def generator_train():
for data in train_data:
x, y = data[0], data[1]
#x_emb = embedding_bert.get_batch_emb(x, doc_len, sen_len, tokenizer, estimator)
doc_seq_len, sen_seq_len = embedding_bert.get_batch_seq([x],
doc_len,
sen_len,
tokenizer,
tol=2)
yield doc_seq_len, sen_seq_len
#print('Epoch:', '%d' % (epoch + 0), 'cost =', '{:.6f}'.format(loss))
# testing
n_batch = len(X_test_doc) // batch_size
#for i in range(n_batch):
for i in tqdm(range(n_batch), total=n_batch):
X_batch, y_batch = utils.get_batch_test(X_test_doc, y_test, i,
batch_size)
X_batch_emb = embedding_bert.get_batch_emb(X_batch, doc_len,
sen_len, tokenizer,
estimator)
#X_batch_emb = X_batch_emb[:, :, 0, :]
#X_batch_seq, _ = embedding_bert.get_batch_seq(X_batch, doc_len, sen_len, tokenizer, tol = 2)
doc_seq_len, sen_seq_len = embedding_bert.get_batch_seq(X_batch,
doc_len,
sen_len,
tokenizer,
tol=2)
#batch_seq_len = np.array([list(x).index(0) + 1 for x in X_batch])
loss, y_pred_val = sess.run(
[cost, y_pred],
feed_dict={
X_emb: X_batch_emb,
y: y_batch,
doc_seq_length: doc_seq_len,
sen_seq_length: sen_seq_len,
is_training: False
}) #seq_length: batch_seq_len
#acc_test += acc
# loss_train = loss_tr * DELTA + loss_train * (1 - DELTA), delta = 0.5??
loss_test += loss
### test # In[19]: X_train[:3] # In[20]: a = embedding_bert.get_batch_emb(X_train[:3], doc_len, sen_len, tokenizer, estimator) b, c = embedding_bert.get_batch_seq(X_train[:3], doc_len, sen_len, tokenizer, tol = 2) # In[ ]: # In[21]: a.shape, b.shape, c.shape # In[22]:
prepare_bert(bert_vocab_file, bert_config_file, init_checkpoint, sen_len, select_layers, batch_size, graph_file) # In[17]: ### test # In[18]: a = embedding_bert.get_batch_emb(X_train[:4], doc_len, sen_len, tokenizer, estimator) b, c = embedding_bert.get_batch_seq(X_train[:4], doc_len, sen_len, tokenizer, tol = 2) # In[19]: a.shape, b.shape, c.shape # In[20]: c # In[22]:
# In[11]:
xx = embedding_bert.get_batch_emb([data_part[0][0]], doc_len, sen_len,
tokenizer, estimator)
# In[12]:
xx.shape
# In[ ]:
# In[9]:
a, b = embedding_bert.get_batch_seq([data_part[0][0], data_part[1][0]],
doc_len,
sen_len,
tokenizer,
tol=2)
a, b
# In[10]:
a[0], b[0]
# In[11]:
xx[0].shape
# In[24]:
embedding_bert.get_batch_seq([data_part[4][0]],