i_leq_j_mask = tf.expand_dims(i_leq_j_mask ,0)
logits = tf.expand_dims(start_logits ,2 ) +tf.expand_dims(end_logits ,1 ) +cross
logits -= 10000 *i_leq_j_mask
logits = tf.reshape(logits, [batch_size ,seq_length *seq_length])
short_p = tf.nn.softmax(logits)
indices = tf.argsort(short_p ,axis=1 ,direction='DESCENDING')[: ,:n_keep]
short_p = tf.gather(short_p ,indices ,batch_dims=1)
return dict(unique_id = unique_id,
ans_logits= ans_logits,
long_p = long_p,
short_p = short_p,
short_p_indices = indices)
test=read_dataset()
predictions={}
for tema in test:
for cosa in tema["paragraphs"]:
text=cosa["context"]
for question in cosa["qas"]:
unique_id=question["id"]
text_tokens = tokenizer.tokenize(text)
question_len=np.array([len(tokenizer.tokenize(question["question"]))])
data_len = np.array([512])
input_word_ids,input_mask,segment_id =convert_two_sentences_to_features(question["question"],text,tokenizer,512)
c=[data_len.reshape(-1,),input_word_ids.reshape(1,-1), question_len.reshape(-1,)]
# writer.write(serialized_features_dataset)
# writer.close()
# filenames = [filename]
# train_dataset=tf.data.TFRecordDataset('x_test.tfrecord').batch(32)
# for serialized_example in train_dataset:
# for elem in serialized_example
# example = tf.train.Example()
# example.ParseFromString(elem)
# x_1 = np.array(example.features.feature['X'].float_list.value)
# y_1 = np.array(example.features.feature['Y'].float_list.value)
# break
#
#
path = read_dataset(mode="train",
dataset="naturalq",
tokenizer=tokenizer,
max_seq_length=max_seq_length,
fragmented=False)
#
# import datetime
# t = datetime.datetime.now().time()
# log_name = "Salida_modelo_{}.txt".format(t)
x, y = crear_batch(path, fragmented=False)
N = len(x)
entrada = {
"questions_id": np.squeeze(x[:N, 3].astype(np.int32)),
"question_input_mask": np.squeeze(x[:N, 4].astype(np.int32)),
"question_segment_id": np.squeeze(x[:N, 5].astype(np.int32)),
"context_id": np.squeeze(x[:N, 0].astype(np.int32)),
"context_input_mask": np.squeeze(x[:N, 1].astype(np.int32)),
"context_segment_id": np.squeeze(x[:N, 2].astype(np.int32))
# writer = tf.data.experimental.TFRecordWriter(filename)
# writer.write(serialized_features_dataset)
# writer.close()
# filenames = [filename]
# train_dataset=tf.data.TFRecordDataset('x_test.tfrecord').batch(32)
# for serialized_example in train_dataset:
# for elem in serialized_example
# example = tf.train.Example()
# example.ParseFromString(elem)
# x_1 = np.array(example.features.feature['X'].float_list.value)
# y_1 = np.array(example.features.feature['Y'].float_list.value)
# break
#
path = read_dataset(mode="train",
tokenizer=tokenizer,
max_seq_length=max_seq_length,
fragmented=False)
#
import time
t = time.time()
log_name = "Salida_modelo_{}.txt".format(t)
x, y = crear_batch(path, fragmented=False)
entrada = {
"questions_id": np.squeeze(x[:, 3]),
"question_input_mask": np.squeeze(x[:, 4]),
"question_segment_id": np.squeeze(x[:, 5]),
"context_id": np.squeeze(x[:, 0]),
"context_input_mask": np.squeeze(x[:, 1]),
"context_segment_id": np.squeeze(x[:, 2])
}
salida = [y[:, 0], y[:, 1]]
device = torch.cuda.current_device()
max_seq_length = 350
model = build_model(max_seq_length)
optimizer = torch.optim.Adam(model.parameters(),
lr=0.00005,
betas=(0.9, 0.98),
eps=1e-9)
model.cuda()
criterion = negLogSum
# Train data
path = read_dataset(mode="train",
dataset="squad",
tokenizer=tokenizer,
max_seq_length=max_seq_length,
fragmented=False,
framework="torch")
x, y = crear_batch(path, fragmented=False)
thing = list(map(list, zip(*x)))
ids, mask = np.squeeze(np.array(thing[0])), np.squeeze(np.array(thing[1]))
train_dataset = SquadDataset(ids, mask, y[:, 0], y[:, 1])
train_dataset.batch(32)
# Test data
path = read_dataset(mode="test",
dataset="squad",
tokenizer=tokenizer,
max_seq_length=max_seq_length,
fragmented=False,