models

import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F

def normalized_columns_initializer(weights, std=1.0):
    out = torch.randn(weights.size())
    out *= std / torch.sqrt(out.pow(2).sum(1, keepdim=True).expand_as(out))

    return out

def weights_init(m):
    classname =m.__class__.__name__

    if classname.find('Conv') != -1:
        weight_shape = list(m.weight.data.size())
        fan_in = np.prod(weight_shape[1:4])
        fan_out = np.prod(weight_shape[2:4]) * weight_shape[0]
        w_bound = np.sqrt(6. / (fan_in  + fan_out))
        m.weight.data.uniform_(-w_bound, w_bound)
        m.bias.data.fill_(0)
    elif classname.find('Linear') != -1:
        weight_shape = list(m.weight.data.size())
        fan_in = weight_shape[1]
        fan_out = weight_shape[0]
        w_bound = np.sqrt(6. / (fan_in + fan_out))
        m.weight.data.uniform_(-w_bound, w_bound)
        m.bias.data.fill_(0)


class LSTM_GA(nn.Module):
    def __init__(self, max_n_steps, vocab_size):
        super(LSTM_GA, self).__init__()

        # Image Processing
        self.conv1 = nn.Conv2d(3, 128, kernel_size=8, stride=4)
        self.conv2 = nn.Conv2d(128, 64, kernel_size=4, stride=2)
        self.conv3 = nn.Conv2d(64, 64, kernel_size=4, stride=2)

        # Instruction Processing
        self.gru_hidden_size = 256
        self.input_size = vocab_size
        self.embedding = nn.Embedding(self.input_size, 32)
        self.gru = nn.GRU(32, self.gru_hidden_size, batch_first=True)

        # Gated-Attention Layers
        self.attn_linear = nn.Linear(self.gru_hidden_size, 64)

        # Time-Embedding Layer, helps in stabilizing value prediction
        self.time_emb_dim = 32
        self.time_emb_layer = nn.Embedding(
            max_n_steps+1,
            self.time_emb_dim
        )

        # A2C-LSTM layers
        self.linear = nn.Linear(64*8*17, 256)
        self.lstm = nn.LSTMCell(256, 256)
        self.critic_linear = nn.Linear(256 + self.time_emb_dim, 1)
        self.actor_linear = nn.Linear(256 + self.time_emb_dim, 3)


        # Initializing weights
        self.apply(weights_init)
        self.actor_linear.weight.data = normalized_columns_initializer(
            self.actor_linear.weight.data, 0.01
        )
        self.actor_linear.bias.data.fill_(0)
        self.critic_linear.weight.data = normalized_columns_initializer(
            self.critic_linear.weight.data, 1.0
        )
        self.critic_linear.bias.data.fill_(0)

        self.lstm.bias_ih.data.fill_(0)
        self.lstm.bias_hh.data.fill_(0)
        # self.train()


    def _format(self, inputs):
        pass

    def forward(self, inputs):
        x, input_inst, (tx, hx, cx) = inputs
        # print(x.size(),)
        n_workers = x.size(0)

        # Get the image representation
        x = F.relu(self.conv1(x), inplace=False)
        x = F.relu(self.conv2(x), inplace=False)
        x_image_rep = F.relu(self.conv3(x), inplace=False)

        # Get the instruction representation
        # encoder_hidden = torch.zeros(1, 1, self.gru_hidden_size)
        ## Check here, we will be sending the entire embedded instruction matrix
        ## and not create it in a loop, we need to send padded instructions
        encoder_hidden = torch.zeros(1, n_workers, self.gru_hidden_size)
        embedded_instruction = self.embedding(input_inst)
        _, encoder_hidden = self.gru(embedded_instruction, encoder_hidden)

        # x_instr_rep = encoder_hidden.view(encoder_hidden.size(1), -1)
        x_instr_rep = encoder_hidden

        # Get the attention vector from the instruction representation
        x_attention = F.sigmoid(self.attn_linear(x_instr_rep))

        # Gated-Attention
        ## Need to change this as the number of instructions or environment has changed
        ## increased to n_workers
        x_attention = x_attention.squeeze(0).unsqueeze(2).unsqueeze(3)
        x_attention = x_attention.expand(n_workers, 64, 8, 17)

        assert x_image_rep.size() == x_attention.size()

        x = x_image_rep * x_attention
        x = x.view(x.size(0), -1)

        # A3C-LSTM
        x = F.relu(self.linear(x), inplace=False)
        new_hx, new_cx = self.lstm(x, (hx, cx))
        time_emb = self.time_emb_layer(tx)
        x = torch.cat((new_hx, time_emb.view(-1, self.time_emb_dim)), 1)

        return  self.actor_linear(x), self.critic_linear(x), (new_hx, new_cx)

    def full_pass(self, inputs):
        '''
        inputs: combination of images_batch, instruction_batch, hx, cx, tx;
            every input is torch tensor of self.device type
        outputs:
            action: either scalar or numpy array depending on number of n_workers (Size: n_workers)
            is_exploratory: boolean array (Size: n_workers)
            logpa: torch tensor of self.device type
            entropy: torch tensor of self.device type
            value: torch tensor of self.device type
            hx: torch tensor of self.device type
            cx: torch tensor of self.device type
        '''

        logits, value, (hx, cx) = self.forward(inputs)
        dist = torch.distributions.Categorical(logits=logits)
        action = dist.sample()
        logpa = dist.log_prob(action).unsqueeze(-1)
        entropy = dist.entropy().unsqueeze(-1)
        action = action.item() if len(action) == 1 else action.detach().cpu().numpy()

        ## Check this once, look at axis
        is_exploratory = action != np.argmax(logits.detach().cpu().numpy(), axis=1)
        return action, is_exploratory, logpa, entropy, value, (hx, cx)

    def select_action(self, inputs):

        ## Check whether (hc, cx) is needed or not; looks like it is needed, but not sure
        ## take the output from forward method
        logits, _, (hx, cx) = self.forward(inputs)

        dist = torch.distributions.Categorical(logits=logits)
        action = dist.sample()
        action = action.item() if len(action) == 1 else action.detach().cpu().numpy()
        return action, (hx, cx)

    def select_greedy_action(self, inputs):
        ## Check whether (hc, cx) is needed or not; looks like it is needed, but not sure
        ## take the output from forward method
        logits, _, (hx, cx) = self.forward(inputs)

        return np.argmax(logits.detach().cpu().numpy()), (hx, cx)

    def evaluate_state(self, inputs):
        _, value, (hx, cx) = self.forward(inputs)
        return value, (hx, cx)