DQN-2018-TAU/double_dqn_learn.py

"""
    This file is copied/apdated from https://github.com/berkeleydeeprlcourse/homework/tree/master/hw3
"""
import sys
import pickle
import numpy as np
from collections import namedtuple
from itertools import count
import random
import gym.spaces
import json

import torch
import torch.autograd as autograd

from utils.replay_buffer import ReplayBuffer
from utils.gym import get_wrapper_by_name

USE_CUDA = torch.cuda.is_available()
print("USE_CUDA=", USE_CUDA)
dtype = torch.cuda.FloatTensor if USE_CUDA else torch.FloatTensor
longType = torch.cuda.LongTensor if USE_CUDA else torch.LongTensor


class Variable(autograd.Variable):
    def __init__(self, data, *args, **kwargs):
        if USE_CUDA:
            data = data.cuda()
        super(Variable, self).__init__(data, *args, **kwargs)


"""
    OptimizerSpec containing following attributes
        constructor: The optimizer constructor ex: RMSprop
        kwargs: {Dict} arguments for constructing optimizer
"""
OptimizerSpec = namedtuple("OptimizerSpec", ["constructor", "kwargs"])


def double_dqn_learning(
        env,
        q_func,
        optimizer_spec,
        exploration,
        stopping_criterion=None,
        replay_buffer_size=1000000,
        batch_size=32,
        gamma=0.99,
        learning_starts=50000,
        learning_freq=4,
        frame_history_len=4,
        target_update_freq=10000,
        save_path=None,
        save_freq=100000,
        log_every_n_steps=3000,
        loss="bellman",
        **kwargs # To avoid complaints of unknown keywords
):
    """
    Similar to dqn_learn.py, but implements double DQN learning instead of normal DQN:
    https://arxiv.org/abs/1509.06461.pdf
    """
    assert type(env.observation_space) == gym.spaces.Box
    assert type(env.action_space) == gym.spaces.Discrete

    ###############
    # BUILD MODEL #
    ###############

    if len(env.observation_space.shape) == 1:
        # This means we are running on low-dimensional observations (e.g. RAM)
        input_arg = env.observation_space.shape[0]
    else:
        img_h, img_w, img_c = env.observation_space.shape
        input_arg = frame_history_len * img_c
    num_actions = env.action_space.n

    def to_pytorch(obs, type=dtype, normalize=True):
        t = torch.from_numpy(obs).type(type)
        if normalize:
            return t / 255.0
        else:
            return t

    def to_pytorch_var(x, grad=False, type=dtype, normalize=True):
        return Variable(to_pytorch(x, type=type, normalize=normalize), requires_grad=grad)

    # Construct an epilson greedy policy with given exploration schedule
    def select_epsilon_greedy_action(model, obs, t):
        sample = random.random()
        eps_threshold = exploration.value(t)
        if sample > eps_threshold:
            obs = to_pytorch(obs).unsqueeze(0)
            # Use volatile = True if variable is only used in inference mode, i.e. don’t save the history
            return model(Variable(obs, volatile=True)).data.max(1)[1].cpu()
        else:
            return torch.IntTensor([[random.randrange(num_actions)]])

    # Initialize target q function and q function, i.e. build the model.
    ######

    # YOUR CODE HERE
    print("Input and output size of network:")
    print(input_arg, num_actions)
    Q = q_func(input_arg, num_actions)
    Q_target = q_func(input_arg, num_actions)
    bellman_l1_loss = torch.nn.SmoothL1Loss(size_average=False)

    if USE_CUDA:
        Q = Q.cuda()
        Q_target = Q_target.cuda()

    Q_target.load_state_dict(Q.state_dict())

    def switch_Q_functions():
        print("Switching Q functions")
        q_state_dict = Q.state_dict()
        q_target_state_dict = Q_target.state_dict()
        Q.load_state_dict(q_target_state_dict)
        Q_target.load_state_dict(q_state_dict)

    ######

    # Construct Q network optimizer function
    optimizer = optimizer_spec.constructor(Q.parameters(), **optimizer_spec.kwargs)

    replay_buffer = ReplayBuffer(replay_buffer_size, frame_history_len)

    start_step = 0
    start_episode = 0
    Statistic = {
        "episode_rewards": []
    }
    def update_stats(t,episode,reward):
        rewards = Statistic["episode_rewards"]
        if len(rewards) == 0 or \
                rewards[-1][1] < episode:
            stat_tuple = (t, episode, reward)
            # print("Updating stats with ", stat_tuple)
            rewards.append(stat_tuple)

    def get_mean_episode_rewards(range):
        return np.mean([r for (_,_,r) in Statistic["episode_rewards"][-range:]])

    if save_path is not None:
        try:
            print("Trying to load state from ", save_path)
            with open(save_path + ".Q.pkl", 'rb') as f:
                Q.load_state_dict(pickle.load(f))
            with open(save_path + ".Q_target.pkl", 'rb') as f:
                Q_target.load_state_dict(pickle.load(f))
            with open(save_path + ".stats.json", 'r') as f:
                saved_stats = json.load(f)
                Statistic = saved_stats["stats"]
                start_step = saved_stats["timestep"]
                start_episode = saved_stats["episode"]
        except Exception as e:
            print("Saved state doesn't exist yet (probably)")
            print(e)

    def save_state(t,episode):
        """
        Saves the current stable network weights, together with the current time step and statistics, for resuming later.
        """
        if save_path is not None:
            print("Saving state")
            with open(save_path + ".Q.pkl", 'wb') as f:
                pickle.dump(Q.state_dict(), f, pickle.HIGHEST_PROTOCOL)
            with open(save_path + ".Q_target.pkl", 'wb') as f:
                pickle.dump(Q.state_dict(), f, pickle.HIGHEST_PROTOCOL)
            with open(save_path + ".stats.json", 'w') as f:
                saved_stats = {
                    "timestep": t,
                    "episode": episode,
                    "stats": Statistic
                }
                json.dump(saved_stats, f)

    ###############
    # RUN ENV     #
    ###############
    num_param_updates = 0
    last_obs = env.reset()

    for t in count(start=start_step):
        ### 1. Check stopping criterion
        if stopping_criterion is not None and stopping_criterion(env):
            break

        ### 2. Step the env and store the transition
        last_frame_idx = replay_buffer.store_frame(last_obs)
        enc_last_obs = replay_buffer.encode_recent_observation()
        action = select_epsilon_greedy_action(Q, enc_last_obs, t)
        new_frame, r, done, _ = env.step(action)
        replay_buffer.store_effect(last_frame_idx, action, r, done)
        if done:
            last_obs = env.reset()
        else:
            last_obs = new_frame

        #####

        ### 3. Perform experience replay and train the network.
        if (t > learning_starts and
                t % learning_freq == 0 and
                replay_buffer.can_sample(batch_size)):

            optimizer.zero_grad()

            obs_batch, act_batch, r_batch, next_obs_batch, done_mask = replay_buffer.sample(batch_size)

            Q_val_batch = Q(to_pytorch_var(obs_batch))
            Q_target_val_batch = Q_target(to_pytorch_var(next_obs_batch)).detach()

            # The following code will take only one cell from each vector of the Q_val_batch tensor.
            # Each vector corresponds to a single output of the Q net, and each cell corresponds to a single action.
            # This means we take only the cells of the actions that we actually took, since all others are irrelevant
            # when calculating the loss.
            act_batch_var = to_pytorch_var(act_batch, type=longType, normalize=False).unsqueeze(1)
            vals_of_actions_taken = Q_val_batch.gather(1, act_batch_var)

            # Here is the difference from normal DQN -
            # we select the maximal value action according to Q, and evaluate it according to Q_target
            Q_next_val_batch = Q(to_pytorch_var(next_obs_batch)).detach()
            _, Q_next_val_max_action = Q_next_val_batch.max(1)
            Q_next_val_max_action = Q_next_val_max_action.unsqueeze(1)
            Q_target_next_val_estimate = Q_target_val_batch.gather(1, Q_next_val_max_action)

            reverse_done_mask = 1 - to_pytorch_var(done_mask, normalize=False).unsqueeze(1)
            Q_target_masked_next_val_estimate = (reverse_done_mask * Q_target_next_val_estimate)
            Q_target_discounted = (gamma * Q_target_masked_next_val_estimate)
            r_batch_var = to_pytorch_var(r_batch, normalize=False).unsqueeze(1)
            Q_target_vals = r_batch_var + Q_target_discounted

            if loss == 'l1':
                bellman_l1_loss(vals_of_actions_taken, Q_target_vals).backward()
            else:
                bellman_error = Q_target_vals - vals_of_actions_taken
                clipped_error = bellman_error.clamp(-1, 1)
                vals_of_actions_taken.backward(-clipped_error)
            optimizer.step()

            num_param_updates += 1
            if num_param_updates % target_update_freq == 0:
                switch_Q_functions()

        ### 4. Log progress and keep track of statistics
        episode_rewards = get_wrapper_by_name(env, "Monitor").get_episode_rewards()
        episode_count = len(episode_rewards)
        total_episodes = start_episode + episode_count
        if episode_count > 0:
            update_stats(t, total_episodes, episode_rewards[-1])

        if t % log_every_n_steps == 0 and t > learning_starts and t > start_step:
            print("Timestep %d" % (t,))
            print("Episode %d" % (total_episodes,))
            print("mean reward (100 episodes) %f" % get_mean_episode_rewards(100))
            print("mean reward (10 episodes) %f" % get_mean_episode_rewards(10))
            print("exploration %f" % exploration.value(t))
            sys.stdout.flush()

        if t % save_freq == 0 and t > start_step:
            save_state(t,total_episodes)