def test_to_leela_weights(weight_file):
n = Network(19, 18, 16, 3)
_, tmp = tempfile.mkstemp()
n.to_leela_weights(tmp)
with open(tmp, 'r') as tmpfile:
tmplines = tmpfile.readlines()
weightlines = weight_file.readlines()
assert len(tmplines) == len(weightlines)
for l1, l2 in zip(tmplines, weightlines):
t1 = torch.tensor([float(n) for n in l1.split()])
t2 = torch.tensor([float(n) for n in l2.split()])
assert t1.size() == t2.size()
os.remove(tmp)
def test_network(board_size, in_channels, residual_channels, residual_layers,
input_size, pol_output_size, val_output_size):
n = Network(board_size, in_channels, residual_channels, residual_layers)
(pol, val), (target_pol,
target_val) = n(torch.randn(*input_size),
torch.randint(19 * 19 + 1, (input_size[0], )),
torch.ones((input_size[0], )))
assert pol.size() == pol_output_size
assert val.size() == val_output_size
assert target_pol.size() == (input_size[0], )
assert target_val.size() == (input_size[0], )
def test_network(
board_size,
in_channels,
residual_channels,
residual_layers,
input_size,
pol_output_size,
val_output_size,
):
n = Network(board_size, in_channels, residual_channels, residual_layers)
pol, val = n(torch.randn(*input_size))
assert pol.size() == pol_output_size
assert val.size() == val_output_size