def forward(self, x):
return torch.mkldnn_convolution(
x,
self.weight,
self.bias,
self.padding,
self.stride,
self.dilation,
self.groups)
def test_conv2d_ext(self, device, dtype):
# this list for save the cases which dh,dw == 0
# and raise error in the end
self.Fail = list()
total_cases = self._collect_cases()
for case in total_cases:
case_name = case['case_name']
bs = case['mb']
group = case['g']
ic, ih, iw = case['ic'], case['ih'], case['iw']
oc = case['oc']
kh, kw = case['kh'], case['kw']
sh, sw = case['sh'], case['sw']
ph, pw = case['ph'], case['pw']
dh, dw = case['dh'], case['dw']
has_bias = case['bias']
if dh == 0 or dw == 0:
self.Fail.append(case_name)
continue
ic_g = ic // group
torch.manual_seed(1)
input = torch.randn((bs, ic, ih, iw),
device=device,
dtype=dtype,
requires_grad=True)
weight = torch.randn(
(oc, ic_g, kh, kw), device=device, dtype=dtype) * 0.01
bias = None if has_bias == 'False' else torch.randn(
(oc), device=device, dtype=dtype)
k = [kh, kw]
s = [sh, sw]
p = [ph, pw]
d = [dh, dw]
thnn_output = self._thnn_conv_group(input, weight, k, bias, s, p,
d, group)
if self.device_type == 'cpu' and torch.backends.mkldnn.is_available(
):
output = torch.mkldnn_convolution(input, weight, bias, p, s, d,
group)
elif self.device_type == 'cuda' and torch.backends.cudnn.is_available(
):
output = torch.cudnn_convolution(input, weight, bias, p, s, d,
group, True, True)
else:
output = torch.conv2d(input, weight, bias, s, p, d, group)
msg = 'device:{}, dtype:{}, group:{}, batchsize:{}' \
'input channel:{}, output channel:{}, ' \
'bias:{}, padding:{}, dilation:{}, stride:{}, ' \
'kernel:{}'
msg = msg.format(device, dtype, group, bs, ic, oc, has_bias, p, d,
s, k)
if self.device_type == 'cuda' and cudnn.is_available():
self.assertEqual(output,
thnn_output,
msg=msg,
atol=1e-2,
rtol=1e-2)
else:
self.assertEqual(output, thnn_output, msg=msg)
if self.Fail != []:
warnings.warn('invalid cases dilation height or weight is 0: ' +
",".join(self.Fail))