dpmodel/pt: Random methods without explicit seeds should be forbidden

[njzjz]

For new backends, random methods without explicit seeds should be forbidden, as they create unreproduced behavior.

dpmodel

deepmd-kit/deepmd/dpmodel/utils/network.py

Lines 173 to 174 in d61b152

	rng = np.random.default_rng()
	self.w = rng.normal(size=(num_in, num_out)).astype(prec)

pt

deepmd-kit/deepmd/pt/model/network/network.py

Lines 72 to 73 in d61b152

	random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
	random_tensor.floor_() # binarize

deepmd-kit/deepmd/pt/optimizer/KFWrapper.py

Line 138 in d61b152

rng = np.random.default_rng()

The following normal methods need to be provided with a torch Generator:

deepmd-kit/deepmd/pt/model/network/mlp.py

Line 68 in d61b152

nn.init.normal_(self.matrix.data, std=stddev / np.sqrt(num_out + num_in))

deepmd-kit/deepmd/pt/model/network/mlp.py

Line 73 in d61b152

nn.init.normal_(self.bias.data, mean=bavg, std=stddev)

deepmd-kit/deepmd/pt/model/network/mlp.py

Line 78 in d61b152

nn.init.normal_(self.idt.data, mean=0.1, std=0.001)

deepmd-kit/deepmd/pt/model/network/network.py

Line 141 in d61b152

nn.init.normal_(self.matrix.data, std=stddev / np.sqrt(num_out + num_in))

deepmd-kit/deepmd/pt/model/network/network.py

Line 143 in d61b152

nn.init.normal_(self.bias.data, mean=bavg, std=stddev)

deepmd-kit/deepmd/pt/model/network/network.py

Line 146 in d61b152

nn.init.normal_(self.idt.data, mean=1.0, std=0.001)

deepmd-kit/deepmd/pt/model/network/network.py

Line 307 in d61b152

nn.init.normal_(self.matrix.data, std=stddev / np.sqrt(num_out + num_in))

deepmd-kit/deepmd/pt/model/network/network.py

Line 310 in d61b152

nn.init.normal_(self.bias.data, mean=bavg, std=stddev)

deepmd-kit/deepmd/pt/model/network/network.py

Line 315 in d61b152

nn.init.normal_(self.idt.data, mean=0.1, std=0.001)

deepmd-kit/deepmd/pt/model/network/network.py

Line 370 in d61b152

nn.init.trunc_normal_(self.weight, mean=0.0, std=std)

deepmd-kit/deepmd/pt/model/network/network.py

Line 383 in d61b152

nn.init.kaiming_normal_(self.weight, nonlinearity="linear")

deepmd-kit/deepmd/pt/model/network/network.py

Line 430 in d61b152

nn.init.normal_(self.weight, mean=0.0, std=std)

deepmd-kit/deepmd/pt/model/task/atten_lcc.py

Line 25 in d61b152

nn.init.normal_(self.energe_agg_factor.weight, 0, 0.01)

I also noticed in some modules, torch.manual_seed is called to change the global seed after the network initlization. I don't see the meaning of them.

[wanghan-iapcm]

i do no think it is a issue at least in the UTs, as they do not rely on the reproducibility of the random numbers.