def test_yolo_layer_eval_2(self):
mask = [6, 7, 8]
anchors = [(10, 13), (16, 30), (33, 23), (30, 61), (62, 45), (59, 119),
(116, 90), (156, 198), (373, 326)]
n_classes = 80
img_size = 416
feature_maps = yolo_feature_maps(img_size)
grids = feature_maps[-1]
head = YoloLayer(img_size=img_size,
grids=grids,
mask=mask,
anchors=anchors,
n_classes=n_classes).eval()
x = head(torch.rand([1, 255, 13, 13]))
meshgrid = torch.meshgrid([torch.arange(13), torch.arange(13)])
meshgrid = torch.stack(tuple(reversed(meshgrid)),
2).view(1, 13 * 13, 2)
meshgrid = torch.cat(
(meshgrid.clone(), meshgrid.clone(), meshgrid.clone()), 1)
self.assertEqual(x.size(), torch.Size([1, 3 * 13 * 13, 85]))
min_center = (meshgrid + 0) * 32
max_center = (meshgrid + 1) * 32
pred_center = x[:, :, :2]
pred_wh = x[:, :, 2:4]
pred_obj = x[:, :, 4]
pred_cls = x[:, :, 5:]
self.assertTrue(torch.all(pred_obj >= 0) and torch.all(pred_obj <= 1))
self.assertTrue(torch.all(pred_cls >= 0) and torch.all(pred_cls <= 1))
self.assertTrue(
torch.all(pred_center >= min_center)
and torch.all(pred_center <= max_center))
def test_yolo_layer_train_2(self):
mask = [6, 7, 8]
anchors = [(10, 13), (16, 30), (33, 23), (30, 61), (62, 45), (59, 119),
(116, 90), (156, 198), (373, 326)]
n_classes = 80
img_size = 416
feature_maps = yolo_feature_maps(img_size)
grids = feature_maps[-1]
head = YoloLayer(img_size=img_size,
grids=grids,
mask=mask,
anchors=anchors,
n_classes=n_classes)
x = head(torch.rand([1, 255, 13, 13]))
self.assertEqual(x.size(), torch.Size([1, 3, 13, 13, 85]))
pred_center = x[:, :, :, :, :2]
pred_wh = x[:, :, :, 2:4]
pred_obj = x[:, :, :, :, 4]
pred_cls = x[:, :, :, :, 5:]
self.assertTrue(torch.all(pred_obj >= 0) and torch.all(pred_obj <= 1))
self.assertTrue(torch.all(pred_cls >= 0) and torch.all(pred_cls <= 1))
self.assertTrue(
torch.all(pred_center >= 0) and torch.all(pred_center <= 1))
self.assertTrue(torch.all(torch.exp(pred_wh) < img_size))