a3c/play.py

import os

import torch
from dotenv import load_dotenv
from huggingface_hub import hf_hub_download

from wordle_env.state import update_from_mask

from .net import GreedyNet
from .utils import v_wrap

load_dotenv()

MODEL_NAME = os.getenv("RS_WORDLE_MODEL_NAME")
HF_MODEL_REPO_NAME = os.getenv("HF_MODEL_REPO_NAME")
MODEL_CHECKPOINT_DIR = "checkpoints"


def get_play_model_path():
    return os.path.join(MODEL_CHECKPOINT_DIR, MODEL_NAME)


def get_net(env, pretrained_model_path):
    n_s = env.observation_space.shape[0]
    n_a = env.action_space.n
    words_list = env.words
    word_width = len(env.words[0])
    net = GreedyNet(n_s, n_a, words_list, word_width)
    if not os.path.exists(pretrained_model_path):
        pretrained_model_path = hf_hub_download(
            HF_MODEL_REPO_NAME, MODEL_NAME, local_dir=MODEL_CHECKPOINT_DIR
        )
    net.load_state_dict(torch.load(pretrained_model_path))
    return net


def get_initial_state(env):
    state = env.reset()
    return state


def suggest(env, words, states, pretrained_model_path) -> str:
    """
    Given a list of words and masks, return the next suggested word

    :param agent:
    :param env:
    :param sequence: History of moves and outcomes until now
    :return:
    """
    env = env.unwrapped
    net = get_net(env, pretrained_model_path)
    state = get_initial_state(env)
    for word, mask in zip(words, states):
        word = word.upper()
        mask = list(map(int, mask))
        state = update_from_mask(state, word, mask)
    return env.words[net.choose_action(v_wrap(state[None, :]))]


def play(env, pretrained_model_path, goal_word=None):
    env = env.unwrapped
    net = get_net(env, pretrained_model_path)
    state = get_initial_state(env)
    if goal_word:
        env.set_goal_word(goal_word)
    outcomes = []
    win = False
    for i in range(env.max_turns):
        action = net.choose_action(v_wrap(state[None, :]))
        state, reward, done, _ = env.step(action)
        outcomes.append(env.words[action])
        if done:
            if reward > 0:
                win = True
            break
    return win, outcomes