class DeepEventEventRelationMentionDecoder:
def __init__(self, paramsFile):
params = wbq_params.read(paramsFile)
print("params: " + str(params))
self.tf_model_prefix=params['TensorFlowEventEventRelationMentionExtractionModelPrefix']
self.tf_model_relation2id=params['TensorFlowEventEventRelationMentionExtractionModelRelation2id']
self.tf_model_word2vec=params['TensorFlowEventEventRelationMentionExtractionModelWord2vec']
self.max_length=120
self.batch_size=160
self.word_embedding_dim=300
self.encoder="pcnn"
self.selector="att"
self.test_data_loader = DataLoader(self.tf_model_word2vec,
self.tf_model_relation2id,
mode=DataLoader.MODE_INSTANCE,
shuffle=False,
max_length=self.max_length,
batch_size=self.batch_size)
self.test_data_loader.create_dataset([])
self.model = MultiModel(None, self.test_data_loader,
max_length=self.max_length,
batch_size=self.batch_size,
word_embedding_dim=self.word_embedding_dim,
encoder=self.encoder,
selector=self.selector)
self.model.load_best_model(self.tf_model_prefix)
self.model.load_id2rel(self.tf_model_relation2id)
def decode(self, serializedInstance):
# we have one and only one instance each time this function is called
# decoding_json = [serializedInstance]
decoding_json = [serializedInstance for _ in range(0, 160)]
self.test_data_loader.create_dataset(decoding_json)
predicted_rels,predicted_prob = self.model.predict(self.test_data_loader)
# we have one and only one instance in this list
label = predicted_rels[0]
if label=="NA":
label="OTHER"
confidence = predicted_prob[0]
print("PREDICTION:\t" + label + "\t" + str(confidence) + "\tserializedInstance: " + str(serializedInstance))
return { 'label': label, 'confidence' : str(confidence) }
def train(argv):
loader = DataLoader(FLAGS.train_dir,
n_cls=FLAGS.ncls,
img_shape=FLAGS.img_shape)
train_ds, val_ds = loader.create_dataset(batch_size=FLAGS.mb_size)
logdir = str(
PurePath(
os.path.join(FLAGS.logdir,
f'{datetime.now().strftime("%Y%m%d-%H%M%S")}')))
initial_epoch = 0
checkpoint_path = str(
PurePath(os.path.join(FLAGS.model_dir, "model-{epoch:04d}.ckpt")))
cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_path,
verbose=1,
save_weights_only=True,
period=FLAGS.save_freq)
model = YOLOv3(img_shape=FLAGS.img_shape,
ncls=FLAGS.ncls,
use_spp=True,
use_pretrained_weights=True)
if FLAGS.restore:
initial_epoch = FLAGS.initial_epoch
path = str(
PurePath(
os.path.join(FLAGS.model_dir,
f"model-{initial_epoch:04d}.ckpt")))
model.load_weights(path)
optimizer = tf.keras.optimizers.Adam(learning_rate=FLAGS.learning_rate)
tb_callback = tf.keras.callbacks.TensorBoard(log_dir=logdir,
write_images=True)
image_cb = callbacks.YOLOCallback(model.serving,
val_data=val_ds,
val_steps=5,
logdir=logdir,
encoder=loader.encoder,
write_every_n_epochs=10)
model.compile(optimizer=optimizer)
model.fit(train_ds,
steps_per_epoch=FLAGS.steps_per_epoch,
validation_data=val_ds,
validation_steps=100,
epochs=FLAGS.max_epochs,
callbacks=[image_cb, tb_callback, cp_callback],
initial_epoch=initial_epoch)
