def padding(X, ph, pw, val=0):
"""Pad X with the given value in 2-D
ph, pw : height and width padding
val : padding value, default 0
"""
assert len(X.shape) >= 2
nh, nw = X.shape[-2], X.shape[-1]
return te.compute(
(*X.shape[0:-2], nh + ph * 2, nw + pw * 2),
lambda *i: te.if_then_else(
te.any(i[-2] < ph, i[-2] >= nh + ph, i[-1] < pw, i[-1] >= nw + pw),
val, X[i[:-2] + (i[-2] - ph, i[-1] - pw)]),
name='PaddedX')
def check_bruteforce(bool_expr, vranges, cond=None):
""" Check that bool_expr holds given the condition cond
for every value of free variables from vranges.
TODO(yzhliu): move to utils
"""
if cond is not None:
bool_expr = te.any(tir.Not(cond), bool_expr)
res = run_expr(bool_expr, vranges)
if not np.all(res):
indices = list(np.argwhere(res == 0)[0])
counterex = [(str(v), i + r.min) for (v, r), i in zip(vranges.items(), indices)]
counterex = sorted(counterex, key=lambda x: x[0])
counterex = ", ".join([v + " = " + str(i) for v, i in counterex])
raise AssertionError("Expression {}\nis not true on {}\n"
"Counterexample: {}"
.format(tir.arith.Analyzer().simplify(bool_expr), vranges, counterex))
def _resize_2d(
indices,
data,
roi,
image_height,
image_width,
target_height,
target_width,
boxes=None,
box_indices=None,
method=None,
extrapolation_value=0.0,
layout="NCHW",
coordinate_transformation_mode="align_corners",
rounding_method="",
alpha=-0.5,
exclude_outside=0,
out_dtype=None,
):
"""Perform resize operation on the data with selected method and options.
Parameters
----------
indices : tuple
The indices of input data
data : tvm.te.Tensor
inputs is a 4-D tensor with shape
[batch, channel, in_height, in_width]
or [batch, in_height, in_width, channel]
roi: Tuple of Float or Expr
The region of interest for cropping the input image. Expected to be of
size 4, and format [start_h, start_w, end_h, end_w].
Only used if coordinate_transformation_mode is tf_crop_and_resize.
image_height : integer
Input image height
image_width : integer
Input image width
target_height : integer
The target resized image height
target_width : integer
The target resized image width
boxes : tvm.te.Tensor, optional
A 2-D tensor of shape [num_boxes, 4]. Each row of the tensor specifies
the coordinates of a box.
method: string, optional
method of interpolation ("nearest", "linear", "bicubic")
box_indices : tvm.te.Tensor, optional
A 1-D tensor of shape [num_boxes], box_indices[i] specifies the data that
the i-th box refers to.
extrapolation_value: float, optional
Value used for extrapolation, when applicable.
layout: string, optional
"NCHW", "NHWC", or "NCHWc".
coordinate_transformation_mode : string, optional
Describes how to transform the coordinate in the resized tensor
to the coordinate in the original tensor.
[half_pixel, align_corners, asymmetric, pytorch_half_pixel,
tf_half_pixel_for_nn, and tf_crop_and_resize].
rounding_method: string, optional
indicates how to find the "nearest" pixel in nearest_neighbor method
[round, floor, ceil]
alpha: float, optional
Bicubic spline coefficient
exclude_outside: bool, optional:
Exclude values outside the image fdor bicubic interpolation
out_dtype: string, optional
Type to return. If left None will be same as input type.
Returns
-------
output : out_dtype
The computed result with type out_dtype
"""
def _cast_output(value, data_dtype="float32", out_dtype=None):
if out_dtype:
dtype = out_dtype
else:
dtype = data_dtype
return value.astype(dtype)
height_use_int_div = False
width_use_int_div = False
if method == "nearest_neighbor" and coordinate_transformation_mode == "asymmetric":
height_use_int_div = can_convert_multiply_to_intdiv(
image_height, target_height)
width_use_int_div = can_convert_multiply_to_intdiv(
image_width, target_width)
n, c, y, x, cc, inum, ic = get_2d_indices(indices, layout)
box_idx = box_indices(n) if box_indices is not None else n
if boxes is not None:
y1, x1 = boxes(n, 0), boxes(n, 1)
y2, x2 = boxes(n, 2), boxes(n, 3)
in_h = (image_height - 1) * (y2 - y1)
in_w = (image_width - 1) * (x2 - x1)
h_scale = in_h.astype("float") / (target_height - 1)
w_scale = in_w.astype("float") / (target_width - 1)
in_y = y1 * (image_height - 1) + h_scale * y
in_x = x1 * (image_width - 1) + w_scale * x
else:
in_x = get_inx(
x,
image_width,
target_width,
coordinate_transformation_mode,
roi[1],
roi[3],
width_use_int_div,
)
in_y = get_inx(
y,
image_height,
target_height,
coordinate_transformation_mode,
roi[0],
roi[2],
height_use_int_div,
)
if method == "nearest_neighbor":
if rounding_method == "":
if coordinate_transformation_mode == "align_corners":
rounding_method = "round"
else:
rounding_method = "floor"
closest_x_index = get_closest_index(in_x, rounding_method, boxes,
width_use_int_div)
closest_y_index = get_closest_index(in_y, rounding_method, boxes,
height_use_int_div)
value = get_2d_pixel(
data,
layout,
image_height,
image_width,
box_idx,
c,
closest_y_index,
closest_x_index,
cc,
inum,
ic,
)
elif method == "linear":
y_int = te.floor(in_y).astype("int32")
x_int = te.floor(in_x).astype("int32")
y_lerp = in_y - y_int
x_lerp = in_x - x_int
p = [[0 for i in range(2)] for j in range(2)]
for j in range(2):
for i in range(2):
p[j][i] = get_2d_pixel(
data,
layout,
image_height,
image_width,
box_idx,
c,
y_int + j,
x_int + i,
cc,
inum,
ic,
)
top = _lerp(*p[0], x_lerp)
bottom = _lerp(*p[1], x_lerp)
value = _lerp(top, bottom, y_lerp)
elif method == "cubic":
xint = te.floor(in_x).astype("int32")
xfract = in_x - te.floor(in_x)
yint = te.floor(in_y).astype("int32")
yfract = in_y - te.floor(in_y)
# Get the surrounding values
p = [[0 for i in range(4)] for j in range(4)]
for j in range(4):
for i in range(4):
p[j][i] = get_2d_pixel(
data,
layout,
image_height,
image_width,
box_idx,
c,
yint + j - 1,
xint + i - 1,
cc,
inum,
ic,
)
wx = _cubic_spline_weights(xfract, alpha)
wy = _cubic_spline_weights(yfract, alpha)
if exclude_outside:
for i in range(4):
wx[i] = te.if_then_else(
te.any(xint - 1 + i < 0, xint + i > image_width), 0.0,
wx[i])
wy[i] = te.if_then_else(
te.any(yint - 1 + i < 0, yint + i > image_height), 0.0,
wy[i])
sum_wx = sum(wx)
sum_wy = sum(wy)
wx = [w / sum_wx for w in wx]
wy = [w / sum_wy for w in wy]
col0 = _cubic_kernel(p[0], wx)
col1 = _cubic_kernel(p[1], wx)
col2 = _cubic_kernel(p[2], wx)
col3 = _cubic_kernel(p[3], wx)
value = _cubic_kernel([col0, col1, col2, col3], wy)
else:
raise ValueError("Unknown resize method:", method)
if coordinate_transformation_mode == "tf_crop_and_resize":
out = tvm.tir.if_then_else(
in_y < 0,
extrapolation_value,
tvm.tir.if_then_else(in_y > image_height - 1, extrapolation_value,
value),
)
# use extrapolation_value if in_x is out of boundary
value = tvm.tir.if_then_else(
in_x < 0,
extrapolation_value,
tvm.tir.if_then_else(in_x > image_width - 1, extrapolation_value,
out),
)
return _cast_output(value, data.dtype, out_dtype=out_dtype)
def _resize_1d(
indices,
data,
roi,
image_width,
target_width,
boxes=None,
box_indices=None,
method=None,
extrapolation_value=0.0,
layout="NCW",
coordinate_transformation_mode="align_corners",
rounding_method="",
alpha=-0.5,
exclude_outside=0,
out_dtype=None,
):
"""Perform resize operation on the data with selected method and options.
Parameters
----------
indices : tuple
The indices of input data
data : tvm.te.Tensor
inputs is a 3-D tensor with shape
[batch, channel, in_width]
or [batch, in_width, channel]
roi: Tuple of Float or Expr
The region of interest for cropping the input image. Expected to be of
size 2, and format [start_w, end_w].
Only used if coordinate_transformation_mode is tf_crop_and_resize.
image_width : integer
Input image width
target_width : integer
The target resized image width
boxes : tvm.te.Tensor, optional
A 2-D tensor of shape [num_boxes, 4]. Each row of the tensor specifies
the coordinates of a box.
box_indices : tvm.te.Tensor, optional
A 1-D tensor of shape [num_boxes], box_indices[i] specifies the data that
the i-th box refers to.
extrapolation_value: float, optional
Value used for extrapolation, when applicable.
layout: string, optional
"NCW", "NWC", or "NCWc".
method: string, optional
method of interpolation ("nearest", "linear", "bicubic")
coordinate_transformation_mode : string, optional
Describes how to transform the coordinate in the resized tensor
to the coordinate in the original tensor.
[half_pixel, align_corners, asymmetric, pytorch_half_pixel,
tf_half_pixel_for_nn, and tf_crop_and_resize].
rounding_method: string, optional
indicates how to find the "nearest" pixel in nearest_neighbor method
[round, floor, ceil]
alpha: float, optional
Bicubic spline coefficient
exclude_outside: bool, optional:
Exclude values outside the image fdor bicubic interpolation
out_dtype: string, optional
Type to return. If left None will be same as input type.
Returns
-------
output : out_dtype
The computed result with type out_dtype
"""
def _cast_output(value, data_dtype="float32", out_dtype=None):
if out_dtype:
dtype = out_dtype
else:
dtype = data_dtype
return value.astype(dtype)
n, c, x, cc, inum, ic = get_1d_indices(indices, layout)
box_idx = box_indices(n) if box_indices is not None else n
if boxes is not None:
# TODO(mbrookhart): Find an example of this
raise NotImplementedError(
"resize1d with image boxes not yet implemented")
in_x = get_inx(
x,
image_width,
target_width,
coordinate_transformation_mode,
roi[0],
roi[1],
)
if method == "nearest_neighbor":
if rounding_method == "":
if coordinate_transformation_mode == "align_corners":
rounding_method = "round"
else:
rounding_method = "floor"
closest_x_index = get_closest_index(in_x, rounding_method, boxes)
value = get_1d_pixel(
data,
layout,
image_width,
box_idx,
c,
closest_x_index,
cc,
inum,
ic,
)
elif method == "linear":
x_int = te.floor(in_x).astype("int32")
x_lerp = in_x - x_int
p = [0 for i in range(2)]
for i in range(2):
p[i] = get_1d_pixel(
data,
layout,
image_width,
box_idx,
c,
x_int + i,
cc,
inum,
ic,
)
value = _lerp(*p, x_lerp)
elif method == "cubic":
xint = te.floor(in_x).astype("int32")
xfract = in_x - te.floor(in_x)
# Get the surrounding values
p = [0 for i in range(4)]
for i in range(4):
p[i] = get_1d_pixel(
data,
layout,
image_width,
box_idx,
c,
xint + i - 1,
cc,
inum,
ic,
)
wx = _cubic_spline_weights(xfract, alpha)
if exclude_outside:
for i in range(4):
wx[i] = te.if_then_else(
te.any(xint - 1 + i < 0, xint + i > image_width), 0.0,
wx[i])
sum_wx = sum(wx)
wx = [w / sum_wx for w in wx]
value = _cubic_kernel(p, wx)
else:
raise ValueError("Unknown resize method:", method)
if coordinate_transformation_mode == "tf_crop_and_resize":
# use extrapolation_value if in_x is out of boundary
value = tvm.tir.if_then_else(
in_x < 0,
extrapolation_value,
tvm.tir.if_then_else(in_x > image_width - 1, extrapolation_value,
value),
)
return _cast_output(value, data.dtype, out_dtype=out_dtype)
def _resize_3d(
indices,
data,
roi,
image_depth,
image_height,
image_width,
target_depth,
target_height,
target_width,
boxes=None,
box_indices=None,
method=None,
extrapolation_value=0.0,
layout="NCHW",
coordinate_transformation_mode="align_corners",
rounding_method="",
alpha=-0.5,
exclude_outside=0,
out_dtype=None,
):
"""Perform resize operation on the data with selected method and options.
Parameters
----------
indices : tuple
The indices of input data
data : tvm.te.Tensor
inputs is a 4-D tensor with shape
[batch, channel, in_height, in_width]
or [batch, in_height, in_width, channel]
roi: Tuple of Float or Expr
The region of interest for cropping the input image. Expected to be of
size 6, and format [start_d, start_h, start_w, end_d, end_h, end_w].
Only used if coordinate_transformation_mode is tf_crop_and_resize.
image_depth : integer
Input image depth
image_height : integer
Input image height
image_width : integer
Input image width
target_depth : integer
The target resized image depth
target_height : integer
The target resized image height
target_width : integer
The target resized image width
boxes : tvm.te.Tensor, optional
A 2-D tensor of shape [num_boxes, 4]. Each row of the tensor specifies
the coordinates of a box.
box_indices : tvm.te.Tensor, optional
A 1-D tensor of shape [num_boxes], box_indices[i] specifies the data that
the i-th box refers to.
method: string, optional
method of interpolation ("nearest", "linear", "bicubic")
extrapolation_value: float, optional
Value used for extrapolation, when applicable.
layout: string, optional
"NCHW", "NHWC", or "NCHWc".
coordinate_transformation_mode : string, optional
Describes how to transform the coordinate in the resized tensor
to the coordinate in the original tensor.
[half_pixel, align_corners, asymmetric, pytorch_half_pixel,
tf_half_pixel_for_nn, and tf_crop_and_resize].
rounding_method: string, optional
indicates how to find the "nearest" pixel in nearest_neighbor method
[round, floor, ceil]
alpha: float, optional
Bicubic spline coefficient
exclude_oiutside: bool, optional:
Exclude values outside the image fdor bicubic interpolation
out_dtype: string, optional
Type to return. If left None will be same as input type.
Returns
-------
output : out_dtype
The computed result with type out_dtype
"""
def _cast_output(value, data_dtype="float32", out_dtype=None):
if out_dtype:
dtype = out_dtype
else:
dtype = data_dtype
return value.astype(dtype)
n, c, z, y, x, cc = get_3d_indices(indices, layout)
box_idx = box_indices(n) if box_indices is not None else n
if boxes is not None:
# TODO(mbrookhart): Find an example of this
raise NotImplementedError(
"resize1d with image boxes not yet implemented")
in_z = get_inx(z, image_depth, target_depth,
coordinate_transformation_mode, roi[2], roi[5])
in_y = get_inx(y, image_height, target_height,
coordinate_transformation_mode, roi[1], roi[4])
in_x = get_inx(x, image_width, target_width,
coordinate_transformation_mode, roi[0], roi[3])
if method == "nearest_neighbor":
if rounding_method == "":
if coordinate_transformation_mode == "align_corners":
rounding_method = "round"
else:
rounding_method = "floor"
closest_z_index = get_closest_index(in_z, rounding_method, boxes)
closest_y_index = get_closest_index(in_y, rounding_method, boxes)
closest_x_index = get_closest_index(in_x, rounding_method, boxes)
value = get_3d_pixel(
data,
layout,
image_depth,
image_height,
image_width,
box_idx,
c,
closest_z_index,
closest_y_index,
closest_x_index,
cc,
)
elif method == "linear":
z_int = te.floor(in_z).astype("int32")
y_int = te.floor(in_y).astype("int32")
x_int = te.floor(in_x).astype("int32")
z_lerp = in_z - z_int
y_lerp = in_y - y_int
x_lerp = in_x - x_int
p = [[[0 for i in range(2)] for j in range(2)] for k in range(2)]
for k in range(2):
for j in range(2):
for i in range(2):
p[k][j][i] = get_3d_pixel(
data,
layout,
image_depth,
image_height,
image_width,
box_idx,
c,
z_int + k,
y_int + j,
x_int + i,
cc,
)
l = [[0 for i in range(2)] for j in range(2)]
for j in range(2):
for i in range(2):
l[j][i] = _lerp(*p[j][i], x_lerp)
top = _lerp(*l[0], y_lerp)
bottom = _lerp(*l[1], y_lerp)
value = _lerp(top, bottom, z_lerp)
elif method == "cubic":
zint = te.floor(in_z).astype("int32")
zfract = in_z - te.floor(in_z)
yint = te.floor(in_y).astype("int32")
yfract = in_y - te.floor(in_y)
xint = te.floor(in_x).astype("int32")
xfract = in_x - te.floor(in_x)
# Get the surrounding values
p = [[[0 for i in range(4)] for j in range(4)] for k in range(4)]
for k in range(4):
for j in range(4):
for i in range(4):
p[k][j][i] = get_3d_pixel(
data,
layout,
image_depth,
image_height,
image_width,
box_idx,
c,
zint + k - 1,
yint + j - 1,
xint + i - 1,
cc,
)
wz = _cubic_spline_weights(zfract, alpha)
wy = _cubic_spline_weights(yfract, alpha)
wx = _cubic_spline_weights(xfract, alpha)
if exclude_outside:
for i in range(4):
wz[i] = te.if_then_else(
te.any(xint - 1 + i < 0, xint + i > image_height), 0.0,
wx[i])
wy[i] = te.if_then_else(
te.any(yint - 1 + i < 0, yint + i > image_height), 0.0,
wy[i])
wx[i] = te.if_then_else(
te.any(xint - 1 + i < 0, xint + i > image_width), 0.0,
wx[i])
sum_wz = sum(wz)
sum_wy = sum(wy)
sum_wx = sum(wx)
wz = [w / sum_wz for w in wz]
wy = [w / sum_wy for w in wy]
wx = [w / sum_wx for w in wx]
l = [[0 for i in range(4)] for j in range(4)]
for j in range(4):
for i in range(4):
l[j][i] = _cubic_kernel(p[j][i], wx)
col0 = _cubic_kernel(l[0], wy)
col1 = _cubic_kernel(l[1], wy)
col2 = _cubic_kernel(l[2], wy)
col3 = _cubic_kernel(l[3], wy)
value = _cubic_kernel([col0, col1, col2, col3], wz)
else:
raise ValueError("Unknown resize method:", method)
if coordinate_transformation_mode == "tf_crop_and_resize":
out = tvm.tir.if_then_else(
in_z < 0,
extrapolation_value,
tvm.tir.if_then_else(in_z > image_depth - 1, extrapolation_value,
value),
)
out = tvm.tir.if_then_else(
in_y < 0,
extrapolation_value,
tvm.tir.if_then_else(in_y > image_height - 1, extrapolation_value,
value),
)
# use extrapolation_value if in_x is out of boundary
value = tvm.tir.if_then_else(
in_x < 0,
extrapolation_value,
tvm.tir.if_then_else(in_x > image_width - 1, extrapolation_value,
out),
)
return _cast_output(value, data.dtype, out_dtype=out_dtype)
def resize_bicubic(
indices,
data,
image_height,
image_width,
target_height,
target_width,
boxes=None,
box_indices=None,
extrapolation_value=None,
layout="NCHW",
coordinate_transformation_mode="align_corners",
out_dtype=None,
alpha=-0.5,
exclude_outside=0,
):
"""Perform resize operation with bicubic method on the data.
More details about Bicubic interpolation please refer to
https://en.wikipedia.org/wiki/Bicubic_interpolation.
This algorithm is doing a bicubic spline interpolation
Parameters
----------
indices : tuple
The indices of input data
data : tvm.te.Tensor
inputs is a 4-D tensor with shape
[batch, channel, in_height, in_width]
or [:batch, in_height, in_width, channel]
image_height : integer
Input image height
image_width : integer
Input image width
target_height : integer
The target resized image height
target_width : integer
The target resized image width
boxes : tvm.te.Tensor, optional
A 2-D tensor of shape [num_boxes, 4]. Each row of the tensor specifies
the coordinates of a box.
box_indices : tvm.te.Tensor, optional
A 1-D tensor of shape [num_boxes], box_indices[i] specifies the data that
the i-th box refers to.
extrapolation_value: float, optional
Value used for extrapolation, when applicable.
layout: string, optional
"NCHW", "NHWC", or "NCHWc".
coordinate_transformation_mode: string, optional
Describes how to transform the coordinate in the resized tensor
to the coordinate in the original tensor.
Refer to the ONNX Resize operator specification for details.
Available options are "half_pixel", "align_corners" and "asymmetric".
out_dtype: string, optional
Type to return. If left None will be same as input type.
alpha: float, optional
Bicubic spline coefficient
Returns
-------
output : out_dtype
The computed result with type out_dtype
"""
def _cast_output(value, data_dtype="float32", out_dtype=None):
if out_dtype:
dtype = out_dtype
else:
dtype = data_dtype
return value.astype(dtype)
n, c, y, x, cc, inum, ic = get_2d_indices(indices, layout)
box_idx = box_indices(n) if box_indices is not None else n
if boxes is not None:
y1, x1 = boxes(n, 0), boxes(n, 1)
y2, x2 = boxes(n, 2), boxes(n, 3)
in_h = (image_height - 1) * (y2 - y1)
in_w = (image_width - 1) * (x2 - x1)
h_scale = in_h.astype("float") / (target_height - 1)
w_scale = in_w.astype("float") / (target_width - 1)
in_y = y1 * (image_height - 1) + h_scale * y
in_x = x1 * (image_width - 1) + w_scale * x
else:
in_y, in_x = get_iny_inx(
y,
x,
image_height,
image_width,
target_height,
target_width,
coordinate_transformation_mode,
)
xint = te.floor(in_x).astype("int32")
xfract = in_x - te.floor(in_x)
yint = te.floor(in_y).astype("int32")
yfract = in_y - te.floor(in_y)
# Get the surrounding values
p = [[0 for i in range(4)] for j in range(4)]
for j in range(4):
for i in range(4):
p[j][i] = get_2d_pixel(
data,
layout,
boxes,
image_height,
image_width,
box_idx,
c,
yint + j - 1,
xint + i - 1,
cc,
inum,
ic,
)
# Interpolate bicubically
def _cubic_spline_weights(t):
t2 = t * t
t3 = t * t * t
w1 = alpha * (t3 - 2 * t2 + t)
w2 = (alpha + 2) * t3 - (3 + alpha) * t2 + 1
w3 = -(alpha + 2) * t3 + (3 + 2 * alpha) * t2 - alpha * t
w4 = -alpha * t3 + alpha * t2
return [w1, w2, w3, w4]
def _cubic_kernel(inputs, w):
return sum([a_i * w_i for a_i, w_i in zip(inputs, w)])
wx = _cubic_spline_weights(xfract)
wy = _cubic_spline_weights(yfract)
if exclude_outside:
for i in range(4):
wx[i] = te.if_then_else(
te.any(xint - 1 + i < 0, xint + i > image_width), 0.0, wx[i])
wy[i] = te.if_then_else(
te.any(yint - 1 + i < 0, yint + i > image_height), 0.0, wy[i])
sum_wx = sum(wx)
sum_wy = sum(wy)
wx = [w / sum_wx for w in wx]
wy = [w / sum_wy for w in wy]
col0 = _cubic_kernel(p[0], wx)
col1 = _cubic_kernel(p[1], wx)
col2 = _cubic_kernel(p[2], wx)
col3 = _cubic_kernel(p[3], wx)
value = _cubic_kernel([col0, col1, col2, col3], wy)
# use extrapolation_value if in_y/in_x is out of boundary
if extrapolation_value is not None:
out = tvm.tir.if_then_else(
in_y < 0,
extrapolation_value,
tvm.tir.if_then_else(in_y > image_height - 1, extrapolation_value,
value),
)
value = tvm.tir.if_then_else(
in_x < 0,
extrapolation_value,
tvm.tir.if_then_else(in_x > image_width - 1, extrapolation_value,
out),
)
return _cast_output(value, data.dtype, out_dtype=out_dtype)
