def bbox2delta_v2(src_boxes,
tgt_boxes,
means=(0.0, 0.0, 0.0, 0.0),
stds=(1.0, 1.0, 1.0, 1.0)):
"""Encode bboxes to deltas.
Modified from ppdet.modeling.bbox_utils.bbox2delta.
Args:
src_boxes (Tensor[..., 4]): base bboxes
tgt_boxes (Tensor[..., 4]): target bboxes
means (list[float]): the mean that will be used to normalize delta
stds (list[float]): the std that will be used to normalize delta
"""
if src_boxes.size == 0:
return paddle.empty_like(src_boxes)
src_w = src_boxes[..., 2] - src_boxes[..., 0]
src_h = src_boxes[..., 3] - src_boxes[..., 1]
src_ctr_x = src_boxes[..., 0] + 0.5 * src_w
src_ctr_y = src_boxes[..., 1] + 0.5 * src_h
tgt_w = tgt_boxes[..., 2] - tgt_boxes[..., 0]
tgt_h = tgt_boxes[..., 3] - tgt_boxes[..., 1]
tgt_ctr_x = tgt_boxes[..., 0] + 0.5 * tgt_w
tgt_ctr_y = tgt_boxes[..., 1] + 0.5 * tgt_h
dx = (tgt_ctr_x - src_ctr_x) / src_w
dy = (tgt_ctr_y - src_ctr_y) / src_h
dw = paddle.log(tgt_w / src_w)
dh = paddle.log(tgt_h / src_h)
deltas = paddle.stack((dx, dy, dw, dh), axis=1) # [n, 4]
means = paddle.to_tensor(means, place=src_boxes.place)
stds = paddle.to_tensor(stds, place=src_boxes.place)
deltas = (deltas - means) / stds
return deltas
def test_static_graph(self):
paddle.enable_static()
dtype = 'float32'
train_program = Program()
startup_program = Program()
with program_guard(train_program, startup_program):
x = np.random.random(self.x_shape).astype(dtype)
data_x = paddle.static.data('x',
shape=self.data_x_shape,
dtype=dtype)
out = paddle.empty_like(data_x)
place = paddle.CUDAPlace(
0) if core.is_compiled_with_cuda() else paddle.CPUPlace()
exe = paddle.static.Executor(place)
res = exe.run(train_program, feed={'x': x}, fetch_list=[out])
self.dst_dtype = dtype
self.dst_shape = x.shape
self.__check_out__(res[0])
paddle.disable_static()
def delta2bbox_v2(rois,
deltas,
means=(0.0, 0.0, 0.0, 0.0),
stds=(1.0, 1.0, 1.0, 1.0),
max_shape=None,
wh_ratio_clip=16.0 / 1000.0,
ctr_clip=None):
"""Transform network output(delta) to bboxes.
Based on https://github.com/open-mmlab/mmdetection/blob/master/mmdet/core/
bbox/coder/delta_xywh_bbox_coder.py
Args:
rois (Tensor): shape [..., 4], base bboxes, typical examples include
anchor and rois
deltas (Tensor): shape [..., 4], offset relative to base bboxes
means (list[float]): the mean that was used to normalize deltas,
must be of size 4
stds (list[float]): the std that was used to normalize deltas,
must be of size 4
max_shape (list[float] or None): height and width of image, will be
used to clip bboxes if not None
wh_ratio_clip (float): to clip delta wh of decoded bboxes
ctr_clip (float or None): whether to clip delta xy of decoded bboxes
"""
if rois.size == 0:
return paddle.empty_like(rois)
means = paddle.to_tensor(means)
stds = paddle.to_tensor(stds)
deltas = deltas * stds + means
dxy = deltas[..., :2]
dwh = deltas[..., 2:]
pxy = (rois[..., :2] + rois[..., 2:]) * 0.5
pwh = rois[..., 2:] - rois[..., :2]
dxy_wh = pwh * dxy
max_ratio = np.abs(np.log(wh_ratio_clip))
if ctr_clip is not None:
dxy_wh = paddle.clip(dxy_wh, max=ctr_clip, min=-ctr_clip)
dwh = paddle.clip(dwh, max=max_ratio)
else:
dwh = dwh.clip(min=-max_ratio, max=max_ratio)
gxy = pxy + dxy_wh
gwh = pwh * dwh.exp()
x1y1 = gxy - (gwh * 0.5)
x2y2 = gxy + (gwh * 0.5)
bboxes = paddle.concat([x1y1, x2y2], axis=-1)
if max_shape is not None:
bboxes[..., 0::2] = bboxes[..., 0::2].clip(min=0, max=max_shape[1])
bboxes[..., 1::2] = bboxes[..., 1::2].clip(min=0, max=max_shape[0])
return bboxes
def _reset_parameters(self):
# init all parameters.
prior_prob = 0.01
bias_value = -math.log((1 - prior_prob) / prior_prob)
for m in self.sublayers():
if isinstance(m, nn.Linear):
init.xavier_normal_(m.weight, reverse=True)
elif not isinstance(m, nn.Embedding) and hasattr(
m, "weight") and m.weight.dim() > 1:
init.xavier_normal_(m.weight, reverse=False)
if hasattr(m, "bias") and m.bias is not None and m.bias.shape[
-1] == self.num_classes:
init.constant_(m.bias, bias_value)
init_bboxes = paddle.empty_like(self.init_proposal_boxes.weight)
init_bboxes[:, :2] = 0.5
init_bboxes[:, 2:] = 1.0
self.init_proposal_boxes.weight.set_value(init_bboxes)
def test_dygraph_api_out(self):
paddle.disable_static()
out = paddle.empty_like(self.x, self.dtype)
self.__check_out__(out.numpy())
paddle.enable_static()
def test_dtype():
x = np.random.random((200, 3)).astype("float64")
dtype = 'uint8'
result = paddle.empty_like(x, dtype=dtype)
def taylor(M: int,
nbar=4,
sll=30,
norm=True,
sym: bool = True,
dtype: str = 'float64') -> Tensor:
"""Compute a Taylor window.
The Taylor window taper function approximates the Dolph-Chebyshev window's
constant sidelobe level for a parameterized number of near-in sidelobes.
Parameters:
M(int): window size
nbar, sil, norm: the window-specific parameter.
sym(bool)：whether to return symmetric window.
The default value is True
dtype(str): the datatype of returned tensor.
Returns:
Tensor: the window tensor
Notes:
This function is consistent with scipy.signal.windows.taylor().
"""
if _len_guards(M):
return paddle.ones((M, ), dtype=dtype)
M, needs_trunc = _extend(M, sym)
# Original text uses a negative sidelobe level parameter and then negates
# it in the calculation of B. To keep consistent with other methods we
# assume the sidelobe level parameter to be positive.
B = 10**(sll / 20)
A = _acosh(B) / math.pi
s2 = nbar**2 / (A**2 + (nbar - 0.5)**2)
ma = paddle.arange(1, nbar, dtype=dtype)
Fm = paddle.empty((nbar - 1, ), dtype=dtype)
signs = paddle.empty_like(ma)
signs[::2] = 1
signs[1::2] = -1
m2 = ma * ma
for mi in range(len(ma)):
numer = signs[mi] * paddle.prod(1 - m2[mi] / s2 / (A**2 +
(ma - 0.5)**2))
if mi == 0:
denom = 2 * paddle.prod(1 - m2[mi] / m2[mi + 1:])
elif mi == len(ma) - 1:
denom = 2 * paddle.prod(1 - m2[mi] / m2[:mi])
else:
denom = 2 * paddle.prod(1 - m2[mi] / m2[:mi]) * paddle.prod(
1 - m2[mi] / m2[mi + 1:])
Fm[mi] = numer / denom
def W(n):
return 1 + 2 * paddle.matmul(
Fm.unsqueeze(0),
paddle.cos(2 * math.pi * ma.unsqueeze(1) * (n - M / 2. + 0.5) / M))
w = W(paddle.arange(0, M, dtype=dtype))
# normalize (Note that this is not described in the original text [1])
if norm:
scale = 1.0 / W((M - 1) / 2)
w *= scale
w = w.squeeze()
return _truncate(w, needs_trunc)
def test_create_process_group_nccl(self):
with _test_eager_guard():
paddle.set_device('gpu:%d' %
paddle.distributed.ParallelEnv().dev_id)
pg = init_process_group()
print("rank:", pg.rank(), "size:", pg.size(), "name:", pg.name())
print("test new group api ok")
# test allreduce sum
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
sum_result = tensor_x + tensor_y
if pg.rank() == 0:
task = dist.all_reduce(tensor_x)
assert np.array_equal(tensor_x, sum_result)
else:
task = dist.all_reduce(tensor_y)
assert np.array_equal(tensor_y, sum_result)
print("test allreduce sum api ok")
# test allreduce max
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
max_result = paddle.maximum(tensor_x, tensor_y)
if pg.rank() == 0:
task = dist.all_reduce(tensor_x,
dist.ReduceOp.MAX,
use_calc_stream=False)
task.wait()
assert np.array_equal(tensor_x, max_result)
else:
task = dist.all_reduce(tensor_y,
dist.ReduceOp.MAX,
use_calc_stream=False)
task.wait()
assert np.array_equal(tensor_y, max_result)
print("test allreduce max api ok")
# test allreduce min
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
min_result = paddle.minimum(tensor_x, tensor_y)
if pg.rank() == 0:
task = dist.all_reduce(tensor_x,
dist.ReduceOp.MIN,
use_calc_stream=False)
task.wait()
assert np.array_equal(tensor_x, min_result)
else:
task = dist.all_reduce(tensor_y,
dist.ReduceOp.MIN,
use_calc_stream=False)
task.wait()
assert np.array_equal(tensor_y, min_result)
print("test allreduce min api ok")
# test allreduce prod
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
prod_result = np.multiply(x, y)
if pg.rank() == 0:
task = dist.all_reduce(tensor_x,
dist.ReduceOp.PROD,
use_calc_stream=False)
task.wait()
assert np.array_equal(tensor_x, prod_result)
else:
task = dist.all_reduce(tensor_y,
dist.ReduceOp.PROD,
use_calc_stream=False)
task.wait()
assert np.array_equal(tensor_y, prod_result)
print("test allreduce prod api ok")
# test broadcast
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
broadcast_result = paddle.assign(tensor_x)
if pg.rank() == 0:
task = dist.broadcast(tensor_x, 0, use_calc_stream=False)
task.synchronize()
paddle.device.cuda.synchronize()
assert task.is_completed()
assert np.array_equal(broadcast_result, tensor_x)
else:
task = dist.broadcast(tensor_y, 0)
paddle.device.cuda.synchronize()
assert np.array_equal(broadcast_result, tensor_y)
print("test broadcast api ok")
# test barrier
# rank 0
if pg.rank() == 0:
dist.barrier()
# rank 1
else:
task = pg.barrier()
task.wait()
print("test barrier api ok\n")
# test allgather
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
out_shape = list(self.shape)
out_shape[0] *= 2
out = np.random.random(out_shape).astype(self.dtype)
tensor_out = paddle.to_tensor(out)
if pg.rank() == 0:
task = pg.all_gather(tensor_x, tensor_out)
task.wait()
paddle.device.cuda.synchronize()
# rank 1
else:
tensor_out_list = [
paddle.empty_like(tensor_x),
paddle.empty_like(tensor_x)
]
task = dist.all_gather(tensor_out_list,
tensor_y,
use_calc_stream=False)
paddle.device.cuda.synchronize()
tensor_out = paddle.concat(tensor_out_list)
out_1 = paddle.slice(tensor_out, [0], [0], [out_shape[0] // 2])
out_2 = paddle.slice(tensor_out, [0], [out_shape[0] // 2],
[out_shape[0]])
assert np.array_equal(tensor_x, out_1)
assert np.array_equal(tensor_y, out_2)
print("test allgather api ok\n")
if pg.rank() == 0:
task = pg.all_gather(tensor_x, tensor_out)
task.wait()
paddle.device.cuda.synchronize()
# rank 1
else:
tensor_out_list = []
task = dist.all_gather(tensor_out_list,
tensor_y,
use_calc_stream=False)
paddle.device.cuda.synchronize()
tensor_out = paddle.concat(tensor_out_list)
out_1 = paddle.slice(tensor_out, [0], [0], [out_shape[0] // 2])
out_2 = paddle.slice(tensor_out, [0], [out_shape[0] // 2],
[out_shape[0]])
assert np.array_equal(tensor_x, out_1)
assert np.array_equal(tensor_y, out_2)
print("test allgather api2 ok\n")
# test alltoall
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
out1 = np.random.random(self.shape).astype(self.dtype)
out2 = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
tensor_out1 = paddle.to_tensor(out1)
tensor_out2 = paddle.to_tensor(out2)
raw_tensor_x_2 = paddle.slice(tensor_x, [0], [self.shape[0] // 2],
[self.shape[0]])
raw_tensor_y_1 = paddle.slice(tensor_y, [0], [0],
[self.shape[0] // 2])
if pg.rank() == 0:
task = pg.alltoall(tensor_x, tensor_out1)
task.wait()
# rank 1
else:
in_1, in_2 = paddle.split(tensor_y, 2)
out_1, out_2 = paddle.split(tensor_out2, 2)
out_tensor_list = [out_1, out_2]
task = dist.alltoall([in_1, in_2], out_tensor_list)
paddle.device.cuda.synchronize()
tensor_out2 = paddle.concat(out_tensor_list)
out1_2 = paddle.slice(tensor_out1, [0], [self.shape[0] // 2],
[self.shape[0]])
out2_1 = paddle.slice(tensor_out2, [0], [0], [self.shape[0] // 2])
if pg.rank() == 0:
assert np.array_equal(out1_2.numpy(), raw_tensor_y_1.numpy())
else:
assert np.array_equal(out2_1, raw_tensor_x_2)
print("test alltoall api ok\n")
x = np.random.random(self.shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
out1 = np.random.random(self.shape).astype(self.dtype)
out2 = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
tensor_out1 = paddle.to_tensor(out1)
tensor_out2 = paddle.to_tensor(out2)
raw_tensor_x_2 = paddle.slice(tensor_x, [0], [self.shape[0] // 2],
[self.shape[0]])
raw_tensor_y_1 = paddle.slice(tensor_y, [0], [0],
[self.shape[0] // 2])
if pg.rank() == 0:
task = pg.alltoall(tensor_x, tensor_out1)
task.wait()
# rank 1
else:
in_1, in_2 = paddle.split(tensor_y, 2)
out_1, out_2 = paddle.split(tensor_out2, 2)
out_tensor_list = []
task = dist.alltoall([in_1, in_2], out_tensor_list)
paddle.device.cuda.synchronize()
tensor_out2 = paddle.concat(out_tensor_list)
out1_2 = paddle.slice(tensor_out1, [0], [self.shape[0] // 2],
[self.shape[0]])
out2_1 = paddle.slice(tensor_out2, [0], [0], [self.shape[0] // 2])
if pg.rank() == 0:
assert np.array_equal(out1_2.numpy(), raw_tensor_y_1.numpy())
else:
assert np.array_equal(out2_1, raw_tensor_x_2)
print("test alltoall api2 ok\n")
# test Reduce
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
sum_result = tensor_x + tensor_y
if pg.rank() == 0:
task = dist.reduce(tensor_x, 0, use_calc_stream=True)
paddle.device.cuda.synchronize()
# rank 1
else:
task = dist.reduce(tensor_y, 0, use_calc_stream=False)
task.wait()
paddle.device.cuda.synchronize()
if pg.rank() == 0:
assert np.array_equal(tensor_x, sum_result)
print("test reduce sum api ok\n")
# test reduce max
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
max_result = paddle.maximum(tensor_x, tensor_y)
if pg.rank() == 0:
task = dist.reduce(tensor_x,
0,
dist.ReduceOp.MAX,
use_calc_stream=False)
task.wait()
assert np.array_equal(tensor_x, max_result)
else:
task = dist.reduce(tensor_y,
0,
dist.ReduceOp.MAX,
use_calc_stream=False)
task.wait()
print("test reduce max api ok")
# test reduce min
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
min_result = paddle.minimum(tensor_x, tensor_y)
if pg.rank() == 0:
task = dist.reduce(tensor_x,
0,
dist.ReduceOp.MIN,
use_calc_stream=False)
task.wait()
assert np.array_equal(tensor_x, min_result)
else:
task = dist.reduce(tensor_y,
0,
dist.ReduceOp.MIN,
use_calc_stream=False)
task.wait()
print("test reduce min api ok")
# test reduce product
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
prod_result = np.multiply(x, y)
if pg.rank() == 0:
task = dist.reduce(tensor_x,
0,
dist.ReduceOp.PROD,
use_calc_stream=False)
task.wait()
assert np.array_equal(tensor_x, prod_result)
else:
task = dist.reduce(tensor_y,
0,
dist.ReduceOp.PROD,
use_calc_stream=False)
task.wait()
print("test reduce prod api ok")
# test Scatter
# rank 0
in_shape = list(self.shape)
in_shape[0] *= 2
x = np.random.random(in_shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
if pg.rank() == 0:
in_1, in_2 = paddle.split(tensor_x, 2)
task = dist.scatter(tensor_y, [in_1, in_2],
0,
use_calc_stream=True)
#task.wait()
paddle.device.cuda.synchronize()
# rank 1
else:
task = dist.scatter(tensor_y, [], 0, use_calc_stream=False)
task.wait()
paddle.device.cuda.synchronize()
out1 = paddle.slice(tensor_x, [0], [0], [self.shape[0]])
out2 = paddle.slice(tensor_x, [0], [self.shape[0]],
[self.shape[0] * 2])
if pg.rank() == 0:
assert np.array_equal(tensor_y, out1)
else:
assert np.array_equal(tensor_y, out2)
print("test scatter api ok\n")
# test send min
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
if pg.rank() == 0:
task = dist.send(tensor_x, 1, use_calc_stream=False)
task.wait()
else:
task = dist.recv(tensor_y, 0, use_calc_stream=False)
task.wait()
assert np.array_equal(tensor_y, tensor_x)
print("test send api ok")
# test send min
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
if pg.rank() == 0:
task = dist.send(tensor_x, 1, use_calc_stream=True)
else:
task = dist.recv(tensor_y, 0, use_calc_stream=True)
assert np.array_equal(tensor_y, tensor_x)
print("test send api ok")
