def test_resize_gpu():
# Test with normal case 3D input float type
data_in_3d = nd.random.uniform(0, 255, (300, 300, 3))
out_nd_3d = transforms.Resize((100, 100))(data_in_3d)
data_in_4d_nchw = nd.moveaxis(nd.expand_dims(data_in_3d, axis=0), 3, 1)
data_expected_3d = (nd.moveaxis(
nd.contrib.BilinearResize2D(data_in_4d_nchw,
height=100,
width=100,
align_corners=False), 1, 3))[0]
assert_almost_equal(out_nd_3d.asnumpy(), data_expected_3d.asnumpy())
# Test with normal case 4D input float type
data_in_4d = nd.random.uniform(0, 255, (2, 300, 300, 3))
out_nd_4d = transforms.Resize((100, 100))(data_in_4d)
data_in_4d_nchw = nd.moveaxis(data_in_4d, 3, 1)
data_expected_4d = nd.moveaxis(
nd.contrib.BilinearResize2D(data_in_4d_nchw,
height=100,
width=100,
align_corners=False), 1, 3)
assert_almost_equal(out_nd_4d.asnumpy(), data_expected_4d.asnumpy())
# Test invalid interp
data_in_3d = nd.random.uniform(0, 255, (300, 300, 3))
invalid_transform = transforms.Resize(-150,
keep_ratio=False,
interpolation=2)
assertRaises(MXNetError, invalid_transform, data_in_3d)
# Credited to Hang Zhang
def py_bilinear_resize_nhwc(x, outputHeight, outputWidth):
batch, inputHeight, inputWidth, channel = x.shape
if outputHeight == inputHeight and outputWidth == inputWidth:
return x
y = np.empty([batch, outputHeight, outputWidth,
channel]).astype('uint8')
rheight = 1.0 * (inputHeight - 1) / (outputHeight -
1) if outputHeight > 1 else 0.0
rwidth = 1.0 * (inputWidth - 1) / (outputWidth -
1) if outputWidth > 1 else 0.0
for h2 in range(outputHeight):
h1r = 1.0 * h2 * rheight
h1 = int(np.floor(h1r))
h1lambda = h1r - h1
h1p = 1 if h1 < (inputHeight - 1) else 0
for w2 in range(outputWidth):
w1r = 1.0 * w2 * rwidth
w1 = int(np.floor(w1r))
w1lambda = w1r - w1
w1p = 1 if w1 < (inputHeight - 1) else 0
for b in range(batch):
for c in range(channel):
y[b][h2][w2][c] = (1-h1lambda)*((1-w1lambda)*x[b][h1][w1][c] + \
w1lambda*x[b][h1][w1+w1p][c]) + \
h1lambda*((1-w1lambda)*x[b][h1+h1p][w1][c] + \
w1lambda*x[b][h1+h1p][w1+w1p][c])
return y
def test_resize():
# Test with normal case 3D input float type
data_in_3d = nd.random.uniform(0, 255, (300, 300, 3))
out_nd_3d = transforms.Resize((100, 100))(data_in_3d)
data_in_4d_nchw = nd.moveaxis(nd.expand_dims(data_in_3d, axis=0), 3, 1)
data_expected_3d = (nd.moveaxis(nd.contrib.BilinearResize2D(data_in_4d_nchw, 100, 100), 1, 3))[0]
assert_almost_equal(out_nd_3d.asnumpy(), data_expected_3d.asnumpy())
# Test with normal case 4D input float type
data_in_4d = nd.random.uniform(0, 255, (2, 300, 300, 3))
out_nd_4d = transforms.Resize((100, 100))(data_in_4d)
data_in_4d_nchw = nd.moveaxis(data_in_4d, 3, 1)
data_expected_4d = nd.moveaxis(nd.contrib.BilinearResize2D(data_in_4d_nchw, 100, 100), 1, 3)
assert_almost_equal(out_nd_4d.asnumpy(), data_expected_4d.asnumpy())
# Test invalid interp
data_in_3d = nd.random.uniform(0, 255, (300, 300, 3))
invalid_transform = transforms.Resize(-150, keep_ratio=False, interpolation=2)
assertRaises(MXNetError, invalid_transform, data_in_3d)
# Credited to Hang Zhang
def py_bilinear_resize_nhwc(x, outputHeight, outputWidth):
batch, inputHeight, inputWidth, channel = x.shape
if outputHeight == inputHeight and outputWidth == inputWidth:
return x
y = np.empty([batch, outputHeight, outputWidth, channel]).astype('uint8')
rheight = 1.0 * (inputHeight - 1) / (outputHeight - 1) if outputHeight > 1 else 0.0
rwidth = 1.0 * (inputWidth - 1) / (outputWidth - 1) if outputWidth > 1 else 0.0
for h2 in range(outputHeight):
h1r = 1.0 * h2 * rheight
h1 = int(np.floor(h1r))
h1lambda = h1r - h1
h1p = 1 if h1 < (inputHeight - 1) else 0
for w2 in range(outputWidth):
w1r = 1.0 * w2 * rwidth
w1 = int(np.floor(w1r))
w1lambda = w1r - w1
w1p = 1 if w1 < (inputHeight - 1) else 0
for b in range(batch):
for c in range(channel):
y[b][h2][w2][c] = (1-h1lambda)*((1-w1lambda)*x[b][h1][w1][c] + \
w1lambda*x[b][h1][w1+w1p][c]) + \
h1lambda*((1-w1lambda)*x[b][h1+h1p][w1][c] + \
w1lambda*x[b][h1+h1p][w1+w1p][c])
return y
# Test with normal case 3D input int8 type
data_in_4d = nd.random.uniform(0, 255, (1, 300, 300, 3)).astype('uint8')
out_nd_3d = transforms.Resize((100, 100))(data_in_4d[0])
assert_almost_equal(out_nd_3d.asnumpy(), py_bilinear_resize_nhwc(data_in_4d.asnumpy(), 100, 100)[0], atol=1.0)
# Test with normal case 4D input int8 type
data_in_4d = nd.random.uniform(0, 255, (2, 300, 300, 3)).astype('uint8')
out_nd_4d = transforms.Resize((100, 100))(data_in_4d)
assert_almost_equal(out_nd_4d.asnumpy(), py_bilinear_resize_nhwc(data_in_4d.asnumpy(), 100, 100), atol=1.0)
def data_xform(data):
"""Move channel axis to the beginning, cast to float32, and normalize to [0, 1]."""
return ndarray.moveaxis(data, 2, 0).astype('float32') / 255