def test_pooling():
A = np.ones((1, 2, INT_OVERFLOW))
A[0][0][2] = 100
A.attach_grad()
with mx.autograd.record():
B = npx.pooling(data=A, kernel=(2), stride=2, pool_type='max')
assert B.shape == (1, 2, HALF_INT_OVERFLOW)
assert B[0][0][1] == 100
B.backward()
assert A.grad.shape == (1, 2, INT_OVERFLOW)
assert A.grad[0][0][0] == 1
def test_pooling():
def test_pooling_large_dim():
A = np.ones((1, 1, INT_OVERFLOW))
assertRaises(MXNetError, npx.pooling, data=A, kernel=(2), stride=(2), \
pool_type='max')
test_pooling_large_dim()
D, H, W = 2**12, 2**10, 2**10
A = np.ones((1, 1, D, H, W))
A[0, 0, 0, 0, 2] = 100
A.attach_grad()
with mx.autograd.record():
B = npx.pooling(data=A, kernel=(2, 2, 2), stride=(2, 2, 2), \
pool_type='max')
assert B.shape == (1, 1, int(D / 2), int(H / 2), int(W / 2))
assert B[0, 0, 0, 0, 1] == 100
B.backward()
assert A.grad.shape == (1, 1, D, H, W)
assert A.grad[0, 0, 0, 0, 0] == 1
def forward(self, input):
_, _, h, w = input.shape
p = [input]
for i in range(self.depth):
hnew = h // (2**i)
wnew = w // (2**i)
kernel = (hnew, wnew)
x = FFx.pooling(input,
kernel=kernel,
stride=kernel,
pool_type='max')
#x = mx.nd.UpSampling(x.as_nd_ndarray(),sample_type='nearest',scale=hnew)
x = mx.contrib.ndarray.BilinearResize2D(x.as_nd_ndarray(),
height=h,
width=w)
x = x.as_np_ndarray()
x = self.convs[i](x)
p += [x]
out = FF.concatenate(p, axis=1)
out = self.conv_norm_final(out)
return out