def group_pixels(ctr, offsets):
"""
Gives each pixel in the image an instance id.
Args:
ctr (Tensor): A Tensor of shape [K, 2] where K is the number of center points. The order of second dim is (y, x).
offsets (Tensor): A Tensor of shape [2, H, W] of raw offset output, where N is the batch size,
for consistent, we only support N=1. The order of second dim is (offset_y, offset_x).
Returns:
Tensor: A Tensor of shape [1, H, W], ins_id is 1, 2, ...
"""
height, width = offsets.shape[-2:]
y_coord = paddle.arange(height, dtype=offsets.dtype).reshape([1, -1, 1])
y_coord = paddle.concat([y_coord] * width, axis=2)
x_coord = paddle.arange(width, dtype=offsets.dtype).reshape([1, 1, -1])
x_coord = paddle.concat([x_coord] * height, axis=1)
coord = paddle.concat([y_coord, x_coord], axis=0)
ctr_loc = coord + offsets
ctr_loc = ctr_loc.reshape((2, height * width)).transpose((1, 0))
# ctr: [K, 2] -> [K, 1, 2]
# ctr_loc = [H*W, 2] -> [1, H*W, 2]
ctr = ctr.unsqueeze(1)
ctr_loc = ctr_loc.unsqueeze(0)
# distance: [K, H*W]
distance = paddle.norm((ctr - ctr_loc).astype('float32'), axis=-1)
# finds center with minimum distance at each location, offset by 1, to reserve id=0 for stuff
instance_id = paddle.argmin(distance, axis=0).reshape(
(1, height, width)) + 1
return instance_id
def run(place):
paddle.disable_static(place)
np.random.seed(2021)
numpy_input = (np.random.random(self.dims)).astype(self.dtype)
tensor_input = paddle.to_tensor(numpy_input)
numpy_output = np.argmin(numpy_input, axis=self.axis)
paddle_output = paddle.argmin(tensor_input, axis=self.axis)
self.assertEqual(np.allclose(numpy_output, paddle_output.numpy()),
True)
paddle.enable_static()
def forward(self, inputs):
return paddle.argmin(inputs)
def test_argmin_dtype_type():
data = paddle.static.data(name="test_argmin",
shape=[10],
dtype="float32")
output = paddle.argmin(x=data, dtype=None)
def test_argmin_axis_type():
data = paddle.static.data(name="test_argmin",
shape=[10],
dtype="float32")
output = paddle.argmin(x=data, axis=1.2)
def test_argmin_x_type():
x2 = [1, 2, 3]
output = paddle.argmin(x=x2)
def forward(self, inputs):
"""
forward
"""
x = paddle.argmin(inputs, axis=self.axis, keepdim=self.keepdim)
return x
def min(self, dim, keepdim=None):
return pd_min(self, dim, keepdim), paddle.argmin(self, dim, keepdim)
