def test_conv2d():
"""
Test the Conv2d module to ensure that it produces the exact same
output as the primary torch implementation, in the same order.
"""
model2 = nn2.Conv2d(1, 32, 3, bias=True)
model = nn.Conv2d(
in_channels=1,
out_channels=32,
kernel_size=3,
stride=1,
padding=0,
dilation=1,
groups=1,
bias=True,
padding_mode="zeros",
)
model2.weight = model.weight
model2.bias = model.bias
data = th.rand(10, 1, 28, 28)
out = model(data)
out2 = model2(data)
assert th.isclose(out, out2, atol=1e-6).all()
def test_conv2d(workers):
"""
Test the nn.Conv2d module to ensure that it produces the exact same
output as the primary torch implementation, in the same order.
"""
torch.manual_seed(
121) # Truncation might not always work so we set the random seed
# Disable mkldnn to avoid rounding errors due to difference in implementation
mkldnn_enabled_init = torch._C._get_mkldnn_enabled()
torch._C._set_mkldnn_enabled(False)
# Direct Import from Syft
model = syft_nn.Conv2d(1, 2, 3, bias=True)
model_1 = nn.Conv2d(1, 2, 3, bias=True)
model.weight = model_1.weight.fix_prec()
model.bias = model_1.bias.fix_prec()
data = torch.rand(10, 1, 28, 28) # eg. mnist data
out = model(data.fix_prec()).float_prec()
out_1 = model_1(data)
assert torch.allclose(out, out_1, atol=1e-2)
# Fixed Precision Tensor
model_2 = model_1.copy().fix_prec()
out_2 = model_2(data.fix_prec()).float_prec()
# Note: absolute tolerance can be reduced by increasing precision_fractional of fix_prec()
assert torch.allclose(out_1, out_2, atol=1e-2)
# Additive Shared Tensor
bob, alice, james = (workers["bob"], workers["alice"], workers["james"])
shared_data = data.fix_prec().share(bob, alice, crypto_provider=james)
mode_3 = model_2.share(bob, alice, crypto_provider=james)
out_3 = mode_3(shared_data).get().float_prec()
assert torch.allclose(out_1, out_3, atol=1e-2)
# Reset mkldnn to the original state
torch._C._set_mkldnn_enabled(mkldnn_enabled_init)
def test_conv2d():
"""
Test the Conv2d module to ensure that it produces the exact same
output as the primary torch implementation, in the same order.
"""
# Disable mkldnn to avoid rounding errors due to difference in implementation
mkldnn_enabled_init = th._C._get_mkldnn_enabled()
th._C._set_mkldnn_enabled(False)
model2 = nn2.Conv2d(1, 32, 3, bias=True)
model = nn.Conv2d(
in_channels=1,
out_channels=32,
kernel_size=3,
stride=1,
padding=0,
dilation=1,
groups=1,
bias=True,
padding_mode="zeros",
)
model2.weight = model.weight
model2.bias = model.bias
data = th.rand(10, 1, 28, 28)
out = model(data)
out2 = model2(data)
# Reset mkldnn to the original state
th._C._set_mkldnn_enabled(mkldnn_enabled_init)
assert th.eq(out, out2).all()