def _create_model(self):
print("\n*** HIERARHICAL SEQ2SEQ translates from {} -> {} -> {} ***".
format(self.feats_dict[self._first][2],
self.feats_dict[self._second][2],
self.feats_dict[self._third][2]))
# --------------------Seq2Seq network params--------------------------------
n_samples = self.train_first.shape[0]
input_length = self.train_first.shape[1]
input_dim = self.train_first.shape[2]
output_length = self.train_second.shape[1]
output_dim = self.train_second.shape[2]
# --------------- MODEL TRANSLATION DEFINITION -----------------------------
print("Creating LEVEL 1 TRANSLATION SEQ2SEQ model ...")
inputs, encoded_seq, decoded_seq, cycled_decoded_seq = \
mctn_model(output_dim=output_dim,
hidden_dim=self.configs['translation1']['hidden_dim'],
output_length=output_length,
input_dim=input_dim,
input_length=input_length,
depth=self.configs['translation1']['depth'],
bidirectional=self.configs['translation1']['is_bidirectional'],
is_cycled=self.is_cycled)
# -------------
input_length_2 = input_length # == encoded_seq.shape[1]
input_dim_2 = self.configs['translation1']['hidden_dim']
output_length_2 = self.train_third.shape[1]
output_dim_2 = self.train_third.shape[2]
print("Creating LEVEL 2 TRANSLATION SEQ2SEQ model ...")
inputs_2, encoded_seq_2, decoded_seq_2 = \
mctn_level2_model(input=encoded_seq,
output_dim=output_dim_2,
hidden_dim=self.configs['translation2']['hidden_dim'],
output_length=output_length_2,
input_dim=input_dim_2,
input_length=input_length_2,
depth=self.configs['translation2']['depth'],
bidirectional=self.configs['translation2'][
'is_bidirectional'])
# ---------------- MODEL REGRESSION DEFINITION -----------------------------
print("Creating REGRESSION model ...")
regression_score = \
create_regression_model(
n_hidden=self.configs['regression']['reg_hidden_dim'],
input=encoded_seq_2,
l2_factor=self.configs['regression']['l2_factor'])
# ------------------ E2E REGRESSION DEFINITION -----------------------------
print("BUILDING A JOINT END-TO-END MODEL")
if self.is_cycled:
outputs = [
decoded_seq, cycled_decoded_seq, decoded_seq_2,
regression_score
]
losses = [
self.configs['translation1']['loss_type'],
self.configs['translation1']['cycle_loss_type'],
self.configs['translation2']['loss_type'],
self.configs['regression']['loss_type']
]
losses_weights = [
self.configs['translation1']['loss_weight'],
self.configs['translation1']['cycle_loss_weight'],
self.configs['translation2']['loss_weight'],
self.configs['regression']['loss_weight']
]
else:
outputs = [decoded_seq, decoded_seq_2, regression_score]
losses = [
self.configs['translation1']['loss_type'],
self.configs['translation2']['loss_type'],
self.configs['regression']['loss_type']
]
losses_weights = [
self.configs['translation1']['loss_weight'],
self.configs['translation2']['loss_weight'],
self.configs['regression']['loss_weight']
]
end2end_model = Model(inputs=inputs, outputs=outputs)
print("Compiling model")
optimizer = get_optimizers(opt=self.configs['general']['optim'],
init_lr=self.configs['general']['init_lr'])
end2end_model.compile(loss=losses,
loss_weights=losses_weights,
optimizer=optimizer)
print("Model summary:")
print(end2end_model.summary())
print("END2END MODEL CREATED!")
self.model = end2end_model
def _create_model(self):
print("\n***SEQ2SEQ translates from {} to {}***".
format(self.feats_dict[self._in][2],
self.feats_dict[self._out][2]))
input_dims = self.input_train.shape
output_dims = self.output_train.shape
print("Input Dims = {}".format(input_dims))
print("Output (Translation) Dims = {}".format(output_dims))
print("Output (Regression) Dims = {}\n".format(self.feats_dict[
'train_labels'].shape))
# --------------------Seq2Seq network params--------------------------------
n_samples = input_dims[0]
input_length = input_dims[1]
input_dim = input_dims[2]
output_length = output_dims[1]
output_dim = output_dims[2]
# --------------- MODEL TRANSLATION DEFINITION -----------------------------
print("Creating TRANSLATION SEQ2SEQ model ...")
inputs, encoded_seq, decoded_seq, cycled_decoded_seq = \
mctn_model(output_dim=output_dim,
hidden_dim=self.configs['translation']['hidden_dim'],
output_length=output_length,
input_dim=input_dim,
input_length=input_length,
depth=self.configs['translation']['depth'],
bidirectional=self.configs['translation']['is_bidirectional'],
is_cycled=self.is_cycled,
dropout=0.8
)
# ---------------- MODEL REGRESSION DEFINITION -----------------------------
print("Creating REGRESSION model ...")
regression_score = \
create_regression_model(
n_hidden=self.configs['regression']['reg_hidden_dim'],
input=encoded_seq,
l2_factor=self.configs['regression']['l2_factor'],
input_hidden=self.configs['translation']['hidden_dim'],
dropout=0.8)
# ------------------ E2E REGRESSION DEFINITION -----------------------------
print("BUILDING A JOINT END-TO-END MODEL")
if self.is_cycled:
outputs = [decoded_seq, cycled_decoded_seq, regression_score]
losses = [self.configs['translation']['loss_type'],
self.configs['translation']['cycle_loss_type'],
self.configs['regression']['loss_type']
]
losses_weights = [self.configs['translation']['loss_weight'],
self.configs['translation']['cycle_loss_weight'],
self.configs['regression']['loss_weight']
]
else:
outputs = [decoded_seq, regression_score]
losses = [self.configs['translation']['loss_type'],
self.configs['regression']['loss_type']
]
losses_weights = [self.configs['translation']['loss_weight'],
self.configs['regression']['loss_weight']
]
end2end_model = Model(inputs=inputs,
outputs=outputs)
print("Compiling model")
optimizer = get_optimizers(opt=self.configs['general']['optim'],
init_lr=self.configs['general']['init_lr'])
end2end_model.compile(loss=losses,
loss_weights=losses_weights,
optimizer=optimizer,
metrics=[ 'categorical_accuracy'])
print("Model summary:")
print(end2end_model.summary())
print("END2END MODEL CREATED!")
self.model = end2end_model
def _create_model(self):
input_dim = 200
output_dim = 200
# --------------------Seq2Seq network params--------------------------------
input_length = 100
output_length = 100
inputs = Input(shape=(input_length, ))
embeddings = Embedding(20000, input_dim)(inputs)
# --------------- MODEL TRANSLATION DEFINITION -----------------------------
print("Creating TRANSLATION SEQ2SEQ model ...")
encoded_seq, decoded_seq, cycled_decoded_seq = \
mctn_model(input=embeddings,output_dim=output_dim,
hidden_dim=self.configs['translation']['hidden_dim'],
output_length=output_length,
input_dim=input_dim,
input_length=input_length,
depth=self.configs['translation']['depth'],
bidirectional=self.configs['translation']['is_bidirectional'],
is_cycled=self.is_cycled,
dropout=0.5
)
# ---------------- MODEL REGRESSION DEFINITION -----------------------------
print("Creating REGRESSION model ...")
regression_score = \
create_regression_model(
n_hidden=self.configs['regression']['reg_hidden_dim'],
input=encoded_seq,
l2_factor=self.configs['regression']['l2_factor'],
input_hidden=self.configs['translation']['hidden_dim'],
dropout=0.5)
# ------------------ E2E REGRESSION DEFINITION -----------------------------
print("BUILDING A JOINT END-TO-END MODEL")
if self.is_cycled:
outputs = [decoded_seq, cycled_decoded_seq, regression_score]
losses = [
self.configs['translation']['loss_type'],
self.configs['translation']['cycle_loss_type'],
self.configs['regression']['loss_type']
]
losses_weights = [
self.configs['translation']['loss_weight'],
self.configs['translation']['cycle_loss_weight'],
self.configs['regression']['loss_weight']
]
else:
outputs = [decoded_seq, regression_score]
losses = [
self.configs['translation']['loss_type'],
self.configs['regression']['loss_type']
]
losses_weights = [
self.configs['translation']['loss_weight'],
self.configs['regression']['loss_weight']
]
end2end_model = Model(inputs=inputs, outputs=outputs)
print("Compiling model")
optimizer = get_optimizers(opt=self.configs['general']['optim'],
init_lr=self.configs['general']['init_lr'])
end2end_model.compile(loss=losses,
loss_weights=losses_weights,
optimizer=optimizer,
metrics=['categorical_accuracy'])
# print("Model summary:")
print(end2end_model.summary())
# print("END2END MODEL CREATED!")
self.model = end2end_model
self.embeding_model = Model(inputs=inputs, outputs=[embeddings])
