def main(save_path, params): nhidden = params['nhidden'] dropout = params['dropout'] word2vec = params['word2vec'] dataset = params['dataset'] nlayers = params['nlayers'] train_emb = params['train_emb'] char_dim = params['char_dim'] use_feat = params['use_feat'] gating_fn = params['gating_fn'] out = 'out' # save settings shutil.copyfile('config.py', '%s/config.py' % save_path) use_chars = char_dim > 0 dp = DataPreprocessor.DataPreprocessor() data = dp.preprocess(dataset, no_training_set=False, use_chars=use_chars) word_dictionary = data.dictionary[0] the_index = word_dictionary['the'] #print('the index : {}'.format(word_dictionary['the'])) idx_to_word = dict([(v, k) for (k, v) in word_dictionary.iteritems()]) words = [idx_to_word[i] for i in sorted(idx_to_word.keys())] print("building minibatch loaders ...") batch_loader_train = MiniBatchLoader.MiniBatchLoader(data.training, BATCH_SIZE, sample=1.0) batch_loader_val = MiniBatchLoader.MiniBatchLoader(data.validation, BATCH_SIZE) print("building network ...") W_init, embed_dim, = Helpers.load_word2vec_embeddings( data.dictionary[0], word2vec) #print('the embedding : {}'.format(W_init[the_index])) #print(W_init[0:5]) print("running GAReader ...") m = GAReader.Model(nlayers, data.vocab_size, data.num_chars, W_init, nhidden, embed_dim, dropout, train_emb, char_dim, use_feat, gating_fn, words).build_network() m.compile(optimizer=tf.keras.optimizers.Adam(lr=LEARNING_RATE, clipnorm=GRAD_CLIP), loss=tf.keras.losses.categorical_crossentropy, metrics=[tf.keras.metrics.categorical_accuracy]) #tf.enable_eager_execution(config=tf.ConfigProto(allow_soft_placement = True)) with tf.Graph().as_default(): with tf.Session(config=tf.ConfigProto( allow_soft_placement=True)) as sess: K.set_session(sess) #with tf.device('/gpu:0:'): tensorboard = TensorBoardCustom(log_dir="logs", words=words) modelcheckpoint = tf.keras.callbacks.ModelCheckpoint( 'output/weights.{epoch:02d}-{val_loss:.2f}.hdf5') writer = tf.summary.FileWriter("logs") def schedule(epoch, lr): if epoch >= 3: return lr * 0.5 else: return lr lrate = LearningRateScheduler(schedule, verbose=1) for epoch in xrange(NUM_EPOCHS): for (inputs, a) in batch_loader_train: [dw, qw, m_dw, m_qw, c, m_c, cl] = inputs m = GAReader.Model(nlayers, data.vocab_size, data.num_chars, W_init, nhidden, embed_dim, dropout, train_emb, char_dim, use_feat, gating_fn, words).build_network() m.compile(optimizer=tf.keras.optimizers.Adam( lr=LEARNING_RATE, clipnorm=GRAD_CLIP), loss=tf.keras.losses.categorical_crossentropy, metrics=[tf.keras.metrics.categorical_accuracy]) #print(dw.shape) #print('dw : {}'.format(dw)) #print('qw : {}'.format(qw)) #print('m_dw : {}'.format(m_dw)) #print('m_qw : {}'.format(m_qw)) #print('c : {}'.format(c)) #print([idx_to_word[i] for i in dw[0, :, 0].tolist()]) train_summary = m.train_on_batch( inputs, to_categorical(a, batch_loader_train.max_num_cand)) print(m.get_weights()[0]) print('epoch: {}, train loss: {}, train acc: {}'.format( epoch, train_summary[0], train_summary[1])) lr = tf.summary.scalar('learning_rate', LEARNING_RATE) summary = tf.summary.merge_all() s = sess.run(summary) writer.add_summary(s) writer.close()
def main(save_path, params): nhidden = params['nhidden'] dropout = params['dropout'] word2vec = params['word2vec'] dataset = params['dataset'] nlayers = params['nlayers'] train_emb = params['train_emb'] char_dim = params['char_dim'] use_feat = params['use_feat'] gating_fn = params['gating_fn'] ent_setup = params['ent_setup'] # ent, ent-anonym, no-ent data_path = params['data_path'] # save settings shutil.copyfile('config.py', '%s/config.py' % save_path) use_chars = char_dim > 0 if dataset == "clicr": dp = DataPreprocessor.DataPreprocessorClicr() data = dp.preprocess( #"/mnt/b5320167-5dbd-4498-bf34-173ac5338c8d/Datasets/bmj_case_reports_data/dataset_json_concept_annotated/", data_path, ent_setup=ent_setup, no_training_set=False, use_chars=use_chars) elif dataset == "clicr_novice": dp = DataPreprocessor.DataPreprocessorNovice() data = dp.preprocess(data_path, ent_setup=ent_setup, no_training_set=False, use_chars=use_chars) else: dp = DataPreprocessor.DataPreprocessor() data = dp.preprocess(data_path, no_training_set=False, use_chars=use_chars) print("building minibatch loaders ...") batch_loader_train = MiniBatchLoader.MiniBatchLoader(data.training, BATCH_SIZE, sample=1.0) batch_loader_val = MiniBatchLoader.MiniBatchLoader(data.validation, BATCH_SIZE) print("building network ...") W_init, embed_dim, = Helpers.load_word2vec_embeddings( data.dictionary[0], word2vec) m = GAReader.Model(nlayers, data.vocab_size, data.num_chars, W_init, nhidden, embed_dim, dropout, train_emb, char_dim, use_feat, gating_fn) print("training ...") num_iter = 0 max_acc = 0. deltas = [] logger = open(save_path + '/log', 'a') if os.path.isfile('%s/best_model.p' % save_path): print('loading previously saved model') m.load_model('%s/best_model.p' % save_path) else: print('saving init model') m.save_model('%s/model_init.p' % save_path) print('loading init model') m.load_model('%s/model_init.p' % save_path) for epoch in range(NUM_EPOCHS): estart = time.time() new_max = False for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames in batch_loader_train: loss, tr_acc, probs = m.train(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl) message = "Epoch %d TRAIN loss=%.4e acc=%.4f elapsed=%.1f" % ( epoch, loss, tr_acc, time.time() - estart) print(message) logger.write(message + '\n') num_iter += 1 if num_iter % VALIDATION_FREQ == 0: total_loss, total_acc, n, n_cand = 0., 0., 0, 0. for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames in batch_loader_val: outs = m.validate(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl) loss, acc, probs = outs[:3] bsize = dw.shape[0] total_loss += bsize * loss total_acc += bsize * acc n += bsize val_acc = total_acc / n if val_acc > max_acc: max_acc = val_acc m.save_model('%s/best_model.p' % save_path) new_max = True message = "Epoch %d VAL loss=%.4e acc=%.4f max_acc=%.4f" % ( epoch, total_loss / n, val_acc, max_acc) print(message) logger.write(message + '\n') # m.save_model('%s/model_%d.p'%(save_path,epoch)) message = "After Epoch %d: Train acc=%.4f, Val acc=%.4f" % ( epoch, tr_acc, val_acc) print(message) logger.write(message + '\n') # learning schedule if epoch >= 2: m.anneal() # stopping criterion if not new_max: break logger.close()
def main(load_path, params, mode='test'): regularizer = params['regularizer'] rlambda = params['lambda'] nhidden = params['nhidden'] dropout = params['dropout'] word2vec = params['word2vec'] dataset = params['dataset'] nlayers = params['nlayers'] train_emb = params['train_emb'] subsample = params['subsample'] base_model = params['model'] char_dim = params['char_dim'] use_feat = params['use_feat'] gating_fn = params['gating_fn'] # load settings shutil.copyfile('%s/config.py' % load_path, 'config.py') dp = DataPreprocessor.DataPreprocessor() data = dp.preprocess(dataset) inv_vocab = data.inv_dictionary print("building minibatch loaders ...") if mode == 'test': batch_loader_test = MiniBatchLoader.MiniBatchLoader( data.test, BATCH_SIZE, data.dictionary) else: batch_loader_test = MiniBatchLoader.MiniBatchLoader( data.validation, BATCH_SIZE, data.dictionary) print("building network ...") W_init, embed_dim = Helpers.load_word2vec_embeddings( data.dictionary[0], word2vec) m = eval(base_model).Model(nlayers, data.vocab_size, data.num_chars, W_init, regularizer, rlambda, nhidden, embed_dim, dropout, train_emb, subsample, char_dim, use_feat, data.dictionary[4]) m.load_model('%s/best_model.p' % load_path) print("testing ...") pr = np.zeros((len(batch_loader_test.questions), batch_loader_test.max_num_cand)).astype('float32') fids, attns = [], [] total_loss, total_acc, n = 0., 0., 0 for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames, match_feat, use_char, use_char_q in batch_loader_test: outs = m.validate(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl, match_feat, use_char, use_char_q) loss, acc, probs = outs[:3] attns += [[fnames[0], probs[0, :]] + [o[0, :, :] for o in outs[3:]] ] # store one attention bsize = dw.shape[0] total_loss += bsize * loss total_acc += bsize * acc pr[n:n + bsize, :] = probs fids += fnames n += bsize logger = open(load_path + '/log', 'a', 0) message = '%s Loss %.4e acc=%.4f' % (mode.upper(), total_loss / n, total_acc / n) print message logger.write(message + '\n') logger.close() np.save('%s/%s.probs' % (load_path, mode), np.asarray(pr)) pkl.dump(attns, open('%s/%s.attns' % (load_path, mode), 'w')) f = open('%s/%s.ids' % (load_path, mode), 'w') for item in fids: f.write(item + '\n') f.close()
def main(load_path, params, mode='test'): nhidden = params['nhidden'] dropout = params['dropout'] word2vec = params['word2vec'] dataset = params['data'] nlayers = params['nlayers'] sub2vec = params['sub2vec'] train_emb = params['train_emb'] sub_dim = params['sub_dim'] use_feat = params['use_feat'] gating_fn = params['gating_fn'] use_subs = sub_dim > 0 dp = DataPreprocessor.DataPreprocessor() data = dp.preprocess(dataset, no_training_set=True, use_subs=use_subs) inv_vocab = data.inv_dictionary print("building minibatch loaders ...") if mode == 'test': batch_loader_test = MiniBatchLoader.MiniBatchLoader( data.test, BATCH_SIZE) else: batch_loader_test = MiniBatchLoader.MiniBatchLoader( data.validation, BATCH_SIZE) print("building network ...") W_init, embed_dim = Helpers.load_word2vec_embeddings( data.dictionary[0], word2vec) S_init, sub_dim = Helpers.load_sub_embeddings(data.dictionary[1], sub2vec) m = model.Model(nlayers, data.vocab_size, data.num_chars, W_init, S_init, nhidden, embed_dim, dropout, train_emb, sub_dim, use_feat, gating_fn, save_attn=True) m.load_model('%s/best_model.p' % load_path) print("testing ...") pr = np.zeros((len(batch_loader_test.questions), batch_loader_test.max_num_cand)).astype('float32') fids, attns = [], [] total_loss, total_acc, n = 0., 0., 0 for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames in batch_loader_test: outs = m.validate(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl) loss, acc, probs = outs[:3] attns += [[fnames[0], probs[0, :]] + [o[0, :, :] for o in outs[3:]] ] # store one attention bsize = dw.shape[0] total_loss += bsize * loss total_acc += bsize * acc fids += fnames n += bsize print("step" + str(n) + ",acc" + str(acc)) logger = open(load_path + '/log', 'a', 0) message = '%s Loss %.4e acc=%.4f' % (mode.upper(), total_loss / n, total_acc / n) print message logger.write(message + '\n') logger.close() np.save('%s/%s.probs' % (load_path, mode), np.asarray(pr)) pkl.dump(attns, open('%s/%s.attns' % (load_path, mode), 'w')) f = open('%s/%s.ids' % (load_path, mode), 'w') for item in fids: f.write(item + '\n') f.close()
def main(save_path, params): regularizer = params['regularizer'] rlambda = params['lambda'] nhidden = params['nhidden'] dropout = params['dropout'] word2vec = params['word2vec'] dataset = params['dataset'] nlayers = params['nlayers'] train_emb = params['train_emb'] subsample = params['subsample'] base_model = params['model'] char_dim = params['char_dim'] use_feat = params['use_feat'] train_cut = params['train_cut'] gating_fn = params['gating_fn'] # save settings shutil.copyfile('config.py','%s/config.py'%save_path) use_chars = char_dim>0 dp = DataPreprocessor.DataPreprocessor() data = dp.preprocess(dataset, use_chars=use_chars) print("building minibatch loaders ...") batch_loader_train = MiniBatchLoader.MiniBatchLoader(data.training, BATCH_SIZE, data.dictionary, sample=train_cut, max_qry_len=85) batch_loader_val = MiniBatchLoader.MiniBatchLoader(data.validation, BATCH_SIZE, data.dictionary, max_qry_len=85) batch_loader_test = MiniBatchLoader.MiniBatchLoader(data.test, BATCH_SIZE, data.dictionary) print("building network ...") W_init, embed_dim, = Helpers.load_word2vec_embeddings(data.dictionary[0], word2vec) m = eval(base_model).Model(nlayers, data.vocab_size, data.num_chars, W_init, regularizer, rlambda, nhidden, embed_dim, dropout, train_emb, subsample, char_dim, use_feat, data.dictionary[4]) print("training ...") num_iter = 0 max_acc = 0. deltas = [] test_acc = 0. logger = open(save_path+'/log','a',0) # if os.path.isfile('%s/best_model.p'%save_path): # print('loading previously saved model') # m.load_model('%s/best_model.p'%save_path) # else: # print('saving init model') # m.save_model('%s/model_init.p'%save_path) # print('loading init model') # m.load_model('%s/model_init.p'%save_path) for epoch in xrange(NUM_EPOCHS): estart = time.time() new_max = False for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames, match_feat, use_char, use_char_q in batch_loader_train: loss, tr_acc, probs = m.train(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl, match_feat, use_char, use_char_q) # message = "Epoch %d TRAIN loss=%.4e acc=%.4f elapsed=%.1f" % ( # epoch, loss, tr_acc, time.time()-estart) # print message # logger.write(message+'\n') if num_iter % VALIDATION_FREQ == 0: total_loss, total_acc, n, n_cand = 0., 0., 0, 0. for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames, match_feat, use_char, use_char_q in batch_loader_val: outs = m.validate(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl, match_feat, use_char, use_char_q) loss, acc, probs = outs[:3] bsize = dw.shape[0] total_loss += bsize*loss total_acc += bsize*acc n += bsize val_acc = total_acc/n if val_acc > max_acc: max_acc = val_acc m.save_model('%s/best_model.p'%save_path) temp_acc, temp_n = 0.0, 0 for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames, match_feat, use_char, use_char_q in batch_loader_test: outs = m.validate(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl, match_feat, use_char, use_char_q) _, acc, _ = outs[:3] bsize = dw.shape[0] temp_acc += bsize * acc temp_n += bsize test_acc = temp_acc / temp_n new_max = True message = "Epoch %d VAL loss=%.4e acc=%.4f max_acc=%.4f test=%.4f" % ( epoch, total_loss/n, val_acc, max_acc, test_acc) print message logger.write(message+'\n') num_iter += 1 m.save_model('%s/model_%d.p'%(save_path,epoch)) message = "After Epoch %d: Train acc=%.4f, Val acc=%.4f" % (epoch, tr_acc, val_acc) print message logger.write(message+'\n') # learning schedule if epoch >=2: m.anneal() # stopping criterion if not new_max: break logger.close()
def main(load_path, params, mode='test'): nhidden = params['nhidden'] dropout = params['dropout'] word2vec = params['word2vec'] dataset = params['dataset'] nlayers = params['nlayers'] train_emb = params['train_emb'] char_dim = params['char_dim'] use_feat = params['use_feat'] gating_fn = params['gating_fn'] ent_setup = params['ent_setup'] data_path = params['data_path'] # save settings shutil.copyfile('config.py', '%s/config_test.py' % load_path) use_chars = char_dim > 0 if dataset == "clicr": dp = DataPreprocessor.DataPreprocessorClicr() #dataset_path = "/mnt/b5320167-5dbd-4498-bf34-173ac5338c8d/Datasets/bmj_case_reports_data/dataset_json_concept_annotated/" #dataset_path = "data/" data = dp.preprocess(data_path, ent_setup=ent_setup, no_training_set=True) elif dataset == "clicr_novice": dp = DataPreprocessor.DataPreprocessorNovice() data = dp.preprocess(data_path, ent_setup=ent_setup, no_training_set=True) else: dp = DataPreprocessor.DataPreprocessor() data = dp.preprocess(data_path, no_training_set=True) inv_vocab = data.inv_dictionary assert os.path.exists(params["test_file"] if mode == "test" else params["validation_file"]) print("building minibatch loaders ...") if mode == 'test': batch_loader_test = MiniBatchLoader.MiniBatchLoader( data.test, BATCH_SIZE) else: batch_loader_test = MiniBatchLoader.MiniBatchLoader( data.validation, BATCH_SIZE) f_to_cand = {i[-1]: i[3] for i in batch_loader_test.questions} print("building network ...") W_init, embed_dim = Helpers.load_word2vec_embeddings( data.dictionary[0], word2vec) m = GAReader.Model(nlayers, data.vocab_size, data.num_chars, W_init, nhidden, embed_dim, dropout, train_emb, char_dim, use_feat, gating_fn, save_attn=False) print("model load path") print('%s/best_model.p' % load_path) m.load_model('%s/best_model.p' % load_path) print("testing ...") pr = np.zeros((len(batch_loader_test.questions), batch_loader_test.max_num_cand)).astype('float32') fids, attns = [], [] pred_ans = {} total_loss, total_acc, n = 0., 0., 0 for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames in batch_loader_test: outs = m.validate(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl) loss, acc, probs = outs[:3] attns += [[fnames[0], probs[0, :]] + [o[0, :, :] for o in outs[3:]] ] # store one attention for f in range(len(fnames)): pred_cand = probs[f].argmax() pred_a_ids = f_to_cand[fnames[f]][pred_cand] pred_a = " ".join([inv_vocab[i] for i in pred_a_ids]) if ent_setup == "ent-anonym" and (dataset == "clicr" or dataset == "clicr_novice"): relabeling_dicts = data.test_relabeling_dicts if mode == 'test' else data.val_relabeling_dicts pred_a = relabeling_dicts[fnames[f]][pred_a] pred_ans[fnames[f]] = pred_a bsize = dw.shape[0] total_loss += bsize * loss total_acc += bsize * acc pr[n:n + bsize, :] = probs fids += fnames n += bsize if (params["dataset"] == "clicr" or params["dataset"] == "clicr_plain" or params["dataset"] == "clicr_novice") \ and (mode == 'test' or mode == 'validation'): print("writing predictions") preds_data = utils.to_output_preds(pred_ans) preds_filepath = load_path + '/{}.preds'.format(mode) utils.write_preds(preds_data, file_name=preds_filepath) utils.external_eval(preds_filepath, preds_filepath + ".scores", params["test_file"] if mode == "test" else params["validation_file"], extended=True) logger = open(load_path + '/log.test', 'a') message = '%s Loss %.4e acc=%.4f' % (mode.upper(), total_loss / n, total_acc / n) print(message) logger.write(message + '\n') logger.close() np.save('%s/%s.probs' % (load_path, mode), np.asarray(pr)) pickle.dump(attns, open('%s/%s.attns' % (load_path, mode), 'wb')) f = open('%s/%s.ids' % (load_path, mode), 'w') for item in fids: f.write(item + '\n') f.close()
def main(save_path, params): nhidden = params['nhidden'] dropout = params['dropout'] word2vec = params['word2vec'] sub2vec = params['sub2vec'] subdict = params['subdic'] dataset = params['data'] nlayers = params['nlayers'] train_emb = params['train_emb'] sub_dim = params['sub_dim'] use_feat = params['use_feat'] gating_fn = params['gating_fn'] # save settings shutil.copyfile('config.py', '%s/config.py' % save_path) use_subs = sub_dim > 0 dp = DataPreprocessor.DataPreprocessor() data = dp.preprocess(dataset, no_training_set=False, use_subs=use_subs, subdict=subdict) print "building minibatch loaders ...", datetime.now().strftime( '%Y-%m-%d %H:%M:%S') batch_loader_train = MiniBatchLoader.MiniBatchLoader(data.training, BATCH_SIZE, sample=1) batch_loader_val = MiniBatchLoader.MiniBatchLoader(data.validation, BATCH_SIZE) print "building network ...", datetime.now().strftime('%Y-%m-%d %H:%M:%S') W_init, embed_dim, = Helpers.load_word2vec_embeddings( data.dictionary[0], word2vec) S_init, sub_dim = Helpers.load_sub_embeddings(data.dictionary[1], sub2vec) m = model.Model(nlayers, data.vocab_size, data.num_chars, W_init, S_init, nhidden, embed_dim, dropout, train_emb, sub_dim, use_feat, gating_fn) print "training ...", datetime.now().strftime('%Y-%m-%d %H:%M:%S') num_iter = 0 max_acc = 0. deltas = [] logger = open(save_path + '/log', 'a', 0) if os.path.isfile('%s/best_model.p' % save_path): print 'loading previously saved model', datetime.now().strftime( '%Y-%m-%d %H:%M:%S') m.load_model('%s/best_model.p' % save_path) print "model loaded" else: print 'saving init model', datetime.now().strftime('%Y-%m-%d %H:%M:%S') m.save_model('%s/model_init.p' % save_path) print 'loading init model', datetime.now().strftime( '%Y-%m-%d %H:%M:%S') m.load_model('%s/model_init.p' % save_path) for epoch in xrange(NUM_EPOCHS): print "epochs training ...", datetime.now().strftime( '%Y-%m-%d %H:%M:%S') estart = time.time() new_max = False for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames in batch_loader_train: loss, tr_acc, probs = m.train(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl) message = "Epoch %d TRAIN loss=%.4e acc=%.4f elapsed=%.1f" % ( epoch, loss, tr_acc, time.time() - estart) print message logger.write(message + '\n') num_iter += 1 if num_iter % VALIDATION_FREQ == 0: total_loss, total_acc, n, n_cand = 0., 0., 0, 0. for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames in batch_loader_val: outs = m.validate(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl) loss, acc, probs = outs[:3] bsize = dw.shape[0] total_loss += bsize * loss total_acc += bsize * acc n += bsize print('validate on ', str(n) + 'validation data') val_acc = total_acc / n if val_acc > max_acc: max_acc = val_acc m.save_model('%s/best_model.p' % save_path) new_max = True message = "Epoch %d VAL loss=%.4e acc=%.4f max_acc=%.4f" % ( epoch, total_loss / n, val_acc, max_acc) print message logger.write(message + '\n') m.save_model('%s/model_%d.p' % (save_path, epoch)) message = "After Epoch %d: Train acc=%.4f, Val acc=%.4f" % ( epoch, tr_acc, val_acc) print message logger.write(message + '\n') # learning schedule if epoch >= 2: m.anneal() # stopping criterion if not new_max: break logger.close()
output_path = sys.argv[2] dataset = sys.argv[3] K = int(sys.argv[4]) top_K = 3 dp = DataPreprocessor.DataPreprocessor() # NOTE: make sure vocab.txt is already there! data = dp.preprocess(DATASET, no_training_set=True) inv_vocab = data.inv_dictionary print("building minibatch loaders ...") if not 'CANDIDATE_SUBSET' in locals(): CANDIDATE_SUBSET = False if dataset == 'validation': batch_loader_test = MiniBatchLoader.MiniBatchLoader( data.validation, 128, shuffle=False, candidate_subset=CANDIDATE_SUBSET) elif dataset == 'test': batch_loader_test = MiniBatchLoader.MiniBatchLoader( data.test, 128, shuffle=False, candidate_subset=CANDIDATE_SUBSET) print("building network ...") m = GAReader.Model(K, data.vocab_size) print("loading model from file...") m.load_model(model_path) print("predicting ...") fid = open(output_path, 'w', 0) pr = []
def main(load_path, params, mode='test'): nhidden = params['nhidden'] dropout = params['dropout'] word2vec = params['word2vec'] dataset = params['dataset'] nlayers = params['nlayers'] train_emb = params['train_emb'] char_dim = params['char_dim'] use_feat = params['use_feat'] gating_fn = params['gating_fn'] dp = DataPreprocessor.DataPreprocessor() data = dp.preprocess(dataset, no_training_set=True) inv_vocab = data.inv_dictionary print("building minibatch loaders ...") if mode == 'test': batch_loader_test = MiniBatchLoader.MiniBatchLoader( data.test, BATCH_SIZE) else: batch_loader_test = MiniBatchLoader.MiniBatchLoader( data.validation, BATCH_SIZE) print("building network ...") W_init, embed_dim = Helpers.load_word2vec_embeddings( data.dictionary[0], word2vec) m = Reader.Model(nlayers, data.vocab_size, data.num_chars, W_init, nhidden, embed_dim, dropout, train_emb, char_dim, use_feat, gating_fn, save_attn=True) m.load_model('%s/best_model.p' % load_path) print("testing ...") pr = np.zeros((len(batch_loader_test.questions), batch_loader_test.max_num_cand)).astype('float32') fids, attns = [], [] total_loss, total_acc, n = 0., 0., 0 result = {} for dw, dt, qw, qt, a, m_dw, m_qw, tt, tm, c, m_c, cl, fnames in batch_loader_test: outs = m.validate(dw, dt, qw, qt, c, a, m_dw, m_qw, tt, tm, m_c, cl) loss, acc, probs = outs[:3] attns += [[fnames[0], probs[0, :]] + [o[0, :, :] for o in outs[3:]] ] # store one attention bsize = dw.shape[0] total_loss += bsize * loss total_acc += bsize * acc pr[n:n + bsize, :] = probs fids += fnames n += bsize answer = probs.argmax(1) for it in range(len(fnames)): tid = fnames[it].split('/')[-1].split('.')[0].strip() result[eval(tid)] = answer[it] print tid, answer[it] print('probs----', probs) #print('a----', a) print('fnames----', fnames) print len(result) with open('raw.txt', 'w') as ff: for i in range(1, 2501): ff.write(str(result[i]) + '\n') logger = open(load_path + '/log', 'a', 0) message = '%s Loss %.4e acc=%.4f' % (mode.upper(), total_loss / n, total_acc / n) print message logger.write(message + '\n') logger.close() np.save('%s/%s.probs' % (load_path, mode), np.asarray(pr)) pkl.dump(attns, open('%s/%s.attns' % (load_path, mode), 'w')) f = open('%s/%s.ids' % (load_path, mode), 'w') for item in fids: f.write(item + '\n') f.close()