Curiosity/DPI/models.py

import numpy as np

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.distributions.normal import Normal


class ObservationEncoder(nn.Module):
    def __init__(self, obs_shape, state_size, num_layers=4, num_filters=32, stride=None):
        super().__init__()  

        assert len(obs_shape) == 3

        self.state_size = state_size

        layers = []
        for i in range(num_layers):
            input_channels = obs_shape[0] if i == 0 else output_channels
            output_channels = num_filters * (2 ** i)
            layers.append(nn.Conv2d(in_channels=input_channels, out_channels= output_channels, kernel_size=4, stride=2))
            layers.append(nn.ReLU())

        self.convs = nn.Sequential(*layers)

        self.fc = nn.Linear(256 * 3 * 3, 2 * state_size)

    def forward(self, x):
        x = self.convs(x)
        x = x.view(x.size(0), -1)
        x = self.fc(x)
        
        # Mean and standard deviation
        mean, std = torch.chunk(x, 2, dim=-1)
        std = F.softplus(std)
        std = torch.clamp(std, min=0.0, max=1e5)

        # Normal Distribution
        dist = self.get_dist(mean, std)

        # Sampling via reparameterization Trick
        x = self.reparameterize(mean, std)

        encoded_output = {"sample": x, "distribution": dist}
        return encoded_output
    
    def reparameterize(self, mu, std):
        eps = torch.randn_like(std)
        return mu + eps * std
    
    def get_dist(self, mean, std):
        distribution = torch.distributions.Normal(mean, std)
        distribution = torch.distributions.independent.Independent(distribution, 1)
        return distribution


class ObservationDecoder(nn.Module):
    def __init__(self, state_size, output_shape):
        super().__init__()

        self.state_size = state_size
        self.output_shape = output_shape
        self.input_size = 256 * 3 * 3
        self.in_channels = [self.input_size, 256, 128, 64]
        self.out_channels = [256, 128, 64, 3]

        if output_shape[1] == 84:
            self.kernels = [5, 7, 5, 6]
            self.output_padding = [1, 1, 1, 0]
        elif output_shape[1] == 64:
            self.kernels = [5, 5, 6, 6]
            self.output_padding = [0, 0, 0, 0]

        self.dense = nn.Linear(state_size, self.input_size)
        
        layers = []
        for i in range(len(self.kernels)):
            layers.append(nn.ConvTranspose2d(in_channels=self.in_channels[i], out_channels=self.out_channels[i], 
                                             kernel_size=self.kernels[i], stride=2, output_padding=self.output_padding[i]))
            if i!=len(self.kernels)-1:
                layers.append(nn.ReLU())

        self.convtranspose = nn.Sequential(*layers)

    def forward(self, features):
        out_batch_shape = features.shape[:-1]
        out = self.dense(features)
        out = torch.reshape(out, [-1, self.input_size, 1, 1])
        out = self.convtranspose(out)
        
        mean = torch.reshape(out, (*out_batch_shape, *self.output_shape))
        out_dist = torch.distributions.independent.Independent(torch.distributions.Normal(mean, 1), len(self.output_shape))
        return out_dist


class ActionDecoder(nn.Module):
    def __init__(self, state_size, hidden_size, action_size, num_layers=5):
        super().__init__()
        self.state_size = state_size    
        self.hidden_size = hidden_size
        self.action_size = action_size
        self.num_layers = num_layers

        self._min_std=torch.Tensor([1e-4])[0]
        self._init_std=torch.Tensor([5])[0]
        self._mean_scale=torch.Tensor([5])[0]

        layers = []
        for i in range(self.num_layers):
            input_channels = state_size if i == 0 else self.hidden_size
            output_channels = self.hidden_size if i!= self.num_layers-1 else 2*action_size
            layers.append(nn.Linear(input_channels, output_channels))
            layers.append(nn.ReLU())
        self.action_model = nn.Sequential(*layers)

    def get_dist(self, mean, std):
        distribution = torch.distributions.Normal(mean, std)
        distribution = torch.distributions.transformed_distribution.TransformedDistribution(distribution, TanhBijector())
        distribution = torch.distributions.independent.Independent(distribution, 1)
        return distribution
    
    def forward(self, features):
        out = self.action_model(features)
        mean, std = torch.chunk(out, 2, dim=-1) 
        
        raw_init_std = torch.log(torch.exp(self._init_std) - 1)
        action_mean = self._mean_scale * torch.tanh(mean / self._mean_scale)
        action_std = F.softplus(std + raw_init_std) + self._min_std

        dist = self.get_dist(action_mean, action_std)
        sample = dist.rsample()
        return sample
        

class ValueModel(nn.Module):
    def __init__(self, state_size, hidden_size, num_layers=4):
        super().__init__()
        self.state_size = state_size
        self.hidden_size = hidden_size
        self.num_layers = num_layers

        layers = []
        for i in range(self.num_layers):
            input_channels = state_size if i == 0 else self.hidden_size
            output_channels = self.hidden_size if i!= self.num_layers-1 else 1
            layers.append(nn.Linear(input_channels, output_channels))
            layers.append(nn.ReLU())
        self.value_model = nn.Sequential(*layers) 

    def forward(self, state):
        value = self.value_model(state)
        value_dist = torch.distributions.independent.Independent(torch.distributions.Normal(value, 1), 1)
        return value_dist
    
        
class TransitionModel(nn.Module):
    def __init__(self, state_size, hidden_size, action_size, history_size):
        super().__init__()

        self.state_size = state_size
        self.hidden_size = hidden_size
        self.action_size = action_size
        self.history_size = history_size
        self.act_fn = nn.ReLU()

        self.fc_state_action = nn.Linear(state_size + action_size, hidden_size)
        self.history_cell = nn.GRUCell(hidden_size + history_size, history_size)
        self.fc_state_prior  = nn.Linear(history_size + state_size + action_size, 2 * state_size)
        self.fc_state_posterior = nn.Linear(history_size + state_size + action_size, 2 * state_size)

    def init_states(self, batch_size, device):
        self.prev_state = torch.zeros(batch_size, self.state_size).to(device)
        self.prev_action = torch.zeros(batch_size, self.action_size).to(device)
        self.prev_history = torch.zeros(batch_size, self.history_size).to(device)
            
    def get_dist(self, mean, std):
        distribution = torch.distributions.Normal(mean, std)
        distribution = torch.distributions.independent.Independent(distribution, 1)
        return distribution

    def imagine_step(self, prev_state, prev_action, prev_history):
        state_action = self.act_fn(self.fc_state_action(torch.cat([prev_state, prev_action], dim=-1)))
        history = self.history_cell(torch.cat([state_action, prev_history], dim=-1), prev_history)

        state_prior = self.fc_state_prior(torch.cat([history, prev_state, prev_action], dim=-1))
        state_prior_mean, state_prior_std = torch.chunk(state_prior, 2, dim=-1)
        state_prior_std = F.softplus(state_prior_std)

        # Normal Distribution
        state_prior_dist = self.get_dist(state_prior_mean, state_prior_std)

        # Sampling via reparameterization Trick
        sample_state_prior = self.reparemeterize(state_prior_mean, state_prior_std)
        prior = {"mean": state_prior_mean, "std": state_prior_std, "sample": sample_state_prior, "history": history, "distribution": state_prior_dist}
        return prior
    
    def stack_states(self, states, dim=0):
        
        s = dict(
            mean = torch.stack([state['mean'] for state in states], dim=dim),
            std  = torch.stack([state['std'] for state in states], dim=dim),
            sample = torch.stack([state['sample'] for state in states], dim=dim),
            history = torch.stack([state['history'] for state in states], dim=dim),)
        dist = dict(distribution = [state['distribution'] for state in states])
        s.update(dist)
        return s
    
    def imagine_rollout(self, state, action, history, horizon):
        imagined_priors = []
        for i in range(horizon):
            prior = self.imagine_step(state, action, history)
            state = prior["sample"]
            history = prior["history"]
            imagined_priors.append(prior)
        imagined_priors = self.stack_states(imagined_priors, dim=0)
        return imagined_priors
    
    def reparemeterize(self, mean, std):
        eps = torch.randn_like(std)
        return mean + eps * std


class TanhBijector(torch.distributions.Transform):
    def __init__(self):
        super().__init__()
        self.bijective = True
        self.domain = torch.distributions.constraints.real
        self.codomain = torch.distributions.constraints.interval(-1.0, 1.0)

    @property
    def sign(self): return 1.

    def _call(self, x): return torch.tanh(x)

    def atanh(self, x):
        return 0.5 * torch.log((1 + x) / (1 - x))

    def _inverse(self, y: torch.Tensor):
        y = torch.where(
            (torch.abs(y) <= 1.),
            torch.clamp(y, -0.99999997, 0.99999997),
            y)
        y = self.atanh(y)
        return y

    def log_abs_det_jacobian(self, x, y):
        #return 2. * (np.log(2) - x - F.softplus(-2. * x))
        return 2.0 * (torch.log(torch.tensor([2.0])) - x - F.softplus(-2.0 * x))
     

class ProjectionHead(nn.Module):
    def __init__(self, state_size, action_size, hidden_size):
        super(ProjectionHead, self).__init__()
        self.state_size = state_size
        self.action_size = action_size
        self.hidden_size = hidden_size

        self.projection_model = nn.Sequential(
            nn.Linear(state_size + action_size, hidden_size),
            nn.LayerNorm(hidden_size),
            nn.ReLU(),
            nn.Linear(hidden_size, hidden_size),
            nn.LayerNorm(hidden_size),
            )
    
    def forward(self, state, action):
        x = torch.cat([state, action], dim=-1)
        x = self.projection_model(x)
        return x
    

class ContrastiveHead(nn.Module):
    def __init__(self, hidden_size, temperature=1):
        super(ContrastiveHead, self).__init__()
        self.hidden_size = hidden_size
        self.temperature = temperature
        self.W = nn.Parameter(torch.rand(self.hidden_size, self.hidden_size))

    def forward(self, z_a, z_pos):
        Wz = torch.matmul(self.W, z_pos.T)  # (z_dim,B)
        logits = torch.matmul(z_a, Wz)  # (B,B)
        logits = logits - torch.max(logits, 1)[0][:, None]
        logits = logits * self.temperature
        return logits
    

class CLUBSample(nn.Module):  # Sampled version of the CLUB estimator
    def __init__(self, last_states, current_states, negative_current_states, predicted_current_states):
        super(CLUBSample, self).__init__()
        self.last_states = last_states
        self.current_states = current_states
        self.negative_current_states = negative_current_states
        self.predicted_current_states = predicted_current_states

    def get_mu_var_samples(self, state_dict):
        dist = state_dict["distribution"]
        sample = dist.sample()  # Use state_dict["sample"] if you want to use the same sample for all the losses
        mu = dist.mean
        var = dist.variance
        return mu, var, sample
        
    def loglikeli(self):
        _, _, pred_sample = self.get_mu_var_samples(self.predicted_current_states)
        mu_curr, var_curr, _ = self.get_mu_var_samples(self.current_states)
        logvar_curr = torch.log(var_curr)    
        return (-(mu_curr - pred_sample)**2 /var_curr-logvar_curr).sum(dim=1).mean(dim=0)

    def forward(self):
        _, _, pred_sample = self.get_mu_var_samples(self.predicted_current_states)
        mu_curr, var_curr, _ = self.get_mu_var_samples(self.current_states)
        mu_neg, var_neg, _ = self.get_mu_var_samples(self.negative_current_states)
        
        pos = (-(mu_curr - pred_sample)**2 /var_curr).sum(dim=1).mean(dim=0)
        neg = (-(mu_neg - pred_sample)**2 /var_neg).sum(dim=1).mean(dim=0)
        upper_bound = pos - neg

        return upper_bound/2

    def learning_loss(self):
        return - self.loglikeli()
    

if "__name__ == __main__":
    pass