) # Note: might need call to model
optimizer_weights_group = f['optimizer_weights']
optimizer_weight_names = [
n.decode('utf8')
for n in optimizer_weights_group.attrs['weight_names']
]
optimizer_weight_values = [
optimizer_weights_group[n]
for n in optimizer_weight_names
]
model.optimizer.set_weights(optimizer_weight_values)
else:
logger.error("No optimizer weights in wieghts file!")
raise Exception()
training_loop(model=model,
train=train,
steps_per_epoch=steps_per_epoch,
save_freq=config['save_freq'],
checkpoint_monitor="val_acc",
epochs=config['n_epochs'],
save_path=save_path,
reload=config['reload'],
valid=valid,
custom_callbacks=callbacks,
verbose=2)
if __name__ == "__main__":
wrap(configs, train, vegab_plugins)
EvaluateOnDataStream(model=model,
data_stream=dev2_stream,
prefix="dev2/"))
callbacks.append(
SaveBestScore(save_path=save_path,
dev1_stream=dev1_stream,
dev2_stream=dev2_stream,
test_stream=test_stream))
# Small hack to make sure threshold fitting works
_evaluate_with_threshold_fitting(epoch=-1,
logs={},
model=model,
val_data_thr=dev1_stream,
val_data=dev2_stream,
test_data=test_stream)
# TODO(kududak): Save best val acc test performanc
training_loop(model=model,
train=endless_data_stream(train_stream),
epochs=config['epochs'],
steps_per_epoch=train_steps,
acc_monitor='dev2/acc_thr',
save_path=save_path,
callbacks=callbacks)
if __name__ == '__main__':
wrap(configs_factorized.config, train, plugins=[MetaSaver()])
EvaluateWithThresholdFitting(model=model,
dev2=dev2_stream,
dev1=dev1_stream,
test=test_stream))
callbacks.append(
EvaluateOnDataStream(model=model,
data_stream=dev1_stream,
prefix="dev1/"))
callbacks.append(
EvaluateOnDataStream(model=model,
data_stream=dev2_stream,
prefix="dev2/"))
callbacks.append(
SaveBestScore(save_path=save_path,
dev1_stream=dev1_stream,
dev2_stream=dev2_stream,
test_stream=test_stream))
# Train
training_loop(model=model,
train=endless_data_stream(train_stream),
epochs=config['epochs'],
steps_per_epoch=train_steps,
acc_monitor='dev2/acc_thr',
save_path=save_path,
callbacks=callbacks)
if __name__ == '__main__':
wrap(configs_dnn_ce.config, train, plugins=[MetaSaver()])