# Copyright (c) Facebook, Inc. and its affiliates.
device = torch.device('cuda:%d' % args.gpu)
coach = ConvRnnCoach.load(args.coach_path).to(device)
coach.max_raw_chars = args.max_raw_chars
executor = Executor.load(args.model_path).to(device)
executor_wrapper = ExecutorWrapper(coach, executor, coach.num_instructions,
args.max_raw_chars, args.cheat,
args.inst_mode)
executor_wrapper.train(False)
game_option = get_game_option(args)
ai1_option, ai2_option = get_ai_options(args, coach.num_instructions)
context, act_dc = create_game(args.num_thread, ai1_option, ai2_option,
game_option)
context.start()
dc = DataChannelManager([act_dc])
result_stat = ResultStat('reward', None)
while not context.terminated():
data = dc.get_input(max_timeout_s=1)
if len(data) == 0:
continue
data = to_device(data['act'], device)
result_stat.feed(data)
reply = executor_wrapper.forward(data)
dc.set_reply('act', reply)
print(result_stat.log(0))
dc.terminate()
result1 = ResultStat('reward', None)
result2 = ResultStat('reward', None)
i = 0
while not context.terminated():
i += 1
if i % 1000 == 0:
print('%d, progress agent1: win %d, loss %d' %
(i, result1.win, result1.loss))
data = dc.get_input(max_timeout_s=1)
if len(data) == 0:
continue
for key in data:
# print(key)
batch = to_device(data[key], device)
if key == 'act1':
result1.feed(batch)
with torch.no_grad():
reply = model1.forward(batch)
elif key == 'act2':
result2.feed(batch)
with torch.no_grad():
reply = model2.forward(batch)
else:
assert False
dc.set_reply(key, reply)
print(result1.log(0))
print(result2.log(0))
dc.terminate()
# Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. # import argparse import os import sys import pprint from set_path import append_sys_path append_sys_path() import torch import tube from pytube import DataChannelManager import minirts import numpy as np import random import pickle from collections import defaultdict from rnn_coach import ConvRnnCoach from onehot_coach import ConvOneHotCoach from rnn_generator import RnnGenerator from itertools import groupby from executor_wrapper import ExecutorWrapper from executor import Executor from common_utils import to_device, ResultStat, Logger from best_models import best_executors, best_coaches from tqdm import tqdm p1dict = defaultdict(list)