decision_transformer.py

# Decision Transformer module adopted from official code of Decision Transformer
# https://github.com/kzl/decision-transformer
# @article{chen2021decisiontransformer,
#   title={Decision Transformer: Reinforcement Learning via Sequence Modeling},
#   author={Lili Chen and Kevin Lu and Aravind Rajeswaran and Kimin Lee and Aditya Grover and Michael Laskin and Pieter Abbeel and Aravind Srinivas and Igor Mordatch},
#   journal={arXiv preprint arXiv:2106.01345},
#   year={2021}
# }
# MIT License
#
# Copyright (c) 2021 Decision Transformer (Decision Transformer: Reinforcement Learning via Sequence Modeling) Authors (https://arxiv.org/abs/2106.01345)
#
# Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.


import numpy as np
import torch
import torch.nn as nn

import transformers

from modules.model import TrajectoryModel
from modules.trajectory_gpt2 import GPT2Model


class DecisionTransformer(TrajectoryModel):

    """
    This model uses GPT to model (Return_1, state_1, action_1, Return_2, state_2, ...)
    """

    def __init__(
            self,
            state_dim,
            act_dim,
            hidden_size,
            max_length=None,
            max_ep_len=4096,
            action_tanh=True,
            state_encoder=None,
            in_shape=None,
            **kwargs
    ):
        super().__init__(state_dim, act_dim, max_length=max_length)

        self.hidden_size = hidden_size
        config = transformers.GPT2Config(
            vocab_size=1,  # doesn't matter -- we don't use the vocab
            n_embd=hidden_size,
            **kwargs
        )

        # note: the only difference between this GPT2Model and the default Huggingface version
        # is that the positional embeddings are removed (since we'll add those ourselves)
        self.transformer = GPT2Model(config)

        self.embed_timestep = nn.Embedding(max_ep_len, hidden_size)
        self.embed_return = torch.nn.Linear(1, hidden_size)
        if state_encoder is None:
            self.embed_state = torch.nn.Linear(self.state_dim, hidden_size)
        else:
            if state_encoder == 'mlp':
                self.embed_state = MLP_encoder(state_dim, hidden_size, hidden_size)
            elif state_encoder == 'cnn':
                self.embed_state = CNN_encoder(in_shape, hidden_size, dropout=0.1)
            else:
                raise NotImplementedError("encoder not implemented")
        self.embed_action = torch.nn.Linear(self.act_dim, hidden_size)

        self.embed_ln = nn.LayerNorm(hidden_size)

        # note: we don't predict states or returns for the paper
        # self.predict_state = torch.nn.Linear(hidden_size, self.state_dim)
        self.predict_action = nn.Sequential(
            *([nn.Linear(hidden_size, self.act_dim)] + ([nn.Tanh()] if action_tanh else []))
        )
        # self.predict_return = torch.nn.Linear(hidden_size, 1)

    def forward(self, states, actions, rewards, returns_to_go, timesteps, attention_mask=None):

        batch_size, seq_length = states.shape[0], states.shape[1]

        if attention_mask is None:
            # attention mask for GPT: 1 if can be attended to, 0 if not
            attention_mask = torch.ones((batch_size, seq_length), dtype=torch.long)

        # embed each modality with a different head
        state_embeddings = self.embed_state(states)
        action_embeddings = self.embed_action(actions)
        returns_embeddings = self.embed_return(returns_to_go)
        time_embeddings = self.embed_timestep(timesteps)

        # time embeddings are treated similar to positional embeddings
        state_embeddings = state_embeddings + time_embeddings
        action_embeddings = action_embeddings + time_embeddings
        returns_embeddings = returns_embeddings + time_embeddings

        # this makes the sequence look like (R_1, s_1, a_1, R_2, s_2, a_2, ...)
        # which works nice in an autoregressive sense since states predict actions
        stacked_inputs = torch.stack(
            (returns_embeddings, state_embeddings, action_embeddings), dim=1
        ).permute(0, 2, 1, 3).reshape(batch_size, 3*seq_length, self.hidden_size)
        stacked_inputs = self.embed_ln(stacked_inputs)

        # to make the attention mask fit the stacked inputs, have to stack it as well
        stacked_attention_mask = torch.stack(
            (attention_mask, attention_mask, attention_mask), dim=1
        ).permute(0, 2, 1).reshape(batch_size, 3*seq_length)

        # we feed in the input embeddings (not word indices as in NLP) to the model
        transformer_outputs = self.transformer(
            inputs_embeds=stacked_inputs,
            attention_mask=stacked_attention_mask,
        )
        x = transformer_outputs['last_hidden_state']

        # reshape x so that the second dimension corresponds to the original
        # returns (0), states (1), or actions (2); i.e. x[:,1,t] is the token for s_t
        x = x.reshape(batch_size, seq_length, 3, self.hidden_size).permute(0, 2, 1, 3)

        # get predictions
        # return_preds = self.predict_return(x[:,2])  # predict next return given state and action
        # state_preds = self.predict_state(x[:,2])    # predict next state given state and action
        action_preds = self.predict_action(x[:,1])  # predict next action given state

        #return state_preds, action_preds, return_preds
        return None, action_preds, None

    def get_action(self, states, actions, rewards, returns_to_go, timesteps, **kwargs):
        # we don't care about the past rewards in this model

        states = states.reshape(1, -1, self.state_dim)
        actions = actions.reshape(1, -1, self.act_dim)
        returns_to_go = returns_to_go.reshape(1, -1, 1)
        timesteps = timesteps.reshape(1, -1)

        if self.max_length is not None:
            states = states[:,-self.max_length:]
            actions = actions[:,-self.max_length:]
            returns_to_go = returns_to_go[:,-self.max_length:]
            timesteps = timesteps[:,-self.max_length:]

            # pad all tokens to sequence length
            attention_mask = torch.cat([torch.zeros(self.max_length-states.shape[1]), torch.ones(states.shape[1])])
            attention_mask = attention_mask.to(dtype=torch.long, device=states.device).reshape(1, -1)
            states = torch.cat(
                [torch.zeros((states.shape[0], self.max_length-states.shape[1], self.state_dim), device=states.device), states],
                dim=1).to(dtype=torch.float32)
            actions = torch.cat(
                [torch.zeros((actions.shape[0], self.max_length - actions.shape[1], self.act_dim),
                             device=actions.device), actions],
                dim=1).to(dtype=torch.float32)
            returns_to_go = torch.cat(
                [torch.zeros((returns_to_go.shape[0], self.max_length-returns_to_go.shape[1], 1), device=returns_to_go.device), returns_to_go],
                dim=1).to(dtype=torch.float32)
            timesteps = torch.cat(
                [torch.zeros((timesteps.shape[0], self.max_length-timesteps.shape[1]), device=timesteps.device), timesteps],
                dim=1
            ).to(dtype=torch.long)
        else:
            attention_mask = None

        _, action_preds, return_preds = self.forward(
            states, actions, None, returns_to_go, timesteps, attention_mask=attention_mask, **kwargs)

        return action_preds[0,-1]


class MLP_encoder(torch.nn.Module):
    def __init__(self, in_dim, out_dim, hidden_size=256, dropout=0.1):
        super().__init__()
        self.in_dim = in_dim
        self.out_dim = out_dim
        self.hidden_size = hidden_size

        self.predict_state = nn.Sequential(
            nn.Linear(in_dim, hidden_size),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_size, hidden_size),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_size, out_dim),
        )

    def forward(self, x):
        return self.predict_state(x)


class CNN_encoder(torch.nn.Module):
    def __init__(self, in_shape, out_dim, dropout=0.1):
        super().__init__()
        """
        take input as 1D vector and reshape to in_shape image shape
        e.g. in_shape = (4, 128, 64) for 4-stacked grayscale images
        output is a 1D vector of length out_dim
        """
        self.in_shape = in_shape
        self.out_dim = out_dim
        channel, width, height = in_shape

        self.cnn = nn.Sequential(
            nn.Conv2d(channel, 32, kernel_size=8, stride=4, padding=0),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Conv2d(32, 64, kernel_size=4, stride=2, padding=0),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Dropout(dropout),
            nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=0),
            nn.ReLU(),
            #nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Flatten(),
        )
        dummy_cnn_output = self.cnn(torch.zeros(1, *in_shape))
        print("CNN output shape: ", dummy_cnn_output.shape, "linearly transformed to", out_dim)
        self.linear = nn.Linear(dummy_cnn_output.shape[1], out_dim)

    def forward(self, x):
        batch_size = x.shape[0]
        context_length = x.shape[1]
        x = x.reshape(-1, *self.in_shape)
        x = self.cnn(x)
        x = self.linear(x)
        return x.view(batch_size, context_length, -1)