def test_arr_4_4_2_filter_3_3():
pad_val = 0.98
arr = np.arange(2*2*2).reshape((2, 2, 2)).astype(DTYPE)
filter_shape = (3, 3)
ref = np.array([[[pad_val, pad_val],
[pad_val, pad_val],
[pad_val, pad_val],
[pad_val, pad_val]],
[[pad_val, pad_val],
[0., 1.],
[2., 3.],
[pad_val, pad_val]],
[[pad_val, pad_val],
[4., 5.],
[6., 7.],
[pad_val, pad_val]],
[[pad_val, pad_val],
[pad_val, pad_val],
[pad_val, pad_val],
[pad_val, pad_val]]]).astype(DTYPE)
res = filter_pad2d(arr, filter_shape, constant=pad_val)
assert_allclose(res, ref, rtol=RTOL, atol=ATOL)
def _process_one_op(self, arr, X, Y,
layer_idx, op_idx, kwargs,
op_params,
op_name, nbh, nbw, stride,
pad_apron=False,
interleave_stride=False):
out_l = []
# -- here we compute the pixel coordinates of
# the central pixel in a patch
hc, wc = nbh / 2, nbw / 2
if pad_apron:
arr = filter_pad2d(arr, (nbh, nbw))
X = np.squeeze(filter_pad2d(X[..., np.newaxis], (nbh, nbw),
constant=-1))
Y = np.squeeze(filter_pad2d(Y[..., np.newaxis], (nbh, nbw),
constant=-1))
if interleave_stride:
for i in xrange(stride):
for j in xrange(stride):
arr_out_ij = self._get_feature_map(arr[i::, j::, ...],
layer_idx, op_idx, kwargs,
op_params, op_name)
X_out_ij = view_as_windows(X[i::, j::],
(nbh, nbw))[::stride, ::stride,
hc, wc]
Y_out_ij = view_as_windows(Y[i::, j::],
(nbh, nbw))[::stride, ::stride,
hc, wc]
out_l += [(arr_out_ij, X_out_ij, Y_out_ij)]
else:
arr_out = self._get_feature_map(arr, layer_idx, op_idx, kwargs,
op_params, op_name)
X_out = view_as_windows(X, (nbh, nbw))[::stride, ::stride, hc, wc]
Y_out = view_as_windows(Y, (nbh, nbw))[::stride, ::stride, hc, wc]
out_l += [(arr_out, X_out, Y_out)]
return out_l
def test_arr_3_3_1_filter_2_2():
pad_val = 1.23
arr = np.arange(3*3*1).reshape((3, 3, 1)).astype(DTYPE)
filter_shape = (2, 2)
ref = np.array([[[pad_val], [pad_val], [pad_val], [pad_val]],
[[pad_val], [0.], [1.], [2.]],
[[pad_val], [3.], [4.], [5.]],
[[pad_val], [6.], [7.], [8.]]])
res = filter_pad2d(arr, filter_shape, constant=pad_val)
assert_allclose(res, ref, rtol=RTOL, atol=ATOL)
def test_arr_3_3_1_filter_2_2():
pad_val = 1.23
arr = np.arange(3 * 3 * 1).reshape((3, 3, 1)).astype(DTYPE)
filter_shape = (2, 2)
ref = np.array([[[pad_val], [pad_val], [pad_val], [pad_val]],
[[pad_val], [0.], [1.], [2.]], [[pad_val], [3.], [4.],
[5.]],
[[pad_val], [6.], [7.], [8.]]])
res = filter_pad2d(arr, filter_shape, constant=pad_val)
assert_allclose(res, ref, rtol=RTOL, atol=ATOL)
def test_arr_4_4_2_filter_3_3():
pad_val = 0.98
arr = np.arange(2 * 2 * 2).reshape((2, 2, 2)).astype(DTYPE)
filter_shape = (3, 3)
ref = np.array([[[pad_val, pad_val], [pad_val, pad_val],
[pad_val, pad_val], [pad_val, pad_val]],
[[pad_val, pad_val], [0., 1.], [2., 3.],
[pad_val, pad_val]],
[[pad_val, pad_val], [4., 5.], [6., 7.],
[pad_val, pad_val]],
[[pad_val, pad_val], [pad_val, pad_val],
[pad_val, pad_val], [pad_val, pad_val]]]).astype(DTYPE)
res = filter_pad2d(arr, filter_shape, constant=pad_val)
assert_allclose(res, ref, rtol=RTOL, atol=ATOL)
def transform(self, arr_in, pad_apron=False, interleave_stride=False):
"""XXX: docstring for transform"""
rcpt_field = self.receptive_field_shape
description = self.description
input_shape = arr_in.shape
assert input_shape[:2] == self.in_shape
assert len(input_shape) == 2 or len(input_shape) == 3
if len(input_shape) == 3:
tmp_out = arr_in
elif len(input_shape) == 2:
tmp_out = arr_in[..., np.newaxis]
else:
raise ValueError("The input array should be 2D or 3D")
# -- first we initialize some variables to be used in
# the processing of the feature maps
h, w = self.in_shape
Y, X = np.mgrid[:h, :w]
tmp_out_l = [(tmp_out, X, Y)]
nbh_nbw_stride_l = self.ops_nbh_nbw_stride
# -- loop over all the SLM operations in order
op_counter = 0
for layer_idx, layer_desc in enumerate(description):
for op_idx, (op_name, op_params) in enumerate(layer_desc):
kwargs = op_params['kwargs']
tmp_l = []
_, nbh, nbw, stride = nbh_nbw_stride_l[op_counter]
for arr, X, Y in tmp_out_l:
tmp_l += self._process_one_op(arr, X, Y,
layer_idx, op_idx, kwargs,
op_params,
op_name, nbh, nbw, stride,
pad_apron=pad_apron,
interleave_stride=interleave_stride)
tmp_out_l = tmp_l
op_counter += 1
# -- now we need to possibly interleave the arrays
# in ``tmp_out_l``
if interleave_stride:
out_shape = (h, w, tmp_out_l[0][0].shape[-1])
arr_out = np.empty(out_shape, dtype=arr_in.dtype)
Y_ref, X_ref = np.mgrid[:h, :w]
Y_int, X_int = np.zeros((h, w), dtype=np.int), \
np.zeros((h, w), dtype=np.int)
if pad_apron:
for arr, Xc, Yc in tmp_out_l:
anchor_h, anchor_w = Yc[0, 0], Xc[0, 0]
stride_h = Yc[1, 0] - Yc[0, 0]
stride_w = Xc[0, 1] - Xc[0, 0]
arr_out[anchor_h::stride_h, anchor_w::stride_w, ...] = arr
X_int[anchor_h::stride_h, anchor_w::stride_w] = Xc
Y_int[anchor_h::stride_h, anchor_w::stride_w] = Yc
assert (X_int == X_ref).all()
assert (Y_int == Y_ref).all()
return arr_out
else:
X_int = filter_pad2d(X_int[..., np.newaxis], rcpt_field,
reverse_padding=True).squeeze()
Y_int = filter_pad2d(Y_int[..., np.newaxis], rcpt_field,
reverse_padding=True).squeeze()
X_ref = filter_pad2d(X_ref[..., np.newaxis], rcpt_field,
reverse_padding=True).squeeze()
Y_ref = filter_pad2d(Y_ref[..., np.newaxis], rcpt_field,
reverse_padding=True).squeeze()
arr_out = filter_pad2d(arr_out, rcpt_field,
reverse_padding=True)
offset_Y = Y_ref.min()
offset_X = X_ref.min()
for arr, Xc, Yc in tmp_out_l:
anchor_h, anchor_w = Yc[0, 0] - offset_Y, \
Xc[0, 0] - offset_X
stride_h = Yc[1, 0] - Yc[0, 0]
stride_w = Xc[0, 1] - Xc[0, 0]
arr_out[anchor_h::stride_h, anchor_w::stride_w, ...] = arr
X_int[anchor_h::stride_h, anchor_w::stride_w] = Xc
Y_int[anchor_h::stride_h, anchor_w::stride_w] = Yc
assert (X_int == X_ref).all()
assert (Y_int == Y_ref).all()
return arr_out
else:
assert len(tmp_out_l) == 1
arr_out, _, _ = tmp_out_l[0]
return arr_out
