手动指定pytorch模型权重值

def adjust_net(order):
    net = torch.nn.Sequential(
        torch.nn.Linear(371, 512),
        torch.nn.LeakyReLU(0.01),
        torch.nn.Linear(512, 512),
        torch.nn.LeakyReLU(0.01),
        torch.nn.Linear(512, 512),
        torch.nn.LeakyReLU(0.01),
        torch.nn.Linear(512, 512),
        torch.nn.LeakyReLU(0.01),
        torch.nn.Linear(512, 512),
        torch.nn.LeakyReLU(0.01),
        torch.nn.Linear(512, 512),
        torch.nn.LeakyReLU(0.01),
        torch.nn.Linear(512, 512),
        torch.nn.LeakyReLU(0.01),
        torch.nn.Linear(512, 371),
    )

    net.load_state_dict(
        torch.load(f'ugb-embedded-230504/ugb_ckpt/model_bidding_sl/{order:02d}.ckpt', map_location='cpu')
    )

    left = order if order > 0 else 1
    right = 8 / 13

    torch.set_grad_enabled(False)

    for layer in net.modules():
        if isinstance(layer, torch.nn.Linear):
            print(order, layer)
            for weight in layer.parameters():
                for j in range(512):
                    for i in range(371):
                        if i < 319:
                            weight[j, i] /= left
                        else:
                            weight[j, i] /= right
                torch.set_grad_enabled(True)
                torch.save(
                    net.state_dict(), f'ugb-embedded-230504/ugb_ckpt/model_bidding_rl/{order:02d}.ckpt'
                )
                return

for order in range(15):
    adjust_net(order)