zero_shot_world_models/replay_buffer.py at main · fmalato/zero_shot_world_models · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
import torch

from typing import NamedTuple


class Transition(NamedTuple):
    observation: torch.Tensor
    action: torch.Tensor
    next_observation: torch.Tensor
    mean: torch.Tensor
    sigma: torch.Tensor


class ReplayBuffer:
    def __init__(
            self,
            capacity: int,
            obs_size: int,
            num_actions: int,
            is_minerl: bool = False
    ):
        self.observations = torch.zeros(size=(capacity, obs_size))
        self.actions = torch.zeros(size=(capacity, 1 if not is_minerl else 10))
        self.next_observations = torch.zeros(size=(capacity, obs_size))
        self.means = torch.zeros(size=(capacity, obs_size))
        self.sigmas = torch.zeros(size=(capacity, obs_size))

        self._current_idx = 0
        self._is_minerl = is_minerl
        self.capacity = capacity
        self.num_actions = num_actions

    def add_transition(self, o, a, o_next, m, l):
        self.observations[self._current_idx] = o
        self.actions[self._current_idx] = a
        self.next_observations[self._current_idx] = o_next
        self.means[self._current_idx] = m
        self.sigmas[self._current_idx] = torch.sqrt(torch.exp(l))

        self._current_idx = (self._current_idx + 1) % self.capacity

    def sample_similar(self, latent: torch.Tensor, action: torch.Tensor = None, k: int = 1):
        distances = torch.cdist(self.observations[:self._current_idx], latent, p=2)
        if action is not None:
            if self._is_minerl:
                ok_indices = torch.argwhere((self.actions[:self._current_idx].squeeze() == action).long().sum(-1) == self.actions.size(-1)).squeeze()
            else:
                ok_indices = torch.argwhere(self.actions[:self._current_idx].squeeze() == action)
            if ok_indices.numel() >= k:
                distances = distances[ok_indices]
        indices = torch.topk(-distances, k=k, dim=0).indices.squeeze()

        data = (
            self.observations[indices],
            self.actions[indices],
            self.next_observations[indices],
            self.means[indices],
            self.sigmas[indices]
        )

        return Transition(*tuple(map(self.to_torch, data))), indices

    def sample_similar_distribution(self, mean: torch.Tensor, logvar: torch.Tensor, action: torch.Tensor = None):
        if self._current_idx < self.capacity:
            stored_dists = torch.distributions.Normal(self.means[:self._current_idx], self.sigmas[:self._current_idx] + 1e-8)
            batched_dist = torch.distributions.Normal(mean.repeat(self._current_idx, 1), torch.sqrt(torch.exp(logvar)).repeat(self._current_idx, 1) + 1e-8)
            distances = torch.distributions.kl.kl_divergence(stored_dists, batched_dist).sum(-1)
        else:
            stored_dists = torch.distributions.Normal(self.means, self.sigmas + 1e-8)
            batched_dist = torch.distributions.Normal(mean.repeat(self.means.size(0), 1), torch.sqrt(torch.exp(logvar)).repeat(self.sigmas.size(0), 1) + 1e-8)
            distances = torch.distributions.kl.kl_divergence(stored_dists, batched_dist).sum(-1)

        # Action conditioning
        if action is not None:
            if self._is_minerl:
                ok_indices = torch.argwhere((self.actions[:self._current_idx].squeeze() == action).long().sum(-1) == self.actions.size(-1)).squeeze()
            else:
                ok_indices = torch.argwhere(self.actions[:self._current_idx].squeeze() == action)
            if ok_indices.size(0) >= 1:
                distances = distances[ok_indices]
        closest_idx = torch.argmin(distances)

        data = (
            self.observations[closest_idx],
            self.actions[closest_idx],
            self.next_observations[closest_idx],
            self.means[closest_idx],
            self.sigmas[closest_idx]
        )

        return Transition(*tuple(map(self.to_torch, data))), closest_idx, distances[closest_idx]

    def get_transition(self, transition_index: torch.Tensor):
        data = (
            self.observations[transition_index],
            self.actions[transition_index],
            self.next_observations[transition_index],
            self.means[transition_index],
            self.sigmas[transition_index]
        )

        return Transition(*tuple(map(self.to_torch, data)))

    def to_torch(self, data):
        return torch.Tensor(data)

    def plan(self, latent, horizon: int = 20, k: int = 10, deterministic: bool = False):
        distances = torch.cdist(self.observations[:self._current_idx], latent, p=2)
        indices = torch.topk(-distances, k=k, dim=0).indices.squeeze()
        action_probs = []
        for i in range(horizon):
            step_acts = self.actions[(indices + i) % self._current_idx]
            action_probs.append(torch.bincount(step_acts.flatten().long(), minlength=self.num_actions) / k)
        action_probs = torch.stack(action_probs)
        if deterministic:
            return list(torch.argmax(action_probs, dim=1).numpy())
        else:
            act_distribution = torch.distributions.Categorical(probs=action_probs)
            return act_distribution.sample()

    def plan_distributional(self, mean: torch.Tensor, logvar: torch.Tensor, horizon: int = 20):
        if self._current_idx < self.capacity:
            stored_dists = torch.distributions.Normal(self.means[:self._current_idx], self.sigmas[:self._current_idx] + 1e-8)
            batched_dist = torch.distributions.Normal(mean.repeat(self._current_idx, 1), torch.sqrt(torch.exp(logvar)).repeat(self._current_idx, 1) + 1e-8)
            distances = torch.distributions.kl.kl_divergence(stored_dists, batched_dist).sum(-1)
        else:
            stored_dists = torch.distributions.Normal(self.means, self.sigmas + 1e-8)
            batched_dist = torch.distributions.Normal(mean.repeat(self.means.size(0), 1), torch.sqrt(torch.exp(logvar)).repeat(self.sigmas.size(0), 1) + 1e-8)
            distances = torch.distributions.kl.kl_divergence(stored_dists, batched_dist).sum(-1)

        closest_idx = torch.argmin(distances)

        return [int(self.actions[closest_idx + i % self._current_idx].item()) for i in range(horizon)]