def test_sum_prediction():
""" Test with a model that predicts sum over four samples """
rng = RandomState(3904890384)
n_samples_in_buffer = 1000
dataset = rng.rand(n_samples_in_buffer*2,5).astype(np.float32)
markers = np.ones((n_samples_in_buffer*2,1)).astype(np.float32)
set_and_markers = np.concatenate((dataset, markers), axis=1)
factor_new=0.001
n_stride = 10
pred_freq = 11
standardized = exponential_running_standardize(dataset,
factor_new=factor_new, init_block_size=n_stride)
model = InputLayer([1,1,4,1])
model = GlobalPoolLayer(model,pool_function=T.sum)
expected = [np.sum(standardized[stop-4:stop], axis=0) for stop in xrange(11, dataset.shape[0], 11)]
expected = np.array(expected)
processor = StandardizeProcessor(factor_new=factor_new,
n_samples_in_buffer=n_samples_in_buffer)
online_model = OnlineModel(model)
online_pred = OnlineCoordinator(processor, online_model, pred_freq=pred_freq,
trainer=NoTrainer())
online_pred.initialize(n_chans=dataset.shape[1])
all_preds = []
for i_start_sample in xrange(0,dataset.shape[0]-n_stride+1,n_stride):
online_pred.receive_samples(set_and_markers[i_start_sample:i_start_sample+n_stride])
if online_pred.has_new_prediction():
pred, _ = online_pred.pop_last_prediction_and_sample_ind()
all_preds.append(pred)
assert np.allclose(np.array(all_preds).squeeze(), expected, rtol=1e-3)
def test_online_predictor():
""" Test whether predictions are done at correct timepoints.
Model actually just returns input """
rng = RandomState(3904890384)
n_samples_in_buffer = 1000
dataset = rng.rand(n_samples_in_buffer*2,5).astype(np.float32)
markers = np.ones((n_samples_in_buffer*2,1)).astype(np.float32)
set_and_markers = np.concatenate((dataset, markers), axis=1)
factor_new=0.001
n_stride = 10
pred_freq = 11
standardized = exponential_running_standardize(dataset,
factor_new=factor_new, init_block_size=n_stride)
model = InputLayer([1,1,1,1])
processor = StandardizeProcessor(factor_new=factor_new,
n_samples_in_buffer=n_samples_in_buffer)
online_model = OnlineModel(model)
online_pred = OnlineCoordinator(processor, online_model, pred_freq=pred_freq,
trainer=NoTrainer())
online_pred.initialize(n_chans=dataset.shape[1])
all_preds = []
for i_start_sample in xrange(0,dataset.shape[0]-n_stride+1,n_stride):
online_pred.receive_samples(set_and_markers[i_start_sample:i_start_sample+n_stride])
if online_pred.has_new_prediction():
pred, _ = online_pred.pop_last_prediction_and_sample_ind()
all_preds.append(pred)
assert np.array_equal(np.array(all_preds).squeeze(), standardized[10::pred_freq])
def main(ui_hostname, ui_port, base_name, params_filename, plot_sensors,
use_ui_server, adapt_model, save_data, n_updates_per_break, batch_size,
learning_rate, n_min_trials, trial_start_offset, break_start_offset,
break_stop_offset,
pred_gap,
incoming_port,load_old_data,use_new_adam_params,
input_time_length,
train_on_breaks,
min_break_samples,
min_trial_samples):
setup_logging()
assert np.little_endian, "Should be in little endian"
train_params = None # for trainer, e.g. adam params
if params_filename is not None:
if params_filename == 'newest':
# sort will already sort temporally with our time string format
all_params_files = sorted(glob(base_name + ".*.model_params.npy"))
assert len(all_params_files) > 0, ("Expect atleast one params file "
"if 'newest' given as argument")
params_filename = all_params_files[-1]
log.info("Loading model params from {:s}".format(params_filename))
params = np.load(params_filename)
train_params_filename = params_filename.replace('model_params.npy',
'trainer_params.npy')
if os.path.isfile(train_params_filename):
if use_new_adam_params:
log.info("Loading trainer params from {:s}".format(train_params_filename))
train_params = np.load(train_params_filename)
else:
log.warn("No train/adam params found, starting optimization params "
"from scratch (model params will be loaded anyways).")
else:
params = np.load(base_name + '.npy')
exp = create_experiment(base_name + '.yaml')
# Possibly change input time length, for exmaple
# if input time length very long during training and should be
# shorter for online
if input_time_length is not None:
log.info("Change input time length to {:d}".format(input_time_length))
set_input_window_length(exp.final_layer, input_time_length)
# probably unnecessary, just for safety
exp.iterator.input_time_length = input_time_length
# Have to set for both exp final layer and actually used model
# as exp final layer might be used for adaptation
# maybe check this all for correctness?
cnt_model = exp.final_layer
set_param_values_backwards_compatible(cnt_model, params)
prediction_model = transform_to_normal_net(cnt_model)
set_param_values_backwards_compatible(prediction_model, params)
data_processor = StandardizeProcessor(factor_new=1e-3)
online_model = OnlineModel(prediction_model)
if adapt_model:
online_trainer = BatchWiseCntTrainer(exp, n_updates_per_break,
batch_size, learning_rate, n_min_trials, trial_start_offset,
break_start_offset=break_start_offset,
break_stop_offset=break_stop_offset,
train_param_values=train_params,
add_breaks=train_on_breaks,
min_break_samples=min_break_samples,
min_trial_samples=min_trial_samples)
else:
log.info("Not adapting model...")
online_trainer = NoTrainer()
coordinator = OnlineCoordinator(data_processor, online_model, online_trainer,
pred_gap=pred_gap)
hostname = ''
server = PredictionServer((hostname, incoming_port), coordinator=coordinator,
ui_hostname=ui_hostname, ui_port=ui_port, plot_sensors=plot_sensors,
use_ui_server=use_ui_server, save_data=save_data,
model_base_name=base_name, adapt_model=adapt_model)
# Compilation takes some time so initialize trainer already
# before waiting in connection in server
online_trainer.initialize()
if adapt_model and load_old_data:
online_trainer.add_data_from_today(data_processor)
log.info("Starting server on port {:d}".format(incoming_port))
server.start()
log.info("Started server")
server.serve_forever()
