model.py

import math

import torch
import torch.nn as nn
import torch.nn.functional as F


class A3Clstm(torch.nn.Module):
    def __init__(self, num_inputs, action_space):
        super(A3Clstm, self).__init__()
        self.conv1 = nn.Conv2d(num_inputs, 32, 5, stride=1, padding=2)
        self.maxp1 = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(32, 32, 5, stride=1, padding=1)
        self.maxp2 = nn.MaxPool2d(2, 2)
        self.conv3 = nn.Conv2d(32, 64, 4, stride=1, padding=1)
        self.maxp3 = nn.MaxPool2d(2, 2)
        self.conv4 = nn.Conv2d(64, 64, 3, stride=1, padding=1)
        self.maxp4 = nn.MaxPool2d(2, 2)

        self.lstm = nn.LSTMCell(1024, 512)
        num_outputs = action_space.n
        self.critic_linear = nn.Linear(512, 1)
        self.adv_linear = nn.Linear(512, num_outputs)
        self.actor_linear = nn.Linear(512, num_outputs)

        self.register_parameter(
            'log_alpha',
            nn.Parameter(torch.full((1, ), math.log(.01),
                                    dtype=torch.float32)))
        with torch.no_grad():
            x = torch.full((num_outputs, ),
                           0.1 / num_outputs,
                           dtype=torch.float32)
            x[0] += 0.9
            self.constraint = torch.distributions.categorical.Categorical(
                probs=x).entropy().item()
        self.thres = torch.log(torch.tensor(1e-3, dtype=torch.float32)).item()

    @torch.no_grad()
    def keep_alpha(self):
        self.log_alpha[0] = max(self.thres, self.log_alpha.item())

    def forward(self, inputs):
        inputs, (hx, cx) = inputs

        x = F.relu(self.maxp1(self.conv1(inputs)))
        x = F.relu(self.maxp2(self.conv2(x)))
        x = F.relu(self.maxp3(self.conv3(x)))
        x = F.relu(self.maxp4(self.conv4(x)))
        x = x.view(x.size(0), -1)

        hx, cx = self.lstm(x, (hx, cx))
        x = hx

        logit = self.actor_linear(x)
        dist = torch.distributions.categorical.Categorical(
                logits=logit)
        adv_value = self.adv_linear(x)
        #adv_value = adv_value - (F.softmax(logit, -1).detach() * adv_value).sum(-1, True) - dist.entropy().unsqueeze(-1).detach() * .01
        #adv_value = adv_value - adv_value.mean(-1, True)
        adv_value = adv_value - (F.softmax(logit, -1) * adv_value).sum(-1, True) - dist.entropy().unsqueeze(-1) * .01
        v_value = self.critic_linear(x)
        q_value = adv_value + v_value
        act = dist.sample().unsqueeze(-1)
        q_value_full = q_value
        q_value = q_value.gather(-1, act)
        #adv_value = adv_value.gather(-1, act)
        return act, None, logit, (adv_value, q_value), (hx, cx)