def get_best_move(board):
moves = board.generate_legal_moves()
moves = list(moves)
bitboards = []
for move in moves:
b = board.copy()
b.push(move)
bitboards.append(get_bitboard(b))
bitboards = np.array(bitboards)
curr_bitboard = get_bitboard(board)
_, features = featurize(featurizer, bitboards, device)
features = features.cpu().detach().numpy()
_, curr_features = featurize(featurizer, curr_bitboard, device)
curr_features = curr_features.cpu().detach().numpy()
to_compare = np.hstack((np.repeat(curr_features, len(moves),
axis=0), features))
scores = compare(comparator, to_compare, device)
scores = scores[:, 0]
best_idx = np.argmax(scores)
board.push(moves[best_idx])
return moves[best_idx]
def step(self, action_dict):
if self.is_render:
self.render()
actions = []
for id in range(4):
if id in action_dict:
actions.append(action_dict[id])
else:
actions.append(0)
obs = {}
rewards = {}
dones = {}
infos = {i - 10: {} for i in self.prev_obs[0]['alive']}
_obs, _reward, _done, _info = self.env.step(actions)
for id in self.prev_obs[0]['alive']:
if _done or self.is_done(id - 10, _obs[0]['alive']):
dones[id - 10] = True
infos[id - 10]["metrics"] = self.stat[id - 10]
dones["__all__"] = _done
for id in range(4):
if self.is_agent_alive(id):
obs[id] = featurize(_obs[id])
rewards[id] = self.reward(id, _obs, _info)
infos[id].update(_info)
self.prev_obs = _obs
return obs, rewards, dones, infos
def reset(self):
self.prev_obs = self.env.reset()
obs = {}
self.reset_stat()
for i in range(4):
if self.is_agent_alive(i):
obs[i] = featurize(self.prev_obs[i])
return obs
def get_all_actions():
feature = featurize(obs[train_idx])
action, _states = model.predict(feature)
action = tuple(action)
some_actions = env.act(obs) # 不包含我的 agent
# 如果其他智能体动作不是元组（只有单一动作），改成元组
for agent_idx in range(3):
if not isinstance(some_actions[agent_idx], tuple):
some_actions[agent_idx] = (some_actions[agent_idx], 0, 0)
some_actions.insert(train_idx, action) # 把我的 agent 的动作也加进来
return some_actions
def process_data(traj_1, traj_2):
"""
Input:
Traj: a list of list, contains one trajectory for one driver
example:[[114.10437, 22.573433, '2016-07-02 00:08:45', 1],
[114.179665, 22.558701, '2016-07-02 00:08:52', 1]]
Output:
Data: any format that can be consumed by your model.
"""
data = []
split1 = [
featurize(sub, torch.device("cuda")) for sub in split_traj(traj_1, 64)
]
split2 = [
featurize(sub, torch.device("cuda")) for sub in split_traj(traj_2, 64)
]
x1 = torch.stack(
[torch.stack(random.choices(split1, k=8)) for _ in range(8)])
x2 = torch.stack(
[torch.stack(random.choices(split2, k=8)) for _ in range(8)])
return x1, x2
def step(self, action_dict):
if self.is_render:
self.render()
actions = []
for id in range(4):
if id in action_dict:
actions.append(action_dict[id])
else:
actions.append(0)
obs = {}
rewards = {}
dones = {}
infos = {}
_obs, _reward, _done, _info = self.env.step(actions)
for id in range(4):
if self.is_done(id, _obs[0]['alive']):
dones[id] = True
if id == 0:
_done = True
_info['result'] = constants.Result.Loss
dones["__all__"] = _done
for id in range(4):
if self.is_agent_alive(id):
obs[id] = featurize(_obs[id])
rewards[id] = self.reward(id, _obs, _info)
infos[id] = _info
self.prev_obs = _obs
return obs, rewards, dones, infos
# Assign best accs
metrics['best_epoch'] = best_epoch
metrics['best_val_acc'] = best_val_acc
metrics['best_val_same_acc'] = best_val_same_acc
metrics['best_test_acc'] = best_test_acc
metrics['best_test_same_acc'] = best_test_same_acc
metrics['best_test_acc_ci'] = best_test_acc_ci
metrics['best_val_ap'] = val_avg_prec
metrics['best_test_ap'] = test_avg_prec
metrics['best_val_same_ap'] = val_same_avg_prec
metrics['best_test_same_ap'] = test_same_avg_prec
metrics['has_same'] = has_same
save_defaultdict_to_fs(metrics,
os.path.join(args.exp_dir, 'metrics.json'))
featurize()
if (not args.skip_eval):
print('====> DONE')
print('====> BEST EPOCH: {}'.format(best_epoch))
print('====> {:>17}\tEpoch: {}\tAccuracy: {:.4f}'.format(
'(best_val)', best_epoch, best_val_acc))
print('====> {:>17}\tEpoch: {}\tAccuracy: {:.4f}'.format(
'(best_val_same)', best_epoch, best_val_same_acc))
print('====> {:>17}\tEpoch: {}\tAccuracy: {:.4f}'.format(
'(best_test)', best_epoch, best_test_acc))
print('====> {:>17}\tEpoch: {}\tAccuracy: {:.4f}'.format(
'(best_test_same)', best_epoch, best_test_same_acc))
print('====>')
print('====> {:>17}\tEpoch: {}\tAccuracy: {:.4f}'.format(
'(best_val_avg)', best_epoch,
(best_val_acc + best_val_same_acc) / 2))
for fold_ix, fold in enumerate(dataset):
start_batch = time.time()
batch = [fold]
fold_mut = copy.deepcopy(fold)
print(len(fold['mutation_data']))
num_mutations = len(fold['mutation_data'])
rocklin_df = pd.DataFrame(columns=['seq', 'stabilityscore', 'neglogp'])
try:
for mut_ix, (mut_seq, effect) in enumerate(fold['mutation_data']):
fold_mut['seq'] = mut_seq
# Get a batch
X, S, mask, lengths = featurize([fold_mut], device)
elapsed_featurize = time.time() - start_batch
log_probs = model(X, S, lengths, mask)
loss, loss_av = _loss(S, log_probs, mask)
neglogp = torch.sum(loss * mask, dim=1) / torch.sum(mask, dim=1)
neglogp = neglogp.cpu().data.numpy().tolist()[0]
print(fold['name'], neglogp, effect, mut_ix, num_mutations)
rocklin_df.loc[mut_ix] = [mut_seq, effect, neglogp]
rocklin_df.to_csv('rocklin/mutations/' + fold['name'] + '_' + args.features + '.tsv', sep='\t')
plt.clf()
plt.scatter(rocklin_df['neglogp'], rocklin_df['stabilityscore'], s=5)
plt.xlabel('Transformer Neglogp')
plt.ylabel('Stability score in experiment')
if __name__ == '__main__':
agent_list = [
agents.RandomAgent(),
agents.StaticAgent(),
agents.StaticAgent(),
agents.StaticAgent()
]
env = pommerman.make(
'PommeTeam-v0',
agent_list,
# '/home/lucius/working/projects/pomme_rllib/resources/one_line_state.json'
)
obs = env.reset()
while True:
features = featurize(obs[0])
for i in range(17):
print(features[i])
print()
actions = env.act(obs)
print(actions)
obs, reward, done, info = env.step(actions)
if done:
break
print(obs)
features = featurize(obs[0])
for i in range(17):
print(features[i])
print()
del matches_hard[("66ef54d76ff989a91d52", "51c9e1dd498e33ecd8670892")]
del matches_hard[("edeba23f215dcc702220", "51a11cbc498e4083823909f1")]
# Extra matches
matches_hard_test = utils.read_matches("matches_test_hard.csv")
matches_easy = utils.read_matches("matches_train.csv")
matches_easy_test = utils.read_matches("matches_test.csv")
# Compiling data sets
try:
f = open("working/locu_classifier.cache", 'rb')
(X_tot, y_tot) = cPickle.load(f)
sys.stderr.write("Loading data from cache.")
except IOError:
sys.stderr.write( "Featurizing easy dataset..." )
(X_easy, index_easy) = utils.featurize(locu_easy, four_easy, utils.sim)
y_easy = utils.get_y(index_easy, matches_easy)
sys.stderr.write( "done.\n" )
sys.stderr.write( "Featurizing easy test dataset..." )
(X_easy_test, index_easy_test) = utils.featurize(locu_easy_test, four_easy_test, utils.sim)
y_easy_test = utils.get_y(index_easy_test, matches_easy_test)
sys.stderr.write( "done.\n" )
sys.stderr.write( "Featurizing hard dataset..." )
(X, index) = utils.featurize(locu, four, utils.sim)
y = utils.get_y(index, matches_hard)
sys.stderr.write( "done.\n" )
sys.stderr.write( "Featurizing hard test dataset..." )
(X_hard_test, index_hard_test) = utils.featurize(locu_test, four_test, utils.sim)
for i in range(len(matching)):
locu = reverse_a[i]
four = reverse_b[matching[i]]
w = weights[i][matching[i]]
res[(locu, four)] = w
return res
# ----------------
# Main
# ----------------
# Load in the json files
model_file = sys.argv[1]
locu_file = sys.argv[2]
four_file = sys.argv[3]
match_out_file = sys.argv[4]
locu = utils.load_json(locu_file)
four = utils.load_json(four_file)
sys.stderr.write( "Featurizing dataset..." )
(X, index) = utils.featurize(locu, four, utils.sim)
sys.stderr.write( "done.\n" )
# Load in model
(model, thresh) = joblib.load(model_file)
## Test on testing
p = model.predict_proba(X)
res = weights_to_matching(p, index)
utils.write_matching(res, thresh, match_out_file)
# fdb733b6
checkpoint = 600
checkpoint_dir = "/home/lucius/ray_results/two_policies_vs_static_agents/PPO_RllibPomme_0_2020-06-09_23-39-347whmqdrs"
ppo_agent.restore("{}/checkpoint_{}/checkpoint-{}".format(
checkpoint_dir, checkpoint, checkpoint))
agent_list = []
for agent_id in range(4):
agent_list.append(agents.StaticAgent())
env = pommerman.make("PommeTeam-v0", agent_list=agent_list)
for i in range(1):
obs = env.reset()
done = False
while not done:
env.render()
actions = env.act(obs)
actions[0] = ppo_agent.compute_action(observation=featurize(obs[0]),
policy_id="policy_0")
actions[2] = ppo_agent.compute_action(observation=featurize(obs[2]),
policy_id="policy_0")
obs, reward, done, info = env.step(actions)
print("reward:", reward)
print("done:", done)
print("info:", info)
print("=========")
env.render(close=True)
# env.close()