path = args.vocab_path if args.vocab_path is not None else args.model_path model_path = args.model_path model_name = args.model_name num_epochs = args.num_epochs batch_size = args.batch_size hidden_size = args.hidden_size arc_space = args.arc_space type_space = args.type_space num_layers = args.num_layers num_filters = args.num_filters learning_rate = args.learning_rate opt = args.opt momentum = 0.9 betas = (0.9, 0.9) eps = args.epsilon decay_rate = args.decay_rate clip = args.clip gamma = args.gamma schedule = args.schedule p_rnn = tuple(args.p_rnn) p_in = args.p_in p_out = args.p_out unk_replace = args.unk_replace punctuation = args.punctuation freeze = args.freeze use_word_emb = not args.no_word word_embedding = args.word_embedding word_path = args.word_path use_char = args.char char_embedding = args.char_embedding char_path = args.char_path attn_on_rnn = args.attn_on_rnn encoder_type = args.encoder_type if attn_on_rnn: assert encoder_type == 'RNN' t_types = (args.adv_training, args.motivate) t_count = sum(1 for tt in t_types if tt) if t_count > 1: assert False, "Only one of: adv_training or motivate can be true" # ------------------- Loading/initializing embeddings -------------------- # use_pos = args.pos pos_dim = args.pos_dim word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path) char_dict = None char_dim = args.char_dim if char_embedding != 'random': char_dict, char_dim = utils.load_embedding_dict( char_embedding, char_path) logger.info("Creating Alphabets") alphabet_path = os.path.join(vocab_path, 'alphabets/') model_name = os.path.join(model_path, model_name) # TODO (WARNING): must build vocabs previously assert os.path.isdir(alphabet_path), "should have build vocabs previously" word_alphabet, char_alphabet, pos_alphabet, type_alphabet, max_sent_length = conllx_data.create_alphabets( alphabet_path, train_path, data_paths=[dev_path, test_path], max_vocabulary_size=50000, embedd_dict=word_dict) max_sent_length = max(max_sent_length, args.position_embed_num) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) # ------------------------------------------------------------------------- # # --------------------- Loading/building the model ------------------------ # logger.info("Reading Data") use_gpu = torch.cuda.is_available() def construct_word_embedding_table(): scale = np.sqrt(3.0 / word_dim) table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.zeros([1, word_dim]).astype( np.float32) if freeze else np.random.uniform( -scale, scale, [1, word_dim]).astype(np.float32) oov = 0 for word, index in word_alphabet.items(): if word in word_dict: embedding = word_dict[word] elif word.lower() in word_dict: embedding = word_dict[word.lower()] else: embedding = np.zeros([1, word_dim]).astype( np.float32) if freeze else np.random.uniform( -scale, scale, [1, word_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('word OOV: %d' % oov) return torch.from_numpy(table) def construct_char_embedding_table(): if char_dict is None: return None scale = np.sqrt(3.0 / char_dim) table = np.empty([num_chars, char_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.random.uniform( -scale, scale, [1, char_dim]).astype(np.float32) oov = 0 for char, index, in char_alphabet.items(): if char in char_dict: embedding = char_dict[char] else: embedding = np.random.uniform(-scale, scale, [1, char_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('character OOV: %d' % oov) return torch.from_numpy(table) word_table = construct_word_embedding_table() if use_word_emb else None char_table = construct_char_embedding_table() if use_char else None def load_model_arguments_from_json(): arguments = json.load(open(pre_model_path, 'r')) return arguments['args'], arguments['kwargs'] window = 3 if obj == 'cross_entropy': if args.pre_model_path and args.pre_model_name: pre_model_name = os.path.join(args.pre_model_path, args.pre_model_name) pre_model_path = pre_model_name + '.arg.json' model_args, kwargs = load_model_arguments_from_json() network = BiRecurrentConvBiAffine(use_gpu=use_gpu, *model_args, **kwargs) network.load_state_dict(torch.load(pre_model_name)) logger.info("Model reloaded from %s" % pre_model_path) # Adjust the word embedding layer if network.embedder.word_embedd is not None: network.embedder.word_embedd = nn.Embedding(num_words, word_dim, _weight=word_table) else: network = BiRecurrentConvBiAffine( word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, num_filters, window, mode, hidden_size, num_layers, num_types, arc_space, type_space, embedd_word=word_table, embedd_char=char_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, biaffine=True, pos=use_pos, char=use_char, train_position=args.train_position, encoder_type=encoder_type, trans_hid_size=args.trans_hid_size, d_k=args.d_k, d_v=args.d_v, num_head=args.num_head, enc_use_neg_dist=args.enc_use_neg_dist, enc_clip_dist=args.enc_clip_dist, position_dim=args.position_dim, max_sent_length=max_sent_length, use_gpu=use_gpu, use_word_emb=use_word_emb, input_concat_embeds=args.input_concat_embeds, input_concat_position=args.input_concat_position, attn_on_rnn=attn_on_rnn, partitioned=args.partitioned, partition_type=args.partition_type, use_all_encoder_layers=args.use_all_encoder_layers, use_bert=args.use_bert) elif obj == 'crf': raise NotImplementedError else: raise RuntimeError('Unknown objective: %s' % obj) # ------------------------------------------------------------------------- # # --------------------- Loading data -------------------------------------- # train_data = dict() dev_data = dict() test_data = dict() num_data = dict() lang_ids = dict() reverse_lang_ids = dict() # ===== the reading ============================================= def _read_one(path, is_train): lang_id = guess_language_id(path) logger.info("Reading: guess that the language of file %s is %s." % (path, lang_id)) one_data = conllx_data.read_data_to_variable( path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=False, volatile=(not is_train), symbolic_root=True, lang_id=lang_id, use_bert=args.use_bert, len_thresh=(args.train_len_thresh if is_train else 100000)) return one_data data_train = _read_one(train_path, True) train_data[args.src_lang] = data_train num_data[args.src_lang] = sum(data_train[1]) lang_ids[args.src_lang] = len(lang_ids) reverse_lang_ids[lang_ids[args.src_lang]] = args.src_lang data_dev = _read_one(dev_path, False) data_test = _read_one(test_path, False) dev_data[args.src_lang] = data_dev test_data[args.src_lang] = data_test # =============================================================== # ===== reading data for adversarial training =================== if t_count > 0: for language in args.aux_lang: aux_train_path = args.data_dir + language + "_train.debug.1_10.conllu" \ if args.debug else args.data_dir + language + '_train.conllu' aux_train_data = _read_one(aux_train_path, True) num_data[language] = sum(aux_train_data[1]) train_data[language] = aux_train_data lang_ids[language] = len(lang_ids) reverse_lang_ids[lang_ids[language]] = language # =============================================================== punct_set = None if punctuation is not None: punct_set = set(punctuation) logger.info("punctuations(%d): %s" % (len(punct_set), ' '.join(punct_set))) def save_args(): arg_path = model_name + '.arg.json' arguments = [ word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, num_filters, window, mode, hidden_size, num_layers, num_types, arc_space, type_space ] kwargs = { 'p_in': p_in, 'p_out': p_out, 'p_rnn': p_rnn, 'biaffine': True, 'pos': use_pos, 'char': use_char, 'train_position': args.train_position, 'encoder_type': args.encoder_type, 'trans_hid_size': args.trans_hid_size, 'd_k': args.d_k, 'd_v': args.d_v, 'num_head': args.num_head, 'enc_use_neg_dist': args.enc_use_neg_dist, 'enc_clip_dist': args.enc_clip_dist, 'position_dim': args.position_dim, 'max_sent_length': max_sent_length, 'use_word_emb': use_word_emb, 'input_concat_embeds': args.input_concat_embeds, 'input_concat_position': args.input_concat_position, 'attn_on_rnn': attn_on_rnn, 'partitioned': args.partitioned, 'partition_type': args.partition_type, 'use_all_encoder_layers': args.use_all_encoder_layers, 'use_bert': args.use_bert } json.dump({ 'args': arguments, 'kwargs': kwargs }, open(arg_path, 'w'), indent=4) if use_word_emb and freeze: freeze_embedding(network.embedder.word_embedd) if args.parallel: network = torch.nn.DataParallel(network) if use_gpu: network = network.cuda() save_args() param_dict = {} encoder = network.module.encoder if args.parallel else network.encoder for name, param in encoder.named_parameters(): if param.requires_grad: param_dict[name] = np.prod(param.size()) total_params = np.sum(list(param_dict.values())) logger.info('Total Encoder Parameters = %d' % total_params) # ------------------------------------------------------------------------- # # ============================================= if args.adv_training: disc_feat_size = network.module.encoder.output_dim if args.parallel else network.encoder.output_dim reverse_grad = args.train_type == 'GR' nclass = len(lang_ids) if args.train_type == 'GR' else 1 kwargs = { 'input_size': disc_feat_size, 'disc_type': args.discriminator, 'train_level': args.train_level, 'train_type': args.train_type, 'reverse_grad': reverse_grad, 'soft_label': True, 'nclass': nclass, 'scale': args.lambdaG, 'use_gpu': use_gpu, 'opt': 'adam', 'lr': 0.001, 'betas': (0.9, 0.999), 'gamma': 0, 'eps': 1e-8, 'momentum': 0, 'clip_disc': args.clip_disc } AdvAgent = Adversarial(**kwargs) if use_gpu: AdvAgent.cuda() elif args.motivate: disc_feat_size = network.module.encoder.output_dim if args.parallel else network.encoder.output_dim nclass = len(lang_ids) kwargs = { 'input_size': disc_feat_size, 'disc_type': args.discriminator, 'train_level': args.train_level, 'nclass': nclass, 'scale': args.lambdaG, 'use_gpu': use_gpu, 'opt': 'adam', 'lr': 0.001, 'betas': (0.9, 0.999), 'gamma': 0, 'eps': 1e-8, 'momentum': 0, 'clip_disc': args.clip_disc } MtvAgent = Motivator(**kwargs) if use_gpu: MtvAgent.cuda() # ============================================= # --------------------- Initializing the optimizer ------------------------ # lr = learning_rate optim = generate_optimizer(opt, lr, network.parameters(), betas, gamma, eps, momentum) opt_info = 'opt: %s, ' % opt if opt == 'adam': opt_info += 'betas=%s, eps=%.1e' % (betas, eps) elif opt == 'sgd': opt_info += 'momentum=%.2f' % momentum elif opt == 'adamax': opt_info += 'betas=%s, eps=%.1e' % (betas, eps) # ============================================= total_data = min(num_data.values()) word_status = 'frozen' if freeze else 'fine tune' char_status = 'enabled' if use_char else 'disabled' pos_status = 'enabled' if use_pos else 'disabled' logger.info( "Embedding dim: word=%d (%s), char=%d (%s), pos=%d (%s)" % (word_dim, word_status, char_dim, char_status, pos_dim, pos_status)) logger.info("CNN: filter=%d, kernel=%d" % (num_filters, window)) logger.info( "RNN: %s, num_layer=%d, hidden=%d, arc_space=%d, type_space=%d" % (mode, num_layers, hidden_size, arc_space, type_space)) logger.info( "train: obj: %s, l2: %f, (#data: %d, batch: %d, clip: %.2f, unk replace: %.2f)" % (obj, gamma, total_data, batch_size, clip, unk_replace)) logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" % (p_in, p_out, p_rnn)) logger.info("decoding algorithm: %s" % decoding) logger.info(opt_info) # ------------------------------------------------------------------------- # # --------------------- Form the mini-batches ----------------------------- # num_batches = total_data // batch_size + 1 aux_lang = [] if t_count > 0: for language in args.aux_lang: aux_lang.extend([language] * num_data[language]) assert num_data[args.src_lang] <= len(aux_lang) # ------------------------------------------------------------------------- # dev_ucorrect = 0.0 dev_lcorrect = 0.0 dev_ucomlpete_match = 0.0 dev_lcomplete_match = 0.0 dev_ucorrect_nopunc = 0.0 dev_lcorrect_nopunc = 0.0 dev_ucomlpete_match_nopunc = 0.0 dev_lcomplete_match_nopunc = 0.0 dev_root_correct = 0.0 best_epoch = 0 if decoding == 'greedy': decode = network.module.decode if args.parallel else network.decode elif decoding == 'mst': decode = network.module.decode_mst if args.parallel else network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) patient = 0 decay = 0 max_decay = args.max_decay double_schedule_decay = args.double_schedule_decay # lrate schedule step_num = 0 use_warmup_schedule = args.use_warmup_schedule if use_warmup_schedule: logger.info("Use warmup lrate for the first epoch, from 0 up to %s." % (lr, )) skip_adv_tuning = 0 loss_fn = network.module.loss if args.parallel else network.loss for epoch in range(1, num_epochs + 1): print( 'Epoch %d (%s, optim: %s, learning rate=%.6f, eps=%.1e, decay rate=%.2f (schedule=%d, patient=%d, decay=%d)): ' % (epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay)) train_err = 0. train_err_arc = 0. train_err_type = 0. train_total = 0. start_time = time.time() num_back = 0 skip_adv_tuning += 1 loss_d_real, loss_d_fake = [], [] acc_d_real, acc_d_fake, = [], [] gen_loss, parsing_loss = [], [] disent_loss = [] if t_count > 0 and skip_adv_tuning > args.delay: batch_size = args.batch_size // 2 num_batches = total_data // batch_size + 1 # ---------------------- Sample the mini-batches -------------------------- # if t_count > 0: sampled_aux_lang = random.sample(aux_lang, num_batches) lang_in_batch = [(args.src_lang, sampled_aux_lang[k]) for k in range(num_batches)] else: lang_in_batch = [(args.src_lang, None) for _ in range(num_batches)] assert len(lang_in_batch) == num_batches # ------------------------------------------------------------------------- # network.train() warmup_factor = (lr + 0.) / num_batches for batch in range(1, num_batches + 1): update_generator = True update_discriminator = False # lrate schedule (before each step) step_num += 1 if use_warmup_schedule and epoch <= 1: cur_lrate = warmup_factor * step_num # set lr for param_group in optim.param_groups: param_group['lr'] = cur_lrate # considering source language as real and auxiliary languages as fake real_lang, fake_lang = lang_in_batch[batch - 1] real_idx, fake_idx = lang_ids.get(real_lang), lang_ids.get( fake_lang, -1) # word, char, pos, heads, types, masks, lengths, bert_inputs = conllx_data.get_batch_variable( train_data[real_lang], batch_size, unk_replace=unk_replace) if use_gpu: word = word.cuda() char = char.cuda() pos = pos.cuda() heads = heads.cuda() types = types.cuda() masks = masks.cuda() lengths = lengths.cuda() if bert_inputs[0] is not None: bert_inputs[0] = bert_inputs[0].cuda() bert_inputs[1] = bert_inputs[1].cuda() bert_inputs[2] = bert_inputs[2].cuda() real_enc = network(word, char, pos, input_bert=bert_inputs, mask=masks, length=lengths, hx=None) # ========== Update the discriminator ========== if t_count > 0 and skip_adv_tuning > args.delay: # fake examples = 0 word_f, char_f, pos_f, heads_f, types_f, masks_f, lengths_f, bert_inputs = conllx_data.get_batch_variable( train_data[fake_lang], batch_size, unk_replace=unk_replace) if use_gpu: word_f = word_f.cuda() char_f = char_f.cuda() pos_f = pos_f.cuda() heads_f = heads_f.cuda() types_f = types_f.cuda() masks_f = masks_f.cuda() lengths_f = lengths_f.cuda() if bert_inputs[0] is not None: bert_inputs[0] = bert_inputs[0].cuda() bert_inputs[1] = bert_inputs[1].cuda() bert_inputs[2] = bert_inputs[2].cuda() fake_enc = network(word_f, char_f, pos_f, input_bert=bert_inputs, mask=masks_f, length=lengths_f, hx=None) # TODO: temporary crack if t_count > 0 and skip_adv_tuning > args.delay: # skip discriminator training for '|n_critic|' iterations if 'n_critic' < 0 if args.n_critic > 0 or (batch - 1) % (-1 * args.n_critic) == 0: update_discriminator = True if update_discriminator: if args.adv_training: real_loss, fake_loss, real_acc, fake_acc = AdvAgent.update( real_enc['output'].detach(), fake_enc['output'].detach(), real_idx, fake_idx) loss_d_real.append(real_loss) loss_d_fake.append(fake_loss) acc_d_real.append(real_acc) acc_d_fake.append(fake_acc) elif args.motivate: real_loss, fake_loss, real_acc, fake_acc = MtvAgent.update( real_enc['output'].detach(), fake_enc['output'].detach(), real_idx, fake_idx) loss_d_real.append(real_loss) loss_d_fake.append(fake_loss) acc_d_real.append(real_acc) acc_d_fake.append(fake_acc) else: raise NotImplementedError() if args.n_critic > 0 and (batch - 1) % args.n_critic != 0: update_generator = False # ============================================== # =========== Update the generator ============= if update_generator: others_loss = None if args.adv_training and skip_adv_tuning > args.delay: # for GAN: L_G= L_parsing - (lambda_G * L_D) # for GR : L_G= L_parsing + L_D others_loss = AdvAgent.gen_loss(real_enc['output'], fake_enc['output'], real_idx, fake_idx) gen_loss.append(others_loss.item()) elif args.motivate and skip_adv_tuning > args.delay: others_loss = MtvAgent.gen_loss(real_enc['output'], fake_enc['output'], real_idx, fake_idx) gen_loss.append(others_loss.item()) optim.zero_grad() loss_arc, loss_type = loss_fn(real_enc['output'], heads, types, mask=masks, length=lengths) loss = loss_arc + loss_type num_inst = word.size( 0) if obj == 'crf' else masks.sum() - word.size(0) train_err += loss.item() * num_inst train_err_arc += loss_arc.item() * num_inst train_err_type += loss_type.item() * num_inst train_total += num_inst parsing_loss.append(loss.item()) if others_loss is not None: loss = loss + others_loss loss.backward() clip_grad_norm_(network.parameters(), clip) optim.step() time_ave = (time.time() - start_time) / batch time_left = (num_batches - batch) * time_ave if (args.adv_training or args.motivate) and skip_adv_tuning > args.delay: logger.info( 'epoch: %d train: %d loss: %.4f, arc: %.4f, type: %.4f, dis_loss: (%.2f, %.2f), dis_acc: (%.2f, %.2f), ' 'gen_loss: %.2f, time: %.2fs' % (epoch, num_batches, train_err / train_total, train_err_arc / train_total, train_err_type / train_total, sum(loss_d_real) / len(loss_d_real), sum(loss_d_fake) / len(loss_d_fake), sum(acc_d_real) / len(acc_d_real), sum(acc_d_fake) / len(acc_d_fake), sum(gen_loss) / len(gen_loss), time.time() - start_time)) else: logger.info( 'epoch: %d train: %d loss: %.4f, arc: %.4f, type: %.4f, time: %.2fs' % (epoch, num_batches, train_err / train_total, train_err_arc / train_total, train_err_type / train_total, time.time() - start_time)) ################# Validation on Dependency Parsing Only ################################# if epoch % args.check_dev != 0: continue with torch.no_grad(): # evaluate performance on dev data network.eval() dev_ucorr = 0.0 dev_lcorr = 0.0 dev_total = 0 dev_ucomlpete = 0.0 dev_lcomplete = 0.0 dev_ucorr_nopunc = 0.0 dev_lcorr_nopunc = 0.0 dev_total_nopunc = 0 dev_ucomlpete_nopunc = 0.0 dev_lcomplete_nopunc = 0.0 dev_root_corr = 0.0 dev_total_root = 0.0 dev_total_inst = 0.0 for lang, data_dev in dev_data.items(): for batch in conllx_data.iterate_batch_variable( data_dev, batch_size): word, char, pos, heads, types, masks, lengths, bert_inputs = batch if use_gpu: word = word.cuda() char = char.cuda() pos = pos.cuda() heads = heads.cuda() types = types.cuda() masks = masks.cuda() lengths = lengths.cuda() if bert_inputs[0] is not None: bert_inputs[0] = bert_inputs[0].cuda() bert_inputs[1] = bert_inputs[1].cuda() bert_inputs[2] = bert_inputs[2].cuda() heads_pred, types_pred = decode( word, char, pos, input_bert=bert_inputs, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.cpu().numpy() pos = pos.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.cpu().numpy() types = types.cpu().numpy() stats, stats_nopunc, stats_root, num_inst = parser.eval( word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root dev_ucorr += ucorr dev_lcorr += lcorr dev_total += total dev_ucomlpete += ucm dev_lcomplete += lcm dev_ucorr_nopunc += ucorr_nopunc dev_lcorr_nopunc += lcorr_nopunc dev_total_nopunc += total_nopunc dev_ucomlpete_nopunc += ucm_nopunc dev_lcomplete_nopunc += lcm_nopunc dev_root_corr += corr_root dev_total_root += total_root dev_total_inst += num_inst print( 'W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 / dev_total, dev_lcorr * 100 / dev_total, dev_ucomlpete * 100 / dev_total_inst, dev_lcomplete * 100 / dev_total_inst)) print( 'Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc, dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc * 100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 / dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst)) print('Root: corr: %d, total: %d, acc: %.2f%%' % (dev_root_corr, dev_total_root, dev_root_corr * 100 / dev_total_root)) if dev_lcorrect_nopunc < dev_lcorr_nopunc or ( dev_lcorrect_nopunc == dev_lcorr_nopunc and dev_ucorrect_nopunc < dev_ucorr_nopunc): dev_ucorrect_nopunc = dev_ucorr_nopunc dev_lcorrect_nopunc = dev_lcorr_nopunc dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc dev_lcomplete_match_nopunc = dev_lcomplete_nopunc dev_ucorrect = dev_ucorr dev_lcorrect = dev_lcorr dev_ucomlpete_match = dev_ucomlpete dev_lcomplete_match = dev_lcomplete dev_root_correct = dev_root_corr best_epoch = epoch patient = 0 state_dict = network.module.state_dict( ) if args.parallel else network.state_dict() torch.save(state_dict, model_name) else: if dev_ucorr_nopunc * 100 / dev_total_nopunc < dev_ucorrect_nopunc * 100 / dev_total_nopunc - 5 or patient >= schedule: state_dict = torch.load(model_name) if args.parallel: network.module.load_state_dict(state_dict) else: network.load_state_dict(state_dict) lr = lr * decay_rate optim = generate_optimizer(opt, lr, network.parameters(), betas, gamma, eps, momentum) if decoding == 'greedy': decode = network.module.decode if args.parallel else network.decode elif decoding == 'mst': decode = network.module.decode_mst if args.parallel else network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) patient = 0 decay += 1 if decay % double_schedule_decay == 0: schedule *= 2 else: patient += 1 print( '----------------------------------------------------------------------------------------------------------------------------' ) print( 'best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (dev_ucorrect, dev_lcorrect, dev_total, dev_ucorrect * 100 / dev_total, dev_lcorrect * 100 / dev_total, dev_ucomlpete_match * 100 / dev_total_inst, dev_lcomplete_match * 100 / dev_total_inst, best_epoch)) print( 'best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc, dev_ucorrect_nopunc * 100 / dev_total_nopunc, dev_lcorrect_nopunc * 100 / dev_total_nopunc, dev_ucomlpete_match_nopunc * 100 / dev_total_inst, dev_lcomplete_match_nopunc * 100 / dev_total_inst, best_epoch)) print( 'best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % (dev_root_correct, dev_total_root, dev_root_correct * 100 / dev_total_root, best_epoch)) print( '----------------------------------------------------------------------------------------------------------------------------' ) if decay == max_decay: break torch.cuda.empty_cache() # release memory that can be released
def main(): parser = argparse.ArgumentParser(description='Tuning with bi-directional RNN-CNN') parser.add_argument('--mode', choices=['RNN', 'LSTM', 'GRU'], help='architecture of rnn', required=True) parser.add_argument('--cuda', action='store_true', help='using GPU') parser.add_argument('--num_epochs', type=int, default=1000, help='Number of training epochs') parser.add_argument('--batch_size', type=int, default=16, help='Number of sentences in each batch') parser.add_argument('--hidden_size', type=int, default=128, help='Number of hidden units in RNN') parser.add_argument('--tag_space', type=int, default=0, help='Dimension of tag space') parser.add_argument('--num_layers', type=int, default=1, help='Number of layers of RNN') parser.add_argument('--num_filters', type=int, default=30, help='Number of filters in CNN') parser.add_argument('--char_dim', type=int, default=30, help='Dimension of Character embeddings') parser.add_argument('--learning_rate', type=float, default=0.1, help='Learning rate') parser.add_argument('--decay_rate', type=float, default=0.1, help='Decay rate of learning rate') parser.add_argument('--gamma', type=float, default=0.0, help='weight for regularization') parser.add_argument('--dropout', choices=['std', 'variational'], help='type of dropout', required=True) parser.add_argument('--p_rnn', nargs=2, type=float, required=True, help='dropout rate for RNN') parser.add_argument('--p_in', type=float, default=0.33, help='dropout rate for input embeddings') parser.add_argument('--p_out', type=float, default=0.33, help='dropout rate for output layer') parser.add_argument('--schedule', type=int, help='schedule for learning rate decay') parser.add_argument('--unk_replace', type=float, default=0., help='The rate to replace a singleton word with UNK') parser.add_argument('--embedding', choices=['glove', 'senna', 'sskip', 'polyglot'], help='Embedding for words', required=True) parser.add_argument('--embedding_dict', help='path for embedding dict') parser.add_argument('--train') # "data/POS-penn/wsj/split1/wsj1.train.original" parser.add_argument('--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original" parser.add_argument('--test') # "data/POS-penn/wsj/split1/wsj1.test.original" args = parser.parse_args() logger = get_logger("POSTagger") mode = args.mode train_path = args.train dev_path = args.dev test_path = args.test num_epochs = args.num_epochs batch_size = args.batch_size hidden_size = args.hidden_size num_filters = args.num_filters learning_rate = args.learning_rate momentum = 0.9 decay_rate = args.decay_rate gamma = args.gamma schedule = args.schedule p_rnn = tuple(args.p_rnn) p_in = args.p_in p_out = args.p_out unk_replace = args.unk_replace embedding = args.embedding embedding_path = args.embedding_dict embedd_dict, embedd_dim = utils.load_embedding_dict(embedding, embedding_path) logger.info("Creating Alphabets") word_alphabet, char_alphabet, pos_alphabet, \ type_alphabet = conllx_data.create_alphabets("data/alphabets/pos/", train_path, data_paths=[dev_path,test_path], max_vocabulary_size=50000, embedd_dict=embedd_dict) logger.info("Word Alphabet Size: %d" % word_alphabet.size()) logger.info("Character Alphabet Size: %d" % char_alphabet.size()) logger.info("POS Alphabet Size: %d" % pos_alphabet.size()) logger.info("Reading Data") device = torch.device('cuda') if args.cuda else torch.device('cpu') data_train = conllx_data.read_data_to_tensor(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, device=device) # data_train = conllx_data.read_data(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet) # num_data = sum([len(bucket) for bucket in data_train]) num_data = sum(data_train[1]) num_labels = pos_alphabet.size() data_dev = conllx_data.read_data_to_tensor(dev_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, device=device) data_test = conllx_data.read_data_to_tensor(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, device=device) def construct_word_embedding_table(): scale = np.sqrt(3.0 / embedd_dim) table = np.empty([word_alphabet.size(), embedd_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.random.uniform(-scale, scale, [1, embedd_dim]).astype(np.float32) oov = 0 for word, index in word_alphabet.items(): if word in embedd_dict: embedding = embedd_dict[word] elif word.lower() in embedd_dict: embedding = embedd_dict[word.lower()] else: embedding = np.random.uniform(-scale, scale, [1, embedd_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('oov: %d' % oov) return torch.from_numpy(table) word_table = construct_word_embedding_table() logger.info("constructing network...") char_dim = args.char_dim window = 3 num_layers = args.num_layers tag_space = args.tag_space initializer = nn.init.xavier_uniform_ if args.dropout == 'std': network = BiRecurrentConv(embedd_dim, word_alphabet.size(), char_dim, char_alphabet.size(), num_filters, window, mode, hidden_size, num_layers, num_labels, tag_space=tag_space, embedd_word=word_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, initializer=initializer) else: network = BiVarRecurrentConv(embedd_dim, word_alphabet.size(), char_dim, char_alphabet.size(), num_filters, window, mode, hidden_size, num_layers, num_labels, tag_space=tag_space, embedd_word=word_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, initializer=initializer) network = network.to(device) lr = learning_rate # optim = Adam(network.parameters(), lr=lr, betas=(0.9, 0.9), weight_decay=gamma) optim = SGD(network.parameters(), lr=lr, momentum=momentum, weight_decay=gamma, nesterov=True) logger.info("Network: %s, num_layer=%d, hidden=%d, filter=%d, tag_space=%d" % (mode, num_layers, hidden_size, num_filters, tag_space)) logger.info("training: l2: %f, (#training data: %d, batch: %d, unk replace: %.2f)" % (gamma, num_data, batch_size, unk_replace)) logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" % (p_in, p_out, p_rnn)) num_batches = num_data / batch_size + 1 dev_correct = 0.0 best_epoch = 0 test_correct = 0.0 test_total = 0 for epoch in range(1, num_epochs + 1): print('Epoch %d (%s(%s), learning rate=%.4f, decay rate=%.4f (schedule=%d)): ' % (epoch, mode, args.dropout, lr, decay_rate, schedule)) train_err = 0. train_corr = 0. train_total = 0. start_time = time.time() num_back = 0 network.train() for batch in range(1, num_batches + 1): word, char, labels, _, _, masks, lengths = conllx_data.get_batch_tensor(data_train, batch_size, unk_replace=unk_replace) optim.zero_grad() loss, corr, _ = network.loss(word, char, labels, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) loss.backward() optim.step() with torch.no_grad(): num_tokens = masks.sum() train_err += loss * num_tokens train_corr += corr train_total += num_tokens time_ave = (time.time() - start_time) / batch time_left = (num_batches - batch) * time_ave # update log if batch % 100 == 0: sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) log_info = 'train: %d/%d loss: %.4f, acc: %.2f%%, time left (estimated): %.2fs' % (batch, num_batches, train_err / train_total, train_corr * 100 / train_total, time_left) sys.stdout.write(log_info) sys.stdout.flush() num_back = len(log_info) sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) print('train: %d loss: %.4f, acc: %.2f%%, time: %.2fs' % (num_batches, train_err / train_total, train_corr * 100 / train_total, time.time() - start_time)) # evaluate performance on dev data with torch.no_grad(): network.eval() dev_corr = 0.0 dev_total = 0 for batch in conllx_data.iterate_batch_tensor(data_dev, batch_size): word, char, labels, _, _, masks, lengths = batch _, corr, preds = network.loss(word, char, labels, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) num_tokens = masks.sum() dev_corr += corr dev_total += num_tokens print('dev corr: %d, total: %d, acc: %.2f%%' % (dev_corr, dev_total, dev_corr * 100 / dev_total)) if dev_correct < dev_corr: dev_correct = dev_corr best_epoch = epoch # evaluate on test data when better performance detected test_corr = 0.0 test_total = 0 for batch in conllx_data.iterate_batch_tensor(data_test, batch_size): word, char, labels, _, _, masks, lengths = batch _, corr, preds = network.loss(word, char, labels, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) num_tokens = masks.sum() test_corr += corr test_total += num_tokens test_correct = test_corr print("best dev corr: %d, total: %d, acc: %.2f%% (epoch: %d)" % (dev_correct, dev_total, dev_correct * 100 / dev_total, best_epoch)) print("best test corr: %d, total: %d, acc: %.2f%% (epoch: %d)" % (test_correct, test_total, test_correct * 100 / test_total, best_epoch)) if epoch % schedule == 0: lr = learning_rate / (1.0 + epoch * decay_rate) optim = SGD(network.parameters(), lr=lr, momentum=momentum, weight_decay=gamma, nesterov=True)
def parser(): try: params = request.json if params is None: return http_sender.send_error_response( http_response_message.ResponseCode.JSON_SYNTAX_ERROR) if 'sentence' in params: query = params['sentence'].strip() logging.info("Input: " + query) # check query null if query.strip() == "": return http_sender.send_error_response( http_response_message.ResponseCode.EMPTY_REQUEST) result_segment = '' words, postags = word_segmentation(query) for index, (word, pos) in enumerate(zip(words, postags)): word = word.replace("_", " ") if pos == 'CH': pos = 'PUNCT' elif pos == 'L': pos = 'DET' elif pos == 'A': pos = 'ADJ' elif pos == 'R': pos = 'ADV' elif pos == 'Np': pos = 'NNP' elif pos == 'M': pos = 'NUM' elif pos == 'E': pos = 'PRE' elif pos == 'P': pos = 'PRO' elif pos == 'Cc': pos = 'CC' elif pos == 'T': pos = 'PART' elif pos == 'Y': pos = 'NNP' elif pos == 'Cb': pos = 'CC' elif pos == 'Eb': pos = 'FW' elif pos == 'Ni': pos = 'Ny' elif pos == 'B': pos = 'NNP' elif pos == 'L': pos = 'DET' elif pos == 'Aux': pos = 'AUX' elif pos == 'NN': pos = 'N' result_segment += str(index + 1) + '\t' + word + '\t' + word.lower() + '\t' + pos + '\t' + pos + '\t' \ + '_' + '\t' + '_' + '\t' + '_' + '\t' + '_' + '\t' + '_' + '\n' result_segment = result_segment.strip() # split data for test test_folder = 'tmp' if not os.path.exists(test_folder): os.mkdir(test_folder) else: for file in os.listdir(test_folder): os.remove(test_folder + '/' + file) output_path = test_folder + '/test.txt' fout = open(output_path, 'w') fout.write(result_segment + '\n') fout.close() alphabet_path = os.path.join(model_path, 'alphabets/') model_name = os.path.join(model_path, 'network.pt') word_alphabet, char_alphabet, pos_alphabet, \ type_alphabet, max_sent_length = conllx_data.create_alphabets(alphabet_path, None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) tokenizer = AutoTokenizer.from_pretrained(phobert_path) model_bert = AutoModel.from_pretrained(phobert_path) processor = DPProcessor() test_path = 'tmp/test.txt' feature_bert_path = 'tmp/phobert_features.pth' train_examples = processor.get_train_examples(test_path) all_lengths = [] for t in train_examples: all_lengths.append(len(t.text_a)) max_seq_len = max(all_lengths) + 1 if max_seq_len > 512: max_seq_len = 512 logger.info("Max sequence length reset to 512") device = torch.device("cuda") model_bert.to(device) train_features = convert_examples_to_features( train_examples, max_seq_len, tokenizer) all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long) all_input_mask = torch.tensor( [f.input_mask for f in train_features], dtype=torch.long) all_segment_ids = torch.tensor( [f.token_type_ids for f in train_features], dtype=torch.long) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids) train_sampler = SequentialSampler(train_data) train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=32) model_bert.eval() to_save = {} # for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")): for step, batch in enumerate(train_dataloader): batch = tuple(t.to(device) for t in batch) input_ids, input_mask, token_type_ids = batch with torch.no_grad(): all_encoder_layers = model_bert( input_ids, attention_mask=input_mask, token_type_ids=token_type_ids) output_ = all_encoder_layers[0] for j in range(len(input_ids)): sent_id = j + step * 32 layer_output = output_[j, :input_mask[j].to('cpu').sum()] to_save[sent_id] = layer_output.detach().cpu().numpy() torch.save(to_save, feature_bert_path) data_test = conllx_data.read_data_to_tensor(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, feature_bert_path, elmo_path, symbolic_root=True, device=device, use_elmo=False, use_bert=False, use_elmo_bert=True, use_test=True) pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) logger.info('model: %s' % model_name) def load_model_arguments_from_json(): arguments = json.load(open(arg_path, 'r')) return arguments['args'], arguments['kwargs'] arg_path = model_name[0:-1] + '.arg.json' if not os.path.isfile(arg_path): arg_path = model_name + '.arg.json' args_, kwargs = load_model_arguments_from_json() network = DeepBiAffineTransform(*args_, **kwargs, use_elmo=False, use_bert=False, use_elmo_bert=True) network.load_state_dict(torch.load(model_name)) if use_gpu: network.cuda() else: network.cpu() network.eval() decode = network.decode_mst out_filename = 'tmp/test' pred_writer.start(out_filename + '_pred.conll') for batch in conllx_data.iterate_batch_tensor(data_test, 1, use_elmo=False, use_bert=False, use_elmo_bert=True): sys.stdout.flush() word, char, pos, heads, types, masks, lengths, elmos, berts = batch heads_pred, types_pred = decode( word, char, pos, elmos, berts, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True) pred_writer.close() sents_gold = result_segment.split('\n') result = '' test_path = 'tmp/test_pred.conll' lines = open(test_path, 'r').readlines() for i, line in enumerate(lines): if line.strip() != '': sent = sents_gold[i] words_gold = sent.split('\t') word = words_gold[1] line = line.strip() words = line.split('\t') line = words[0] + '\t' + word + '\t' + word.lower( ) + '\t' + words[4] + '\t' + words[4] + '\t_\t' + words[ 6] + '\t' + words[7] + '\t_\t_' + '\n' if line != '': result += line + '\n' result = result.strip() logging.info("Result: " + str(result)) return http_sender.send_http_result(result) else: return http_sender.send_error_response( http_response_message.ResponseCode.INPUT_FORMAT_ERROR) except BadRequest: return http_sender.send_error_response( http_response_message.ResponseCode.JSON_SYNTAX_ERROR)
def biaffine(model_path, model_name, test_path, punct_set, use_gpu, logger, args): alphabet_path = os.path.join(model_path, 'alphabets/') model_name = os.path.join(model_path, model_name) word_alphabet, char_alphabet, pos_alphabet, type_alphabet, max_sent_length = conllx_data.create_alphabets(alphabet_path, None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None) # word_alphabet, char_alphabet, pos_alphabet, type_alphabet = create_alphabets(alphabet_path, # None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) decoding = args.decode out_filename = args.out_filename constraints_method = args.constraints_method constraintFile = args.constraint_file ratioFile = args.ratio_file tolerance = args.tolerance gamma = args.gamma the_language = args.mt_log[9:11] mt_log = open(args.mt_log, 'a') summary_log = open(args.summary_log, 'a') logger.info('use gpu: %s, decoding: %s' % (use_gpu, decoding)) # extra_embeds_arr = augment_with_extra_embedding(word_alphabet, args.extra_embed, args.extra_embed_src, test_path, logger) # ===== the reading def _read_one(path, is_train): lang_id = guess_language_id(path) logger.info("Reading: guess that the language of file %s is %s." % (path, lang_id)) one_data = conllx_data.read_data_to_variable(path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=(not is_train), symbolic_root=True, lang_id=lang_id) return one_data data_test = _read_one(test_path, False) # data_test = conllx_data.read_data_to_variable(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, # use_gpu=use_gpu, volatile=True, symbolic_root=True) pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) logger.info('model: %s' % model_name) def load_model_arguments_from_json(): arguments = json.load(open(arg_path, 'r')) return arguments['args'], arguments['kwargs'] arg_path = model_name + '.arg.json' args, kwargs = load_model_arguments_from_json() network = BiRecurrentConvBiAffine(use_gpu=use_gpu, *args, **kwargs) network.load_state_dict(torch.load(model_name)) # augment_network_embed(word_alphabet.size(), network, extra_embeds_arr) if use_gpu: network.cuda() else: network.cpu() network.eval() if decoding == 'greedy': decode = network.decode elif decoding == 'mst': decode = network.decode_mst elif decoding == 'proj': decode = network.decode_proj else: raise ValueError('Unknown decoding algorithm: %s' % decoding) # pred_writer.start('tmp/analyze_pred_%s' % str(uid)) # gold_writer.start('tmp/analyze_gold_%s' % str(uid)) # pred_writer.start(model_path + out_filename + '_pred') # gold_writer.start(model_path + out_filename + '_gold') pred_writer.start(out_filename + '_pred') gold_writer.start(out_filename + '_gold') sent = 0 start_time = time.time() constraints = [] mt_log.write("=====================%s, Ablation 2================\n"%(constraints_method)) summary_log.write("==========================%s, Ablation 2=============\n"%(constraints_method)) if ratioFile == 'WALS': import pickle as pk cFile = open(constraintFile, 'rb') WALS_data = pk.load(cFile) for idx in ['85A', '87A', '89A']: constraint = Constraint(0,0,0) extra_const = constraint.load_WALS(idx, WALS_data[the_language][idx], pos_alphabet, method=constraints_method) constraints.append(constraint) if extra_const: constraints.append(extra_const) constraint = Constraint(0,0,0) extra_const = constraint.load_WALS_unary(WALS_data[the_language], pos_alphabet, method=constraints_method) if extra_const: constraints.append(extra_const) constraints.append(constraint) elif ratioFile == 'None': summary_log.write("=================No it is baseline================\n") mt_log.write("==================No it is baseline==============\n") else: cFile = open(constraintFile, 'r') for line in cFile: if len(line.strip()) < 2: break pos1, pos2 = line.strip().split('\t') constraint = Constraint(0,0,0) constraint.load(pos1, pos2, ratioFile, pos_alphabet) constraints.append(constraint) test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete_match = 0.0 test_lcomplete_match = 0.0 test_total = 0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_match_nopunc = 0.0 test_lcomplete_match_nopunc = 0.0 test_total_nopunc = 0 test_total_inst = 0 test_root_correct = 0.0 test_total_root = 0 arc_list = [] type_list = [] length_list = [] pos_list = [] for batch in conllx_data.iterate_batch_variable(data_test, 1): word, char, pos, heads, types, masks, lengths = batch out_arc, out_type, length = network.pretrain_constraint(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) arc_list += list(out_arc) type_list += list(out_type) length_list += list(length) pos_list += list(pos) if constraints_method == 'binary': train_constraints = network.binary_constraints if constraints_method == 'Lagrange': train_constraints = network.Lagrange_constraints if constraints_method == 'PR': train_constraints = network.PR_constraints train_constraints(arc_list, type_list, length_list, pos_list, constraints, tolerance, mt_log, gamma=gamma) for batch in conllx_data.iterate_batch_variable(data_test, 1): #sys.stdout.write('%d, ' % sent) #sys.stdout.flush() sent += 1 word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS, constraints=constraints, method=constraints_method, gamma=gamma) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.data.cpu().numpy() types = types.data.cpu().numpy() pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True) gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True) stats, stats_nopunc, stats_root, num_inst = parser.eval(word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root test_ucorrect += ucorr test_lcorrect += lcorr test_total += total test_ucomlpete_match += ucm test_lcomplete_match += lcm test_ucorrect_nopunc += ucorr_nopunc test_lcorrect_nopunc += lcorr_nopunc test_total_nopunc += total_nopunc test_ucomlpete_match_nopunc += ucm_nopunc test_lcomplete_match_nopunc += lcm_nopunc test_root_correct += corr_root test_total_root += total_root test_total_inst += num_inst print('\ntime: %.2fs' % (time.time() - start_time)) print('test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % ( test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 / test_total, test_lcorrect * 100 / test_total, test_ucomlpete_match * 100 / test_total_inst, test_lcomplete_match * 100 / test_total_inst)) print('test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % ( test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc, test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc, test_ucomlpete_match_nopunc * 100 / test_total_inst, test_lcomplete_match_nopunc * 100 / test_total_inst)) print('test Root: corr: %d, total: %d, acc: %.2f%%' % ( test_root_correct, test_total_root, test_root_correct * 100 / test_total_root)) mt_log.write('uas: %.2f, las: %.2f\n'%(test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc)) summary_log.write('%s: %.2f %.2f\n'%(the_language, test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc)) pred_writer.close() gold_writer.close()
def main(): parser = argparse.ArgumentParser( description='NER with bi-directional RNN-CNN') parser.add_argument('--config', type=str, help='config file', required=True) parser.add_argument('--num_epochs', type=int, default=100, help='Number of training epochs') parser.add_argument('--batch_size', type=int, default=16, help='Number of sentences in each batch') parser.add_argument('--loss_type', choices=['sentence', 'token'], default='sentence', help='loss type (default: sentence)') parser.add_argument('--optim', choices=['sgd', 'adam'], help='type of optimizer', required=True) parser.add_argument('--learning_rate', type=float, default=0.1, help='Learning rate') parser.add_argument('--lr_decay', type=float, default=0.999995, help='Decay rate of learning rate') parser.add_argument('--amsgrad', action='store_true', help='AMS Grad') parser.add_argument('--grad_clip', type=float, default=0, help='max norm for gradient clip (default 0: no clip') parser.add_argument('--warmup_steps', type=int, default=0, metavar='N', help='number of steps to warm up (default: 0)') parser.add_argument('--weight_decay', type=float, default=0.0, help='weight for l2 norm decay') parser.add_argument('--unk_replace', type=float, default=0., help='The rate to replace a singleton word with UNK') parser.add_argument('--embedding', choices=['glove', 'senna', 'sskip', 'polyglot'], help='Embedding for words', required=True) parser.add_argument('--embedding_dict', help='path for embedding dict') parser.add_argument('--train', help='path for training file.', required=True) parser.add_argument('--dev', help='path for dev file.', required=True) parser.add_argument('--test', help='path for test file.', required=True) parser.add_argument('--model_path', help='path for saving model file.', required=True) args = parser.parse_args() logger = get_logger("POS") args.cuda = torch.cuda.is_available() device = torch.device('cuda', 0) if args.cuda else torch.device('cpu') train_path = args.train dev_path = args.dev test_path = args.test num_epochs = args.num_epochs batch_size = args.batch_size optim = args.optim learning_rate = args.learning_rate lr_decay = args.lr_decay amsgrad = args.amsgrad warmup_steps = args.warmup_steps weight_decay = args.weight_decay grad_clip = args.grad_clip loss_ty_token = args.loss_type == 'token' unk_replace = args.unk_replace model_path = args.model_path model_name = os.path.join(model_path, 'model.pt') embedding = args.embedding embedding_path = args.embedding_dict print(args) embedd_dict, embedd_dim = utils.load_embedding_dict( embedding, embedding_path) logger.info("Creating Alphabets") alphabet_path = os.path.join(model_path, 'alphabets') word_alphabet, char_alphabet, pos_alphabet, type_alphabet = conllx_data.create_alphabets( alphabet_path, train_path, data_paths=[dev_path, test_path], embedd_dict=embedd_dict, max_vocabulary_size=50000) logger.info("Word Alphabet Size: %d" % word_alphabet.size()) logger.info("Character Alphabet Size: %d" % char_alphabet.size()) logger.info("POS Alphabet Size: %d" % pos_alphabet.size()) logger.info("Reading Data") data_train = conllx_data.read_bucketed_data(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet) num_data = sum(data_train[1]) num_labels = pos_alphabet.size() data_dev = conllx_data.read_data(dev_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet) data_test = conllx_data.read_data(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet) def construct_word_embedding_table(): scale = np.sqrt(3.0 / embedd_dim) table = np.empty([word_alphabet.size(), embedd_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.random.uniform( -scale, scale, [1, embedd_dim]).astype(np.float32) oov = 0 for word, index in word_alphabet.items(): if word in embedd_dict: embedding = embedd_dict[word] elif word.lower() in embedd_dict: embedding = embedd_dict[word.lower()] else: embedding = np.random.uniform( -scale, scale, [1, embedd_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('oov: %d' % oov) return torch.from_numpy(table) word_table = construct_word_embedding_table() logger.info("constructing network...") hyps = json.load(open(args.config, 'r')) json.dump(hyps, open(os.path.join(model_path, 'config.json'), 'w'), indent=2) dropout = hyps['dropout'] crf = hyps['crf'] bigram = hyps['bigram'] assert embedd_dim == hyps['embedd_dim'] char_dim = hyps['char_dim'] mode = hyps['rnn_mode'] hidden_size = hyps['hidden_size'] out_features = hyps['out_features'] num_layers = hyps['num_layers'] p_in = hyps['p_in'] p_out = hyps['p_out'] p_rnn = hyps['p_rnn'] activation = hyps['activation'] if dropout == 'std': if crf: network = BiRecurrentConvCRF(embedd_dim, word_alphabet.size(), char_dim, char_alphabet.size(), mode, hidden_size, out_features, num_layers, num_labels, embedd_word=word_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, bigram=bigram, activation=activation) else: network = BiRecurrentConv(embedd_dim, word_alphabet.size(), char_dim, char_alphabet.size(), mode, hidden_size, out_features, num_layers, num_labels, embedd_word=word_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, activation=activation) elif dropout == 'variational': if crf: network = BiVarRecurrentConvCRF(embedd_dim, word_alphabet.size(), char_dim, char_alphabet.size(), mode, hidden_size, out_features, num_layers, num_labels, embedd_word=word_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, bigram=bigram, activation=activation) else: network = BiVarRecurrentConv(embedd_dim, word_alphabet.size(), char_dim, char_alphabet.size(), mode, hidden_size, out_features, num_layers, num_labels, embedd_word=word_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, activation=activation) else: raise ValueError('Unkown dropout type: {}'.format(dropout)) network = network.to(device) optimizer, scheduler = get_optimizer(network.parameters(), optim, learning_rate, lr_decay, amsgrad, weight_decay, warmup_steps) model = "{}-CNN{}".format(mode, "-CRF" if crf else "") logger.info("Network: %s, num_layer=%d, hidden=%d, act=%s" % (model, num_layers, hidden_size, activation)) logger.info( "training: l2: %f, (#training data: %d, batch: %d, unk replace: %.2f)" % (weight_decay, num_data, batch_size, unk_replace)) logger.info("dropout(in, out, rnn): %s(%.2f, %.2f, %s)" % (dropout, p_in, p_out, p_rnn)) print('# of Parameters: %d' % (sum([param.numel() for param in network.parameters()]))) best_corr = 0.0 best_total = 0.0 test_corr = 0.0 test_total = 0.0 best_epoch = 0 patient = 0 num_batches = num_data // batch_size + 1 result_path = os.path.join(model_path, 'tmp') if not os.path.exists(result_path): os.makedirs(result_path) for epoch in range(1, num_epochs + 1): start_time = time.time() train_loss = 0. num_insts = 0 num_words = 0 num_back = 0 network.train() lr = scheduler.get_lr()[0] print('Epoch %d (%s, lr=%.6f, lr decay=%.6f, amsgrad=%s, l2=%.1e): ' % (epoch, optim, lr, lr_decay, amsgrad, weight_decay)) if args.cuda: torch.cuda.empty_cache() gc.collect() for step, data in enumerate( iterate_data(data_train, batch_size, bucketed=True, unk_replace=unk_replace, shuffle=True)): optimizer.zero_grad() words = data['WORD'].to(device) chars = data['CHAR'].to(device) labels = data['POS'].to(device) masks = data['MASK'].to(device) nbatch = words.size(0) nwords = masks.sum().item() loss_total = network.loss(words, chars, labels, mask=masks).sum() if loss_ty_token: loss = loss_total.div(nwords) else: loss = loss_total.div(nbatch) loss.backward() if grad_clip > 0: clip_grad_norm_(network.parameters(), grad_clip) optimizer.step() scheduler.step() with torch.no_grad(): num_insts += nbatch num_words += nwords train_loss += loss_total.item() # update log if step % 100 == 0: torch.cuda.empty_cache() sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) curr_lr = scheduler.get_lr()[0] log_info = '[%d/%d (%.0f%%) lr=%.6f] loss: %.4f (%.4f)' % ( step, num_batches, 100. * step / num_batches, curr_lr, train_loss / num_insts, train_loss / num_words) sys.stdout.write(log_info) sys.stdout.flush() num_back = len(log_info) sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) print('total: %d (%d), loss: %.4f (%.4f), time: %.2fs' % (num_insts, num_words, train_loss / num_insts, train_loss / num_words, time.time() - start_time)) print('-' * 100) # evaluate performance on dev data with torch.no_grad(): dev_corr, dev_total = eval(data_dev, network, device) print('Dev corr: %d, total: %d, acc: %.2f%%' % (dev_corr, dev_total, dev_corr * 100 / dev_total)) if best_corr < dev_corr: torch.save(network.state_dict(), model_name) best_corr = dev_corr best_total = dev_total best_epoch = epoch # evaluate on test data when better performance detected test_corr, test_total = eval(data_test, network, device) print('test corr: %d, total: %d, acc: %.2f%%' % (test_corr, test_total, test_corr * 100 / test_total)) patient = 0 else: patient += 1 print('-' * 100) print( "Best dev corr: %d, total: %d, acc: %.2f%% (epoch: %d (%d))" % (best_corr, best_total, best_corr * 100 / best_total, best_epoch, patient)) print( "Best test corr: %d, total: %d, acc: %.2f%% (epoch: %d (%d))" % (test_corr, test_total, test_corr * 100 / test_total, best_epoch, patient)) print('=' * 100) if patient > 4: logger.info('reset optimizer momentums') scheduler.reset_state() patient = 0
def biaffine(model_path, model_name, pre_model_path, pre_model_name, use_gpu, logger, args): alphabet_path = os.path.join(pre_model_path, 'alphabets/') logger.info("Alphabet Path: %s" % alphabet_path) pre_model_name = os.path.join(pre_model_path, pre_model_name) model_name = os.path.join(model_path, model_name) # Load pre-created alphabets word_alphabet, char_alphabet, pos_alphabet, type_alphabet, max_sent_length = conllx_data.create_alphabets( alphabet_path, None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) logger.info('use gpu: %s' % (use_gpu)) if args.test_lang: extra_embed = args.embed_dir + ("wiki.multi.%s.vec" % args.test_lang) extra_word_dict, _ = load_embedding_dict('word2vec', extra_embed) test_path = args.data_dir + args.test_lang + '_test.conllu' extra_embeds_arr = augment_with_extra_embedding(word_alphabet, extra_word_dict, test_path, logger) else: extra_embeds_arr = [] for language in args.langs: extra_embed = args.embed_dir + ("wiki.multi.%s.vec" % language) extra_word_dict, _ = load_embedding_dict('word2vec', extra_embed) test_path = args.data_dir + language + '_train.conllu' embeds_arr1 = augment_with_extra_embedding(word_alphabet, extra_word_dict, test_path, logger) test_path = args.data_dir + language + '_dev.conllu' embeds_arr2 = augment_with_extra_embedding(word_alphabet, extra_word_dict, test_path, logger) test_path = args.data_dir + language + '_test.conllu' embeds_arr3 = augment_with_extra_embedding(word_alphabet, extra_word_dict, test_path, logger) extra_embeds_arr.extend(embeds_arr1 + embeds_arr2 + embeds_arr3) # ------------------------------------------------------------------------- # # --------------------- Loading model ------------------------------------- # def load_model_arguments_from_json(): arguments = json.load(open(arg_path, 'r')) return arguments['args'], arguments['kwargs'] arg_path = pre_model_name + '.arg.json' margs, kwargs = load_model_arguments_from_json() network = BiRecurrentConvBiAffine(use_gpu=use_gpu, *margs, **kwargs) network.load_state_dict(torch.load(pre_model_name)) args.use_bert = kwargs.get('use_bert', False) # augment_network_embed(word_alphabet.size(), network, extra_embeds_arr) network.eval() logger.info('model: %s' % pre_model_name) # Freeze the network for p in network.parameters(): p.requires_grad = False nclass = args.nclass classifier = nn.Sequential( nn.Linear(network.encoder.output_dim, 512), nn.Linear(512, nclass) ) if use_gpu: network.cuda() classifier.cuda() else: network.cpu() classifier.cpu() batch_size = args.batch_size # ===== the reading def _read_one(path, is_train=False, max_size=None): lang_id = guess_language_id(path) logger.info("Reading: guess that the language of file %s is %s." % (path, lang_id)) one_data = conllx_data.read_data_to_variable(path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=(not is_train), use_bert=args.use_bert, symbolic_root=True, lang_id=lang_id, max_size=max_size) return one_data def compute_accuracy(data, lang_idx): total_corr, total = 0, 0 classifier.eval() with torch.no_grad(): for batch in conllx_data.iterate_batch_variable(data, batch_size): word, char, pos, _, _, masks, lengths, bert_inputs = batch if use_gpu: word = word.cuda() char = char.cuda() pos = pos.cuda() masks = masks.cuda() lengths = lengths.cuda() if bert_inputs[0] is not None: bert_inputs[0] = bert_inputs[0].cuda() bert_inputs[1] = bert_inputs[1].cuda() bert_inputs[2] = bert_inputs[2].cuda() output = network.forward(word, char, pos, input_bert=bert_inputs, mask=masks, length=lengths, hx=None) output = output['output'].detach() if args.train_level == 'word': output = classifier(output) output = output.contiguous().view(-1, output.size(2)) else: output = torch.mean(output, dim=1) output = classifier(output) preds = output.max(1)[1].cpu() labels = torch.LongTensor([lang_idx]) labels = labels.expand(*preds.size()) n_correct = preds.eq(labels).sum().item() total_corr += n_correct total += output.size(0) return {'total_corr': total_corr, 'total': total} if args.test_lang: classifier.load_state_dict(torch.load(model_name)) path = args.data_dir + args.test_lang + '_train.conllu' test_data = _read_one(path) # TODO: fixed indexing is not GOOD lang_idx = 0 if args.test_lang == args.src_lang else 1 result = compute_accuracy(test_data, lang_idx) accuracy = (result['total_corr'] * 100.0) / result['total'] logger.info('[Classifier performance] Language: %s || accuracy: %.2f%%' % (args.test_lang, accuracy)) else: # if output directory doesn't exist, create it if not os.path.exists(args.model_path): os.makedirs(args.model_path) # --------------------- Loading data -------------------------------------- # train_data = dict() dev_data = dict() test_data = dict() num_data = dict() lang_ids = dict() reverse_lang_ids = dict() # loading language data for language in args.langs: lang_ids[language] = len(lang_ids) reverse_lang_ids[lang_ids[language]] = language train_path = args.data_dir + language + '_train.conllu' # Utilize at most 10000 examples tmp_data = _read_one(train_path, max_size=10000) num_data[language] = sum(tmp_data[1]) train_data[language] = tmp_data dev_path = args.data_dir + language + '_dev.conllu' tmp_data = _read_one(dev_path) dev_data[language] = tmp_data test_path = args.data_dir + language + '_test.conllu' tmp_data = _read_one(test_path) test_data[language] = tmp_data # ------------------------------------------------------------------------- # optim = torch.optim.Adam(classifier.parameters(), lr=0.001) criterion = nn.CrossEntropyLoss() def compute_loss(lang_name, land_idx): word, char, pos, _, _, masks, lengths, bert_inputs = conllx_data.get_batch_variable(train_data[lang_name], batch_size, unk_replace=0.5) if use_gpu: word = word.cuda() char = char.cuda() pos = pos.cuda() masks = masks.cuda() lengths = lengths.cuda() if bert_inputs[0] is not None: bert_inputs[0] = bert_inputs[0].cuda() bert_inputs[1] = bert_inputs[1].cuda() bert_inputs[2] = bert_inputs[2].cuda() output = network.forward(word, char, pos, input_bert=bert_inputs, mask=masks, length=lengths, hx=None) output = output['output'].detach() if args.train_level == 'word': output = classifier(output) output = output.contiguous().view(-1, output.size(2)) else: output = torch.mean(output, dim=1) output = classifier(output) labels = torch.empty(output.size(0)).fill_(land_idx).type_as(output).long() loss = criterion(output, labels) return loss # ---------------------- Form the mini-batches -------------------------- # num_batches = 0 batch_lang_labels = [] for lang in args.langs: nbatches = num_data[lang] // batch_size + 1 batch_lang_labels.extend([lang] * nbatches) num_batches += nbatches assert len(batch_lang_labels) == num_batches # ------------------------------------------------------------------------- # best_dev_accuracy = 0 patience = 0 for epoch in range(1, args.num_epochs + 1): # shuffling the data lang_in_batch = copy.copy(batch_lang_labels) random.shuffle(lang_in_batch) classifier.train() for batch in range(1, num_batches + 1): lang_name = lang_in_batch[batch - 1] lang_id = lang_ids.get(lang_name) loss = compute_loss(lang_name, lang_id) loss.backward() optim.step() # Validation avg_acc = dict() for dev_lang in dev_data.keys(): lang_idx = lang_ids.get(dev_lang) result = compute_accuracy(dev_data[dev_lang], lang_idx) accuracy = (result['total_corr'] * 100.0) / result['total'] avg_acc[dev_lang] = accuracy acc = ', '.join('%s: %.2f' % (key, val) for (key, val) in avg_acc.items()) logger.info('Epoch: %d, Performance[%s]' % (epoch, acc)) avg_acc = sum(avg_acc.values()) / len(avg_acc) if best_dev_accuracy < avg_acc: best_dev_accuracy = avg_acc patience = 0 state_dict = classifier.state_dict() torch.save(state_dict, model_name) else: patience += 1 if patience >= 5: break # Testing logger.info('Testing model %s' % pre_model_name) total_corr, total = 0, 0 for test_lang in UD_languages: if test_lang in test_data: lang_idx = lang_ids.get(test_lang) result = compute_accuracy(test_data[test_lang], lang_idx) accuracy = (result['total_corr'] * 100.0) / result['total'] print('[LANG]: %s, [ACC]: %.2f' % (test_lang.upper(), accuracy)) total_corr += result['total_corr'] total += result['total'] print('[Avg. Performance]: %.2f' % ((total_corr * 100.0) / total))
def main(): args_parser = argparse.ArgumentParser( description='Tuning with graph-based parsing') args_parser.add_argument('--seed', type=int, default=1234, help='random seed for reproducibility') args_parser.add_argument('--mode', choices=['RNN', 'LSTM', 'GRU', 'FastLSTM'], help='architecture of rnn', required=True) args_parser.add_argument('--num_epochs', type=int, default=1000, help='Number of training epochs') args_parser.add_argument('--batch_size', type=int, default=64, help='Number of sentences in each batch') args_parser.add_argument('--hidden_size', type=int, default=256, help='Number of hidden units in RNN') args_parser.add_argument('--arc_space', type=int, default=128, help='Dimension of tag space') args_parser.add_argument('--type_space', type=int, default=128, help='Dimension of tag space') args_parser.add_argument('--num_layers', type=int, default=1, help='Number of layers of encoder.') args_parser.add_argument('--num_filters', type=int, default=50, help='Number of filters in CNN') args_parser.add_argument('--pos', action='store_true', help='use part-of-speech embedding.') args_parser.add_argument('--char', action='store_true', help='use character embedding and CNN.') args_parser.add_argument('--pos_dim', type=int, default=50, help='Dimension of POS embeddings') args_parser.add_argument('--char_dim', type=int, default=50, help='Dimension of Character embeddings') args_parser.add_argument('--opt', choices=['adam', 'sgd', 'adamax'], help='optimization algorithm') args_parser.add_argument('--objective', choices=['cross_entropy', 'crf'], default='cross_entropy', help='objective function of training procedure.') args_parser.add_argument('--decode', choices=['mst', 'greedy'], default='mst', help='decoding algorithm') args_parser.add_argument('--learning_rate', type=float, default=0.01, help='Learning rate') # args_parser.add_argument('--decay_rate', type=float, default=0.05, help='Decay rate of learning rate') args_parser.add_argument('--clip', type=float, default=5.0, help='gradient clipping') args_parser.add_argument('--gamma', type=float, default=0.0, help='weight for regularization') args_parser.add_argument('--epsilon', type=float, default=1e-8, help='epsilon for adam or adamax') args_parser.add_argument('--p_rnn', nargs='+', type=float, required=True, help='dropout rate for RNN') args_parser.add_argument('--p_in', type=float, default=0.33, help='dropout rate for input embeddings') args_parser.add_argument('--p_out', type=float, default=0.33, help='dropout rate for output layer') # args_parser.add_argument('--schedule', type=int, help='schedule for learning rate decay') args_parser.add_argument( '--unk_replace', type=float, default=0., help='The rate to replace a singleton word with UNK') args_parser.add_argument('--punctuation', nargs='+', type=str, help='List of punctuations') args_parser.add_argument( '--word_embedding', choices=['word2vec', 'glove', 'senna', 'sskip', 'polyglot'], help='Embedding for words', required=True) args_parser.add_argument('--word_path', help='path for word embedding dict') args_parser.add_argument( '--freeze', action='store_true', help='frozen the word embedding (disable fine-tuning).') args_parser.add_argument('--char_embedding', choices=['random', 'polyglot'], help='Embedding for characters', required=True) args_parser.add_argument('--char_path', help='path for character embedding dict') args_parser.add_argument( '--train') # "data/POS-penn/wsj/split1/wsj1.train.original" args_parser.add_argument( '--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original" args_parser.add_argument( '--test') # "data/POS-penn/wsj/split1/wsj1.test.original" args_parser.add_argument('--vocab_path', help='path for prebuilt alphabets.', default=None) args_parser.add_argument('--model_path', help='path for saving model file.', required=True) args_parser.add_argument('--model_name', help='name for saving model file.', required=True) # args_parser.add_argument('--no_word', action='store_true', help='do not use word embedding.') # # lrate schedule with warmup in the first iter. args_parser.add_argument('--use_warmup_schedule', action='store_true', help="Use warmup lrate schedule.") args_parser.add_argument('--decay_rate', type=float, default=0.75, help='Decay rate of learning rate') args_parser.add_argument('--max_decay', type=int, default=9, help='Number of decays before stop') args_parser.add_argument('--schedule', type=int, help='schedule for learning rate decay') args_parser.add_argument('--double_schedule_decay', type=int, default=5, help='Number of decays to double schedule') args_parser.add_argument( '--check_dev', type=int, default=5, help='Check development performance in every n\'th iteration') # Tansformer encoder args_parser.add_argument('--no_CoRNN', action='store_true', help='do not use context RNN.') args_parser.add_argument( '--trans_hid_size', type=int, default=1024, help='#hidden units in point-wise feed-forward in transformer') args_parser.add_argument( '--d_k', type=int, default=64, help='d_k for multi-head-attention in transformer encoder') args_parser.add_argument( '--d_v', type=int, default=64, help='d_v for multi-head-attention in transformer encoder') args_parser.add_argument('--multi_head_attn', action='store_true', help='use multi-head-attention.') args_parser.add_argument('--num_head', type=int, default=8, help='Value of h in multi-head attention') # - positional args_parser.add_argument( '--enc_use_neg_dist', action='store_true', help="Use negative distance for enc's relational-distance embedding.") args_parser.add_argument( '--enc_clip_dist', type=int, default=0, help="The clipping distance for relative position features.") args_parser.add_argument('--position_dim', type=int, default=50, help='Dimension of Position embeddings.') args_parser.add_argument( '--position_embed_num', type=int, default=200, help= 'Minimum value of position embedding num, which usually is max-sent-length.' ) args_parser.add_argument('--train_position', action='store_true', help='train positional encoding for transformer.') # args_parser.add_argument( '--train_len_thresh', type=int, default=100, help='In training, discard sentences longer than this.') # args = args_parser.parse_args() # fix data-prepare seed random.seed(1234) np.random.seed(1234) # model's seed torch.manual_seed(args.seed) logger = get_logger("GraphParser") mode = args.mode obj = args.objective decoding = args.decode train_path = args.train dev_path = args.dev test_path = args.test model_path = args.model_path model_name = args.model_name num_epochs = args.num_epochs batch_size = args.batch_size hidden_size = args.hidden_size arc_space = args.arc_space type_space = args.type_space num_layers = args.num_layers num_filters = args.num_filters learning_rate = args.learning_rate opt = args.opt momentum = 0.9 betas = (0.9, 0.9) eps = args.epsilon decay_rate = args.decay_rate clip = args.clip gamma = args.gamma schedule = args.schedule p_rnn = tuple(args.p_rnn) p_in = args.p_in p_out = args.p_out unk_replace = args.unk_replace punctuation = args.punctuation freeze = args.freeze word_embedding = args.word_embedding word_path = args.word_path use_char = args.char char_embedding = args.char_embedding char_path = args.char_path use_pos = args.pos pos_dim = args.pos_dim word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path) char_dict = None char_dim = args.char_dim if char_embedding != 'random': char_dict, char_dim = utils.load_embedding_dict( char_embedding, char_path) # vocab_path = args.vocab_path if args.vocab_path is not None else args.model_path logger.info("Creating Alphabets") alphabet_path = os.path.join(vocab_path, 'alphabets/') model_name = os.path.join(model_path, model_name) # todo(warn): exactly same for loading vocabs word_alphabet, char_alphabet, pos_alphabet, type_alphabet, max_sent_length = conllx_data.create_alphabets( alphabet_path, train_path, data_paths=[dev_path, test_path], max_vocabulary_size=50000, embedd_dict=word_dict) max_sent_length = max(max_sent_length, args.position_embed_num) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) logger.info("Reading Data") use_gpu = torch.cuda.is_available() # ===== the reading def _read_one(path, is_train): lang_id = guess_language_id(path) logger.info("Reading: guess that the language of file %s is %s." % (path, lang_id)) one_data = conllx_data.read_data_to_variable( path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=(not is_train), symbolic_root=True, lang_id=lang_id, len_thresh=(args.train_len_thresh if is_train else 100000)) return one_data data_train = _read_one(train_path, True) num_data = sum(data_train[1]) data_dev = _read_one(dev_path, False) data_test = _read_one(test_path, False) # ===== punct_set = None if punctuation is not None: punct_set = set(punctuation) logger.info("punctuations(%d): %s" % (len(punct_set), ' '.join(punct_set))) def construct_word_embedding_table(): scale = np.sqrt(3.0 / word_dim) table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.zeros([1, word_dim]).astype( np.float32) if freeze else np.random.uniform( -scale, scale, [1, word_dim]).astype(np.float32) oov = 0 for word, index in word_alphabet.items(): if word in word_dict: embedding = word_dict[word] elif word.lower() in word_dict: embedding = word_dict[word.lower()] else: embedding = np.zeros([1, word_dim]).astype( np.float32) if freeze else np.random.uniform( -scale, scale, [1, word_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('word OOV: %d' % oov) return torch.from_numpy(table) def construct_char_embedding_table(): if char_dict is None: return None scale = np.sqrt(3.0 / char_dim) table = np.empty([num_chars, char_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.random.uniform( -scale, scale, [1, char_dim]).astype(np.float32) oov = 0 for char, index, in char_alphabet.items(): if char in char_dict: embedding = char_dict[char] else: embedding = np.random.uniform(-scale, scale, [1, char_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('character OOV: %d' % oov) return torch.from_numpy(table) word_table = construct_word_embedding_table() char_table = construct_char_embedding_table() window = 3 if obj == 'cross_entropy': network = BiRecurrentConvBiAffine( word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, num_filters, window, mode, hidden_size, num_layers, num_types, arc_space, type_space, embedd_word=word_table, embedd_char=char_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, biaffine=True, pos=use_pos, char=use_char, train_position=args.train_position, use_con_rnn=(not args.no_CoRNN), trans_hid_size=args.trans_hid_size, d_k=args.d_k, d_v=args.d_v, multi_head_attn=args.multi_head_attn, num_head=args.num_head, enc_use_neg_dist=args.enc_use_neg_dist, enc_clip_dist=args.enc_clip_dist, position_dim=args.position_dim, max_sent_length=max_sent_length, use_gpu=use_gpu, no_word=args.no_word) elif obj == 'crf': raise NotImplementedError else: raise RuntimeError('Unknown objective: %s' % obj) def save_args(): arg_path = model_name + '.arg.json' arguments = [ word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, num_filters, window, mode, hidden_size, num_layers, num_types, arc_space, type_space ] kwargs = { 'p_in': p_in, 'p_out': p_out, 'p_rnn': p_rnn, 'biaffine': True, 'pos': use_pos, 'char': use_char, 'train_position': args.train_position, 'use_con_rnn': (not args.no_CoRNN), 'trans_hid_size': args.trans_hid_size, 'd_k': args.d_k, 'd_v': args.d_v, 'multi_head_attn': args.multi_head_attn, 'num_head': args.num_head, 'enc_use_neg_dist': args.enc_use_neg_dist, 'enc_clip_dist': args.enc_clip_dist, 'position_dim': args.position_dim, 'max_sent_length': max_sent_length, 'no_word': args.no_word } json.dump({ 'args': arguments, 'kwargs': kwargs }, open(arg_path, 'w'), indent=4) if freeze: network.word_embedd.freeze() if use_gpu: network.cuda() save_args() pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) def generate_optimizer(opt, lr, params): params = filter(lambda param: param.requires_grad, params) if opt == 'adam': return Adam(params, lr=lr, betas=betas, weight_decay=gamma, eps=eps) elif opt == 'sgd': return SGD(params, lr=lr, momentum=momentum, weight_decay=gamma, nesterov=True) elif opt == 'adamax': return Adamax(params, lr=lr, betas=betas, weight_decay=gamma, eps=eps) else: raise ValueError('Unknown optimization algorithm: %s' % opt) lr = learning_rate optim = generate_optimizer(opt, lr, network.parameters()) opt_info = 'opt: %s, ' % opt if opt == 'adam': opt_info += 'betas=%s, eps=%.1e' % (betas, eps) elif opt == 'sgd': opt_info += 'momentum=%.2f' % momentum elif opt == 'adamax': opt_info += 'betas=%s, eps=%.1e' % (betas, eps) word_status = 'frozen' if freeze else 'fine tune' char_status = 'enabled' if use_char else 'disabled' pos_status = 'enabled' if use_pos else 'disabled' logger.info( "Embedding dim: word=%d (%s), char=%d (%s), pos=%d (%s)" % (word_dim, word_status, char_dim, char_status, pos_dim, pos_status)) logger.info("CNN: filter=%d, kernel=%d" % (num_filters, window)) logger.info( "RNN: %s, num_layer=%d, hidden=%d, arc_space=%d, type_space=%d" % (mode, num_layers, hidden_size, arc_space, type_space)) logger.info( "train: obj: %s, l2: %f, (#data: %d, batch: %d, clip: %.2f, unk replace: %.2f)" % (obj, gamma, num_data, batch_size, clip, unk_replace)) logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" % (p_in, p_out, p_rnn)) logger.info("decoding algorithm: %s" % decoding) logger.info(opt_info) num_batches = num_data / batch_size + 1 dev_ucorrect = 0.0 dev_lcorrect = 0.0 dev_ucomlpete_match = 0.0 dev_lcomplete_match = 0.0 dev_ucorrect_nopunc = 0.0 dev_lcorrect_nopunc = 0.0 dev_ucomlpete_match_nopunc = 0.0 dev_lcomplete_match_nopunc = 0.0 dev_root_correct = 0.0 best_epoch = 0 test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete_match = 0.0 test_lcomplete_match = 0.0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_match_nopunc = 0.0 test_lcomplete_match_nopunc = 0.0 test_root_correct = 0.0 test_total = 0 test_total_nopunc = 0 test_total_inst = 0 test_total_root = 0 if decoding == 'greedy': decode = network.decode elif decoding == 'mst': decode = network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) patient = 0 decay = 0 max_decay = args.max_decay double_schedule_decay = args.double_schedule_decay # lrate schedule step_num = 0 use_warmup_schedule = args.use_warmup_schedule warmup_factor = (lr + 0.) / num_batches if use_warmup_schedule: logger.info("Use warmup lrate for the first epoch, from 0 up to %s." % (lr, )) # for epoch in range(1, num_epochs + 1): print( 'Epoch %d (%s, optim: %s, learning rate=%.6f, eps=%.1e, decay rate=%.2f (schedule=%d, patient=%d, decay=%d)): ' % (epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay)) train_err = 0. train_err_arc = 0. train_err_type = 0. train_total = 0. start_time = time.time() num_back = 0 network.train() for batch in range(1, num_batches + 1): # lrate schedule (before each step) step_num += 1 if use_warmup_schedule and epoch <= 1: cur_lrate = warmup_factor * step_num # set lr for param_group in optim.param_groups: param_group['lr'] = cur_lrate # word, char, pos, heads, types, masks, lengths = conllx_data.get_batch_variable( data_train, batch_size, unk_replace=unk_replace) optim.zero_grad() loss_arc, loss_type = network.loss(word, char, pos, heads, types, mask=masks, length=lengths) loss = loss_arc + loss_type loss.backward() clip_grad_norm(network.parameters(), clip) optim.step() num_inst = word.size( 0) if obj == 'crf' else masks.data.sum() - word.size(0) train_err += loss.data[0] * num_inst train_err_arc += loss_arc.data[0] * num_inst train_err_type += loss_type.data[0] * num_inst train_total += num_inst time_ave = (time.time() - start_time) / batch time_left = (num_batches - batch) * time_ave # update log if batch % 10 == 0: sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) log_info = 'train: %d/%d loss: %.4f, arc: %.4f, type: %.4f, time left: %.2fs' % ( batch, num_batches, train_err / train_total, train_err_arc / train_total, train_err_type / train_total, time_left) sys.stdout.write(log_info) sys.stdout.flush() num_back = len(log_info) sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) print( 'train: %d loss: %.4f, arc: %.4f, type: %.4f, time: %.2fs' % (num_batches, train_err / train_total, train_err_arc / train_total, train_err_type / train_total, time.time() - start_time)) ################################################################################################ if epoch % args.check_dev != 0: continue # evaluate performance on dev data network.eval() pred_filename = 'tmp/%spred_dev%d' % (str(uid), epoch) pred_writer.start(pred_filename) gold_filename = 'tmp/%sgold_dev%d' % (str(uid), epoch) gold_writer.start(gold_filename) dev_ucorr = 0.0 dev_lcorr = 0.0 dev_total = 0 dev_ucomlpete = 0.0 dev_lcomplete = 0.0 dev_ucorr_nopunc = 0.0 dev_lcorr_nopunc = 0.0 dev_total_nopunc = 0 dev_ucomlpete_nopunc = 0.0 dev_lcomplete_nopunc = 0.0 dev_root_corr = 0.0 dev_total_root = 0.0 dev_total_inst = 0.0 for batch in conllx_data.iterate_batch_variable(data_dev, batch_size): word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode( word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.data.cpu().numpy() types = types.data.cpu().numpy() pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True) gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True) stats, stats_nopunc, stats_root, num_inst = parser.eval( word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root dev_ucorr += ucorr dev_lcorr += lcorr dev_total += total dev_ucomlpete += ucm dev_lcomplete += lcm dev_ucorr_nopunc += ucorr_nopunc dev_lcorr_nopunc += lcorr_nopunc dev_total_nopunc += total_nopunc dev_ucomlpete_nopunc += ucm_nopunc dev_lcomplete_nopunc += lcm_nopunc dev_root_corr += corr_root dev_total_root += total_root dev_total_inst += num_inst pred_writer.close() gold_writer.close() print( 'W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 / dev_total, dev_lcorr * 100 / dev_total, dev_ucomlpete * 100 / dev_total_inst, dev_lcomplete * 100 / dev_total_inst)) print( 'Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc, dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc * 100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 / dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst)) print('Root: corr: %d, total: %d, acc: %.2f%%' % (dev_root_corr, dev_total_root, dev_root_corr * 100 / dev_total_root)) if dev_lcorrect_nopunc < dev_lcorr_nopunc or ( dev_lcorrect_nopunc == dev_lcorr_nopunc and dev_ucorrect_nopunc < dev_ucorr_nopunc): dev_ucorrect_nopunc = dev_ucorr_nopunc dev_lcorrect_nopunc = dev_lcorr_nopunc dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc dev_lcomplete_match_nopunc = dev_lcomplete_nopunc dev_ucorrect = dev_ucorr dev_lcorrect = dev_lcorr dev_ucomlpete_match = dev_ucomlpete dev_lcomplete_match = dev_lcomplete dev_root_correct = dev_root_corr best_epoch = epoch patient = 0 # torch.save(network, model_name) torch.save(network.state_dict(), model_name) pred_filename = 'tmp/%spred_test%d' % (str(uid), epoch) pred_writer.start(pred_filename) gold_filename = 'tmp/%sgold_test%d' % (str(uid), epoch) gold_writer.start(gold_filename) test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete_match = 0.0 test_lcomplete_match = 0.0 test_total = 0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_match_nopunc = 0.0 test_lcomplete_match_nopunc = 0.0 test_total_nopunc = 0 test_total_inst = 0 test_root_correct = 0.0 test_total_root = 0 for batch in conllx_data.iterate_batch_variable( data_test, batch_size): word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode( word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.data.cpu().numpy() types = types.data.cpu().numpy() pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True) gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True) stats, stats_nopunc, stats_root, num_inst = parser.eval( word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root test_ucorrect += ucorr test_lcorrect += lcorr test_total += total test_ucomlpete_match += ucm test_lcomplete_match += lcm test_ucorrect_nopunc += ucorr_nopunc test_lcorrect_nopunc += lcorr_nopunc test_total_nopunc += total_nopunc test_ucomlpete_match_nopunc += ucm_nopunc test_lcomplete_match_nopunc += lcm_nopunc test_root_correct += corr_root test_total_root += total_root test_total_inst += num_inst pred_writer.close() gold_writer.close() else: if dev_ucorr_nopunc * 100 / dev_total_nopunc < dev_ucorrect_nopunc * 100 / dev_total_nopunc - 5 or patient >= schedule: # network = torch.load(model_name) network.load_state_dict(torch.load(model_name)) lr = lr * decay_rate optim = generate_optimizer(opt, lr, network.parameters()) if decoding == 'greedy': decode = network.decode elif decoding == 'mst': decode = network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) patient = 0 decay += 1 if decay % double_schedule_decay == 0: schedule *= 2 else: patient += 1 print( '----------------------------------------------------------------------------------------------------------------------------' ) print( 'best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (dev_ucorrect, dev_lcorrect, dev_total, dev_ucorrect * 100 / dev_total, dev_lcorrect * 100 / dev_total, dev_ucomlpete_match * 100 / dev_total_inst, dev_lcomplete_match * 100 / dev_total_inst, best_epoch)) print( 'best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc, dev_ucorrect_nopunc * 100 / dev_total_nopunc, dev_lcorrect_nopunc * 100 / dev_total_nopunc, dev_ucomlpete_match_nopunc * 100 / dev_total_inst, dev_lcomplete_match_nopunc * 100 / dev_total_inst, best_epoch)) print('best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % (dev_root_correct, dev_total_root, dev_root_correct * 100 / dev_total_root, best_epoch)) print( '----------------------------------------------------------------------------------------------------------------------------' ) print( 'best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 / test_total, test_lcorrect * 100 / test_total, test_ucomlpete_match * 100 / test_total_inst, test_lcomplete_match * 100 / test_total_inst, best_epoch)) print( 'best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc, test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc, test_ucomlpete_match_nopunc * 100 / test_total_inst, test_lcomplete_match_nopunc * 100 / test_total_inst, best_epoch)) print('best test Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % (test_root_correct, test_total_root, test_root_correct * 100 / test_total_root, best_epoch)) print( '============================================================================================================================' ) if decay == max_decay: break
def biaffine(model_path, model_name, test_path, punct_set, use_gpu, logger, args): alphabet_path = os.path.join(model_path, 'alphabets/') model_name = os.path.join(model_path, model_name) word_alphabet, char_alphabet, pos_alphabet, \ type_alphabet = conllx_data.create_alphabets(alphabet_path, None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) decoding = args.decode logger.info('use gpu: %s, decoding: %s' % (use_gpu, decoding)) data_test = conllx_data.read_data_to_tensor(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=True, symbolic_root=True) pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) logger.info('model: %s' % model_name) def load_model_arguments_from_json(): arguments = json.load(open(arg_path, 'r')) return arguments['args'], arguments['kwargs'] arg_path = model_name + '.arg.json' args, kwargs = load_model_arguments_from_json() network = BiRecurrentConvBiAffine(*args, **kwargs) network.load_state_dict(torch.load(model_name)) if use_gpu: network.cuda() else: network.cpu() network.eval() test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete_match = 0.0 test_lcomplete_match = 0.0 test_total = 0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_match_nopunc = 0.0 test_lcomplete_match_nopunc = 0.0 test_total_nopunc = 0 test_total_inst = 0 test_root_correct = 0.0 test_total_root = 0 if decoding == 'greedy': decode = network.decode elif decoding == 'mst': decode = network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) pred_writer.start('tmp/analyze_pred_%s' % str(uid)) gold_writer.start('tmp/analyze_gold_%s' % str(uid)) sent = 0 start_time = time.time() for batch in conllx_data.iterate_batch_tensor(data_test, 1): sys.stdout.write('%d, ' % sent) sys.stdout.flush() sent += 1 word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode( word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.data.cpu().numpy() types = types.data.cpu().numpy() pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True) gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True) stats, stats_nopunc, stats_root, num_inst = parser.eval( word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root test_ucorrect += ucorr test_lcorrect += lcorr test_total += total test_ucomlpete_match += ucm test_lcomplete_match += lcm test_ucorrect_nopunc += ucorr_nopunc test_lcorrect_nopunc += lcorr_nopunc test_total_nopunc += total_nopunc test_ucomlpete_match_nopunc += ucm_nopunc test_lcomplete_match_nopunc += lcm_nopunc test_root_correct += corr_root test_total_root += total_root test_total_inst += num_inst pred_writer.close() gold_writer.close() print('\ntime: %.2fs' % (time.time() - start_time)) print( 'test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 / test_total, test_lcorrect * 100 / test_total, test_ucomlpete_match * 100 / test_total_inst, test_lcomplete_match * 100 / test_total_inst)) print( 'test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc, test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc, test_ucomlpete_match_nopunc * 100 / test_total_inst, test_lcomplete_match_nopunc * 100 / test_total_inst)) print('test Root: corr: %d, total: %d, acc: %.2f%%' % (test_root_correct, test_total_root, test_root_correct * 100 / test_total_root))
def run_biaffine(model_path, model_name, test_path, punct_set, use_gpu, logger, args): alphabet_path = os.path.join(model_path, 'alphabets/') model_name = os.path.join(model_path, model_name) word_alphabet, char_alphabet, pos_alphabet, \ type_alphabet = conllx_data.create_alphabets( alphabet_path, None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None ) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) decoding = args.decode logger.info('use gpu: %s, decoding: %s' % (use_gpu, decoding)) device = torch.device('cuda') if use_gpu else torch.device('cpu') data_test = aida_data.read_data_to_tensor(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, symbolic_root=True, device=device) pred_writer = AIDAWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) logger.info('model: %s' % model_name) def load_model_arguments_from_json(): arguments = json.load(open(arg_path, 'r')) return arguments['args'], arguments['kwargs'] arg_path = model_name + '.arg.json' model_args, model_kwargs = load_model_arguments_from_json() network = BiRecurrentConvBiAffine(*model_args, **model_kwargs) if torch.cuda.is_available(): map_location = lambda storage, loc: storage.cuda() else: map_location = 'cpu' network.load_state_dict(torch.load(model_name, map_location=map_location)) if use_gpu: network.cuda() else: network.cpu() network.eval() if decoding == 'greedy': decode = network.decode elif decoding == 'mst': decode = network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) pred_writer.start(args.output_path) sent = 0 start_time = time.time() with torch.no_grad(): for batch in aida_data.iterate_batch_tensor(data_test, 1): sys.stdout.write('Processing sentence: %d\n' % sent) sys.stdout.flush() sent += 1 word, char, pos, _, _, masks, lengths, segment_ids_words = batch heads_pred, types_pred = decode( word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() segment_ids, segment_words = zip(*segment_ids_words) pred_writer.write(segment_ids, segment_words, word, pos, heads_pred, types_pred, lengths, symbolic_root=True) pred_writer.close() print('\ntime: %.2fs' % (time.time() - start_time))
def main(): args_parser = argparse.ArgumentParser(description='Tuning with graph-based parsing') args_parser.add_argument('--mode', choices=['RNN', 'LSTM', 'GRU', 'FastLSTM'], help='architecture of rnn', required=True) args_parser.add_argument('--num_epochs', type=int, default=200, help='Number of training epochs') args_parser.add_argument('--batch_size', type=int, default=64, help='Number of sentences in each batch') args_parser.add_argument('--hidden_size', type=int, default=256, help='Number of hidden units in RNN') args_parser.add_argument('--arc_space', type=int, default=128, help='Dimension of tag space') args_parser.add_argument('--type_space', type=int, default=128, help='Dimension of tag space') args_parser.add_argument('--num_layers', type=int, default=1, help='Number of layers of RNN') args_parser.add_argument('--num_filters', type=int, default=50, help='Number of filters in CNN') args_parser.add_argument('--pos', action='store_true', help='use part-of-speech embedding.') args_parser.add_argument('--pos_dim', type=int, default=50, help='Dimension of POS embeddings') args_parser.add_argument('--char_dim', type=int, default=50, help='Dimension of Character embeddings') args_parser.add_argument('--objective', choices=['cross_entropy', 'crf'], default='cross_entropy', help='objective function of training procedure.') args_parser.add_argument('--decode', choices=['mst', 'greedy'], help='decoding algorithm', required=True) args_parser.add_argument('--learning_rate', type=float, default=0.01, help='Learning rate') args_parser.add_argument('--decay_rate', type=float, default=0.05, help='Decay rate of learning rate') args_parser.add_argument('--gamma', type=float, default=0.0, help='weight for regularization') args_parser.add_argument('--p_rnn', nargs=2, type=float, required=True, help='dropout rate for RNN') args_parser.add_argument('--p_in', type=float, default=0.33, help='dropout rate for input embeddings') args_parser.add_argument('--p_out', type=float, default=0.33, help='dropout rate for output layer') args_parser.add_argument('--schedule', type=int, help='schedule for learning rate decay') args_parser.add_argument('--unk_replace', type=float, default=0., help='The rate to replace a singleton word with UNK') args_parser.add_argument('--punctuation', nargs='+', type=str, help='List of punctuations') args_parser.add_argument('--word_embedding', choices=['glove', 'senna', 'sskip', 'polyglot'], help='Embedding for words', required=True) args_parser.add_argument('--word_path', help='path for word embedding dict') args_parser.add_argument('--char_embedding', choices=['random', 'polyglot'], help='Embedding for characters', required=True) args_parser.add_argument('--char_path', help='path for character embedding dict') args_parser.add_argument('--train') # "data/POS-penn/wsj/split1/wsj1.train.original" args_parser.add_argument('--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original" args_parser.add_argument('--test') # "data/POS-penn/wsj/split1/wsj1.test.original" args_parser.add_argument('--model_path', help='path for saving model file.', required=True) args = args_parser.parse_args() print("*** Model UID: %s ***" % uid) logger = get_logger("GraphParser") mode = args.mode obj = args.objective decoding = args.decode train_path = args.train dev_path = args.dev test_path = args.test model_path = args.model_path num_epochs = args.num_epochs batch_size = args.batch_size hidden_size = args.hidden_size arc_space = args.arc_space type_space = args.type_space num_layers = args.num_layers num_filters = args.num_filters learning_rate = args.learning_rate momentum = 0.9 betas = (0.9, 0.9) decay_rate = args.decay_rate gamma = args.gamma schedule = args.schedule p_rnn = tuple(args.p_rnn) p_in = args.p_in p_out = args.p_out unk_replace = args.unk_replace punctuation = args.punctuation word_embedding = args.word_embedding word_path = args.word_path char_embedding = args.char_embedding char_path = args.char_path use_pos = args.pos pos_dim = args.pos_dim word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path) char_dict = None char_dim = args.char_dim if char_embedding != 'random': char_dict, char_dim = utils.load_embedding_dict(char_embedding, char_path) logger.info("Creating Alphabets") alphabet_path = os.path.join(model_path, 'alphabets/') word_alphabet, char_alphabet, pos_alphabet, type_alphabet = conllx_data.create_alphabets(alphabet_path, train_path, data_paths=[dev_path, test_path], max_vocabulary_size=50000, embedd_dict=word_dict) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) logger.info("Reading Data") use_gpu = torch.cuda.is_available() data_train = conllx_data.read_data_to_variable(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, symbolic_root=True) # data_train = conllx_data.read_data(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet) # num_data = sum([len(bucket) for bucket in data_train]) num_data = sum(data_train[1]) data_dev = conllx_data.read_data_to_variable(dev_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=True, symbolic_root=True) data_test = conllx_data.read_data_to_variable(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=True, symbolic_root=True) punct_set = None if punctuation is not None: punct_set = set(punctuation) logger.info("punctuations(%d): %s" % (len(punct_set), ' '.join(punct_set))) def construct_word_embedding_table(): scale = np.sqrt(3.0 / word_dim) table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.random.uniform(-scale, scale, [1, word_dim]).astype(np.float32) oov = 0 for word, index in word_alphabet.items(): if word in word_dict: embedding = word_dict[word] elif word.lower() in word_dict: embedding = word_dict[word.lower()] else: embedding = np.random.uniform(-scale, scale, [1, word_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('word OOV: %d' % oov) return torch.from_numpy(table) def construct_char_embedding_table(): if char_dict is None: return None scale = np.sqrt(3.0 / char_dim) table = np.empty([num_chars, char_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.random.uniform(-scale, scale, [1, char_dim]).astype(np.float32) oov = 0 for char, index, in char_alphabet.items(): if char in char_dict: embedding = char_dict[char] else: embedding = np.random.uniform(-scale, scale, [1, char_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('character OOV: %d' % oov) return torch.from_numpy(table) word_table = construct_word_embedding_table() char_table = construct_char_embedding_table() window = 3 if obj == 'cross_entropy': network = BiRecurrentConvBiAffine(word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, num_filters, window, mode, hidden_size, num_layers, num_types, arc_space, type_space, embedd_word=word_table, embedd_char=char_table, p_in=p_in, p_out=p_out, p_rnn=p_rnn, biaffine=True, pos=use_pos) elif obj == 'crf': raise NotImplementedError else: raise RuntimeError('Unknown objective: %s' % obj) if use_gpu: network.cuda() pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) adam_epochs = 50 adam_rate = 0.001 if adam_epochs > 0: lr = adam_rate opt = 'adam' optim = Adam(network.parameters(), lr=adam_rate, betas=betas, weight_decay=gamma) else: opt = 'sgd' lr = learning_rate optim = SGD(network.parameters(), lr=lr, momentum=momentum, weight_decay=gamma, nesterov=True) logger.info("Embedding dim: word=%d, char=%d, pos=%d (%s)" % (word_dim, char_dim, pos_dim, use_pos)) logger.info("Network: %s, num_layer=%d, hidden=%d, filter=%d, arc_space=%d, type_space=%d" % ( mode, num_layers, hidden_size, num_filters, arc_space, type_space)) logger.info("train: obj: %s, l2: %f, (#data: %d, batch: %d, dropout(in, out, rnn): (%.2f, %.2f, %s), unk replace: %.2f)" % ( obj, gamma, num_data, batch_size, p_in, p_out, p_rnn, unk_replace)) logger.info("decoding algorithm: %s" % decoding) num_batches = num_data / batch_size + 1 dev_ucorrect = 0.0 dev_lcorrect = 0.0 dev_ucomlpete_match = 0.0 dev_lcomplete_match = 0.0 dev_ucorrect_nopunc = 0.0 dev_lcorrect_nopunc = 0.0 dev_ucomlpete_match_nopunc = 0.0 dev_lcomplete_match_nopunc = 0.0 dev_root_correct = 0.0 best_epoch = 0 test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete_match = 0.0 test_lcomplete_match = 0.0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_match_nopunc = 0.0 test_lcomplete_match_nopunc = 0.0 test_root_correct = 0.0 test_total = 0 test_total_nopunc = 0 test_total_inst = 0 test_total_root = 0 if decoding == 'greedy': decode = network.decode elif decoding == 'mst': decode = network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) for epoch in range(1, num_epochs + 1): print('Epoch %d (%s, optim: %s, learning rate=%.4f, decay rate=%.4f (schedule=%d)): ' % ( epoch, mode, opt, lr, decay_rate, schedule)) train_err = 0. train_err_arc = 0. train_err_type = 0. train_total = 0. start_time = time.time() num_back = 0 network.train() for batch in range(1, num_batches + 1): word, char, pos, heads, types, masks, lengths = conllx_data.get_batch_variable(data_train, batch_size, unk_replace=unk_replace) optim.zero_grad() loss_arc, loss_type = network.loss(word, char, pos, heads, types, mask=masks, length=lengths) loss = loss_arc + loss_type loss.backward() optim.step() num_inst = word.size(0) if obj == 'crf' else masks.data.sum() - word.size(0) train_err += loss.data[0] * num_inst train_err_arc += loss_arc.data[0] * num_inst train_err_type += loss_type.data[0] * num_inst train_total += num_inst time_ave = (time.time() - start_time) / batch time_left = (num_batches - batch) * time_ave # update log if batch % 10 == 0: sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) log_info = 'train: %d/%d loss: %.4f, arc: %.4f, type: %.4f, time left (estimated): %.2fs' % ( batch, num_batches, train_err / train_total, train_err_arc / train_total, train_err_type / train_total, time_left) sys.stdout.write(log_info) sys.stdout.flush() num_back = len(log_info) sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) print('train: %d loss: %.4f, arc: %.4f, type: %.4f, time: %.2fs' % ( num_batches, train_err / train_total, train_err_arc / train_total, train_err_type / train_total, time.time() - start_time)) # evaluate performance on dev data network.eval() pred_filename = 'tmp/%spred_dev%d' % (str(uid), epoch) pred_writer.start(pred_filename) gold_filename = 'tmp/%sgold_dev%d' % (str(uid), epoch) gold_writer.start(gold_filename) print('[%s] Epoch %d complete' % (time.strftime("%Y-%m-%d %H:%M:%S"), epoch)) dev_ucorr = 0.0 dev_lcorr = 0.0 dev_total = 0 dev_ucomlpete = 0.0 dev_lcomplete = 0.0 dev_ucorr_nopunc = 0.0 dev_lcorr_nopunc = 0.0 dev_total_nopunc = 0 dev_ucomlpete_nopunc = 0.0 dev_lcomplete_nopunc = 0.0 dev_root_corr = 0.0 dev_total_root = 0.0 dev_total_inst = 0.0 for batch in conllx_data.iterate_batch_variable(data_dev, batch_size): word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.data.cpu().numpy() types = types.data.cpu().numpy() pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True) gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True) stats, stats_nopunc, stats_root, num_inst = parser.eval(word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root dev_ucorr += ucorr dev_lcorr += lcorr dev_total += total dev_ucomlpete += ucm dev_lcomplete += lcm dev_ucorr_nopunc += ucorr_nopunc dev_lcorr_nopunc += lcorr_nopunc dev_total_nopunc += total_nopunc dev_ucomlpete_nopunc += ucm_nopunc dev_lcomplete_nopunc += lcm_nopunc dev_root_corr += corr_root dev_total_root += total_root dev_total_inst += num_inst pred_writer.close() gold_writer.close() print('W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % ( dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 / dev_total, dev_lcorr * 100 / dev_total, dev_ucomlpete * 100 / dev_total_inst, dev_lcomplete * 100 / dev_total_inst)) print('Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % ( dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc, dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc * 100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 / dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst)) print('Root: corr: %d, total: %d, acc: %.2f%%' %( dev_root_corr, dev_total_root, dev_root_corr * 100 / dev_total_root)) if dev_ucorrect_nopunc <= dev_ucorr_nopunc: dev_ucorrect_nopunc = dev_ucorr_nopunc dev_lcorrect_nopunc = dev_lcorr_nopunc dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc dev_lcomplete_match_nopunc = dev_lcomplete_nopunc dev_ucorrect = dev_ucorr dev_lcorrect = dev_lcorr dev_ucomlpete_match = dev_ucomlpete dev_lcomplete_match = dev_lcomplete dev_root_correct = dev_root_corr best_epoch = epoch pred_filename = 'tmp/%spred_test%d' % (str(uid), epoch) pred_writer.start(pred_filename) gold_filename = 'tmp/%sgold_test%d' % (str(uid), epoch) gold_writer.start(gold_filename) test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete_match = 0.0 test_lcomplete_match = 0.0 test_total = 0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_match_nopunc = 0.0 test_lcomplete_match_nopunc = 0.0 test_total_nopunc = 0 test_total_inst = 0 test_root_correct = 0.0 test_total_root = 0 for batch in conllx_data.iterate_batch_variable(data_test, batch_size): word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.data.cpu().numpy() types = types.data.cpu().numpy() pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True) gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True) stats, stats_nopunc, stats_root, num_inst = parser.eval(word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root test_ucorrect += ucorr test_lcorrect += lcorr test_total += total test_ucomlpete_match += ucm test_lcomplete_match += lcm test_ucorrect_nopunc += ucorr_nopunc test_lcorrect_nopunc += lcorr_nopunc test_total_nopunc += total_nopunc test_ucomlpete_match_nopunc += ucm_nopunc test_lcomplete_match_nopunc += lcm_nopunc test_root_correct += corr_root test_total_root += total_root test_total_inst += num_inst pred_writer.close() gold_writer.close() print('----------------------------------------------------------------------------------------------------------------------------') print('best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( dev_ucorrect, dev_lcorrect, dev_total, dev_ucorrect * 100 / dev_total, dev_lcorrect * 100 / dev_total, dev_ucomlpete_match * 100 / dev_total_inst, dev_lcomplete_match * 100 / dev_total_inst, best_epoch)) print('best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc, dev_ucorrect_nopunc * 100 / dev_total_nopunc, dev_lcorrect_nopunc * 100 / dev_total_nopunc, dev_ucomlpete_match_nopunc * 100 / dev_total_inst, dev_lcomplete_match_nopunc * 100 / dev_total_inst, best_epoch)) print('best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % ( dev_root_correct, dev_total_root, dev_root_correct * 100 / dev_total_root, best_epoch)) print('----------------------------------------------------------------------------------------------------------------------------') print('best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 / test_total, test_lcorrect * 100 / test_total, test_ucomlpete_match * 100 / test_total_inst, test_lcomplete_match * 100 / test_total_inst, best_epoch)) print('best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc, test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc, test_ucomlpete_match_nopunc * 100 / test_total_inst, test_lcomplete_match_nopunc * 100 / test_total_inst, best_epoch)) print('best test Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % ( test_root_correct, test_total_root, test_root_correct * 100 / test_total_root, best_epoch)) print('============================================================================================================================') if epoch % schedule == 0: # lr = lr * decay_rate if epoch < adam_epochs: opt = 'adam' lr = adam_rate / (1.0 + epoch * decay_rate) optim = Adam(network.parameters(), lr=lr, betas=betas, weight_decay=gamma) else: opt = 'sgd' lr = learning_rate / (1.0 + (epoch - adam_epochs) * decay_rate) optim = SGD(network.parameters(), lr=lr, momentum=momentum, weight_decay=gamma, nesterov=True)
def main(): args_parser = argparse.ArgumentParser(description='Tuning with graph-based parsing') args_parser.add_argument('--schedule', type=int, help='schedule for learning rate decay') args_parser.add_argument('--unk_replace', type=float, default=0., help='The rate to replace a singleton word with UNK') args_parser.add_argument('--freeze', action='store_true', help='frozen the word embedding (disable fine-tuning).') args = args_parser.parse_args() logger = get_logger("GraphParser") mode = "FastLSTM" #fast lstm here obj = "cross_entropy" decoding = "mst" #mst decode here train_path = "data/train.stanford.conll" dev_path = "data/dev.stanford.conll" test_path = "data/test.stanford.conll" model_path = "models/parsing/biaffine/" model_name = 'network.pt' num_epochs = 80 batch_size = 32 hidden_size = 512 arc_space = 512 type_space = 128 num_layers = 10 num_filters = 1 learning_rate = 0.001 opt = "adam" #default adam momentum = 0.9 betas = (0.9, 0.9) eps = 1e-4 decay_rate = 0.75 clip = 5 #what is clip gamma = 0 schedule = 10 #?What is this? p_rnn = (0.05,0.05) p_in = 0.33 p_out = 0.33 unk_replace = args.unk_replace# ?what is this? punctuation = ['.','``', "''", ':', ','] freeze = args.freeze word_embedding = 'glove' word_path = "data/glove.6B.100d.txt" use_char = False char_embedding = None #char_path = args.char_path use_pos = True pos_dim = 100 word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path) char_dict = None char_dim = 0 logger.info("Creating Alphabets") alphabet_path = os.path.join(model_path, 'alphabets/') model_name = os.path.join(model_path, model_name) word_alphabet, char_alphabet, pos_alphabet, type_alphabet = conllx_data.create_alphabets(alphabet_path, train_path, data_paths=[dev_path, test_path], max_vocabulary_size=50000, embedd_dict=word_dict) num_words = word_alphabet.size() num_chars = char_alphabet.size() num_pos = pos_alphabet.size() num_types = type_alphabet.size() #print(word_alphabet.instance2index) logger.info("Word Alphabet Size: %d" % num_words) logger.info("Character Alphabet Size: %d" % num_chars) logger.info("POS Alphabet Size: %d" % num_pos) logger.info("Type Alphabet Size: %d" % num_types) logger.info("Reading Data") use_gpu = torch.cuda.is_available() print(use_gpu) data_train = conllx_data.read_data_to_variable(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, symbolic_root=True) # data_train = conllx_data.read_data(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet) # num_data = sum([len(bucket) for bucket in data_train]) num_data = sum(data_train[1]) """ print("bucket_size") print(data_train[1]) print("___________________________________data_train") print(data_train[0]) """ data_dev = conllx_data.read_data_to_variable(dev_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=True, symbolic_root=True) data_test = conllx_data.read_data_to_variable(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=True, symbolic_root=True) punct_set = None if punctuation is not None: punct_set = set(punctuation) logger.info("punctuations(%d): %s" % (len(punct_set), ' '.join(punct_set))) def construct_word_embedding_table(): scale = np.sqrt(3.0 / word_dim) table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32) table[conllx_data.UNK_ID, :] = np.zeros([1, word_dim]).astype(np.float32) if freeze else np.random.uniform(-scale, scale, [1, word_dim]).astype(np.float32) oov = 0 for word, index in word_alphabet.items(): if word in word_dict: embedding = word_dict[word] elif word.lower() in word_dict: embedding = word_dict[word.lower()] else: embedding = np.zeros([1, word_dim]).astype(np.float32) if freeze else np.random.uniform(-scale, scale, [1, word_dim]).astype(np.float32) oov += 1 table[index, :] = embedding print('word OOV: %d' % oov) return torch.from_numpy(table) word_table = construct_word_embedding_table() window = 3 if obj == 'cross_entropy': network = BiRecurrentConvBiAffine(word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, num_filters, window, mode, hidden_size, num_layers, num_types, arc_space, type_space, embedd_word=word_table, embedd_char=None, p_in=p_in, p_out=p_out, p_rnn=p_rnn, biaffine=True, pos=use_pos, char=use_char) def save_args(): arg_path = model_name + '.arg.json' arguments = [word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, num_filters, window, mode, hidden_size, num_layers, num_types, arc_space, type_space] kwargs = {'p_in': p_in, 'p_out': p_out, 'p_rnn': p_rnn, 'biaffine': True, 'pos': use_pos, 'char': use_char} json.dump({'args': arguments, 'kwargs': kwargs}, open(arg_path, 'w'), indent=4) if freeze: network.word_embedd.freeze() if use_gpu: network.cuda() save_args() #pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) #gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet) ##print parameters: print("number of parameters") num_param = sum([param.nelement() for param in network.parameters()]) print(num_param) def generate_optimizer(opt, lr, params): params = filter(lambda param: param.requires_grad, params) if opt == 'adam': return Adam(params, lr=lr, betas=betas, weight_decay=gamma, eps=eps) lr = learning_rate optim = generate_optimizer(opt, lr, network.parameters()) opt_info = 'opt: %s, ' % opt if opt == 'adam': opt_info += 'betas=%s, eps=%.1e' % (betas, eps) word_status = 'frozen' if freeze else 'fine tune' char_status = 'enabled' if use_char else 'disabled' pos_status = 'enabled' if use_pos else 'disabled' logger.info("Embedding dim: word=%d (%s), char=%d (%s), pos=%d (%s)" % (word_dim, word_status, char_dim, char_status, pos_dim, pos_status)) logger.info("CNN: filter=%d, kernel=%d" % (num_filters, window)) logger.info("RNN: %s, num_layer=%d, hidden=%d, arc_space=%d, type_space=%d" % (mode, num_layers, hidden_size, arc_space, type_space)) logger.info("train: obj: %s, l2: %f, (#data: %d, batch: %d, clip: %.2f, unk replace: %.2f)" % (obj, gamma, num_data, batch_size, clip, unk_replace)) logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" % (p_in, p_out, p_rnn)) logger.info("decoding algorithm: %s" % decoding) logger.info(opt_info) #logger.info("Attention") num_batches = num_data / batch_size + 1 dev_ucorrect = 0.0 dev_lcorrect = 0.0 dev_ucomlpete_match = 0.0 dev_lcomplete_match = 0.0 dev_ucorrect_nopunc = 0.0 dev_lcorrect_nopunc = 0.0 dev_ucomlpete_match_nopunc = 0.0 dev_lcomplete_match_nopunc = 0.0 dev_root_correct = 0.0 best_epoch = 0 test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete_match = 0.0 test_lcomplete_match = 0.0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_match_nopunc = 0.0 test_lcomplete_match_nopunc = 0.0 test_root_correct = 0.0 test_total = 0 test_total_nopunc = 0 test_total_inst = 0 test_total_root = 0 if decoding == 'greedy': decode = network.decode elif decoding == 'mst': decode = network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) patient = 0 decay = 0 max_decay = 9 double_schedule_decay = 5 f = open("testout.csv", "wt") writer = csv.writer(f) writer.writerow(('train', 'dev')) for epoch in range(1, num_epochs + 1): print(epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay) print('Epoch %d (%s, optim: %s, learning rate=%.6f, eps=%.1e, decay rate=%.2f (schedule=%d, patient=%d, decay=%d)): ' % (epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay)) train_err = 0. train_err_arc = 0. train_err_type = 0. train_total = 0. start_time = time.time() num_back = 0 network.train() for batch in range(1, num_batches + 1): word, char, pos, heads, types, masks, lengths = conllx_data.get_batch_variable(data_train, batch_size, unk_replace=unk_replace) optim.zero_grad() loss_arc, loss_type = network.loss(word, char, pos, heads, types, mask=masks, length=lengths) loss = loss_arc + loss_type loss.backward() clip_grad_norm(network.parameters(), clip) optim.step() num_inst = word.size(0) if obj == 'crf' else masks.data.sum() - word.size(0) train_err += loss.data[0] * num_inst train_err_arc += loss_arc.data[0] * num_inst train_err_type += loss_type.data[0] * num_inst train_total += num_inst #bp() time_ave = (time.time() - start_time) / batch time_left = (num_batches - batch) * time_ave # update log if batch % 10 == 0: sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) log_info = 'train: %d/%d loss: %.4f, arc: %.4f, type: %.4f, time left: %.2fs' % (batch, num_batches, train_err / train_total, train_err_arc / train_total, train_err_type / train_total, time_left) sys.stdout.write(log_info) sys.stdout.flush() num_back = len(log_info) sys.stdout.write("\b" * num_back) sys.stdout.write(" " * num_back) sys.stdout.write("\b" * num_back) print('train: %d loss: %.4f, arc: %.4f, type: %.4f, time: %.2fs' % (num_batches, train_err / train_total, train_err_arc / train_total, train_err_type / train_total, time.time() - start_time)) # evaluate performance on dev data network.eval() dev_ucorr = 0.0 dev_lcorr = 0.0 dev_total = 0 dev_ucomlpete = 0.0 dev_lcomplete = 0.0 dev_ucorr_nopunc = 0.0 dev_lcorr_nopunc = 0.0 dev_total_nopunc = 0 dev_ucomlpete_nopunc = 0.0 dev_lcomplete_nopunc = 0.0 dev_root_corr = 0.0 dev_total_root = 0.0 dev_total_inst = 0.0 t_ucorr = 0.0 t_lcorr = 0.0 t_total = 0 t_ucomlpete = 0.0 t_lcomplete = 0.0 t_ucorr_nopunc = 0.0 t_lcorr_nopunc = 0.0 t_total_nopunc = 0 t_ucomlpete_nopunc = 0.0 t_lcomplete_nopunc = 0.0 t_root_corr = 0.0 t_total_root = 0.0 t_total_inst = 0.0 list_iter = iter(conllx_data.iterate_batch_variable(data_train, batch_size)) for batch in list_iter: word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.data.cpu().numpy() types = types.data.cpu().numpy() stats, stats_nopunc, stats_root, num_inst = parser.eval(word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root #print(t_ucorr) t_ucorr += ucorr t_lcorr += lcorr t_total += total t_ucomlpete += ucm t_lcomplete += lcm t_ucorr_nopunc += ucorr_nopunc t_lcorr_nopunc += lcorr_nopunc t_total_nopunc += total_nopunc t_ucomlpete_nopunc += ucm_nopunc t_lcomplete_nopunc += lcm_nopunc t_root_corr += corr_root t_total_root += total_root t_total_inst += num_inst for _ in range(10): next(list_iter, None) for batch in conllx_data.iterate_batch_variable(data_dev, batch_size): word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.data.cpu().numpy() types = types.data.cpu().numpy() stats, stats_nopunc, stats_root, num_inst = parser.eval(word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root dev_ucorr += ucorr dev_lcorr += lcorr dev_total += total dev_ucomlpete += ucm dev_lcomplete += lcm dev_ucorr_nopunc += ucorr_nopunc dev_lcorr_nopunc += lcorr_nopunc dev_total_nopunc += total_nopunc dev_ucomlpete_nopunc += ucm_nopunc dev_lcomplete_nopunc += lcm_nopunc dev_root_corr += corr_root dev_total_root += total_root dev_total_inst += num_inst writer.writerow((t_ucorr_nopunc*100/t_total_nopunc,dev_ucorr_nopunc*100/dev_total_nopunc)) f.flush() #pred_writer.close() #gold_writer.close() print('Train Wo Punct:%.2f%%'% (t_ucorr_nopunc*100/t_total_nopunc)) print('W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % ( dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 / dev_total, dev_lcorr * 100 / dev_total, dev_ucomlpete * 100 / dev_total_inst, dev_lcomplete * 100 / dev_total_inst)) print('Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % ( dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc, dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc * 100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 / dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst)) print('Root: corr: %d, total: %d, acc: %.2f%%' %(dev_root_corr, dev_total_root, dev_root_corr * 100 / dev_total_root)) if dev_lcorrect_nopunc< dev_lcorr_nopunc or (dev_lcorrect_nopunc == dev_lcorr_nopunc and dev_ucorrect_nopunc < dev_ucorr_nopunc): dev_ucorrect_nopunc = dev_ucorr_nopunc dev_lcorrect_nopunc = dev_lcorr_nopunc dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc dev_lcomplete_match_nopunc = dev_lcomplete_nopunc dev_ucorrect = dev_ucorr dev_lcorrect = dev_lcorr dev_ucomlpete_match = dev_ucomlpete dev_lcomplete_match = dev_lcomplete dev_root_correct = dev_root_corr best_epoch = epoch patient = 0 # torch.save(network, model_name) torch.save(network.state_dict(), model_name) #pred_filename = 'tmp/%spred_test%d' % (str(uid), epoch) #pred_writer.start(pred_filename) #gold_filename = 'tmp/%sgold_test%d' % (str(uid), epoch) #gold_writer.start(gold_filename) test_ucorrect = 0.0 test_lcorrect = 0.0 test_ucomlpete_match = 0.0 test_lcomplete_match = 0.0 test_total = 0 test_ucorrect_nopunc = 0.0 test_lcorrect_nopunc = 0.0 test_ucomlpete_match_nopunc = 0.0 test_lcomplete_match_nopunc = 0.0 test_total_nopunc = 0 test_total_inst = 0 test_root_correct = 0.0 test_total_root = 0 for batch in conllx_data.iterate_batch_variable(data_test, batch_size): word, char, pos, heads, types, masks, lengths = batch heads_pred, types_pred = decode(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS) word = word.data.cpu().numpy() pos = pos.data.cpu().numpy() lengths = lengths.cpu().numpy() heads = heads.data.cpu().numpy() types = types.data.cpu().numpy() #pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True) #gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True) stats, stats_nopunc, stats_root, num_inst = parser.eval(word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True) ucorr, lcorr, total, ucm, lcm = stats ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc corr_root, total_root = stats_root test_ucorrect += ucorr test_lcorrect += lcorr test_total += total test_ucomlpete_match += ucm test_lcomplete_match += lcm test_ucorrect_nopunc += ucorr_nopunc test_lcorrect_nopunc += lcorr_nopunc test_total_nopunc += total_nopunc test_ucomlpete_match_nopunc += ucm_nopunc test_lcomplete_match_nopunc += lcm_nopunc test_root_correct += corr_root test_total_root += total_root test_total_inst += num_inst #pred_writer.close() #gold_writer.close() else: if dev_ucorr_nopunc * 100 / dev_total_nopunc < dev_ucorrect_nopunc * 100 / dev_total_nopunc - 5 or patient >= schedule: # network = torch.load(model_name) network.load_state_dict(torch.load(model_name)) lr = lr * decay_rate optim = generate_optimizer(opt, lr, network.parameters()) if decoding == 'greedy': decode = network.decode elif decoding == 'mst': decode = network.decode_mst else: raise ValueError('Unknown decoding algorithm: %s' % decoding) patient = 0 decay += 1 if decay % double_schedule_decay == 0: schedule *= 2 else: patient += 1 print('----------------------------------------------------------------------------------------------------------------------------') print('best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( dev_ucorrect, dev_lcorrect, dev_total, dev_ucorrect * 100 / dev_total, dev_lcorrect * 100 / dev_total, dev_ucomlpete_match * 100 / dev_total_inst, dev_lcomplete_match * 100 / dev_total_inst, best_epoch)) print('best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc, dev_ucorrect_nopunc * 100 / dev_total_nopunc, dev_lcorrect_nopunc * 100 / dev_total_nopunc, dev_ucomlpete_match_nopunc * 100 / dev_total_inst, dev_lcomplete_match_nopunc * 100 / dev_total_inst, best_epoch)) print('best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % ( dev_root_correct, dev_total_root, dev_root_correct * 100 / dev_total_root, best_epoch)) print('----------------------------------------------------------------------------------------------------------------------------') print('best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 / test_total, test_lcorrect * 100 / test_total, test_ucomlpete_match * 100 / test_total_inst, test_lcomplete_match * 100 / test_total_inst, best_epoch)) print('best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % ( test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc, test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc, test_ucomlpete_match_nopunc * 100 / test_total_inst, test_lcomplete_match_nopunc * 100 / test_total_inst, best_epoch)) print('best test Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % ( test_root_correct, test_total_root, test_root_correct * 100 / test_total_root, best_epoch)) print('============================================================================================================================') if decay == max_decay: break