# (n_scat_nodes) means 1 if data not transformed
if isinstance(data, np.ndarray):
data = np.reshape(data, (n_data_total, -1, data.shape[-1]))
elif isinstance(data, list):
data = [np.reshape(data_slice, (-1, data_slice.shape[-1])) for data_slice in data]
else:
raise ValueError("Invalid type of data given")
# take out only a fraction of the test data
data = data[idx_start:idx_end]
labels = labels[idx_start:idx_end]
n_data_total = len(data)
input_size = data[0].shape[0]
output_size = meta['output_size']
dataset = nu.TimeSeriesDataset(data, labels, transform=nu.ToTensor())
dataloader = DataLoader(dataset, sampler=SequentialSampler(range(n_data_total)),
batch_size=batch_size, collate_fn=nu.collate_fn, num_workers=0)
if device == 'cpu':
rnn = nu.RNN(input_size=meta['input_size'], hidden_size=meta['hidden_size'],
output_size=meta['output_size'], n_layers=meta['n_layers'], bidirectional=meta['bidirectional'])
else:
rnn = nu.RNN(input_size=meta['input_size'], hidden_size=meta['hidden_size'],
output_size=meta['output_size'], n_layers=meta['n_layers'], bidirectional=meta['bidirectional']).cuda()
rnn.load_state_dict(meta['model'][idx_min_loss_epoch])
del meta
#criterion = nn.CrossEntropyLoss(reduction='sum') if classifier else nn.MSELoss(reduction='sum')
#metric = 'cross_entropy_mean' if classifier else 'rmse'
loss_sum = {}
loss_metric = {}
input = data.reshape([n_data_total, n_features, data_len])
input = torch.tensor(input, dtype=torch.get_default_dtype())
# following is shaped (n_labels, n_conditions)
labels = np.array(list(product(*samples['labels'])),
dtype='float32').swapaxes(0, 1)
for idx_label in range(n_labels):
net = nu.RNN(input_size=meta['input_size'],
hidden_size=meta['hidden_size'][idx_label],
output_size=1,
n_layers=meta['n_layers'],
bidirectional=meta['bidirectional'])
net.load_state_dict(
meta['weights'][idx_label][idx_epochs[idx_label]])
dataset = nu.TimeSeriesDataset(input, np.zeros((n_data_total, )))
dataloader = DataLoader(dataset,
batch_size=batch_size,
sampler=SequentialSampler(
range(n_data_total)))
output = []
for idx_batch, batch in enumerate(dataloader):
output.append(
net(batch['data'].permute([2, 0, 1])).detach().numpy()[:,
0])
output = np.concatenate(output, axis=0)
if transformed:
output = output.reshape(samples['data'].shape[:-3])
output_mean = output.mean(axis=-1)
output_std = output.std(axis=-1)