-
Notifications
You must be signed in to change notification settings - Fork 76
/
Copy pathevaluate_elos.py
174 lines (138 loc) · 6.1 KB
/
evaluate_elos.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
import numpy as np
import torch
import time
import pufferlib
import random
import glob
import os
from pufferlib.policy_ranker import update_elos
from pufferlib.environments.ocean.environment import env_creator
from pufferlib.environments.ocean.torch import MOBA, Recurrent
import pufferlib.cleanrl
def load_policies(checkpoint_dir, n, map_location='cuda'):
paths = glob.glob(f'{checkpoint_dir}/model_*.pt', recursive=True)
# Sample with replacement if not enough models
if len(paths) < n:
samples = random.choices(paths, k=n)
else:
samples = random.sample(paths, n)
names = [path.split('/')[-1] for path in samples]
return {name: torch.load(path, map_location=map_location)
for name, path in zip(names, samples)}
def rollout(envs, policy, opponents, num_games, timeout=180, render=False):
obs, _ = envs.reset()
# Double reset clears randomizations
obs, _ = envs.reset()
#cenv = envs.c_envs[0]
start = time.time()
step = 0
num_envs = len(envs.c_envs)
num_opponents = len(opponents)
envs_per_opponent = num_envs // num_opponents
my_states = [None for _ in range(num_opponents)]
opp_states = [None for _ in range(num_opponents)]
prev_radiant_victories = [c.radiant_victories for c in envs.c_envs]
prev_dire_victories = [c.dire_victories for c in envs.c_envs]
scores = []
atn_shape = (10*num_envs, len(envs.action_space.nvec))
actions = torch.zeros(atn_shape, dtype=torch.int64).cuda()
actions_struct = actions.view(num_opponents, envs_per_opponent, 2, 5, len(envs.action_space.nvec))
slice_idxs = torch.arange(10*num_envs).reshape(num_opponents, envs_per_opponent, 2, 5).cuda()
flat_teams = np.random.randint(0, 2, num_envs)
team_assignments = torch.from_numpy(flat_teams.reshape(num_opponents, envs_per_opponent)).cuda()
arange = torch.arange(envs_per_opponent).cuda()
games_played = 0
while games_played < num_games and time.time() - start < timeout:
#if render and step % 10 == 0:
# env.render()
step += 1
with torch.no_grad():
obs = torch.as_tensor(obs).cuda()
for i in range(num_opponents):
idxs = slice_idxs[i]
teams = team_assignments[i]
my_obs = obs[idxs[arange, teams]].view(5*envs_per_opponent, -1)
opp_obs = obs[idxs[arange, 1 - teams]].view(5*envs_per_opponent, -1)
if hasattr(policy, 'lstm'):
my_actions, _, _, _, my_states[i] = policy(my_obs, my_states[i])
opp_atn, _, _, _, opp_states[i] = opponents[i](opp_obs, opp_states[i])
else:
my_actions, _, _, _ = policy(my_obs)
opp_atn, _, _, _ = opponents[i](opp_obs)
actions_struct[i, arange, teams] = my_actions.view(envs_per_opponent, 5, -1)
actions_struct[i, arange, 1 - teams] = opp_atn.view(envs_per_opponent, 5, -1)
obs, reward, done, truncated, info = envs.step(actions.cpu().numpy())
for i in range(num_envs):
c = envs.c_envs[i]
opp_idx = i // envs_per_opponent
if c.radiant_victories > prev_radiant_victories[i]:
prev_radiant_victories[i] = c.radiant_victories
scores.append((opp_idx, flat_teams[i] == 0))
games_played += 1
print('Radiant Victory')
elif c.dire_victories > prev_dire_victories[i]:
prev_dire_victories[i] = c.dire_victories
scores.append((opp_idx, flat_teams[i] == 1))
games_played += 1
print('Dire Victory')
return scores
def calc_elo(checkpoint, checkpoint_dir, elos, num_envs=128, num_games=128, num_opponents=8, k=24.0):
print(f'Calculating ELO for {checkpoint}')
make_env = env_creator('moba')
envs = make_env(num_envs=num_envs)
policy = torch.load(os.path.join(checkpoint_dir, checkpoint), map_location='cuda')
print(f'Loaded policy {checkpoint}')
paths = glob.glob(f'{checkpoint_dir}/model_*.pt', recursive=True)
names = [path.split('/')[-1] for path in paths]
print(f'Loaded {len(paths)} models')
paths.remove(f'{checkpoint_dir}/{checkpoint}')
print(f'Removed {checkpoint} from paths')
elos[checkpoint] = 1000
# Sample with replacement if not enough models
print(f'Sampling {num_opponents} opponents')
n_models = len(paths)
if n_models < num_opponents:
idxs = random.choices(range(n_models), k=num_opponents)
else:
idxs = random.sample(range(n_models), num_opponents)
print(f'Sampled {num_opponents} opponents')
opponent_names = [names[i] for i in idxs]
opponents = [torch.load(paths[i], map_location='cuda') for i in idxs]
print(f'Loaded {num_opponents} opponents')
results = rollout(envs, policy, opponents, num_games=num_games, render=False)
print(f'Finished {num_games} games')
for game in results:
opponent, win = game
if win:
score = np.array([1, 0])
else:
score = np.array([0, 1])
opp_name = opponent_names[opponent]
elo_pair = np.array([elos[checkpoint], elos[opp_name]])
elo_pair = update_elos(elo_pair, score, k=24.0)
elos[checkpoint] = elo_pair[0]
#elos[opp_name] = elo_pair[1]
print(f'Finished calculating ELO for {checkpoint}')
for k, v in elos.items():
print(f'{k}: {v}')
return elos
'''
for game in range(1000):
opponent, name = load_policy(checkpoint_dir)
print(f'Game: {game} Opponent: {name}')
scores = rollout(env, policy, opponent, render=False)
if scores is None:
continue
elo_pair = np.array([elos['mine'], elos[name]])
elo_pair = update_elos(elo_pair, scores, k=24.0)
elos['mine'] = elo_pair[0]
elos[name] = elo_pair[1]
for k, v in elos.items():
print(f'{k}: {v}')
print()
'''
if __name__ == '__main__':
checkpoint_dir = 'moba_elo'
checkpoint = 'model_0.pt'
elos = {'model_random.pt': 1000}
calc_elo(checkpoint, checkpoint_dir, elos, num_games=16)