def stack_lstm_default_state(batch_size, hidden_size, num_layers,
bidirectional):
"""init default input."""
num_directions = 2 if bidirectional else 1
h_list = c_list = []
for _ in range(num_layers):
h_list.append(ts.zeros((num_directions, batch_size, hidden_size)))
c_list.append(ts.zeros((num_directions, batch_size, hidden_size)))
h, c = tuple(h_list), tuple(c_list)
return h, c
def _initialize_weights(self):
self.init_parameters_data()
for _, m in self.cells_and_names():
if isinstance(m, layers.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.set_data(
Tensor(
np.random.normal(0, np.sqrt(
2. / n), m.weight.data.shape).astype("float32")))
if m.bias is not None:
m.bias.set_data(ts.zeros(m.bias.data.shape))
elif isinstance(m, layers.BatchNorm2d):
m.gamma.set_data(ts.ones(m.gamma.data.shape))
m.beta.set_data(ts.zeros(m.beta.data.shape))
def eval_process(args_opt, cityscapes_dir, result_dir):
CS = CityScapes()
hist_perframe = ts.zeros((CS.class_num, CS.class_num)).asnumpy()
cityscapes = generate_image_list(cityscapes_dir)
args_opt.dataset_size = len(cityscapes)
reporter = GanReporter(args_opt)
reporter.start_eval()
for i, img_path in enumerate(cityscapes):
if i % 100 == 0:
reporter.info('Evaluating: %d/%d' % (i, len(cityscapes)))
img_name = os.path.split(img_path)[1]
ids1 = CS.get_id(os.path.join(cityscapes_dir, img_name))
ids2 = CS.get_id(os.path.join(result_dir, img_name))
hist_perframe += fast_hist(ids1.flatten(), ids2.flatten(),
CS.class_num)
mean_pixel_acc, mean_class_acc, mean_class_iou, per_class_acc, per_class_iou = get_scores(
hist_perframe)
reporter.info(
"mean_pixel_acc:{}, mean_class_acc:{}, mean_class_iou: {}".format(
mean_pixel_acc, mean_class_acc, mean_class_iou))
reporter.info("************ Per class numbers below ************")
for i, cl in enumerate(CS.classes):
while len(cl) < 15:
cl = cl + ' '
reporter.info("{}: acc = {}, iou = {}".format(cl, per_class_acc[i],
per_class_iou[i]))
reporter.end_eval()
def _initialize_weights(self):
self.init_parameters_data()
for _, m in self.cells_and_names():
if isinstance(m, layers.Dense):
m.weight.set_data(
Tensor(
np.random.normal(
0, 0.01, m.weight.data.shape).astype("float32")))
if m.bias is not None:
m.bias.set_data(ts.zeros(m.bias.data.shape))
def lstm_default_state(batch_size, hidden_size, num_layers, bidirectional):
"""init default input."""
num_directions = 2 if bidirectional else 1
h = ts.zeros((num_layers * num_directions, batch_size, hidden_size))
c = ts.zeros((num_layers * num_directions, batch_size, hidden_size))
return h, c