def train(train_path, validation_path, dictionary_path, model_path, reload_state=False, dim_word=100, # word vector dimensionality dim=1000, # the number of LSTM units encoder='lstm', patience=10, max_epochs=5000, dispFreq=100, decay_c=0., alpha_c=0., diag_c=0., lrate=0.01, n_words=100000, maxlen=100, # maximum length of the description optimizer='rmsprop', batch_size = 16, valid_batch_size = 16, validFreq=1000, saveFreq=1000, # save the parameters after every saveFreq updates sampleFreq=100, # generate some text samples after every sampleFreq updates profile=False): # Model options model_options = locals().copy() worddicts = dict() worddicts_r = dict() with open(dictionary_path, 'rb') as f: for (i, line) in enumerate(f): word = line.strip() code = i + 2 worddicts_r[code] = word worddicts[word] = code # reload options if reload_state and os.path.exists(model_path): with open('%s.pkl' % model_path, 'rb') as f: models_options = pkl.load(f) print '### Loading data.' train = TextIterator(train_path, worddicts, n_words_source=n_words, batch_size=batch_size, maxlen=maxlen) valid = TextIterator(validation_path, worddicts, n_words_source=n_words, batch_size=valid_batch_size, maxlen=maxlen) print '### Building neural network.' rnnlm = RNNLM(model_options) trainer = ModelTrainer(rnnlm, optimizer, model_options) sampler = TextSampler(rnnlm, model_options) print '### Training neural network.' best_params = None bad_count = 0 if validFreq == -1: validFreq = len(train[0])/batch_size if saveFreq == -1: saveFreq = len(train[0])/batch_size if sampleFreq == -1: sampleFreq = len(train[0])/batch_size uidx = 0 estop = False for eidx in xrange(max_epochs): n_samples = 0 for x in train: n_samples += len(x) uidx += 1 x, x_mask = prepare_data(x, maxlen=maxlen, n_words=n_words) if x == None: print 'Minibatch with zero sample under length ', maxlen uidx -= 1 continue ud_start = time.time() cost = trainer.f_grad_shared(x, x_mask) trainer.f_update(lrate) ud = time.time() - ud_start if numpy.isnan(cost) or numpy.isinf(cost): print 'NaN detected' return 1., 1., 1. if numpy.mod(uidx, dispFreq) == 0: print 'Epoch ', eidx, 'Update ', uidx, 'Cost ', cost, 'UD ', ud if numpy.mod(uidx, saveFreq) == 0: # Save the best parameters, or the current state if best_params # is None. rnnlm.save_params(best_params) # Save the training options. pkl.dump(model_options, open('%s.pkl' % model_path, 'wb')) if numpy.mod(uidx, sampleFreq) == 0: # FIXME: random selection? for jj in xrange(5): sample, score = sampler.generate() print 'Sample ', jj, ': ', ss = sample for vv in ss: if vv == 0: break if vv in worddicts_r: print worddicts_r[vv], else: print 'UNK', print if numpy.mod(uidx, validFreq) == 0: valid_errs = pred_probs(f_log_probs, prepare_data, model_options, valid) valid_err = valid_errs.mean() rnnlm.error_history.append(valid_err) if uidx == 0 or valid_err <= numpy.array(error_history).min(): best_params = rnnlm.get_param_values() bad_counter = 0 if len(rnnlm.error_history) > patience and valid_err >= numpy.array(rnnlm.error_history)[:-patience].min(): bad_counter += 1 if bad_counter > patience: print 'Early Stop!' estop = True break if numpy.isnan(valid_err): import ipdb; ipdb.set_trace() print 'Valid ', valid_err print 'Seen %d samples'%n_samples if estop: break if best_params is not None: rnnlm.set_param_values(best_params) valid_err = pred_probs(f_log_probs, prepare_data, model_options, valid).mean() print 'Valid ', valid_err params = copy.copy(best_params) numpy.savez(model_path, zipped_params=best_params, error_history=rnnlm.error_history, **params) return valid_err
def train( train_path, validation_path, dictionary_path, model_path, reload_state=False, dim_word=100, # word vector dimensionality dim=1000, # the number of LSTM units encoder='lstm', patience=10, max_epochs=5000, dispFreq=100, decay_c=0., alpha_c=0., diag_c=0., lrate=0.01, n_words=100000, maxlen=100, # maximum length of the description optimizer='rmsprop', batch_size=16, valid_batch_size=16, validFreq=1000, saveFreq=1000, # save the parameters after every saveFreq updates sampleFreq=100, # generate some text samples after every sampleFreq updates profile=False): # Model options model_options = locals().copy() worddicts = dict() worddicts_r = dict() with open(dictionary_path, 'rb') as f: for (i, line) in enumerate(f): word = line.strip() code = i + 2 worddicts_r[code] = word worddicts[word] = code # reload options if reload_state and os.path.exists(model_path): with open('%s.pkl' % model_path, 'rb') as f: models_options = pkl.load(f) print '### Loading data.' train = TextIterator(train_path, worddicts, n_words_source=n_words, batch_size=batch_size, maxlen=maxlen) valid = TextIterator(validation_path, worddicts, n_words_source=n_words, batch_size=valid_batch_size, maxlen=maxlen) print '### Building neural network.' rnnlm = RNNLM(model_options) trainer = ModelTrainer(rnnlm, optimizer, model_options) sampler = TextSampler(rnnlm, model_options) print '### Training neural network.' best_params = None bad_count = 0 if validFreq == -1: validFreq = len(train[0]) / batch_size if saveFreq == -1: saveFreq = len(train[0]) / batch_size if sampleFreq == -1: sampleFreq = len(train[0]) / batch_size uidx = 0 estop = False for eidx in xrange(max_epochs): n_samples = 0 for x in train: n_samples += len(x) uidx += 1 x, x_mask = prepare_data(x, maxlen=maxlen, n_words=n_words) if x == None: print 'Minibatch with zero sample under length ', maxlen uidx -= 1 continue ud_start = time.time() cost = trainer.f_grad_shared(x, x_mask) trainer.f_update(lrate) ud = time.time() - ud_start if numpy.isnan(cost) or numpy.isinf(cost): print 'NaN detected' return 1., 1., 1. if numpy.mod(uidx, dispFreq) == 0: print 'Epoch ', eidx, 'Update ', uidx, 'Cost ', cost, 'UD ', ud if numpy.mod(uidx, saveFreq) == 0: # Save the best parameters, or the current state if best_params # is None. rnnlm.save_params(best_params) # Save the training options. pkl.dump(model_options, open('%s.pkl' % model_path, 'wb')) if numpy.mod(uidx, sampleFreq) == 0: # FIXME: random selection? for jj in xrange(5): sample, score = sampler.generate() print 'Sample ', jj, ': ', ss = sample for vv in ss: if vv == 0: break if vv in worddicts_r: print worddicts_r[vv], else: print 'UNK', print if numpy.mod(uidx, validFreq) == 0: valid_errs = pred_probs(f_log_probs, prepare_data, model_options, valid) valid_err = valid_errs.mean() rnnlm.error_history.append(valid_err) if uidx == 0 or valid_err <= numpy.array(error_history).min(): best_params = rnnlm.get_param_values() bad_counter = 0 if len(rnnlm.error_history ) > patience and valid_err >= numpy.array( rnnlm.error_history)[:-patience].min(): bad_counter += 1 if bad_counter > patience: print 'Early Stop!' estop = True break if numpy.isnan(valid_err): import ipdb ipdb.set_trace() print 'Valid ', valid_err print 'Seen %d samples' % n_samples if estop: break if best_params is not None: rnnlm.set_param_values(best_params) valid_err = pred_probs(f_log_probs, prepare_data, model_options, valid).mean() print 'Valid ', valid_err params = copy.copy(best_params) numpy.savez(model_path, zipped_params=best_params, error_history=rnnlm.error_history, **params) return valid_err