def test_resnext101_feats(benchmark, model, dev, normalize, dets, padding,
streams):
im_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])
frames = th.rand((streams, 3, 720, 1280), dtype=th.float32)
H, W = frames.shape[-2:]
width = dets[:, 2] - dets[:, 0] + 1
height = dets[:, 3] - dets[:, 1] + 1
paddingW = width * padding[0]
paddingH = height * padding[1]
dets[:, 0] -= paddingW
dets[:, 1] -= paddingH
dets[:, 2] += paddingW
dets[:, 3] += paddingH
dets = clip_boxes_to_image(dets, (H, W))
batch = []
for s in range(streams):
for box in dets.round().long():
x1, y1, x2, y2 = box.tolist()
batch.append(im_transform(frames[s, :, y1:y2, x1:x2]))
batch = th.stack(batch).to(dev)
th.cuda.synchronize()
feats = benchmark(call_backbone, model, batch)
assert feats.shape[0] == batch.shape[0]
def test_convnext_small_spatial_feats(benchmark, model, normalize, dev,
batch_size):
spatial_transform = transforms.Compose([normalize])
batch = []
for n in range(batch_size):
frame = th.rand((3, 720, 1280), dtype=th.float32)
frame = spatial_transform(frame).to(dev)
batch.append(frame)
batch = th.stack(batch)
th.cuda.synchronize()
spatial_feats = benchmark(call_backbone_spatial, model, batch)
assert spatial_feats.shape[0] == batch_size
def transform(shape):
return T.Compose([T.Resize(shape[1:]), T.ToTensor()])
def transform(shape):
import ml.vision.transforms as T
return T.Compose([T.Resize(shape[1:]), T.ToTensor()])