helper.py

import random
import collections
import torch
import numpy as np
from IPython.display import HTML
from base64 import b64encode

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

"""
memory to save the state, action, reward sequence from the current episode. 
"""
class Memory:
    def __init__(self, len):
        self.rewards = collections.deque(maxlen=len)
        self.state = collections.deque(maxlen=len)
        self.action = collections.deque(maxlen=len)
        self.is_done = collections.deque(maxlen=len)

    def update(self, state, action, reward, done):
        # if the episode is finished we do not save to new state. Otherwise we have more states per episode than rewards
        # and actions whcih leads to a mismatch when we sample from memory.
        if not done:
            self.state.append(state)
        self.action.append(action)
        self.rewards.append(reward)
        self.is_done.append(done)

    def sample(self, batch_size):
        """
        sample "batch_size" many (state, action, reward, next state, is_done) datapoints.
        """
        n = len(self.is_done)
        idx = random.sample(range(0, n-1), batch_size)

        return torch.Tensor(self.state)[idx].to(device), torch.Tensor(self.action)[idx].to(device), \
               torch.Tensor(self.state)[1+np.array(idx)].to(device), torch.Tensor(self.rewards)[idx].to(device), \
               torch.Tensor(self.is_done)[idx].to(device)

    def reset(self):
        self.rewards.clear()
        self.state.clear()
        self.action.clear()
        self.is_done.clear()

def select_action(model, env, state):
    state = torch.Tensor(state).to(device)
    with torch.no_grad():
        values = model(state)
    action = values.cpu().numpy()

    return action

def update_parameters(current_model, target_model):
    target_model.load_state_dict(current_model.state_dict())

def show_video(video_path, video_width = 400):
   
  video_file = open(video_path, "r+b").read()
 
  video_url = f"data:video/mp4;base64,{b64encode(video_file).decode()}"
  return HTML(f"""<video autoplay="true" loop="true" width={video_width} controls><source src="{video_url}"></video>""")