def evaluate_lenet5(learning_rate=0.02, n_epochs=100, emb_size=300, batch_size=10, filter_size=[3, 5], maxSentLen=40, hidden_size=[300, 300]): model_options = locals().copy() print "model options", model_options seed = 1234 np.random.seed(seed) rng = np.random.RandomState( seed) #random seed, control the model generates the same results srng = T.shared_randomstreams.RandomStreams(rng.randint(seed)) all_sentences, all_masks, all_labels, word2id = load_BBN_multi_labels_dataset( maxlen=maxSentLen ) #minlen, include one label, at least one word in the sentence train_sents = np.asarray(all_sentences[0], dtype='int32') train_masks = np.asarray(all_masks[0], dtype=theano.config.floatX) train_labels = np.asarray(all_labels[0], dtype='int32') train_size = len(train_labels) dev_sents = np.asarray(all_sentences[1], dtype='int32') dev_masks = np.asarray(all_masks[1], dtype=theano.config.floatX) dev_labels = np.asarray(all_labels[1], dtype='int32') dev_size = len(dev_labels) test_sents = np.asarray(all_sentences[2], dtype='int32') test_masks = np.asarray(all_masks[2], dtype=theano.config.floatX) test_labels = np.asarray(all_labels[2], dtype='int32') test_size = len(test_labels) vocab_size = len(word2id) + 1 # add one zero pad index rand_values = rng.normal( 0.0, 0.01, (vocab_size, emb_size)) #generate a matrix by Gaussian distribution rand_values[0] = np.array(np.zeros(emb_size), dtype=theano.config.floatX) id2word = {y: x for x, y in word2id.iteritems()} word2vec = load_word2vec() rand_values = load_word2vec_to_init(rand_values, id2word, word2vec) embeddings = theano.shared( value=np.array(rand_values, dtype=theano.config.floatX), borrow=True ) #wrap up the python variable "rand_values" into theano variable #now, start to build the input form of the model sents_id_matrix = T.imatrix('sents_id_matrix') sents_mask = T.fmatrix('sents_mask') labels = T.imatrix('labels') #batch*12 ###################### # BUILD ACTUAL MODEL # ###################### print '... building the model' common_input = embeddings[sents_id_matrix.flatten()].reshape( (batch_size, maxSentLen, emb_size)).dimshuffle( 0, 2, 1) #the input format can be adapted into CNN or GRU or LSTM conv_W, conv_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) conv_W2, conv_b2 = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[1])) NN_para = [conv_W, conv_b, conv_W2, conv_b2] conv_model = Conv_with_Mask( rng, input_tensor3=common_input, mask_matrix=sents_mask, image_shape=(batch_size, 1, emb_size, maxSentLen), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero sent_embeddings = conv_model.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size conv_model2 = Conv_with_Mask( rng, input_tensor3=common_input, mask_matrix=sents_mask, image_shape=(batch_size, 1, emb_size, maxSentLen), filter_shape=(hidden_size[0], 1, emb_size, filter_size[1]), W=conv_W2, b=conv_b2 ) #mutiple mask with the conv_out to set the features by UNK to zero sent_embeddings2 = conv_model2.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size LR_input = T.concatenate([sent_embeddings, sent_embeddings2], axis=1) LR_input_size = hidden_size[0] * 2 #classification layer, it is just mapping from a feature vector of size "hidden_size" to a vector of only two values: positive, negative U_a = create_ensemble_para( rng, 12, LR_input_size) # the weight matrix hidden_size*2 LR_b = theano.shared(value=np.zeros((12, ), dtype=theano.config.floatX), name='LR_b', borrow=True) #bias for each target class LR_para = [U_a, LR_b] layer_LR = LogisticRegression( rng, input=LR_input, n_in=LR_input_size, n_out=12, W=U_a, b=LR_b ) #basically it is a multiplication between weight matrix and input feature vector score_matrix = T.nnet.sigmoid(layer_LR.before_softmax) #batch * 12 prob_pos = T.where(labels < 1, 1.0 - score_matrix, score_matrix) loss = -T.mean(T.log(prob_pos)) # loss=layer_LR.negative_log_likelihood(labels) #for classification task, we usually used negative log likelihood as loss, the lower the better. params = [embeddings ] + NN_para + LR_para # put all model parameters together cost = loss #+Div_reg*diversify_reg#+L2_weight*L2_reg updates = Gradient_Cost_Para(cost, params, learning_rate) ''' testing ''' binarize_prob = T.where(score_matrix > 0.5, 1, 0) #train_model = theano.function([sents_id_matrix, sents_mask, labels], cost, updates=updates, on_unused_input='ignore') train_model = theano.function([sents_id_matrix, sents_mask, labels], cost, updates=updates, allow_input_downcast=True, on_unused_input='ignore') # dev_model = theano.function([sents_id_matrix, sents_mask, labels], layer_LR.errors(labels), allow_input_downcast=True, on_unused_input='ignore') test_model = theano.function([sents_id_matrix, sents_mask], binarize_prob, allow_input_downcast=True, on_unused_input='ignore') ############### # TRAIN MODEL # ############### print '... training' # early-stopping parameters patience = 50000000000 # look as this many examples regardless start_time = time.time() mid_time = start_time past_time = mid_time epoch = 0 done_looping = False n_train_batches = train_size / batch_size train_batch_start = list( np.arange(n_train_batches) * batch_size) + [train_size - batch_size] # n_dev_batches=dev_size/batch_size # dev_batch_start=list(np.arange(n_dev_batches)*batch_size)+[dev_size-batch_size] n_test_batches = test_size / batch_size test_batch_start = list( np.arange(n_test_batches) * batch_size) + [test_size - batch_size] # max_acc_dev=0.0 max_meanf1_test = 0.0 max_weightf1_test = 0.0 train_indices = range(train_size) while epoch < n_epochs: epoch = epoch + 1 random.Random(100).shuffle(train_indices) iter_accu = 0 cost_i = 0.0 for batch_id in train_batch_start: #for each batch # iter means how many batches have been run, taking into loop iter = (epoch - 1) * n_train_batches + iter_accu + 1 iter_accu += 1 train_id_batch = train_indices[batch_id:batch_id + batch_size] cost_i += train_model(train_sents[train_id_batch], train_masks[train_id_batch], train_labels[train_id_batch]) #after each 1000 batches, we test the performance of the model on all test data if iter % 20 == 0: print 'Epoch ', epoch, 'iter ' + str( iter) + ' average cost: ' + str(cost_i / iter), 'uses ', ( time.time() - past_time) / 60.0, 'min' past_time = time.time() error_sum = 0.0 all_pred_labels = [] all_gold_labels = [] for test_batch_id in test_batch_start: # for each test batch pred_labels = test_model( test_sents[test_batch_id:test_batch_id + batch_size], test_masks[test_batch_id:test_batch_id + batch_size]) gold_labels = test_labels[test_batch_id:test_batch_id + batch_size] all_pred_labels.append(pred_labels) all_gold_labels.append(gold_labels) all_pred_labels = np.concatenate(all_pred_labels) all_gold_labels = np.concatenate(all_gold_labels) test_mean_f1, test_weight_f1 = average_f1_two_array_by_col( all_pred_labels, all_gold_labels) if test_weight_f1 > max_weightf1_test: max_weightf1_test = test_weight_f1 if test_mean_f1 > max_meanf1_test: max_meanf1_test = test_mean_f1 print '\t\t\t\t\t\t\t\tcurrent f1s:', test_mean_f1, test_weight_f1, '\t\tmax_f1:', max_meanf1_test, max_weightf1_test print 'Epoch ', epoch, 'uses ', (time.time() - mid_time) / 60.0, 'min' mid_time = time.time() #print 'Batch_size: ', update_freq end_time = time.time() print >> sys.stderr, ('The code for file ' + os.path.split(__file__)[1] + ' ran for %.2fm' % ((end_time - start_time) / 60.)) return max_acc_test
def evaluate_lenet5(learning_rate=0.02, n_epochs=100, emb_size=300, batch_size=50, filter_size=[3], sent_len=40, claim_len=40, cand_size=10, hidden_size=[300, 300], max_pred_pick=5): model_options = locals().copy() print "model options", model_options pred_id2label = {1: 'SUPPORTS', 0: 'REFUTES', 2: 'NOT ENOUGH INFO'} seed = 1234 np.random.seed(seed) rng = np.random.RandomState( seed) #random seed, control the model generates the same results srng = T.shared_randomstreams.RandomStreams(rng.randint(seed)) "load raw data" train_sents, train_sent_masks, train_sent_labels, train_claims, train_claim_mask, train_labels, word2id = load_fever_train( sent_len, claim_len, cand_size) train_3th_sents, train_3th_sent_masks, train_3th_sent_labels, train_3th_claims, train_3th_claim_mask, train_3th_labels, word2id = load_fever_train_NoEnoughInfo( sent_len, claim_len, cand_size, word2id) test_sents, test_sent_masks, test_sent_labels, test_claims, test_claim_mask, test_sent_names, test_ground_names, test_labels, word2id = load_fever_dev( sent_len, claim_len, cand_size, word2id) test_3th_sents, test_3th_sent_masks, test_3th_sent_labels, test_3th_claims, test_3th_claim_mask, test_3th_labels, word2id = load_fever_dev_NoEnoughInfo( sent_len, claim_len, cand_size, word2id) train_sents = np.asarray(train_sents, dtype='int32') train_3th_sents = np.asarray(train_3th_sents, dtype='int32') joint_train_sents = np.concatenate((train_sents, train_3th_sents)) test_sents = np.asarray(test_sents, dtype='int32') test_3th_sents = np.asarray(test_3th_sents, dtype='int32') joint_test_sents = np.concatenate((test_sents, test_3th_sents)) train_sent_masks = np.asarray(train_sent_masks, dtype=theano.config.floatX) train_3th_sent_masks = np.asarray(train_3th_sent_masks, dtype=theano.config.floatX) joint_train_sent_masks = np.concatenate( (train_sent_masks, train_3th_sent_masks)) test_sent_masks = np.asarray(test_sent_masks, dtype=theano.config.floatX) test_3th_sent_masks = np.asarray(test_3th_sent_masks, dtype=theano.config.floatX) joint_test_sent_masks = np.concatenate( (test_sent_masks, test_3th_sent_masks)) train_sent_labels = np.asarray(train_sent_labels, dtype='int32') train_3th_sent_labels = np.asarray(train_3th_sent_labels, dtype='int32') joint_train_sent_labels = np.concatenate( (train_sent_labels, train_3th_sent_labels)) test_sent_labels = np.asarray(test_sent_labels, dtype='int32') test_3th_sent_labels = np.asarray(test_3th_sent_labels, dtype='int32') joint_test_sent_labels = np.concatenate( (test_sent_labels, test_3th_sent_labels)) train_claims = np.asarray(train_claims, dtype='int32') train_3th_claims = np.asarray(train_3th_claims, dtype='int32') joint_train_claims = np.concatenate((train_claims, train_3th_claims)) test_claims = np.asarray(test_claims, dtype='int32') test_3th_claims = np.asarray(test_3th_claims, dtype='int32') joint_test_claims = np.concatenate((test_claims, test_3th_claims)) train_claim_mask = np.asarray(train_claim_mask, dtype=theano.config.floatX) train_3th_claim_mask = np.asarray(train_3th_claim_mask, dtype=theano.config.floatX) joint_train_claim_mask = np.concatenate( (train_claim_mask, train_3th_claim_mask)) test_claim_mask = np.asarray(test_claim_mask, dtype=theano.config.floatX) test_3th_claim_mask = np.asarray(test_3th_claim_mask, dtype=theano.config.floatX) joint_test_claim_mask = np.concatenate( (test_claim_mask, test_3th_claim_mask)) train_labels = np.asarray(train_labels, dtype='int32') train_3th_labels = np.asarray(train_3th_labels, dtype='int32') joint_train_labels = np.concatenate((train_labels, train_3th_labels)) test_labels = np.asarray(test_labels, dtype='int32') test_3th_labels = np.asarray(test_3th_labels, dtype='int32') joint_test_labels = np.concatenate((test_labels, test_3th_labels)) joint_train_size = len(joint_train_claims) joint_test_size = len(joint_test_claims) train_size = len(train_claims) test_size = len(test_claims) test_3th_size = len(test_3th_claims) vocab_size = len(word2id) + 1 print 'joint_train size: ', joint_train_size, ' joint_test size: ', joint_test_size print 'train size: ', train_size, ' test size: ', test_size print 'vocab size: ', vocab_size rand_values = rng.normal( 0.0, 0.01, (vocab_size, emb_size)) #generate a matrix by Gaussian distribution id2word = {y: x for x, y in word2id.iteritems()} word2vec = load_word2vec() rand_values = load_word2vec_to_init(rand_values, id2word, word2vec) init_embeddings = theano.shared( value=np.array(rand_values, dtype=theano.config.floatX), borrow=True ) #wrap up the python variable "rand_values" into theano variable "now, start to build the input form of the model" sents_ids = T.itensor3() #(batch, cand_size, sent_len) sents_mask = T.ftensor3() sents_labels = T.imatrix() #(batch, cand_size) claim_ids = T.imatrix() #(batch, claim_len) claim_mask = T.fmatrix() joint_sents_ids = T.itensor3() #(batch, cand_size, sent_len) joint_sents_mask = T.ftensor3() joint_sents_labels = T.imatrix() #(batch, cand_size) joint_claim_ids = T.imatrix() #(batch, claim_len) joint_claim_mask = T.fmatrix() joint_labels = T.ivector() ###################### # BUILD ACTUAL MODEL # ###################### print '... building the model' embed_input_sents = init_embeddings[sents_ids.flatten( )].reshape((batch_size * cand_size, sent_len, emb_size)).dimshuffle( 0, 2, 1 ) #embed_input(init_embeddings, sents_ids_l)#embeddings[sents_ids_l.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1) #the input format can be adapted into CNN or GRU or LSTM embed_input_claim = init_embeddings[claim_ids.flatten()].reshape( (batch_size, claim_len, emb_size)).dimshuffle(0, 2, 1) conv_W, conv_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) att_conv_W, att_conv_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) conv_W_context, conv_b_context = create_conv_para( rng, filter_shape=(hidden_size[0], 1, emb_size, 1)) NN_para = [conv_W, conv_b, att_conv_W, att_conv_b, conv_W_context] conv_model_sents = Conv_with_Mask( rng, input_tensor3=embed_input_sents, mask_matrix=sents_mask.reshape( (sents_mask.shape[0] * sents_mask.shape[1], sents_mask.shape[2])), image_shape=(batch_size * cand_size, 1, emb_size, sent_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero sent_embeddings = conv_model_sents.maxpool_vec #(batch_size*cand_size, hidden_size) # each sentence then have an embedding of length hidden_size batch_sent_emb = sent_embeddings.reshape( (batch_size, cand_size, hidden_size[0])) conv_model_claims = Conv_with_Mask( rng, input_tensor3=embed_input_claim, mask_matrix=claim_mask, image_shape=(batch_size, 1, emb_size, claim_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero claim_embeddings = conv_model_claims.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size batch_claim_emb = T.repeat(claim_embeddings.dimshuffle(0, 'x', 1), cand_size, axis=1) # concate_claim_sent = T.concatenate([batch_claim_emb,batch_sent_emb ], axis=2) # concate_2_matrix = concate_claim_sent.reshape((batch_size*cand_size, hidden_size[0]*2)) concate_claim_sent = T.concatenate([ batch_claim_emb, batch_sent_emb, T.sum(batch_claim_emb * batch_sent_emb, axis=2).dimshuffle(0, 1, 'x') ], axis=2) concate_2_matrix = concate_claim_sent.reshape( (batch_size * cand_size, hidden_size[0] * 2 + 1)) LR_input = concate_2_matrix LR_input_size = hidden_size[0] * 2 + 1 #classification layer, it is just mapping from a feature vector of size "hidden_size" to a vector of only two values: positive, negative U_a = create_ensemble_para( rng, 1, LR_input_size) # the weight matrix hidden_size*2 # LR_b = theano.shared(value=np.zeros((8,),dtype=theano.config.floatX),name='LR_b', borrow=True) #bias for each target class LR_para = [U_a] # layer_LR=LogisticRegression(rng, input=LR_input, n_in=LR_input_size, n_out=8, W=U_a, b=LR_b) #basically it is a multiplication between weight matrix and input feature vector score_matrix = T.nnet.sigmoid(LR_input.dot(U_a)) #batch * 12 inter_matrix = score_matrix.reshape((batch_size, cand_size)) # inter_sent_claim = T.batched_dot(batch_sent_emb, batch_claim_emb) #(batch_size, cand_size, 1) # inter_matrix = T.nnet.sigmoid(inter_sent_claim.reshape((batch_size, cand_size))) ''' maybe 1.0-inter_matrix can be rewritten into 1/e^(inter_matrix) ''' # prob_pos = T.where( sents_labels < 1, 1.0-inter_matrix, inter_matrix) # loss = -T.mean(T.log(prob_pos)) #f1 as loss batch_overlap = T.sum(sents_labels * inter_matrix, axis=1) batch_recall = batch_overlap / T.sum(sents_labels, axis=1) batch_precision = batch_overlap / T.sum(inter_matrix, axis=1) batch_f1 = 2.0 * batch_recall * batch_precision / (batch_recall + batch_precision) loss = -T.mean(T.log(batch_f1)) # loss = T.nnet.nnet.binary_crossentropy(inter_matrix, sents_labels).mean() ''' training task2, predict 3 labels ''' joint_embed_input_sents = init_embeddings[joint_sents_ids.flatten( )].reshape((batch_size * cand_size, sent_len, emb_size)).dimshuffle( 0, 2, 1 ) #embed_input(init_embeddings, sents_ids_l)#embeddings[sents_ids_l.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1) #the input format can be adapted into CNN or GRU or LSTM joint_embed_input_claim = init_embeddings[ joint_claim_ids.flatten()].reshape( (batch_size, claim_len, emb_size)).dimshuffle(0, 2, 1) joint_conv_model_sents = Conv_with_Mask( rng, input_tensor3=joint_embed_input_sents, mask_matrix=joint_sents_mask.reshape( (joint_sents_mask.shape[0] * joint_sents_mask.shape[1], joint_sents_mask.shape[2])), image_shape=(batch_size * cand_size, 1, emb_size, sent_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero joint_sent_embeddings = joint_conv_model_sents.maxpool_vec #(batch_size*cand_size, hidden_size) # each sentence then have an embedding of length hidden_size joint_batch_sent_emb = joint_sent_embeddings.reshape( (batch_size, cand_size, hidden_size[0])) joint_premise_emb = T.sum(joint_batch_sent_emb * joint_sents_labels.dimshuffle(0, 1, 'x'), axis=1) #(batch, hidden_size) joint_conv_model_claims = Conv_with_Mask( rng, input_tensor3=joint_embed_input_claim, mask_matrix=joint_claim_mask, image_shape=(batch_size, 1, emb_size, claim_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero joint_claim_embeddings = joint_conv_model_claims.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size joint_premise_hypo_emb = T.concatenate( [joint_premise_emb, joint_claim_embeddings], axis=1) #(batch, 2*hidden_size) ''' attentive conv in task2 ''' joint_sents_tensor3 = joint_embed_input_sents.dimshuffle(0, 2, 1).reshape( (batch_size, cand_size * sent_len, emb_size)) joint_sents_dot = T.batched_dot( joint_sents_tensor3, joint_sents_tensor3.dimshuffle( 0, 2, 1)) #(batch_size, cand_size*sent_len, cand_size*sent_len) joint_sents_dot_2_matrix = T.nnet.softmax( joint_sents_dot.reshape( (batch_size * cand_size * sent_len, cand_size * sent_len))) joint_sents_context = T.batched_dot( joint_sents_dot_2_matrix.reshape( (batch_size, cand_size * sent_len, cand_size * sent_len)), joint_sents_tensor3) #(batch_size, cand_size*sent_len, emb_size) joint_add_sents_context = joint_embed_input_sents + joint_sents_context.reshape( (batch_size * cand_size, sent_len, emb_size) ).dimshuffle( 0, 2, 1 ) #T.concatenate([joint_embed_input_sents, joint_sents_context.reshape((batch_size*cand_size, sent_len, emb_size)).dimshuffle(0,2,1)], axis=1) #(batch_size*cand_size, 2*emb_size, sent_len) attentive_conv_layer = Attentive_Conv_for_Pair_easy_version( rng, input_tensor3= joint_add_sents_context, #batch_size*cand_size, 2*emb_size, sent_len input_tensor3_r=T.repeat(joint_embed_input_claim, cand_size, axis=0), mask_matrix=joint_sents_mask.reshape( (joint_sents_mask.shape[0] * joint_sents_mask.shape[1], joint_sents_mask.shape[2])), mask_matrix_r=T.repeat(joint_claim_mask, cand_size, axis=0), image_shape=(batch_size * cand_size, 1, emb_size, sent_len), image_shape_r=(batch_size * cand_size, 1, emb_size, claim_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), filter_shape_context=(hidden_size[0], 1, emb_size, 1), W=att_conv_W, b=att_conv_b, W_context=conv_W_context, b_context=conv_b_context) attentive_sent_embeddings_l = attentive_conv_layer.attentive_maxpool_vec_l.reshape( (batch_size, cand_size, hidden_size[0])) #(batch_size*cand_size, hidden_size) attentive_sent_embeddings_r = attentive_conv_layer.attentive_maxpool_vec_r.reshape( (batch_size, cand_size, hidden_size[0])) masked_sents_attconv = attentive_sent_embeddings_l * joint_sents_labels.dimshuffle( 0, 1, 'x') masked_claim_attconv = attentive_sent_embeddings_r * joint_sents_labels.dimshuffle( 0, 1, 'x') fine_max = T.concatenate([ T.max(masked_sents_attconv, axis=1), T.max(masked_claim_attconv, axis=1) ], axis=1) #(batch, 2*hidden) # fine_sum = T.concatenate([T.sum(masked_sents_attconv, axis=1),T.sum(masked_claim_attconv, axis=1)],axis=1) #(batch, 2*hidden) "Logistic Regression layer" joint_LR_input = T.concatenate([joint_premise_hypo_emb, fine_max], axis=1) joint_LR_input_size = 2 * hidden_size[0] + 2 * hidden_size[0] joint_U_a = create_ensemble_para(rng, 3, joint_LR_input_size) # (input_size, 3) joint_LR_b = theano.shared(value=np.zeros((3, ), dtype=theano.config.floatX), name='LR_b', borrow=True) #bias for each target class joint_LR_para = [joint_U_a, joint_LR_b] joint_layer_LR = LogisticRegression( rng, input=joint_LR_input, n_in=joint_LR_input_size, n_out=3, W=joint_U_a, b=joint_LR_b ) #basically it is a multiplication between weight matrix and input feature vector joint_loss = joint_layer_LR.negative_log_likelihood( joint_labels ) #for classification task, we usually used negative log likelihood as loss, the lower the better. ''' testing ''' # binarize_prob = T.where( inter_matrix > 0.5, 1, 0) #(batch_size, cand_size masked_inter_matrix = inter_matrix * sents_labels #(batch, cand_size) test_premise_emb = T.sum(batch_sent_emb * masked_inter_matrix.dimshuffle(0, 1, 'x'), axis=1) test_premise_hypo_emb = T.concatenate([test_premise_emb, claim_embeddings], axis=1) #fine-maxsum sents_tensor3 = embed_input_sents.dimshuffle(0, 2, 1).reshape( (batch_size, cand_size * sent_len, emb_size)) sents_dot = T.batched_dot(sents_tensor3, sents_tensor3.dimshuffle( 0, 2, 1)) #(batch_size, cand_size*sent_len, cand_size*sent_len) sents_dot_2_matrix = T.nnet.softmax( sents_dot.reshape( (batch_size * cand_size * sent_len, cand_size * sent_len))) sents_context = T.batched_dot( sents_dot_2_matrix.reshape( (batch_size, cand_size * sent_len, cand_size * sent_len)), sents_tensor3) #(batch_size, cand_size*sent_len, emb_size) add_sents_context = embed_input_sents + sents_context.reshape( (batch_size * cand_size, sent_len, emb_size) ).dimshuffle( 0, 2, 1 ) #T.concatenate([embed_input_sents, sents_context.reshape((batch_size*cand_size, sent_len, emb_size)).dimshuffle(0,2,1)], axis=1) #(batch_size*cand_size, 2*emb_size, sent_len) test_attentive_conv_layer = Attentive_Conv_for_Pair_easy_version( rng, input_tensor3= add_sents_context, #batch_size*cand_size, 2*emb_size, sent_len input_tensor3_r=T.repeat(embed_input_claim, cand_size, axis=0), mask_matrix=sents_mask.reshape( (sents_mask.shape[0] * sents_mask.shape[1], sents_mask.shape[2])), mask_matrix_r=T.repeat(claim_mask, cand_size, axis=0), image_shape=(batch_size * cand_size, 1, emb_size, sent_len), image_shape_r=(batch_size * cand_size, 1, emb_size, claim_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), filter_shape_context=(hidden_size[0], 1, emb_size, 1), W=att_conv_W, b=att_conv_b, W_context=conv_W_context, b_context=conv_b_context) # attentive_sent_embeddings_l = attentive_conv_layer.attentive_maxpool_vec_l #(batch_size*cand_size, hidden_size) # attentive_sent_embeddings_r = attentive_conv_layer.attentive_maxpool_vec_r test_attentive_sent_embeddings_l = test_attentive_conv_layer.attentive_maxpool_vec_l.reshape( (batch_size, cand_size, hidden_size[0])) #(batch_size*cand_size, hidden_size) test_attentive_sent_embeddings_r = test_attentive_conv_layer.attentive_maxpool_vec_r.reshape( (batch_size, cand_size, hidden_size[0])) test_masked_sents_attconv = test_attentive_sent_embeddings_l * masked_inter_matrix.dimshuffle( 0, 1, 'x') test_masked_claim_attconv = test_attentive_sent_embeddings_r * masked_inter_matrix.dimshuffle( 0, 1, 'x') test_fine_max = T.concatenate([ T.max(test_masked_sents_attconv, axis=1), T.max(test_masked_claim_attconv, axis=1) ], axis=1) #(batch, 2*hidden) # test_fine_sum = T.concatenate([T.sum(test_masked_sents_attconv, axis=1),T.sum(test_masked_claim_attconv, axis=1)],axis=1) #(batch, 2*hidden) test_LR_input = T.concatenate([test_premise_hypo_emb, test_fine_max], axis=1) test_LR_input_size = joint_LR_input_size test_layer_LR = LogisticRegression( rng, input=test_LR_input, n_in=test_LR_input_size, n_out=3, W=joint_U_a, b=joint_LR_b ) #basically it is a multiplication between weight matrix and input feature vector params = [init_embeddings] + NN_para + LR_para + joint_LR_para cost = loss + joint_loss "Use AdaGrad to update parameters" updates = Gradient_Cost_Para(cost, params, learning_rate) train_model = theano.function([ sents_ids, sents_mask, sents_labels, claim_ids, claim_mask, joint_sents_ids, joint_sents_mask, joint_sents_labels, joint_claim_ids, joint_claim_mask, joint_labels ], cost, updates=updates, allow_input_downcast=True, on_unused_input='ignore') # dev_model = theano.function([sents_ids_l, sents_mask_l, sents_ids_r, sents_mask_r, labels], layer_LR.errors(labels), allow_input_downcast=True, on_unused_input='ignore') test_model = theano.function([ sents_ids, sents_mask, sents_labels, claim_ids, claim_mask, joint_labels ], [ inter_matrix, test_layer_LR.errors(joint_labels), test_layer_LR.y_pred ], allow_input_downcast=True, on_unused_input='ignore') ############### # TRAIN MODEL # ############### print '... training' # early-stopping parameters patience = 50000000000 # look as this many examples regardless start_time = time.time() mid_time = start_time past_time = mid_time epoch = 0 done_looping = False joint_n_train_batches = joint_train_size / batch_size joint_train_batch_start = list( np.arange(joint_n_train_batches) * batch_size) + [joint_train_size - batch_size] n_train_batches = train_size / batch_size train_batch_start = list( np.arange(n_train_batches) * batch_size) + [train_size - batch_size] n_test_batches = test_size / batch_size test_batch_start = list( np.arange(n_test_batches) * batch_size) + [test_size - batch_size] n_test_3th_batches = test_3th_size / batch_size test_3th_batch_start = list(np.arange(n_test_3th_batches) * batch_size) + [test_3th_size - batch_size] max_acc = 0.0 max_test_f1 = 0.0 max_acc_full_evi = 0.0 cost_i = 0.0 joint_train_indices = range(joint_train_size) train_indices = range(train_size) while epoch < n_epochs: epoch = epoch + 1 random.Random(100).shuffle( joint_train_indices ) #shuffle training set for each new epoch, is supposed to promote performance, but not garrenteed random.Random(100).shuffle(train_indices) iter_accu = 0 for joint_batch_id in joint_train_batch_start: #for each batch # iter means how many batches have been run, taking into loop iter = (epoch - 1) * joint_n_train_batches + iter_accu + 1 iter_accu += 1 joint_train_id_batch = joint_train_indices[ joint_batch_id:joint_batch_id + batch_size] for i in range(3): batch_id = random.choice(train_batch_start) train_id_batch = train_indices[batch_id:batch_id + batch_size] cost_i += train_model( train_sents[train_id_batch], train_sent_masks[train_id_batch], train_sent_labels[train_id_batch], train_claims[train_id_batch], train_claim_mask[train_id_batch], #joint_sents_ids,joint_sents_mask,joint_sents_labels, joint_claim_ids, joint_claim_mask, joint_labels joint_train_sents[joint_train_id_batch], joint_train_sent_masks[joint_train_id_batch], joint_train_sent_labels[joint_train_id_batch], joint_train_claims[joint_train_id_batch], joint_train_claim_mask[joint_train_id_batch], joint_train_labels[joint_train_id_batch]) #after each 1000 batches, we test the performance of the model on all test data # if (epoch==1 and iter%1000==0) or (epoch>=2 and iter%5==0): if iter % 100 == 0: print 'Epoch ', epoch, 'iter ' + str( iter) + ' average cost: ' + str(cost_i / iter), 'uses ', ( time.time() - past_time) / 60.0, 'min' past_time = time.time() f1_sum = 0.0 error_sum = 0.0 full_evi = 0 predictions = [] for test_batch_id in test_batch_start: # for each test batch batch_prob, error_i, pred_i = test_model( test_sents[test_batch_id:test_batch_id + batch_size], test_sent_masks[test_batch_id:test_batch_id + batch_size], test_sent_labels[test_batch_id:test_batch_id + batch_size], test_claims[test_batch_id:test_batch_id + batch_size], test_claim_mask[test_batch_id:test_batch_id + batch_size], test_labels[test_batch_id:test_batch_id + batch_size]) error_sum += error_i batch_sent_labels = test_sent_labels[ test_batch_id:test_batch_id + batch_size] batch_sent_names = test_sent_names[ test_batch_id:test_batch_id + batch_size] batch_ground_names = test_ground_names[ test_batch_id:test_batch_id + batch_size] batch_ground_labels = test_labels[ test_batch_id:test_batch_id + batch_size] for i in range(batch_size): instance_i = {} instance_i['label'] = pred_id2label.get( batch_ground_labels[i]) instance_i['predicted_label'] = pred_id2label.get( pred_i[i]) pred_sent_names = [] gold_sent_names = batch_ground_names[i] zipped = [(batch_prob[i, k], batch_sent_labels[i][k], batch_sent_names[i][k]) for k in range(cand_size)] sorted_zip = sorted(zipped, key=lambda x: x[0], reverse=True) for j in range(cand_size): triple = sorted_zip[j] if triple[1] == 1.0: ''' we should consider a rank, instead of binary if triple[0] >0.5: can control the recall, influence the strict_acc ''' if triple[0] > 0.5: # pred_sent_names.append(batch_sent_names[i][j]) pred_sent_names.append(triple[2]) # if len(pred_sent_names) == max_pred_pick: # break instance_i['predicted_evidence'] = pred_sent_names # print 'pred_sent_names:',pred_sent_names # print 'gold_sent_names:',gold_sent_names new_gold_names = [] for gold_name in gold_sent_names: new_gold_names.append([None, None] + gold_name) instance_i['evidence'] = [new_gold_names] predictions.append(instance_i) strict_score, label_accuracy, precision, recall, f1 = fever_score( predictions) print 'strict_score, label_accuracy, precision, recall, f1: ', strict_score, label_accuracy, precision, recall, f1 # test_f1=f1_sum/(len(test_batch_start)*batch_size) for test_batch_id in test_3th_batch_start: # for each test batch _, error_i, pred_i = test_model( test_3th_sents[test_batch_id:test_batch_id + batch_size], test_3th_sent_masks[test_batch_id:test_batch_id + batch_size], test_3th_sent_labels[test_batch_id:test_batch_id + batch_size], test_3th_claims[test_batch_id:test_batch_id + batch_size], test_3th_claim_mask[test_batch_id:test_batch_id + batch_size], test_3th_labels[test_batch_id:test_batch_id + batch_size]) for i in range(batch_size): instance_i = {} instance_i['label'] = pred_id2label.get(2) instance_i['predicted_label'] = pred_id2label.get( pred_i[i]) instance_i['predicted_evidence'] = [] instance_i['evidence'] = [] predictions.append(instance_i) strict_score, label_accuracy, precision, recall, f1 = fever_score( predictions) print 'strict_score, label_accuracy, precision, recall, f1: ', strict_score, label_accuracy, precision, recall, f1 print 'Epoch ', epoch, 'uses ', (time.time() - mid_time) / 60.0, 'min' mid_time = time.time() #print 'Batch_size: ', update_freq end_time = time.time() print >> sys.stderr, ('The code for file ' + os.path.split(__file__)[1] + ' ran for %.2fm' % ((end_time - start_time) / 60.)) return max_acc_test
def evaluate_lenet5(learning_rate=0.01, n_epochs=4, emb_size=300, batch_size=50, describ_max_len=20, type_size=12, filter_size=[3, 5], maxSentLen=200, hidden_size=[300, 300]): model_options = locals().copy() print "model options", model_options emb_root = '/save/wenpeng/datasets/LORELEI/multi-lingual-emb/2018-il9-il10/multi-emb/' test_file_path = '/save/wenpeng/datasets/LORELEI/il9/il9-setE-as-test-input_ner_filtered_w2.txt' output_file_path = '/save/wenpeng/datasets/LORELEI/il9/il9_system_output_concMT_BBN_NI_epoch4.json' seed = 1234 np.random.seed(seed) rng = np.random.RandomState( seed) #random seed, control the model generates the same results srng = T.shared_randomstreams.RandomStreams(rng.randint(seed)) word2id = {} # all_sentences, all_masks, all_labels, all_other_labels, word2id=load_BBN_il5Trans_il5_dataset(maxlen=maxSentLen) #minlen, include one label, at least one word in the sentence train_p1_sents, train_p1_masks, train_p1_labels, word2id = load_trainingData_types( word2id, maxSentLen) train_p2_sents, train_p2_masks, train_p2_labels, train_p2_other_labels, word2id = load_trainingData_types_plus_others( word2id, maxSentLen) test_sents, test_masks, test_lines, word2id = load_official_testData_il_and_MT( word2id, maxSentLen, test_file_path) label_sent, label_mask = load_SF_type_descriptions(word2id, type_size, describ_max_len) label_sent = np.asarray(label_sent, dtype='int32') label_mask = np.asarray(label_mask, dtype=theano.config.floatX) train_p1_sents = np.asarray(train_p1_sents, dtype='int32') train_p1_masks = np.asarray(train_p1_masks, dtype=theano.config.floatX) train_p1_labels = np.asarray(train_p1_labels, dtype='int32') train_p1_size = len(train_p1_labels) train_p2_sents = np.asarray(train_p2_sents, dtype='int32') train_p2_masks = np.asarray(train_p2_masks, dtype=theano.config.floatX) train_p2_labels = np.asarray(train_p2_labels, dtype='int32') train_p2_other_labels = np.asarray(train_p2_other_labels, dtype='int32') train_p2_size = len(train_p2_labels) ''' combine train_p1 and train_p2 ''' train_sents = np.concatenate([train_p1_sents, train_p2_sents], axis=0) train_masks = np.concatenate([train_p1_masks, train_p2_masks], axis=0) train_labels = np.concatenate([train_p1_labels, train_p2_labels], axis=0) train_size = train_p1_size + train_p2_size test_sents = np.asarray(test_sents, dtype='int32') test_masks = np.asarray(test_masks, dtype=theano.config.floatX) # test_labels=np.asarray(all_labels[2], dtype='int32') test_size = len(test_sents) vocab_size = len(word2id) + 1 # add one zero pad index rand_values = rng.normal( 0.0, 0.01, (vocab_size, emb_size)) #generate a matrix by Gaussian distribution rand_values[0] = np.array(np.zeros(emb_size), dtype=theano.config.floatX) id2word = {y: x for x, y in word2id.iteritems()} word2vec = load_fasttext_multiple_word2vec_given_file([ emb_root + '100k-ENG-multicca.300.ENG.txt', emb_root + '100k-IL9-multicca.d300.IL9.txt' ], 300) rand_values = load_word2vec_to_init(rand_values, id2word, word2vec) embeddings = theano.shared( value=np.array(rand_values, dtype=theano.config.floatX), borrow=True ) #wrap up the python variable "rand_values" into theano variable #now, start to build the input form of the model sents_id_matrix = T.imatrix('sents_id_matrix') sents_mask = T.fmatrix('sents_mask') labels = T.imatrix('labels') #batch*12 other_labels = T.imatrix() #batch*4 des_id_matrix = T.imatrix() des_mask = T.fmatrix() ###################### # BUILD ACTUAL MODEL # ###################### print '... building the model' common_input = embeddings[sents_id_matrix.flatten()].reshape( (batch_size, maxSentLen, emb_size)).dimshuffle( 0, 2, 1) #the input format can be adapted into CNN or GRU or LSTM bow_emb = T.sum(common_input * sents_mask.dimshuffle(0, 'x', 1), axis=2) repeat_common_input = T.repeat( normalize_tensor3_colwise(common_input), type_size, axis=0) #(batch_size*type_size, emb_size, maxsentlen) des_input = embeddings[des_id_matrix.flatten()].reshape( (type_size, describ_max_len, emb_size)).dimshuffle(0, 2, 1) bow_des = T.sum(des_input * des_mask.dimshuffle(0, 'x', 1), axis=2) #(tyope_size, emb_size) repeat_des_input = T.tile( normalize_tensor3_colwise(des_input), (batch_size, 1, 1)) #(batch_size*type_size, emb_size, maxsentlen) conv_W, conv_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) conv_W2, conv_b2 = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[1])) multiCNN_para = [conv_W, conv_b, conv_W2, conv_b2] conv_att_W, conv_att_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) conv_W_context, conv_b_context = create_conv_para( rng, filter_shape=(hidden_size[0], 1, emb_size, 1)) conv_att_W2, conv_att_b2 = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[1])) conv_W_context2, conv_b_context2 = create_conv_para( rng, filter_shape=(hidden_size[0], 1, emb_size, 1)) ACNN_para = [ conv_att_W, conv_att_b, conv_W_context, conv_att_W2, conv_att_b2, conv_W_context2 ] ''' multi-CNN ''' conv_model = Conv_with_Mask( rng, input_tensor3=common_input, mask_matrix=sents_mask, image_shape=(batch_size, 1, emb_size, maxSentLen), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero sent_embeddings = conv_model.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size conv_model2 = Conv_with_Mask( rng, input_tensor3=common_input, mask_matrix=sents_mask, image_shape=(batch_size, 1, emb_size, maxSentLen), filter_shape=(hidden_size[0], 1, emb_size, filter_size[1]), W=conv_W2, b=conv_b2 ) #mutiple mask with the conv_out to set the features by UNK to zero sent_embeddings2 = conv_model2.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size ''' GRU ''' U1, W1, b1 = create_GRU_para(rng, emb_size, hidden_size[0]) GRU_NN_para = [ U1, W1, b1 ] #U1 includes 3 matrices, W1 also includes 3 matrices b1 is bias # gru_input = common_input.dimshuffle((0,2,1)) #gru requires input (batch_size, emb_size, maxSentLen) gru_layer = GRU_Batch_Tensor_Input_with_Mask(common_input, sents_mask, hidden_size[0], U1, W1, b1) gru_sent_embeddings = gru_layer.output_sent_rep # (batch_size, hidden_size) ''' ACNN ''' attentive_conv_layer = Attentive_Conv_for_Pair( rng, origin_input_tensor3=common_input, origin_input_tensor3_r=common_input, input_tensor3=common_input, input_tensor3_r=common_input, mask_matrix=sents_mask, mask_matrix_r=sents_mask, image_shape=(batch_size, 1, emb_size, maxSentLen), image_shape_r=(batch_size, 1, emb_size, maxSentLen), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), filter_shape_context=(hidden_size[0], 1, emb_size, 1), W=conv_att_W, b=conv_att_b, W_context=conv_W_context, b_context=conv_b_context) sent_att_embeddings = attentive_conv_layer.attentive_maxpool_vec_l attentive_conv_layer2 = Attentive_Conv_for_Pair( rng, origin_input_tensor3=common_input, origin_input_tensor3_r=common_input, input_tensor3=common_input, input_tensor3_r=common_input, mask_matrix=sents_mask, mask_matrix_r=sents_mask, image_shape=(batch_size, 1, emb_size, maxSentLen), image_shape_r=(batch_size, 1, emb_size, maxSentLen), filter_shape=(hidden_size[0], 1, emb_size, filter_size[1]), filter_shape_context=(hidden_size[0], 1, emb_size, 1), W=conv_att_W2, b=conv_att_b2, W_context=conv_W_context2, b_context=conv_b_context2) sent_att_embeddings2 = attentive_conv_layer2.attentive_maxpool_vec_l ''' cross-DNN-dataless ''' #first map label emb into hidden space HL_layer_1_W, HL_layer_1_b = create_HiddenLayer_para( rng, emb_size, hidden_size[0]) HL_layer_1_params = [HL_layer_1_W, HL_layer_1_b] HL_layer_1 = HiddenLayer(rng, input=bow_des, n_in=emb_size, n_out=hidden_size[0], W=HL_layer_1_W, b=HL_layer_1_b, activation=T.tanh) des_rep_hidden = HL_layer_1.output #(type_size, hidden_size) dot_dnn_dataless_1 = T.tanh(sent_embeddings.dot( des_rep_hidden.T)) #(batch_size, type_size) dot_dnn_dataless_2 = T.tanh(sent_embeddings2.dot(des_rep_hidden.T)) ''' dataless cosine ''' cosine_scores = normalize_matrix_rowwise(bow_emb).dot( normalize_matrix_rowwise(bow_des).T) cosine_score_matrix = T.nnet.sigmoid( cosine_scores) #(batch_size, type_size) ''' dataless top-30 fine grained cosine ''' fine_grained_cosine = T.batched_dot( repeat_common_input.dimshuffle(0, 2, 1), repeat_des_input) #(batch_size*type_size,maxsentlen,describ_max_len) fine_grained_cosine_to_matrix = fine_grained_cosine.reshape( (batch_size * type_size, maxSentLen * describ_max_len)) sort_fine_grained_cosine_to_matrix = T.sort(fine_grained_cosine_to_matrix, axis=1) top_k_simi = sort_fine_grained_cosine_to_matrix[:, -30:] # (batch_size*type_size, 5) max_fine_grained_cosine = T.mean(top_k_simi, axis=1) top_k_cosine_scores = max_fine_grained_cosine.reshape( (batch_size, type_size)) top_k_score_matrix = T.nnet.sigmoid(top_k_cosine_scores) acnn_LR_input = T.concatenate([ dot_dnn_dataless_1, dot_dnn_dataless_2, cosine_score_matrix, top_k_score_matrix, sent_embeddings, sent_embeddings2, gru_sent_embeddings, sent_att_embeddings, sent_att_embeddings2, bow_emb ], axis=1) acnn_LR_input_size = hidden_size[0] * 5 + emb_size + 4 * type_size #classification layer, it is just mapping from a feature vector of size "hidden_size" to a vector of only two values: positive, negative acnn_U_a, acnn_LR_b = create_LR_para(rng, acnn_LR_input_size, 12) acnn_LR_para = [acnn_U_a, acnn_LR_b] acnn_layer_LR = LogisticRegression( rng, input=acnn_LR_input, n_in=acnn_LR_input_size, n_out=12, W=acnn_U_a, b=acnn_LR_b ) #basically it is a multiplication between weight matrix and input feature vector acnn_score_matrix = T.nnet.sigmoid( acnn_layer_LR.before_softmax) #batch * 12 acnn_prob_pos = T.where(labels < 1, 1.0 - acnn_score_matrix, acnn_score_matrix) acnn_loss = -T.mean(T.log(acnn_prob_pos)) acnn_other_U_a, acnn_other_LR_b = create_LR_para(rng, acnn_LR_input_size, 16) acnn_other_LR_para = [acnn_other_U_a, acnn_other_LR_b] acnn_other_layer_LR = LogisticRegression(rng, input=acnn_LR_input, n_in=acnn_LR_input_size, n_out=16, W=acnn_other_U_a, b=acnn_other_LR_b) acnn_other_prob_matrix = T.nnet.softmax( acnn_other_layer_LR.before_softmax.reshape((batch_size * 4, 4))) acnn_other_prob_tensor3 = acnn_other_prob_matrix.reshape( (batch_size, 4, 4)) acnn_other_prob = acnn_other_prob_tensor3[ T.repeat(T.arange(batch_size), 4), T.tile(T.arange(4), (batch_size)), other_labels.flatten()] acnn_other_field_loss = -T.mean(T.log(acnn_other_prob)) params = multiCNN_para + GRU_NN_para + ACNN_para + acnn_LR_para + HL_layer_1_params # put all model parameters together cost = acnn_loss + 1e-4 * ((conv_W**2).sum() + (conv_W2**2).sum() + (conv_att_W**2).sum() + (conv_att_W2**2).sum()) updates = Gradient_Cost_Para(cost, params, learning_rate) other_paras = params + acnn_other_LR_para cost_other = cost + acnn_other_field_loss other_updates = Gradient_Cost_Para(cost_other, other_paras, learning_rate) ''' testing ''' ensemble_NN_scores = acnn_score_matrix #T.max(T.concatenate([att_score_matrix.dimshuffle('x',0,1), score_matrix.dimshuffle('x',0,1), acnn_score_matrix.dimshuffle('x',0,1)],axis=0),axis=0) # ''' # majority voting, does not work # ''' # binarize_NN = T.where(ensemble_NN_scores > 0.5, 1, 0) # binarize_dataless = T.where(cosine_score_matrix > 0.5, 1, 0) # binarize_dataless_finegrained = T.where(top_k_score_matrix > 0.5, 1, 0) # binarize_conc = T.concatenate([binarize_NN.dimshuffle('x',0,1), binarize_dataless.dimshuffle('x',0,1),binarize_dataless_finegrained.dimshuffle('x',0,1)],axis=0) # sum_binarize_conc = T.sum(binarize_conc,axis=0) # binarize_prob = T.where(sum_binarize_conc > 0.0, 1, 0) # ''' # sum up prob, works # ''' # ensemble_scores_1 = 0.6*ensemble_NN_scores+0.4*top_k_score_matrix # binarize_prob = T.where(ensemble_scores_1 > 0.3, 1, 0) ''' sum up prob, works ''' ensemble_scores = ensemble_NN_scores #0.6*ensemble_NN_scores+0.4*0.5*(cosine_score_matrix+top_k_score_matrix) binarize_prob = T.where(ensemble_scores > 0.3, 1, 0) ''' test for other fields ''' sum_tensor3 = acnn_other_prob_tensor3 #(batch, 4, 3) #train_model = theano.function([sents_id_matrix, sents_mask, labels], cost, updates=updates, on_unused_input='ignore') train_p1_model = theano.function( [sents_id_matrix, sents_mask, labels, des_id_matrix, des_mask], cost, updates=updates, allow_input_downcast=True, on_unused_input='ignore') train_p2_model = theano.function([ sents_id_matrix, sents_mask, labels, des_id_matrix, des_mask, other_labels ], cost_other, updates=other_updates, allow_input_downcast=True, on_unused_input='ignore') test_model = theano.function( [sents_id_matrix, sents_mask, des_id_matrix, des_mask], [binarize_prob, ensemble_scores, sum_tensor3], allow_input_downcast=True, on_unused_input='ignore') ############### # TRAIN MODEL # ############### print '... training' # early-stopping parameters patience = 50000000000 # look as this many examples regardless start_time = time.time() mid_time = start_time past_time = mid_time epoch = 0 done_looping = False n_train_batches = train_size / batch_size train_batch_start = list( np.arange(n_train_batches) * batch_size) + [train_size - batch_size] n_train_p2_batches = train_p2_size / batch_size train_p2_batch_start = list(np.arange(n_train_p2_batches) * batch_size) + [train_p2_size - batch_size] n_test_batches = test_size / batch_size n_test_remain = test_size % batch_size test_batch_start = list( np.arange(n_test_batches) * batch_size) + [test_size - batch_size] train_p2_batch_start_set = set(train_p2_batch_start) # max_acc_dev=0.0 # max_meanf1_test=0.0 # max_weightf1_test=0.0 train_indices = range(train_size) train_p2_indices = range(train_p2_size) cost_i = 0.0 other_cost_i = 0.0 min_mean_frame = 100.0 while epoch < n_epochs: epoch = epoch + 1 random.Random(100).shuffle(train_indices) random.Random(100).shuffle(train_p2_indices) iter_accu = 0 for batch_id in train_batch_start: #for each batch # iter means how many batches have been run, taking into loop iter = (epoch - 1) * n_train_batches + iter_accu + 1 iter_accu += 1 train_id_batch = train_indices[batch_id:batch_id + batch_size] cost_i += train_p1_model(train_sents[train_id_batch], train_masks[train_id_batch], train_labels[train_id_batch], label_sent, label_mask) if batch_id in train_p2_batch_start_set: train_p2_id_batch = train_p2_indices[batch_id:batch_id + batch_size] other_cost_i += train_p2_model( train_p2_sents[train_p2_id_batch], train_p2_masks[train_p2_id_batch], train_p2_labels[train_p2_id_batch], label_sent, label_mask, train_p2_other_labels[train_p2_id_batch]) # else: # random_batch_id = random.choice(train_p2_batch_start) # train_p2_id_batch = train_p2_indices[random_batch_id:random_batch_id+batch_size] # other_cost_i+=train_p2_model( # train_p2_sents[train_p2_id_batch], # train_p2_masks[train_p2_id_batch], # train_p2_labels[train_p2_id_batch], # label_sent, # label_mask, # train_p2_other_labels[train_p2_id_batch] # ) #after each 1000 batches, we test the performance of the model on all test data if iter % 20 == 0: print 'Epoch ', epoch, 'iter ' + str( iter) + ' average cost: ' + str(cost_i / iter), str( other_cost_i / iter), 'uses ', (time.time() - past_time) / 60.0, 'min' past_time = time.time() pred_types = [] pred_confs = [] pred_others = [] for i, test_batch_id in enumerate( test_batch_start): # for each test batch pred_types_i, pred_conf_i, pred_fields_i = test_model( test_sents[test_batch_id:test_batch_id + batch_size], test_masks[test_batch_id:test_batch_id + batch_size], label_sent, label_mask) if i < len(test_batch_start) - 1: pred_types.append(pred_types_i) pred_confs.append(pred_conf_i) pred_others.append(pred_fields_i) else: pred_types.append(pred_types_i[-n_test_remain:]) pred_confs.append(pred_conf_i[-n_test_remain:]) pred_others.append(pred_fields_i[-n_test_remain:]) pred_types = np.concatenate(pred_types, axis=0) pred_confs = np.concatenate(pred_confs, axis=0) pred_others = np.concatenate(pred_others, axis=0) mean_frame = generate_2018_official_output( test_lines, output_file_path, pred_types, pred_confs, pred_others, min_mean_frame) if mean_frame < min_mean_frame: min_mean_frame = mean_frame print '\t\t\t test over, min_mean_frame:', min_mean_frame print 'Epoch ', epoch, 'uses ', (time.time() - mid_time) / 60.0, 'min' mid_time = time.time() #print 'Batch_size: ', update_freq end_time = time.time() print >> sys.stderr, ('The code for file ' + os.path.split(__file__)[1] + ' ran for %.2fm' % ((end_time - start_time) / 60.))
def evaluate_lenet5(claim, title2sentlist, title2wordlist, word2id): learning_rate = 0.02 n_epochs = 100 emb_size = 300 batch_size = 1 #50 filter_size = [3] sent_len = 40 claim_len = 40 cand_size = 10 hidden_size = [300, 300] max_pred_pick = 5 # model_options = locals().copy() # print("model options", model_options) # print('title2sentlist len', len(title2sentlist)) # print('title2wordlist len', len(title2wordlist)) pred_id2label = {1: 'SUPPORTS', 0: 'REFUTES', 2: 'NOT ENOUGH INFO'} seed = 1234 np.random.seed(seed) rng = np.random.RandomState( seed) #random seed, control the model generates the same results srng = T.shared_randomstreams.RandomStreams(rng.randint(seed)) claim_idlist, claim_masklist, sent_ins_ids, sent_ins_mask, sent_cand_list = claim_input_2_theano_input( claim, word2id, claim_len, sent_len, cand_size, title2sentlist, title2wordlist) test_claims = np.asarray([claim_idlist], dtype='int32') test_claim_mask = np.asarray([claim_masklist], dtype=theano.config.floatX) test_sents = np.asarray([sent_ins_ids], dtype='int32') test_sent_masks = np.asarray([sent_ins_mask], dtype=theano.config.floatX) vocab_size = len(word2id) + 1 rand_values = rng.normal( 0.0, 0.01, (vocab_size, emb_size)) #generate a matrix by Gaussian distribution # id2word = {y:x for x,y in word2id.items()} # word2vec=load_word2vec() # rand_values=load_word2vec_to_init(rand_values, id2word, word2vec) init_embeddings = theano.shared( value=np.array(rand_values, dtype=theano.config.floatX), borrow=True ) #wrap up the python variable "rand_values" into theano variable ''' the first block for evidence identification in two classes (support & reject) the second block for textual entailment: given evidence labels, predict the claim labels ''' sents_ids = T.itensor3() #(batch, cand_size, sent_len) sents_mask = T.ftensor3() # sents_labels=T.imatrix() #(batch, cand_size) claim_ids = T.imatrix() #(batch, claim_len) claim_mask = T.fmatrix() # joint_sents_ids=T.itensor3() #(batch, cand_size, sent_len) # joint_sents_mask=T.ftensor3() # # joint_sents_labels=T.imatrix() #(batch, cand_size) # joint_claim_ids = T.imatrix() #(batch, claim_len) # joint_claim_mask = T.fmatrix() # joint_labels=T.ivector() ###################### # BUILD ACTUAL MODEL # ###################### print('... building the model') embed_input_sents = init_embeddings[sents_ids.flatten( )].reshape((batch_size * cand_size, sent_len, emb_size)).dimshuffle( 0, 2, 1 ) #embed_input(init_embeddings, sents_ids_l)#embeddings[sents_ids_l.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1) #the input format can be adapted into CNN or GRU or LSTM embed_input_claim = init_embeddings[claim_ids.flatten()].reshape( (batch_size, claim_len, emb_size)).dimshuffle(0, 2, 1) "shared parameters" conv_W, conv_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) "tasl 1 parameters" task1_att_conv_W, task1_att_conv_b = create_conv_para( rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) task1_conv_W_context, task1_conv_b_context = create_conv_para( rng, filter_shape=(hidden_size[0], 1, emb_size, 1)) "task 2 parameters" att_conv_W, att_conv_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) conv_W_context, conv_b_context = create_conv_para( rng, filter_shape=(hidden_size[0], 1, emb_size, 1)) NN_para = [ conv_W, conv_b, task1_att_conv_W, task1_att_conv_b, att_conv_W, att_conv_b, task1_conv_W_context, conv_W_context ] conv_model_sents = Conv_with_Mask( rng, input_tensor3=embed_input_sents, mask_matrix=sents_mask.reshape( (sents_mask.shape[0] * sents_mask.shape[1], sents_mask.shape[2])), image_shape=(batch_size * cand_size, 1, emb_size, sent_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero sent_embeddings = conv_model_sents.maxpool_vec #(batch_size*cand_size, hidden_size) # each sentence then have an embedding of length hidden_size batch_sent_emb = sent_embeddings.reshape( (batch_size, cand_size, hidden_size[0])) conv_model_claims = Conv_with_Mask( rng, input_tensor3=embed_input_claim, mask_matrix=claim_mask, image_shape=(batch_size, 1, emb_size, claim_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero claim_embeddings = conv_model_claims.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size batch_claim_emb = T.repeat(claim_embeddings.dimshuffle(0, 'x', 1), cand_size, axis=1) ''' attentive conv for task1 ''' task1_attentive_conv_layer = Attentive_Conv_for_Pair_easy_version( rng, input_tensor3= embed_input_sents, #batch_size*cand_size, emb_size, sent_len input_tensor3_r=T.repeat(embed_input_claim, cand_size, axis=0), mask_matrix=sents_mask.reshape( (sents_mask.shape[0] * sents_mask.shape[1], sents_mask.shape[2])), mask_matrix_r=T.repeat(claim_mask, cand_size, axis=0), image_shape=(batch_size * cand_size, 1, emb_size, sent_len), image_shape_r=(batch_size * cand_size, 1, emb_size, claim_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), filter_shape_context=(hidden_size[0], 1, emb_size, 1), W=task1_att_conv_W, b=task1_att_conv_b, W_context=task1_conv_W_context, b_context=task1_conv_b_context) task1_attentive_sent_embeddings_l = task1_attentive_conv_layer.attentive_maxpool_vec_l #(batch_size*cand_size, hidden_size) task1_attentive_sent_embeddings_r = task1_attentive_conv_layer.attentive_maxpool_vec_r concate_claim_sent = T.concatenate([ batch_claim_emb, batch_sent_emb, T.sum(batch_claim_emb * batch_sent_emb, axis=2).dimshuffle(0, 1, 'x') ], axis=2) concate_2_matrix = concate_claim_sent.reshape( (batch_size * cand_size, hidden_size[0] * 2 + 1)) "to score each evidence sentence, we use the output of attentiveConv, as well as the output of standard CNN" LR_input = T.concatenate([ concate_2_matrix, task1_attentive_sent_embeddings_l, task1_attentive_sent_embeddings_r ], axis=1) LR_input_size = hidden_size[0] * 2 + 1 + hidden_size[0] * 2 # LR_input = concate_2_matrix # LR_input_size = hidden_size[0]*2+1 #classification layer, it is just mapping from a feature vector of size "hidden_size" to a vector of only two values: positive, negative U_a = create_ensemble_para( rng, 1, LR_input_size) # the weight matrix hidden_size*2 # LR_b = theano.shared(value=np.zeros((8,),dtype=theano.config.floatX),name='LR_b', borrow=True) #bias for each target class LR_para = [U_a] # layer_LR=LogisticRegression(rng, input=LR_input, n_in=LR_input_size, n_out=8, W=U_a, b=LR_b) #basically it is a multiplication between weight matrix and input feature vector score_matrix = T.nnet.sigmoid(LR_input.dot(U_a)) #batch * 12 inter_matrix = score_matrix.reshape((batch_size, cand_size)) # inter_sent_claim = T.batched_dot(batch_sent_emb, batch_claim_emb) #(batch_size, cand_size, 1) # inter_matrix = T.nnet.sigmoid(inter_sent_claim.reshape((batch_size, cand_size))) ''' maybe 1.0-inter_matrix can be rewritten into 1/e^(inter_matrix) ''' binarize_prob = T.where(inter_matrix > 0.5, 1, 0) #(batch_size, cand_size) sents_labels = inter_matrix * binarize_prob ''' training task2, predict 3 labels ''' # joint_embed_input_sents=init_embeddings[joint_sents_ids.flatten()].reshape((batch_size*cand_size, sent_len, emb_size)).dimshuffle(0,2,1)#embed_input(init_embeddings, sents_ids_l)#embeddings[sents_ids_l.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1) #the input format can be adapted into CNN or GRU or LSTM # joint_embed_input_claim=init_embeddings[joint_claim_ids.flatten()].reshape((batch_size,claim_len, emb_size)).dimshuffle(0,2,1) # joint_conv_model_sents = Conv_with_Mask(rng, input_tensor3=joint_embed_input_sents, # mask_matrix = joint_sents_mask.reshape((joint_sents_mask.shape[0]*joint_sents_mask.shape[1],joint_sents_mask.shape[2])), # image_shape=(batch_size*cand_size, 1, emb_size, sent_len), # filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b) #mutiple mask with the conv_out to set the features by UNK to zero # joint_sent_embeddings=joint_conv_model_sents.maxpool_vec #(batch_size*cand_size, hidden_size) # each sentence then have an embedding of length hidden_size # joint_batch_sent_emb = joint_sent_embeddings.reshape((batch_size, cand_size, hidden_size[0])) # "??? use joint_sents_labels means the evidence labels are not provided by task 1?" # joint_premise_emb = T.sum(joint_batch_sent_emb*joint_sents_labels.dimshuffle(0,1,'x'), axis=1) #(batch, hidden_size) premise_emb = T.sum(batch_sent_emb * sents_labels.dimshuffle(0, 1, 'x'), axis=1) # joint_conv_model_claims = Conv_with_Mask(rng, input_tensor3=joint_embed_input_claim, # mask_matrix = joint_claim_mask, # image_shape=(batch_size, 1, emb_size, claim_len), # filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b) #mutiple mask with the conv_out to set the features by UNK to zero # joint_claim_embeddings=joint_conv_model_claims.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size premise_hypo_emb = T.concatenate([premise_emb, claim_embeddings], axis=1) #(batch, 2*hidden_size) ''' attentive conv in task2 ''' sents_tensor3 = embed_input_sents.dimshuffle(0, 2, 1).reshape( (batch_size, cand_size * sent_len, emb_size)) sents_dot = T.batched_dot(sents_tensor3, sents_tensor3.dimshuffle( 0, 2, 1)) #(batch_size, cand_size*sent_len, cand_size*sent_len) sents_dot_2_matrix = T.nnet.softmax( sents_dot.reshape( (batch_size * cand_size * sent_len, cand_size * sent_len))) sents_context = T.batched_dot( sents_dot_2_matrix.reshape( (batch_size, cand_size * sent_len, cand_size * sent_len)), sents_tensor3) #(batch_size, cand_size*sent_len, emb_size) add_sents_context = embed_input_sents + sents_context.reshape( (batch_size * cand_size, sent_len, emb_size) ).dimshuffle( 0, 2, 1 ) #T.concatenate([joint_embed_input_sents, joint_sents_context.reshape((batch_size*cand_size, sent_len, emb_size)).dimshuffle(0,2,1)], axis=1) #(batch_size*cand_size, 2*emb_size, sent_len) attentive_conv_layer = Attentive_Conv_for_Pair_easy_version( rng, input_tensor3= add_sents_context, #batch_size*cand_size, 2*emb_size, sent_len input_tensor3_r=T.repeat(embed_input_claim, cand_size, axis=0), mask_matrix=sents_mask.reshape( (sents_mask.shape[0] * sents_mask.shape[1], sents_mask.shape[2])), mask_matrix_r=T.repeat(claim_mask, cand_size, axis=0), image_shape=(batch_size * cand_size, 1, emb_size, sent_len), image_shape_r=(batch_size * cand_size, 1, emb_size, claim_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), filter_shape_context=(hidden_size[0], 1, emb_size, 1), W=att_conv_W, b=att_conv_b, W_context=conv_W_context, b_context=conv_b_context) attentive_sent_embeddings_l = attentive_conv_layer.attentive_maxpool_vec_l.reshape( (batch_size, cand_size, hidden_size[0])) #(batch_size*cand_size, hidden_size) attentive_sent_embeddings_r = attentive_conv_layer.attentive_maxpool_vec_r.reshape( (batch_size, cand_size, hidden_size[0])) masked_sents_attconv = attentive_sent_embeddings_l * sents_labels.dimshuffle( 0, 1, 'x') masked_claim_attconv = attentive_sent_embeddings_r * sents_labels.dimshuffle( 0, 1, 'x') fine_max = T.concatenate([ T.max(masked_sents_attconv, axis=1), T.max(masked_claim_attconv, axis=1) ], axis=1) #(batch, 2*hidden) # fine_sum = T.concatenate([T.sum(masked_sents_attconv, axis=1),T.sum(masked_claim_attconv, axis=1)],axis=1) #(batch, 2*hidden) "Logistic Regression layer" joint_LR_input = T.concatenate([premise_hypo_emb, fine_max], axis=1) joint_LR_input_size = 2 * hidden_size[0] + 2 * hidden_size[0] joint_U_a = create_ensemble_para(rng, 3, joint_LR_input_size) # (input_size, 3) joint_LR_b = theano.shared(value=np.zeros((3, ), dtype=theano.config.floatX), name='LR_b', borrow=True) #bias for each target class joint_LR_para = [joint_U_a, joint_LR_b] joint_layer_LR = LogisticRegression( rng, input=joint_LR_input, n_in=joint_LR_input_size, n_out=3, W=joint_U_a, b=joint_LR_b ) #basically it is a multiplication between weight matrix and input feature vector # joint_loss=joint_layer_LR.negative_log_likelihood(joint_labels) #for classification task, we usually used negative log likelihood as loss, the lower the better. params = [init_embeddings] + NN_para + LR_para + joint_LR_para print('initialze model parameters...') load_model_from_file( '/home1/w/wenpeng/dataset/FEVER/model_para_0.9936287838053803', params) # train_model = theano.function([sents_ids,sents_mask,sents_labels,claim_ids,claim_mask,joint_sents_ids,joint_sents_mask,joint_sents_labels, joint_claim_ids, joint_claim_mask, joint_labels], cost, updates=updates, allow_input_downcast=True, on_unused_input='ignore') test_model = theano.function( [sents_ids, sents_mask, claim_ids, claim_mask], [inter_matrix, binarize_prob, joint_layer_LR.y_pred], allow_input_downcast=True, on_unused_input='ignore') # dev_model = theano.function([sents_ids,sents_mask, claim_ids,claim_mask], [binarize_prob,joint_layer_LR.y_pred], allow_input_downcast=True, on_unused_input='ignore') ############### # TRAIN MODEL # ############### print('... testing') # early-stopping parameters batch_score_vec, batch_binary_vec, pred_i = test_model( test_sents, test_sent_masks, test_claims, test_claim_mask) sorted_indices = np.argsort(batch_score_vec[0])[::-1] #descending order selected_sents = [] for index in sorted_indices: if batch_binary_vec[0][index] == 1: selected_sents.append(sent_cand_list[index]) if len(selected_sents) == 5: break # for i, indicator in enumerate(list(batch_binary_vec[0])): # if indicator == 1: # selected_sents.append(sent_cand_list[i]) return pred_id2label.get( pred_i[0]) + '"<p>"' + '"<br />"'.join(selected_sents) + '"<p/>"'
def evaluate_lenet5(learning_rate=0.01, n_epochs=100, emb_size=40, batch_size=50, describ_max_len=20, type_size=12, filter_size=[3, 5], maxSentLen=100, hidden_size=[300, 300]): model_options = locals().copy() print "model options", model_options emb_root = '/save/wenpeng/datasets/LORELEI/multi-lingual-emb/' seed = 1234 np.random.seed(seed) rng = np.random.RandomState( seed) #random seed, control the model generates the same results srng = T.shared_randomstreams.RandomStreams(rng.randint(seed)) all_sentences, all_masks, all_labels, word2id = load_BBN_multi_labels_dataset( maxlen=maxSentLen ) #minlen, include one label, at least one word in the sentence label_sent, label_mask = load_SF_type_descriptions(word2id, type_size, describ_max_len) label_sent = np.asarray(label_sent, dtype='int32') label_mask = np.asarray(label_mask, dtype=theano.config.floatX) train_sents = np.asarray(all_sentences[0], dtype='int32') train_masks = np.asarray(all_masks[0], dtype=theano.config.floatX) train_labels = np.asarray(all_labels[0], dtype='int32') train_size = len(train_labels) dev_sents = np.asarray(all_sentences[1], dtype='int32') dev_masks = np.asarray(all_masks[1], dtype=theano.config.floatX) dev_labels = np.asarray(all_labels[1], dtype='int32') dev_size = len(dev_labels) test_sents = np.asarray(all_sentences[2], dtype='int32') test_masks = np.asarray(all_masks[2], dtype=theano.config.floatX) test_labels = np.asarray(all_labels[2], dtype='int32') test_size = len(test_labels) vocab_size = len(word2id) + 1 # add one zero pad index rand_values = rng.normal( 0.0, 0.01, (vocab_size, emb_size)) #generate a matrix by Gaussian distribution rand_values[0] = np.array(np.zeros(emb_size), dtype=theano.config.floatX) id2word = {y: x for x, y in word2id.iteritems()} word2vec = load_fasttext_multiple_word2vec_given_file([ emb_root + 'IL5-cca-wiki-lorelei-d40.eng.vec', emb_root + 'IL5-cca-wiki-lorelei-d40.IL5.vec' ], 40) rand_values = load_word2vec_to_init(rand_values, id2word, word2vec) embeddings = theano.shared( value=np.array(rand_values, dtype=theano.config.floatX), borrow=True ) #wrap up the python variable "rand_values" into theano variable #now, start to build the input form of the model sents_id_matrix = T.imatrix('sents_id_matrix') sents_mask = T.fmatrix('sents_mask') labels = T.imatrix('labels') #batch*12 des_id_matrix = T.imatrix() des_mask = T.fmatrix() ###################### # BUILD ACTUAL MODEL # ###################### print '... building the model' common_input = embeddings[sents_id_matrix.flatten()].reshape( (batch_size, maxSentLen, emb_size)).dimshuffle( 0, 2, 1) #the input format can be adapted into CNN or GRU or LSTM bow_emb = T.sum(common_input * sents_mask.dimshuffle(0, 'x', 1), axis=2) # bow_mean_emb = bow_emb/T.sum(sents_mask,axis=1).dimshuffle(0,'x') des_input = embeddings[des_id_matrix.flatten()].reshape( (type_size, describ_max_len, emb_size)).dimshuffle(0, 2, 1) bow_des = T.sum(des_input * des_mask.dimshuffle(0, 'x', 1), axis=2) #(tyope_size, emb_size) conv_W, conv_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) conv_W2, conv_b2 = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[1])) multiCNN_para = [conv_W, conv_b, conv_W2, conv_b2] conv_att_W, conv_att_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) conv_W_context, conv_b_context = create_conv_para( rng, filter_shape=(hidden_size[0], 1, emb_size, 1)) conv_att_W2, conv_att_b2 = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[1])) conv_W_context2, conv_b_context2 = create_conv_para( rng, filter_shape=(hidden_size[0], 1, emb_size, 1)) ACNN_para = [ conv_att_W, conv_att_b, conv_W_context, conv_att_W2, conv_att_b2, conv_W_context2 ] # NN_para = multiCNN_para+ACNN_para conv_model = Conv_with_Mask( rng, input_tensor3=common_input, mask_matrix=sents_mask, image_shape=(batch_size, 1, emb_size, maxSentLen), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero sent_embeddings = conv_model.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size conv_model2 = Conv_with_Mask( rng, input_tensor3=common_input, mask_matrix=sents_mask, image_shape=(batch_size, 1, emb_size, maxSentLen), filter_shape=(hidden_size[0], 1, emb_size, filter_size[1]), W=conv_W2, b=conv_b2 ) #mutiple mask with the conv_out to set the features by UNK to zero sent_embeddings2 = conv_model2.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size LR_input = T.concatenate([sent_embeddings, sent_embeddings2, bow_emb], axis=1) LR_input_size = hidden_size[0] * 2 + emb_size #classification layer, it is just mapping from a feature vector of size "hidden_size" to a vector of only two values: positive, negative U_a = create_ensemble_para( rng, 12, LR_input_size) # the weight matrix hidden_size*2 LR_b = theano.shared(value=np.zeros((12, ), dtype=theano.config.floatX), name='LR_b', borrow=True) #bias for each target class LR_para = [U_a, LR_b] layer_LR = LogisticRegression( rng, input=LR_input, n_in=LR_input_size, n_out=12, W=U_a, b=LR_b ) #basically it is a multiplication between weight matrix and input feature vector score_matrix = T.nnet.sigmoid(layer_LR.before_softmax) #batch * 12 prob_pos = T.where(labels < 1, 1.0 - score_matrix, score_matrix) loss = -T.mean(T.log(prob_pos)) ''' GRU ''' U1, W1, b1 = create_GRU_para(rng, emb_size, hidden_size[0]) GRU_NN_para = [ U1, W1, b1 ] #U1 includes 3 matrices, W1 also includes 3 matrices b1 is bias # gru_input = common_input.dimshuffle((0,2,1)) #gru requires input (batch_size, emb_size, maxSentLen) gru_layer = GRU_Batch_Tensor_Input_with_Mask(common_input, sents_mask, hidden_size[0], U1, W1, b1) gru_sent_embeddings = gru_layer.output_sent_rep # (batch_size, hidden_size) LR_att_input = T.concatenate([gru_sent_embeddings, bow_emb], axis=1) LR_att_input_size = hidden_size[0] + emb_size #classification layer, it is just mapping from a feature vector of size "hidden_size" to a vector of only two values: positive, negative U_att_a = create_ensemble_para( rng, 12, LR_att_input_size) # the weight matrix hidden_size*2 LR_att_b = theano.shared(value=np.zeros((12, ), dtype=theano.config.floatX), name='LR_b', borrow=True) #bias for each target class LR_att_para = [U_att_a, LR_att_b] layer_att_LR = LogisticRegression( rng, input=LR_att_input, n_in=LR_att_input_size, n_out=12, W=U_att_a, b=LR_att_b ) #basically it is a multiplication between weight matrix and input feature vector att_score_matrix = T.nnet.sigmoid(layer_att_LR.before_softmax) #batch * 12 att_prob_pos = T.where(labels < 1, 1.0 - att_score_matrix, att_score_matrix) att_loss = -T.mean(T.log(att_prob_pos)) ''' ACNN ''' attentive_conv_layer = Attentive_Conv_for_Pair( rng, origin_input_tensor3=common_input, origin_input_tensor3_r=common_input, input_tensor3=common_input, input_tensor3_r=common_input, mask_matrix=sents_mask, mask_matrix_r=sents_mask, image_shape=(batch_size, 1, emb_size, maxSentLen), image_shape_r=(batch_size, 1, emb_size, maxSentLen), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), filter_shape_context=(hidden_size[0], 1, emb_size, 1), W=conv_att_W, b=conv_att_b, W_context=conv_W_context, b_context=conv_b_context) sent_att_embeddings = attentive_conv_layer.attentive_maxpool_vec_l attentive_conv_layer2 = Attentive_Conv_for_Pair( rng, origin_input_tensor3=common_input, origin_input_tensor3_r=common_input, input_tensor3=common_input, input_tensor3_r=common_input, mask_matrix=sents_mask, mask_matrix_r=sents_mask, image_shape=(batch_size, 1, emb_size, maxSentLen), image_shape_r=(batch_size, 1, emb_size, maxSentLen), filter_shape=(hidden_size[0], 1, emb_size, filter_size[1]), filter_shape_context=(hidden_size[0], 1, emb_size, 1), W=conv_att_W2, b=conv_att_b2, W_context=conv_W_context2, b_context=conv_b_context2) sent_att_embeddings2 = attentive_conv_layer2.attentive_maxpool_vec_l acnn_LR_input = T.concatenate( [sent_att_embeddings, sent_att_embeddings2, bow_emb], axis=1) acnn_LR_input_size = hidden_size[0] * 2 + emb_size #classification layer, it is just mapping from a feature vector of size "hidden_size" to a vector of only two values: positive, negative acnn_U_a = create_ensemble_para( rng, 12, acnn_LR_input_size) # the weight matrix hidden_size*2 acnn_LR_b = theano.shared(value=np.zeros((12, ), dtype=theano.config.floatX), name='LR_b', borrow=True) #bias for each target class acnn_LR_para = [acnn_U_a, acnn_LR_b] acnn_layer_LR = LogisticRegression( rng, input=acnn_LR_input, n_in=acnn_LR_input_size, n_out=12, W=acnn_U_a, b=acnn_LR_b ) #basically it is a multiplication between weight matrix and input feature vector acnn_score_matrix = T.nnet.sigmoid( acnn_layer_LR.before_softmax) #batch * 12 acnn_prob_pos = T.where(labels < 1, 1.0 - acnn_score_matrix, acnn_score_matrix) acnn_loss = -T.mean(T.log(acnn_prob_pos)) ''' dataless cosine ''' cosine_scores = normalize_matrix_rowwise(bow_emb).dot( normalize_matrix_rowwise(bow_des).T) cosine_score_matrix = T.nnet.sigmoid( cosine_scores) #(batch_size, type_size) params = multiCNN_para + LR_para + GRU_NN_para + LR_att_para + ACNN_para + acnn_LR_para # put all model parameters together cost = loss + att_loss + acnn_loss + 1e-4 * ((conv_W**2).sum() + (conv_W2**2).sum()) updates = Gradient_Cost_Para(cost, params, learning_rate) ''' testing ''' ensemble_NN_scores = T.max(T.concatenate([ att_score_matrix.dimshuffle('x', 0, 1), score_matrix.dimshuffle('x', 0, 1), acnn_score_matrix.dimshuffle('x', 0, 1) ], axis=0), axis=0) ensemble_scores = 0.5 * ensemble_NN_scores + 0.5 * cosine_score_matrix binarize_prob = T.where(ensemble_scores > 0.3, 1, 0) #train_model = theano.function([sents_id_matrix, sents_mask, labels], cost, updates=updates, on_unused_input='ignore') train_model = theano.function( [sents_id_matrix, sents_mask, labels, des_id_matrix, des_mask], cost, updates=updates, allow_input_downcast=True, on_unused_input='ignore') # dev_model = theano.function([sents_id_matrix, sents_mask, labels], layer_LR.errors(labels), allow_input_downcast=True, on_unused_input='ignore') test_model = theano.function( [sents_id_matrix, sents_mask, des_id_matrix, des_mask], binarize_prob, allow_input_downcast=True, on_unused_input='ignore') ############### # TRAIN MODEL # ############### print '... training' # early-stopping parameters patience = 50000000000 # look as this many examples regardless start_time = time.time() mid_time = start_time past_time = mid_time epoch = 0 done_looping = False n_train_batches = train_size / batch_size train_batch_start = list( np.arange(n_train_batches) * batch_size) + [train_size - batch_size] # n_dev_batches=dev_size/batch_size # dev_batch_start=list(np.arange(n_dev_batches)*batch_size)+[dev_size-batch_size] n_test_batches = test_size / batch_size test_batch_start = list( np.arange(n_test_batches) * batch_size) + [test_size - batch_size] # max_acc_dev=0.0 max_meanf1_test = 0.0 max_weightf1_test = 0.0 train_indices = range(train_size) cost_i = 0.0 while epoch < n_epochs: epoch = epoch + 1 random.Random(100).shuffle(train_indices) iter_accu = 0 for batch_id in train_batch_start: #for each batch # iter means how many batches have been run, taking into loop iter = (epoch - 1) * n_train_batches + iter_accu + 1 iter_accu += 1 train_id_batch = train_indices[batch_id:batch_id + batch_size] cost_i += train_model(train_sents[train_id_batch], train_masks[train_id_batch], train_labels[train_id_batch], label_sent, label_mask) #after each 1000 batches, we test the performance of the model on all test data if iter % 20 == 0: print 'Epoch ', epoch, 'iter ' + str( iter) + ' average cost: ' + str(cost_i / iter), 'uses ', ( time.time() - past_time) / 60.0, 'min' past_time = time.time() error_sum = 0.0 all_pred_labels = [] all_gold_labels = [] for test_batch_id in test_batch_start: # for each test batch pred_labels = test_model( test_sents[test_batch_id:test_batch_id + batch_size], test_masks[test_batch_id:test_batch_id + batch_size], label_sent, label_mask) gold_labels = test_labels[test_batch_id:test_batch_id + batch_size] # print 'pred_labels:', pred_labels # print 'gold_labels;', gold_labels all_pred_labels.append(pred_labels) all_gold_labels.append(gold_labels) all_pred_labels = np.concatenate(all_pred_labels) all_gold_labels = np.concatenate(all_gold_labels) test_mean_f1, test_weight_f1 = average_f1_two_array_by_col( all_pred_labels, all_gold_labels) if test_weight_f1 > max_weightf1_test: max_weightf1_test = test_weight_f1 if test_mean_f1 > max_meanf1_test: max_meanf1_test = test_mean_f1 print '\t\t\t\t\t\t\t\tcurrent f1s:', test_mean_f1, test_weight_f1, '\t\tmax_f1:', max_meanf1_test, max_weightf1_test print 'Epoch ', epoch, 'uses ', (time.time() - mid_time) / 60.0, 'min' mid_time = time.time() #print 'Batch_size: ', update_freq end_time = time.time() print >> sys.stderr, ('The code for file ' + os.path.split(__file__)[1] + ' ran for %.2fm' % ((end_time - start_time) / 60.))
def evaluate_lenet5(learning_rate=0.02, n_epochs=100, emb_size=300, batch_size=50, filter_size=[3], sent_len=40, claim_len=40, cand_size=10, hidden_size=[300, 300], max_pred_pick=5): model_options = locals().copy() print "model options", model_options pred_id2label = {1: 'SUPPORTS', 0: 'REFUTES', 2: 'NOT ENOUGH INFO'} root = '/save/wenpeng/datasets/FEVER/' seed = 1234 np.random.seed(seed) rng = np.random.RandomState( seed) #random seed, control the model generates the same results srng = T.shared_randomstreams.RandomStreams(rng.randint(seed)) "load raw data" vocabfile = codecs.open(root + 'word2id.txt', 'r', 'utf-8') word2id = json.loads(vocabfile.read()) # co=0 # for line in vocabfile: # word2id = json.loads(line) # co+=1 # print 'co: ', co # word2id = json.load(open(root+'word2id.json')) #json.loads(vocabfile) vocabfile.close() print 'load word2id over' # train_sents, train_sent_masks, train_sent_labels, train_claims, train_claim_mask, train_labels, word2id = load_fever_train(sent_len, claim_len, cand_size) # train_3th_sents, train_3th_sent_masks, train_3th_sent_labels, train_3th_claims, train_3th_claim_mask, train_3th_labels, word2id = load_fever_train_NoEnoughInfo(sent_len, claim_len, cand_size, word2id) all_sentences_l, all_masks_l, all_sentences_r, all_masks_r, all_labels, _ = load_SciTailV1_dataset( sent_len, word2id) # all_sentences_l, all_masks_l, all_sentences_r, all_masks_r, all_labels, _ = load_RTE_dataset_as_test(sent_len, word2id) # dev_sents_l=np.asarray(all_sentences_l[1], dtype='int32') test_sents_l = np.asarray(all_sentences_l[2], dtype='int32') # dev_masks_l=np.asarray(all_masks_l[1], dtype=theano.config.floatX) test_masks_l = np.asarray(all_masks_l[2], dtype=theano.config.floatX) # dev_sents_r=np.asarray(all_sentences_r[1] , dtype='int32') test_sents_r = np.asarray(all_sentences_r[2], dtype='int32') # dev_masks_r=np.asarray(all_masks_r[1], dtype=theano.config.floatX) test_masks_r = np.asarray(all_masks_r[2], dtype=theano.config.floatX) # dev_labels_store=np.asarray(all_labels[1], dtype='int32') test_labels_store = np.asarray(all_labels[2], dtype='int32') # dev_size=len(dev_labels_store) test_size = len(test_labels_store) vocab_size = len(word2id) + 1 print 'vocab size: ', vocab_size rand_values = rng.normal( 0.0, 0.01, (vocab_size, emb_size)) #generate a matrix by Gaussian distribution # id2word = {y:x for x,y in word2id.iteritems()} # word2vec=load_word2vec() # rand_values=load_word2vec_to_init(rand_values, id2word, word2vec) init_embeddings = theano.shared( value=np.array(rand_values, dtype=theano.config.floatX), borrow=True ) #wrap up the python variable "rand_values" into theano variable "now, start to build the input form of the model" sents_ids = T.imatrix() #(batch, cand_size, sent_len) sents_mask = T.fmatrix() # sents_labels=T.imatrix() #(batch, cand_size) # claim_ids = T.imatrix() #(batch, claim_len) # claim_mask = T.fmatrix() # joint_sents_ids=T.itensor3() #(batch, cand_size, sent_len) # joint_sents_mask=T.ftensor3() # joint_sents_labels=T.imatrix() #(batch, cand_size) claim_ids = T.imatrix() #(batch, claim_len) claim_mask = T.fmatrix() labels = T.ivector() # test_premise_ids = T.imatrix() # test_premise_matrix = T.fmatrix() # test_hypo_ids = T.imatrix() # test_hypo_matrix = T.fmatrix() # test_scitail_minibatch_labels = T.ivector() ###################### # BUILD ACTUAL MODEL # ###################### print '... building the model' embed_input_sents = init_embeddings[sents_ids.flatten( )].reshape((batch_size, sent_len, emb_size)).dimshuffle( 0, 2, 1 ) #embed_input(init_embeddings, sents_ids_l)#embeddings[sents_ids_l.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1) #the input format can be adapted into CNN or GRU or LSTM embed_input_claim = init_embeddings[claim_ids.flatten()].reshape( (batch_size, sent_len, emb_size)).dimshuffle(0, 2, 1) conv_W, conv_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) # task1_att_conv_W, task1_att_conv_b=create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) # task1_conv_W_context, task1_conv_b_context=create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, 1)) att_conv_W, att_conv_b = create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) conv_W_context, conv_b_context = create_conv_para( rng, filter_shape=(hidden_size[0], 1, emb_size, 1)) NN_para = [ conv_W, conv_b, att_conv_W, att_conv_b, conv_W_context, conv_b_context ] ''' training task2, predict 3 labels ''' joint_embed_input_sents = init_embeddings[sents_ids.flatten( )].reshape((batch_size, sent_len, emb_size)).dimshuffle( 0, 2, 1 ) #embed_input(init_embeddings, sents_ids_l)#embeddings[sents_ids_l.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1) #the input format can be adapted into CNN or GRU or LSTM joint_embed_input_claim = init_embeddings[claim_ids.flatten()].reshape( (batch_size, sent_len, emb_size)).dimshuffle(0, 2, 1) joint_conv_model_sents = Conv_with_Mask( rng, input_tensor3=joint_embed_input_sents, mask_matrix=sents_mask, image_shape=(batch_size, 1, emb_size, sent_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero joint_premise_emb = joint_conv_model_sents.maxpool_vec #(batch_size*cand_size, hidden_size) # each sentence then have an embedding of length hidden_size # joint_batch_sent_emb = joint_sent_embeddings.reshape((batch_size, cand_size, hidden_size[0])) # joint_premise_emb = T.sum(joint_batch_sent_emb*joint_sents_labels.dimshuffle(0,1,'x'), axis=1) #(batch, hidden_size) joint_conv_model_claims = Conv_with_Mask( rng, input_tensor3=joint_embed_input_claim, mask_matrix=claim_mask, image_shape=(batch_size, 1, emb_size, claim_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b ) #mutiple mask with the conv_out to set the features by UNK to zero joint_claim_embeddings = joint_conv_model_claims.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size joint_premise_hypo_emb = T.concatenate( [joint_premise_emb, joint_claim_embeddings], axis=1) #(batch, 2*hidden_size) ''' attentive conv in task2 ''' # joint_sents_tensor3 = joint_embed_input_sents.dimshuffle(0,2,1).reshape((batch_size, cand_size*sent_len, emb_size)) # joint_sents_dot = T.batched_dot(joint_sents_tensor3, joint_sents_tensor3.dimshuffle(0,2,1)) #(batch_size, cand_size*sent_len, cand_size*sent_len) # joint_sents_dot_2_matrix = T.nnet.softmax(joint_sents_dot.reshape((batch_size*cand_size*sent_len, cand_size*sent_len))) # joint_sents_context = T.batched_dot(joint_sents_dot_2_matrix.reshape((batch_size, cand_size*sent_len, cand_size*sent_len)), joint_sents_tensor3) #(batch_size, cand_size*sent_len, emb_size) # joint_add_sents_context = joint_embed_input_sents+joint_sents_context.reshape((batch_size*cand_size, sent_len, emb_size)).dimshuffle(0,2,1)#T.concatenate([joint_embed_input_sents, joint_sents_context.reshape((batch_size*cand_size, sent_len, emb_size)).dimshuffle(0,2,1)], axis=1) #(batch_size*cand_size, 2*emb_size, sent_len) attentive_conv_layer = Attentive_Conv_for_Pair_easy_version( rng, input_tensor3= joint_embed_input_sents, #batch_size*cand_size, 2*emb_size, sent_len input_tensor3_r=joint_embed_input_claim, mask_matrix=sents_mask, mask_matrix_r=claim_mask, image_shape=(batch_size, 1, emb_size, sent_len), image_shape_r=(batch_size, 1, emb_size, sent_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), filter_shape_context=(hidden_size[0], 1, emb_size, 1), W=att_conv_W, b=att_conv_b, W_context=conv_W_context, b_context=conv_b_context) attentive_sent_embeddings_l = attentive_conv_layer.attentive_maxpool_vec_l #(batch_size*cand_size, hidden_size) attentive_sent_embeddings_r = attentive_conv_layer.attentive_maxpool_vec_r "Logistic Regression layer" joint_LR_input = T.concatenate([ joint_premise_hypo_emb, attentive_sent_embeddings_l, attentive_sent_embeddings_r ], axis=1) joint_LR_input_size = 2 * hidden_size[0] + 2 * hidden_size[0] joint_U_a = create_ensemble_para(rng, 3, joint_LR_input_size) # (input_size, 3) joint_LR_b = theano.shared(value=np.zeros((3, ), dtype=theano.config.floatX), name='LR_b', borrow=True) #bias for each target class joint_LR_para = [joint_U_a, joint_LR_b] joint_layer_LR = LogisticRegression( rng, input=joint_LR_input, n_in=joint_LR_input_size, n_out=3, W=joint_U_a, b=joint_LR_b ) #basically it is a multiplication between weight matrix and input feature vector # joint_loss=joint_layer_LR.negative_log_likelihood(joint_labels) #for classification task, we usually used negative log likelihood as loss, the lower the better. ''' testing joint_sents_ids=T.itensor3() #(batch, cand_size, sent_len) joint_sents_mask=T.ftensor3() joint_sents_labels=T.imatrix() #(batch, cand_size) joint_claim_ids = T.imatrix() #(batch, claim_len) joint_claim_mask = T.fmatrix() joint_labels=T.ivector() ''' pred_minibatch_labels = joint_layer_LR.y_pred pred_minibatch_labels_2_2classes = T.where(pred_minibatch_labels > 1, 0, pred_minibatch_labels) pred_minibatch_error = T.mean( T.neq(pred_minibatch_labels_2_2classes, labels)) params = [init_embeddings] + NN_para + joint_LR_para load_model_from_file(root + 'para_for_test_scitail', params) # train_model = theano.function([sents_ids,sents_mask,sents_labels,claim_ids,claim_mask,joint_sents_ids,joint_sents_mask,joint_sents_labels, joint_claim_ids, joint_claim_mask, joint_labels], cost, updates=updates, allow_input_downcast=True, on_unused_input='ignore') test_model = theano.function( [sents_ids, sents_mask, claim_ids, claim_mask, labels], pred_minibatch_error, allow_input_downcast=True, on_unused_input='ignore') # dev_model = theano.function([joint_sents_ids,joint_sents_mask,joint_sents_labels, joint_claim_ids, joint_claim_mask, joint_labels], pred_minibatch_error, allow_input_downcast=True, on_unused_input='ignore') # test_model = theano.function([sents_ids,sents_mask,sents_labels, claim_ids,claim_mask, joint_labels], [inter_matrix,test_layer_LR.errors(joint_labels), test_layer_LR.y_pred], allow_input_downcast=True, on_unused_input='ignore') # dev_model = theano.function([sents_ids,sents_mask,sents_labels, claim_ids,claim_mask, joint_labels], [inter_matrix,test_layer_LR.errors(joint_labels), test_layer_LR.y_pred], allow_input_downcast=True, on_unused_input='ignore') ############### # TRAIN MODEL # ############### print '... testing' # early-stopping parameters patience = 50000000000 # look as this many examples regardless start_time = time.time() mid_time = start_time past_time = mid_time epoch = 0 done_looping = False # joint_n_train_batches=joint_train_size/batch_size # joint_train_batch_start=list(np.arange(joint_n_train_batches)*batch_size)+[joint_train_size-batch_size] # n_train_batches=train_size/batch_size # train_batch_start=list(np.arange(n_train_batches)*batch_size)+[train_size-batch_size] # n_dev_batches=dev_size/batch_size # dev_batch_start=list(np.arange(n_dev_batches)*batch_size)+[dev_size-batch_size] n_test_batches = test_size / batch_size remain_test_batches = test_size % batch_size test_batch_start = list( np.arange(n_test_batches) * batch_size) + [test_size - batch_size] max_acc_dev = 0.0 max_acc_test = 0.0 cost_i = 0.0 error_sum = 0.0 for idd, test_batch_id in enumerate( test_batch_start): # for each test batch error_i = test_model( test_sents_l[test_batch_id:test_batch_id + batch_size], test_masks_l[test_batch_id:test_batch_id + batch_size], test_sents_r[test_batch_id:test_batch_id + batch_size], test_masks_r[test_batch_id:test_batch_id + batch_size], test_labels_store[test_batch_id:test_batch_id + batch_size]) error_sum += error_i test_acc = 1.0 - error_sum / (len(test_batch_start)) print '\tcurrent test_acc:', test_acc
def evaluate_lenet5(learning_rate=0.02, n_epochs=100, emb_size=300, batch_size=50, filter_size=[3], sent_len=40, claim_len=20, cand_size=10,hidden_size=[300,300], max_pred_pick=5): model_options = locals().copy() print "model options", model_options seed=1234 np.random.seed(seed) rng = np.random.RandomState(seed) #random seed, control the model generates the same results srng = T.shared_randomstreams.RandomStreams(rng.randint(seed)) "load raw data" train_sents, train_sent_masks, train_sent_labels, train_claims, train_claim_mask, _, word2id = load_fever_train(sent_len, claim_len, cand_size) test_sents, test_sent_masks, test_sent_labels, test_claims, test_claim_mask, test_sent_names,test_ground_names,_, word2id = load_fever_dev(sent_len, claim_len, cand_size, word2id) train_sents=np.asarray(train_sents, dtype='int32') # dev_sents_l=np.asarray(all_sentences_l[1], dtype='int32') test_sents=np.asarray(test_sents, dtype='int32') train_sent_masks=np.asarray(train_sent_masks, dtype=theano.config.floatX) # dev_masks_l=np.asarray(all_masks_l[1], dtype=theano.config.floatX) test_sent_masks=np.asarray(test_sent_masks, dtype=theano.config.floatX) train_sent_labels=np.asarray(train_sent_labels, dtype='int32') # dev_sents_r=np.asarray(all_sentences_r[1] , dtype='int32') # test_sent_labels=np.asarray(test_sent_labels, dtype='int32') train_claims=np.asarray(train_claims, dtype='int32') # dev_sents_r=np.asarray(all_sentences_r[1] , dtype='int32') test_claims=np.asarray(test_claims, dtype='int32') train_claim_mask=np.asarray(train_claim_mask, dtype=theano.config.floatX) # dev_masks_r=np.asarray(all_masks_r[1], dtype=theano.config.floatX) test_claim_mask=np.asarray(test_claim_mask, dtype=theano.config.floatX) # train_labels_store=np.asarray(all_labels[0], dtype='int32') # dev_labels_store=np.asarray(all_labels[1], dtype='int32') # test_labels_store=np.asarray(all_labels[2], dtype='int32') train_size=len(train_claims) # dev_size=len(dev_labels_store) test_size=len(test_claims) print 'train size: ', train_size, ' test size: ', test_size vocab_size=len(word2id)+1 rand_values=rng.normal(0.0, 0.01, (vocab_size, emb_size)) #generate a matrix by Gaussian distribution id2word = {y:x for x,y in word2id.iteritems()} word2vec=load_word2vec() rand_values=load_word2vec_to_init(rand_values, id2word, word2vec) init_embeddings=theano.shared(value=np.array(rand_values,dtype=theano.config.floatX), borrow=True) #wrap up the python variable "rand_values" into theano variable "now, start to build the input form of the model" sents_ids=T.itensor3() #(batch, cand_size, sent_len) sents_mask=T.ftensor3() sents_labels=T.imatrix() #(batch, cand_size) claim_ids = T.imatrix() #(batch, claim_len) claim_mask = T.imatrix() # labels=T.ivector() ###################### # BUILD ACTUAL MODEL # ###################### print '... building the model' embed_input_sents=init_embeddings[sents_ids.flatten()].reshape((batch_size*cand_size, sent_len, emb_size)).dimshuffle(0,2,1)#embed_input(init_embeddings, sents_ids_l)#embeddings[sents_ids_l.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1) #the input format can be adapted into CNN or GRU or LSTM embed_input_claim=init_embeddings[claim_ids.flatten()].reshape((batch_size,claim_len, emb_size)).dimshuffle(0,2,1) conv_W, conv_b=create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[0])) # conv_W2, conv_b2=create_conv_para(rng, filter_shape=(hidden_size[0], 1, emb_size, filter_size[1])) NN_para = [conv_W, conv_b] conv_model_sents = Conv_with_Mask(rng, input_tensor3=embed_input_sents, mask_matrix = sents_mask.reshape((sents_mask.shape[0]*sents_mask.shape[1],sents_mask.shape[2])), image_shape=(batch_size*cand_size, 1, emb_size, sent_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b) #mutiple mask with the conv_out to set the features by UNK to zero sent_embeddings=conv_model_sents.maxpool_vec #(batch_size*cand_size, hidden_size) # each sentence then have an embedding of length hidden_size batch_sent_emb = sent_embeddings.reshape((batch_size, cand_size, hidden_size[0])) conv_model_claims = Conv_with_Mask(rng, input_tensor3=embed_input_claim, mask_matrix = claim_mask, image_shape=(batch_size, 1, emb_size, claim_len), filter_shape=(hidden_size[0], 1, emb_size, filter_size[0]), W=conv_W, b=conv_b) #mutiple mask with the conv_out to set the features by UNK to zero claim_embeddings=conv_model_claims.maxpool_vec #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size batch_claim_emb = T.repeat(claim_embeddings.dimshuffle(0,'x', 1), cand_size, axis=1) concate_claim_sent = T.concatenate([batch_claim_emb,batch_sent_emb ], axis=2) concate_2_matrix = concate_claim_sent.reshape((batch_size*cand_size, hidden_size[0]*2)) LR_input = concate_2_matrix#T.concatenate([sent_embeddings,sent_embeddings2], axis=1) LR_input_size = hidden_size[0]*2 #classification layer, it is just mapping from a feature vector of size "hidden_size" to a vector of only two values: positive, negative U_a = create_ensemble_para(rng, 1, LR_input_size) # the weight matrix hidden_size*2 # LR_b = theano.shared(value=np.zeros((8,),dtype=theano.config.floatX),name='LR_b', borrow=True) #bias for each target class LR_para=[U_a] # layer_LR=LogisticRegression(rng, input=LR_input, n_in=LR_input_size, n_out=8, W=U_a, b=LR_b) #basically it is a multiplication between weight matrix and input feature vector score_matrix = T.nnet.sigmoid(concate_2_matrix.dot(U_a)) #batch * 12 inter_matrix = score_matrix.reshape((batch_size, cand_size)) # inter_sent_claim = T.batched_dot(batch_sent_emb, batch_claim_emb) #(batch_size, cand_size, 1) # inter_matrix = T.nnet.sigmoid(inter_sent_claim.reshape((batch_size, cand_size))) ''' maybe 1.0-inter_matrix can be rewritten into 1/e^(inter_matrix) ''' prob_pos = T.where( sents_labels < 1, 1.0-inter_matrix, inter_matrix) loss = -T.mean(T.log(prob_pos)) # # "Logistic Regression layer" # LR_input = T.concatenate([attentive_sent_embeddings_l,attentive_sent_embeddings_r,attentive_sent_embeddings_l+attentive_sent_embeddings_r,attentive_sent_embeddings_l*attentive_sent_embeddings_r],axis=1) # LR_input_size=4*hidden_size[0] # # U_a = create_ensemble_para(rng, 3, LR_input_size) # (input_size, 3) # LR_b = theano.shared(value=np.zeros((3,),dtype=theano.config.floatX),name='LR_b', borrow=True) #bias for each target class # LR_para=[U_a, LR_b] # # layer_LR=LogisticRegression(rng, input=normalize_matrix_col_wise(LR_input), n_in=LR_input_size, n_out=3, W=U_a, b=LR_b) #basically it is a multiplication between weight matrix and input feature vector # loss=layer_LR.negative_log_likelihood(labels) #for classification task, we usually used negative log likelihood as loss, the lower the better. ''' testing ''' binarize_prob = T.where( inter_matrix > 0.5, 1, 0) #(batch_size, cand_size params = [init_embeddings]+NN_para+LR_para cost=loss "Use AdaGrad to update parameters" updates = Gradient_Cost_Para(cost,params, learning_rate) train_model = theano.function([sents_ids,sents_mask,sents_labels,claim_ids,claim_mask], cost, updates=updates, allow_input_downcast=True, on_unused_input='ignore') # dev_model = theano.function([sents_ids_l, sents_mask_l, sents_ids_r, sents_mask_r, labels], layer_LR.errors(labels), allow_input_downcast=True, on_unused_input='ignore') test_model = theano.function([sents_ids,sents_mask,claim_ids,claim_mask], inter_matrix, allow_input_downcast=True, on_unused_input='ignore') ############### # TRAIN MODEL # ############### print '... training' # early-stopping parameters patience = 50000000000 # look as this many examples regardless start_time = time.time() mid_time = start_time past_time= mid_time epoch = 0 done_looping = False n_train_batches=train_size/batch_size train_batch_start=list(np.arange(n_train_batches)*batch_size)+[train_size-batch_size] # n_dev_batches=dev_size/batch_size # dev_batch_start=list(np.arange(n_dev_batches)*batch_size)+[dev_size-batch_size] n_test_batches=test_size/batch_size test_batch_start=list(np.arange(n_test_batches)*batch_size)+[test_size-batch_size] max_acc_dev=0.0 max_test_f1=0.0 cost_i=0.0 train_indices = range(train_size) while epoch < n_epochs: epoch = epoch + 1 random.Random(100).shuffle(train_indices) #shuffle training set for each new epoch, is supposed to promote performance, but not garrenteed iter_accu=0 for batch_id in train_batch_start: #for each batch # iter means how many batches have been run, taking into loop iter = (epoch - 1) * n_train_batches + iter_accu +1 iter_accu+=1 train_id_batch = train_indices[batch_id:batch_id+batch_size] ''' train_sents, train_sent_masks, train_sent_labels, train_claims, train_claim_mask sents_ids,sents_mask,sents_labels,claim_ids,claim_mask ''' cost_i+= train_model( train_sents[train_id_batch], train_sent_masks[train_id_batch], train_sent_labels[train_id_batch], train_claims[train_id_batch], train_claim_mask[train_id_batch]) #after each 1000 batches, we test the performance of the model on all test data # if (epoch==1 and iter%1000==0) or (epoch>=2 and iter%5==0): if iter%10==0: print 'Epoch ', epoch, 'iter '+str(iter)+' average cost: '+str(cost_i/iter), 'uses ', (time.time()-past_time)/60.0, 'min' past_time = time.time() ''' test test_sents, test_sent_masks, test_sent_labels, test_claims, test_claim_mask, sents_ids,sents_mask,claim_ids,claim_mask ''' f1_sum=0.0 for test_batch_id in test_batch_start: # for each test batch batch_prob=test_model( test_sents[test_batch_id:test_batch_id+batch_size], test_sent_masks[test_batch_id:test_batch_id+batch_size], test_claims[test_batch_id:test_batch_id+batch_size], test_claim_mask[test_batch_id:test_batch_id+batch_size]) batch_sent_labels = test_sent_labels[test_batch_id:test_batch_id+batch_size] batch_sent_names = test_sent_names[test_batch_id:test_batch_id+batch_size] batch_ground_names = test_ground_names[test_batch_id:test_batch_id+batch_size] for i in range(batch_size): pred_sent_names = [] gold_sent_names = batch_ground_names[i] zipped=[(batch_prob[i,k],batch_sent_labels[i][k],batch_sent_names[i][k]) for k in range(cand_size)] sorted_zip = sorted(zipped, key=lambda x: x[0], reverse=True) # print 'sorted_zip:', sorted_zip # exit(0) for j in range(cand_size): triple = sorted_zip[j] if triple[1] == 1.0: ''' we should consider a rank, instead of binary ''' if triple[0] >0.5: pred_sent_names.append(batch_sent_names[i][j]) if len(pred_sent_names) == max_pred_pick: break f1_i = compute_f1_two_list_names(pred_sent_names, gold_sent_names) f1_sum+=f1_i test_f1=f1_sum/(len(test_batch_start)*batch_size) if test_f1 > max_test_f1: max_test_f1=test_f1 print '\t\tcurrent test_f1:', test_f1,' ; ','\t\t\t\t\tmax_test_f1:', max_test_f1 print 'Epoch ', epoch, 'uses ', (time.time()-mid_time)/60.0, 'min' mid_time = time.time() #print 'Batch_size: ', update_freq end_time = time.time() print >> sys.stderr, ('The code for file ' + os.path.split(__file__)[1] + ' ran for %.2fm' % ((end_time - start_time) / 60.)) return max_acc_test
def __init__(self, rng, origin_input_tensor3, origin_input_tensor3_r, input_tensor3, input_tensor3_r, #our method cm_origin_input_tensor3, cm_origin_input_tensor3_r, cm_input_tensor3, cm_input_tensor3_r, ########### mask_matrix, mask_matrix_r, filter_shape, filter_shape_context, image_shape, image_shape_r, W, b, W_posi, b_posi, W_context, b_context, posi_emb_matrix, posi_emb_size, K_ratio): batch_size = origin_input_tensor3.shape[0] hidden_size = origin_input_tensor3.shape[1] l_len = origin_input_tensor3.shape[2] r_len = origin_input_tensor3_r.shape[2] #construct interaction matrix input_tensor3 = input_tensor3*mask_matrix.dimshuffle(0,'x',1) input_tensor3_r = input_tensor3_r*mask_matrix_r.dimshuffle(0,'x',1) #(batch, hidden, r_len) dot_tensor3 = T.batched_dot(input_tensor3.dimshuffle(0,2,1),input_tensor3_r) #(batch, l_len, r_len) # our method cm_input_tensor3 = cm_input_tensor3 * mask_matrix.dimshuffle(0,'x',1) cm_input_tensor3_r = cm_input_tensor3_r * mask_matrix_r.dimshuffle(0,'x',1) #(batch, hidden, r_len) cm_dot_tensor3 = T.batched_dot(cm_input_tensor3.dimshuffle(0,2,1), cm_input_tensor3_r) #(batch, l_len, r_len) new_dot_tensor3 = dot_tensor3 + 0.01 * cm_dot_tensor3 ''' try to get position shift of best match ''' aligned_posi_l = T.argmax(dot_tensor3, axis=2).flatten() posi_emb_tensor3_l = posi_emb_matrix[aligned_posi_l].reshape((batch_size,dot_tensor3.shape[1],posi_emb_size)).dimshuffle(0,2,1) #(batch, emb_size, l_len) aligned_posi_r = T.argmax(dot_tensor3, axis=1).flatten() posi_emb_tensor3_r = posi_emb_matrix[aligned_posi_r].reshape((batch_size,dot_tensor3.shape[2],posi_emb_size)).dimshuffle(0,2,1) #(batch, emb_size, r_len) l_max_cos = (1.0 - T.min(selu(cm_dot_tensor3), axis=2))/(1.0 + 0.5 * T.max(selu(dot_tensor3), axis=2))#1.0/T.exp(T.max(T.nnet.sigmoid(dot_tensor3), axis=2)) #(batch, l_len) r_max_cos = (1.0 - T.min(selu(cm_dot_tensor3), axis=1))/(1.0 + 0.5 * T.max(selu(dot_tensor3), axis=1))#1.0/T.exp(T.max(T.nnet.sigmoid(dot_tensor3), axis=1)) #(batch, r_len) ''' another interaction matrix ''' dot_matrix_for_right = T.nnet.softmax(new_dot_tensor3.reshape((batch_size*l_len, r_len))) #(batch*l_len, r_len) dot_tensor3_for_right = dot_matrix_for_right.reshape((batch_size, l_len, r_len))#(batch, l_len, r_len) weighted_sum_r = T.batched_dot(dot_tensor3_for_right, input_tensor3_r.dimshuffle(0,2,1)).dimshuffle(0,2,1)*mask_matrix.dimshuffle(0,'x',1) #(batch,hidden, l_len) dot_matrix_for_left = T.nnet.softmax(new_dot_tensor3.dimshuffle(0,2,1).reshape((batch_size*r_len, l_len))) #(batch*r_len, l_len) dot_tensor3_for_left = dot_matrix_for_left.reshape((batch_size, r_len, l_len))#(batch, r_len, l_len) weighted_sum_l = T.batched_dot(dot_tensor3_for_left, input_tensor3.dimshuffle(0,2,1)).dimshuffle(0,2,1)*mask_matrix_r.dimshuffle(0,'x',1) #(batch,hidden, r_len) #convolve left, weighted sum r biased_conv_model_l = Conv_with_Mask(rng, input_tensor3=origin_input_tensor3*l_max_cos.dimshuffle(0,'x',1), mask_matrix = mask_matrix, image_shape=image_shape, filter_shape=filter_shape, W=W, b=b) biased_temp_conv_output_l = biased_conv_model_l.naked_conv_out self.conv_out_l = biased_conv_model_l.masked_conv_out self.maxpool_vec_l = biased_conv_model_l.maxpool_vec conv_model_l = Conv_with_Mask(rng, input_tensor3=T.concatenate([origin_input_tensor3,posi_emb_tensor3_l],axis=1), mask_matrix = mask_matrix, image_shape=(image_shape[0], image_shape[1], image_shape[2]+posi_emb_size, image_shape[3]), filter_shape=(filter_shape[0],filter_shape[1],filter_shape[2]+posi_emb_size,filter_shape[3]), W=W_posi, b=b_posi) temp_conv_output_l = conv_model_l.naked_conv_out conv_model_weighted_r = Conv_with_Mask(rng, input_tensor3=weighted_sum_r, mask_matrix = mask_matrix, image_shape=image_shape, filter_shape=filter_shape_context, W=W_context, b=b_context) # note that b_context is not used temp_conv_output_weighted_r = conv_model_weighted_r.naked_conv_out ''' combine ''' mask_for_conv_output_l=T.repeat(mask_matrix.dimshuffle(0,'x',1), filter_shape[0], axis=1) #(batch_size, emb_size, maxSentLen-filter_size+1) mask_for_conv_output_l=(1.0-mask_for_conv_output_l)*(mask_for_conv_output_l-10) self.biased_conv_attend_out_l = T.tanh(biased_temp_conv_output_l+ temp_conv_output_weighted_r+ b.dimshuffle('x', 0, 'x'))*mask_matrix.dimshuffle(0,'x',1) self.biased_attentive_sumpool_vec_l=T.sum(self.biased_conv_attend_out_l, axis=2) self.biased_attentive_meanpool_vec_l=self.biased_attentive_sumpool_vec_l/T.sum(mask_matrix,axis=1).dimshuffle(0,'x') masked_biased_conv_output_l=self.biased_conv_attend_out_l+mask_for_conv_output_l #mutiple mask with the conv_out to set the features by UNK to zero self.biased_attentive_maxpool_vec_l=T.max(masked_biased_conv_output_l, axis=2) #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size self.conv_attend_out_l = T.tanh(temp_conv_output_l+ temp_conv_output_weighted_r+ b_posi.dimshuffle('x', 0, 'x'))*mask_matrix.dimshuffle(0,'x',1) masked_conv_output_l=self.conv_attend_out_l+mask_for_conv_output_l #mutiple mask with the conv_out to set the features by UNK to zero self.attentive_maxpool_vec_l=T.max(masked_conv_output_l, axis=2) #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size #convolve right, weighted sum l biased_conv_model_r = Conv_with_Mask(rng, input_tensor3=origin_input_tensor3_r*r_max_cos.dimshuffle(0,'x',1), mask_matrix = mask_matrix_r, image_shape=image_shape_r, filter_shape=filter_shape, W=W, b=b) biased_temp_conv_output_r = biased_conv_model_r.naked_conv_out self.conv_out_r = biased_conv_model_r.masked_conv_out self.maxpool_vec_r = biased_conv_model_r.maxpool_vec conv_model_r = Conv_with_Mask(rng, input_tensor3=T.concatenate([origin_input_tensor3_r,posi_emb_tensor3_r],axis=1), mask_matrix = mask_matrix_r, image_shape=(image_shape_r[0],image_shape_r[1],image_shape_r[2]+posi_emb_size,image_shape_r[3]), filter_shape=(filter_shape[0],filter_shape[1],filter_shape[2]+posi_emb_size,filter_shape[3]), W=W_posi, b=b_posi) temp_conv_output_r = conv_model_r.naked_conv_out conv_model_weighted_l = Conv_with_Mask(rng, input_tensor3=weighted_sum_l, mask_matrix = mask_matrix_r, image_shape=image_shape_r, filter_shape=filter_shape_context, W=W_context, b=b_context) # note that b_context is not used temp_conv_output_weighted_l = conv_model_weighted_l.naked_conv_out ''' combine ''' mask_for_conv_output_r=T.repeat(mask_matrix_r.dimshuffle(0,'x',1), filter_shape[0], axis=1) #(batch_size, emb_size, maxSentLen-filter_size+1) mask_for_conv_output_r=(1.0-mask_for_conv_output_r)*(mask_for_conv_output_r-10) self.biased_conv_attend_out_r = T.tanh(biased_temp_conv_output_r+ temp_conv_output_weighted_l+ b.dimshuffle('x', 0, 'x'))*mask_matrix_r.dimshuffle(0,'x',1) self.biased_attentive_sumpool_vec_r=T.sum(self.biased_conv_attend_out_r, axis=2) self.biased_attentive_meanpool_vec_r=self.biased_attentive_sumpool_vec_r/T.sum(mask_matrix_r,axis=1).dimshuffle(0,'x') self.conv_attend_out_r = T.tanh(temp_conv_output_r+ temp_conv_output_weighted_l+ b_posi.dimshuffle('x', 0, 'x'))*mask_matrix_r.dimshuffle(0,'x',1) masked_biased_conv_output_r=self.biased_conv_attend_out_r+mask_for_conv_output_r #mutiple mask with the conv_out to set the features by UNK to zero self.biased_attentive_maxpool_vec_r=T.max(masked_biased_conv_output_r, axis=2) #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size masked_conv_output_r=self.conv_attend_out_r+mask_for_conv_output_r #mutiple mask with the conv_out to set the features by UNK to zero self.attentive_maxpool_vec_r=T.max(masked_conv_output_r, axis=2) #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size