def gather_plot_values(dir_games_all, data_store, config, model):
mu_all = []
var_all = []
for dir_game in dir_games_all:
if dir_game == '.DS_Store':
continue
state_trace_length, state_input, reward, action, team_id, player_index = get_icehockey_game_data(
data_store=data_store,
dir_game=dir_game,
config=config,
player_id_cluster_dir=None)
action_seq = transfer2seq(data=action,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
team_id_seq = transfer2seq(data=team_id,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
player_index_seq = transfer2seq(data=player_index,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
input_data = np.concatenate([state_input, action_seq], axis=2)
[mu, var] = sess.run(
[model.mu_out, model.var_out],
feed_dict={
model.rnn_input_ph: input_data,
model.trace_lengths_ph: state_trace_length
})
mu_all.append(mu)
var_all.append(var)
return mu_all, var_all
def gathering_running_and_run(dir_game, config, player_id_cluster_dir, data_store,
model, sess, training_flag, game_number, validate_cvrnn_flag=False,
validate_td_flag=False, validate_variance_flag=False,
output_decoder_all=None,
target_data_all=None, selection_matrix_all=None,
q_values_all=None, pretrain_flag=None):
if validate_variance_flag:
match_q_values_players_dict = {}
for i in range(config.Learn.player_cluster_number):
match_q_values_players_dict.update({i: []})
else:
match_q_values_players_dict = None
state_trace_length, state_input, reward, action, team_id, player_index = get_icehockey_game_data(
data_store=data_store, dir_game=dir_game, config=config, player_id_cluster_dir=player_id_cluster_dir)
action_seq = transfer2seq(data=action, trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
team_id_seq = transfer2seq(data=team_id, trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
player_index_seq = transfer2seq(data=player_index, trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
# reward_count = sum(reward)
# print ("reward number" + str(reward_count))
if len(state_input) != len(reward) or len(state_trace_length) != len(reward):
raise Exception('state length does not equal to reward length')
train_len = len(state_input)
train_number = 0
batch_number = 0
s_t0 = state_input[train_number]
train_number += 1
while True:
# try:
batch_return, \
train_number, \
s_tl, \
print_flag = get_together_training_batch(s_t0=s_t0,
state_input=state_input,
reward=reward,
player_index=player_index_seq,
train_number=train_number,
train_len=train_len,
state_trace_length=state_trace_length,
action=action_seq,
team_id=team_id_seq,
# win_info=None,
config=config)
# get the batch variables
# s_t0, s_t1, r_t0_combine, s_length_t0, s_length_t1, action_id_t0, action_id_t1, team_id_t0,
# team_id_t1, 0, 0
s_t0_batch = [d[0] for d in batch_return]
s_t1_batch = [d[1] for d in batch_return]
r_t_batch = [d[2] for d in batch_return]
trace_t0_batch = [d[3] for d in batch_return]
trace_t1_batch = [d[4] for d in batch_return]
action_id_t0 = [d[5] for d in batch_return]
action_id_t1 = [d[6] for d in batch_return]
team_id_t0_batch = [d[7] for d in batch_return]
team_id_t1_batch = [d[8] for d in batch_return]
player_id_t0_batch = [d[9] for d in batch_return]
player_id_t1_batch = [d[10] for d in batch_return]
if config.Learn.apply_pid:
input_data_t0 = np.concatenate([np.asarray(player_id_t0_batch), np.asarray(s_t0_batch),
np.asarray(action_id_t0)], axis=2)
input_data_t1 = np.concatenate([np.asarray(player_id_t1_batch), np.asarray(s_t1_batch),
np.asarray(action_id_t1)], axis=2)
else:
input_data_t0 = np.concatenate([np.asarray(s_t0_batch), np.asarray(action_id_t0)], axis=2)
input_data_t1 = np.concatenate([np.asarray(s_t1_batch), np.asarray(action_id_t1)], axis=2)
trace_lengths_t0 = trace_t0_batch
trace_lengths_t1 = trace_t1_batch
for i in range(0, len(batch_return)):
terminal = batch_return[i][-2]
cut = batch_return[i][-1]
if training_flag:
# print (len(state_input) / (config.Learn.batch_size*10))
print_flag = True if batch_number % (len(state_input) / (config.Learn.batch_size * 10)) == 0 else False
train_td_model(model, sess, config, input_data_t0, trace_lengths_t0,
input_data_t1, trace_lengths_t1, r_t_batch, terminal, cut,
pretrain_flag, print_flag)
else:
pass
batch_number += 1
s_t0 = s_tl
if terminal:
break
return [output_decoder_all, target_data_all, selection_matrix_all, q_values_all, match_q_values_players_dict]
def run_network(sess, model, config, training_dir_games_all,
testing_dir_games_all, model_data_store_dir,
player_id_cluster_dir, source_data_store_dir,
save_network_dir):
game_number = 0
converge_flag = False
saver = tf.train.Saver(max_to_keep=300)
while True:
game_diff_record_dict = {}
iteration_now = game_number / config.Learn.number_of_total_game + 1
game_diff_record_dict.update({"Iteration": iteration_now})
if converge_flag:
break
elif game_number >= len(
training_dir_games_all) * config.Learn.iterate_num:
break
# else:
# converge_flag = True
for dir_game in training_dir_games_all:
if dir_game == '.DS_Store':
continue
game_number += 1
game_cost_record = []
state_trace_length, state_input, reward, action, team_id, player_index = get_icehockey_game_data(
data_store=model_data_store_dir,
dir_game=dir_game,
config=config,
player_id_cluster_dir=player_id_cluster_dir)
action_seq = transfer2seq(data=action,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
team_id_seq = transfer2seq(data=team_id,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
if config.Learn.predict_target == 'ActionGoal' or config.Learn.predict_target == 'Action':
player_index_seq = transfer2seq(
data=player_index,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
else:
player_index_seq = player_index
if config.Learn.predict_target == 'ActionGoal':
actions_all = read_feature_within_events(
directory=dir_game,
data_path=source_data_store_dir,
feature_name='name')
next_goal_label = []
data_length = state_trace_length.shape[0]
win_info = []
new_reward = []
new_action_seq = []
new_state_input = []
new_state_trace_length = []
new_team_id_seq = []
for action_index in range(0, data_length):
action = actions_all[action_index]
if 'shot' in action:
if action_index + 1 == data_length:
continue
new_reward.append(reward[action_index])
new_action_seq.append(action_seq[action_index])
new_state_input.append(state_input[action_index])
new_state_trace_length.append(
state_trace_length[action_index])
new_team_id_seq.append(team_id_seq[action_index])
if actions_all[action_index + 1] == 'goal':
# print(actions_all[action_index+1])
next_goal_label.append([1, 0])
else:
# print(actions_all[action_index + 1])
next_goal_label.append([0, 1])
reward = np.asarray(new_reward)
action_seq = np.asarray(new_action_seq)
state_input = np.asarray(new_state_input)
state_trace_length = np.asarray(new_state_trace_length)
team_id_seq = np.asarray(new_team_id_seq)
win_info = np.asarray(next_goal_label)
elif config.Learn.predict_target == 'Action':
win_info = action[1:, :]
reward = reward[:-1]
action_seq = action_seq[:-1, :, :]
state_input = state_input[:-1, :, :]
state_trace_length = state_trace_length[:-1]
team_id_seq = team_id_seq[:-1, :, :]
else:
win_info = None
print("\n training file" + str(dir_game))
# reward_count = sum(reward)
# print ("reward number" + str(reward_count))
if len(state_input) != len(reward) or len(
state_trace_length) != len(reward):
raise Exception('state length does not equal to reward length')
train_len = len(state_input)
train_number = 0
s_t0 = state_input[train_number]
train_number += 1
while True:
# try:
batch_return, \
train_number, \
s_tl, \
print_flag = get_together_training_batch(s_t0=s_t0,
state_input=state_input,
reward=reward,
player_index=player_index_seq,
train_number=train_number,
train_len=train_len,
state_trace_length=state_trace_length,
action=action_seq,
team_id=team_id_seq,
win_info=win_info,
config=config)
# get the batch variables
# s_t0, s_t1, r_t0_combine, s_length_t0, s_length_t1, action_id_t0, action_id_t1, team_id_t0,
# team_id_t1, 0, 0
s_t0_batch = [d[0] for d in batch_return]
s_t1_batch = [d[1] for d in batch_return]
r_t_batch = [d[2] for d in batch_return]
trace_t0_batch = [d[3] for d in batch_return]
trace_t1_batch = [d[4] for d in batch_return]
action_id_t0 = [d[5] for d in batch_return]
action_id_t1 = [d[6] for d in batch_return]
team_id_t0_batch = [d[7] for d in batch_return]
team_id_t1_batch = [d[8] for d in batch_return]
player_id_t0_batch = [d[9] for d in batch_return]
player_id_t1_batch = [d[10] for d in batch_return]
win_info_t0_batch = [d[11] for d in batch_return]
if config.Learn.predict_target == 'ActionGoal':
target_data = np.asarray(win_info_t0_batch)
m2balanced = True
elif config.Learn.predict_target == 'Action':
target_data = np.asarray(win_info_t0_batch)
m2balanced = True
elif config.Learn.predict_target == 'PlayerLocalId':
config.Learn.apply_pid = False
target_data = np.asarray(player_id_t0_batch)
m2balanced = False
else:
raise ValueError('unknown predict_target {0}'.format(
config.Learn.predict_target))
if config.Learn.apply_pid:
input_data = np.concatenate([
np.asarray(action_id_t0),
np.asarray(s_t0_batch),
np.asarray(player_id_t0_batch)
],
axis=2)
else:
input_data = np.concatenate(
[np.asarray(action_id_t0),
np.asarray(s_t0_batch)],
axis=2)
trace_lengths = trace_t0_batch
for i in range(0, len(batch_return)):
terminal = batch_return[i][-2]
# cut = batch_return[i][8]
if config.Learn.apply_stochastic:
for i in range(len(input_data)):
if m2balanced:
cache_label = 0 if target_data[i][0] == 0 else 1
BalancedMemoryBuffer.push([
input_data[i], target_data[i], trace_lengths[i]
],
cache_label=cache_label)
else:
MemoryBuffer.push([
input_data[i], target_data[i], trace_lengths[i]
])
if game_number <= 10:
s_t0 = s_tl
if terminal:
break
else:
continue
if m2balanced:
sampled_data = BalancedMemoryBuffer.sample(
batch_size=config.Learn.batch_size)
else:
sampled_data = MemoryBuffer.sample(
batch_size=config.Learn.batch_size)
sample_input_data = np.asarray(
[sampled_data[j][0] for j in range(len(sampled_data))])
sample_target_data = np.asarray(
[sampled_data[j][1] for j in range(len(sampled_data))])
sample_trace_lengths = np.asarray(
[sampled_data[j][2] for j in range(len(sampled_data))])
else:
sample_input_data = input_data
sample_target_data = target_data
sample_trace_lengths = trace_lengths
train_model(model, sess, config, sample_input_data,
sample_target_data, sample_trace_lengths, terminal)
s_t0 = s_tl
if terminal:
# save progress after a game
# model.saver.save(sess, saved_network + '/' + config.learn.sport + '-game-',
# global_step=game_number)
# v_diff_record_average = sum(v_diff_record) / len(v_diff_record)
# game_diff_record_dict.update({dir_game: v_diff_record_average})
break
save_model(game_number, saver, sess, save_network_dir, config)
if game_number % 100 == 1:
validation_model(testing_dir_games_all, model_data_store_dir,
config, sess, model, player_id_cluster_dir,
source_data_store_dir)
def validation_model(testing_dir_games_all, data_store, config, sess, model,
player_id_cluster_dir, source_data_store_dir):
output_decoder_all = None
target_data_all = None
selection_matrix_all = None
print('validating model')
game_number = 0
for dir_game in testing_dir_games_all:
if dir_game == '.DS_Store':
continue
game_number += 1
state_trace_length, state_input, reward, action, team_id, player_index = get_icehockey_game_data(
data_store=data_store,
dir_game=dir_game,
config=config,
player_id_cluster_dir=player_id_cluster_dir)
action_seq = transfer2seq(data=action,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
team_id_seq = transfer2seq(data=team_id,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
if config.Learn.predict_target == 'ActionGoal' or config.Learn.predict_target == 'Action':
player_index_seq = transfer2seq(
data=player_index,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
else:
player_index_seq = player_index
if config.Learn.predict_target == 'ActionGoal':
actions_all = read_feature_within_events(
directory=dir_game,
data_path=source_data_store_dir,
feature_name='name')
next_goal_label = []
data_length = state_trace_length.shape[0]
win_info = []
new_reward = []
new_action_seq = []
new_state_input = []
new_state_trace_length = []
new_team_id_seq = []
for action_index in range(0, data_length):
action = actions_all[action_index]
if 'shot' in action:
new_reward.append(reward[action_index])
new_action_seq.append(action_seq[action_index])
new_state_input.append(state_input[action_index])
new_state_trace_length.append(
state_trace_length[action_index])
new_team_id_seq.append(team_id_seq[action_index])
if action_index + 1 == data_length:
continue
if actions_all[action_index + 1] == 'goal':
# print(actions_all[action_index+1])
next_goal_label.append([1, 0])
else:
# print(actions_all[action_index + 1])
next_goal_label.append([0, 1])
reward = np.asarray(new_reward)
action_seq = np.asarray(new_action_seq)
state_input = np.asarray(new_state_input)
state_trace_length = np.asarray(new_state_trace_length)
team_id_seq = np.asarray(new_team_id_seq)
win_info = np.asarray(next_goal_label)
elif config.Learn.predict_target == 'Action':
win_info = action[1:, :]
reward = reward[:-1]
action_seq = action_seq[:-1, :, :]
state_input = state_input[:-1, :, :]
state_trace_length = state_trace_length[:-1]
team_id_seq = team_id_seq[:-1, :, :]
else:
win_info = None
# print ("\n training file" + str(dir_game))
# reward_count = sum(reward)
# print ("reward number" + str(reward_count))
if len(state_input) != len(reward) or len(state_trace_length) != len(
reward):
raise Exception('state length does not equal to reward length')
train_len = len(state_input)
train_number = 0
s_t0 = state_input[train_number]
train_number += 1
while True:
# try:
batch_return, \
train_number, \
s_tl, \
print_flag = get_together_training_batch(s_t0=s_t0,
state_input=state_input,
reward=reward,
player_index=player_index_seq,
train_number=train_number,
train_len=train_len,
state_trace_length=state_trace_length,
action=action_seq,
team_id=team_id_seq,
win_info=win_info,
config=config)
# get the batch variables
# s_t0, s_t1, r_t0_combine, s_length_t0, s_length_t1, action_id_t0, action_id_t1, team_id_t0,
# team_id_t1, 0, 0
s_t0_batch = [d[0] for d in batch_return]
s_t1_batch = [d[1] for d in batch_return]
r_t_batch = [d[2] for d in batch_return]
trace_t0_batch = [d[3] for d in batch_return]
trace_t1_batch = [d[4] for d in batch_return]
action_id_t0 = [d[5] for d in batch_return]
action_id_t1 = [d[6] for d in batch_return]
team_id_t0_batch = [d[7] for d in batch_return]
team_id_t1_batch = [d[8] for d in batch_return]
player_id_t0_batch = [d[9] for d in batch_return]
player_id_t1_batch = [d[10] for d in batch_return]
win_info_t0_batch = [d[11] for d in batch_return]
trace_lengths = trace_t0_batch
if config.Learn.predict_target == 'ActionGoal':
target_data = np.asarray(win_info_t0_batch)
elif config.Learn.predict_target == 'Action':
target_data = np.asarray(win_info_t0_batch)
elif config.Learn.predict_target == 'PlayerLocalId':
config.Learn.apply_pid = False
target_data = np.asarray(player_id_t0_batch)
else:
raise ValueError('unknown predict_target {0}'.format(
config.Learn.predict_target))
if config.Learn.apply_pid:
input_data = np.concatenate([
np.asarray(action_id_t0),
np.asarray(s_t0_batch),
np.asarray(player_id_t0_batch)
],
axis=2)
else:
input_data = np.concatenate(
[np.asarray(action_id_t0),
np.asarray(s_t0_batch)], axis=2)
for i in range(0, len(batch_return)):
terminal = batch_return[i][-2]
# cut = batch_return[i][8]
[output_prob] = sess.run(
[model.read_out],
feed_dict={
model.rnn_input_ph: input_data,
model.y_ph: target_data,
model.trace_lengths_ph: trace_lengths
})
if output_decoder_all is None:
output_decoder_all = output_prob
target_data_all = target_data
else:
output_decoder_all = np.concatenate(
[output_decoder_all, output_prob], axis=0)
target_data_all = np.concatenate(
[target_data_all, target_data], axis=0)
s_t0 = s_tl
if terminal:
# save progress after a game
# model.saver.save(sess, saved_network + '/' + config.learn.sport + '-game-',
# global_step=game_number)
# v_diff_record_average = sum(v_diff_record) / len(v_diff_record)
# game_diff_record_dict.update({dir_game: v_diff_record_average})
break
acc = compute_acc(output_prob=output_decoder_all,
target_label=target_data_all,
if_print=True)
print("testing acc is {0}".format(str(acc)))
def gathering_data_and_run(
dir_game,
config,
player_id_cluster_dir,
data_store,
source_data_dir,
model,
sess,
training_flag,
game_number,
validate_cvrnn_flag=False,
validate_td_flag=False,
validate_diff_flag=False,
validate_variance_flag=False,
validate_predict_flag=False,
output_decoder_all=None,
target_data_all=None,
selection_matrix_all=None,
q_values_all=None,
output_label_all=None,
real_label_all=None,
pred_target_data_all=None,
pred_output_prob_all=None,
training_file=None,
):
if validate_variance_flag:
match_q_values_players_dict = {}
for i in range(config.Learn.player_cluster_number):
match_q_values_players_dict.update({i: []})
else:
match_q_values_players_dict = None
state_trace_length, state_input, reward, action, team_id, player_index = get_icehockey_game_data(
data_store=data_store,
dir_game=dir_game,
config=config,
player_id_cluster_dir=player_id_cluster_dir)
action_seq = transfer2seq(data=action,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
reward_seq = transfer2seq(data=np.expand_dims(reward, axis=-1),
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
team_id_seq = transfer2seq(data=team_id,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
player_index_seq = transfer2seq(data=player_index,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
# win_one_hot = compute_game_win_vec(rewards=reward)
score_diff = compute_game_score_diff_vec(rewards=reward)
score_difference_game = read_feature_within_events(
dir_game,
source_data_dir,
'scoreDifferential',
transfer_home_number=True,
data_store=data_store)
state_reward_input = np.concatenate(
[reward_seq, state_input],
axis=-1) # concatenate the sequence of state and reward.
add_pred_flag = False
if config.Arch.Predict.predict_target == 'ActionGoal':
add_pred_flag = True
actions_all = read_feature_within_events(directory=dir_game,
data_path=source_data_dir,
feature_name='name')
next_goal_label = []
data_length = state_trace_length.shape[0]
new_reward = []
new_action_seq = []
new_state_input = []
new_state_trace_length = []
new_team_id_seq = []
new_player_index_seq = []
for action_index in range(0, data_length):
action = actions_all[action_index]
if 'shot' in action:
if action_index + 1 == data_length:
continue
new_reward.append(reward[action_index])
new_action_seq.append(action_seq[action_index])
new_state_input.append(state_reward_input[action_index])
new_state_trace_length.append(state_trace_length[action_index])
new_team_id_seq.append(team_id_seq[action_index])
new_player_index_seq.append(player_index_seq[action_index])
if actions_all[action_index + 1] == 'goal':
# print(actions_all[action_index+1])
next_goal_label.append([1, 0])
else:
# print(actions_all[action_index + 1])
next_goal_label.append([0, 1])
pred_target = next_goal_label
elif config.Arch.Predict.predict_target == 'Action':
add_pred_flag = True
pred_target = action[1:, :]
new_reward = reward[:-1]
new_action_seq = action_seq[:-1, :, :]
new_state_input = state_reward_input[:-1, :, :]
new_state_trace_length = state_trace_length[:-1]
new_team_id_seq = team_id_seq[:-1, :, :]
new_player_index_seq = player_index_seq[:-1, :, :]
else:
# raise ValueError()
add_pred_flag = False
if add_pred_flag:
if training_flag:
train_mask = np.asarray([[[1]] * config.Learn.max_seq_length] *
len(new_state_input))
else:
train_mask = np.asarray([[[0]] * config.Learn.max_seq_length] *
len(new_state_input))
if config.Learn.predict_target == 'PlayerLocalId':
pred_input_data = np.concatenate([
np.asarray(new_player_index_seq),
np.asarray(new_team_id_seq),
np.asarray(new_state_input),
np.asarray(new_action_seq), train_mask
],
axis=2)
pred_target_data = np.asarray(np.asarray(pred_target))
pred_trace_lengths = new_state_trace_length
pred_selection_matrix = generate_selection_matrix(
new_state_trace_length,
max_trace_length=config.Learn.max_seq_length)
else:
pred_input_data = np.concatenate([
np.asarray(new_player_index_seq),
np.asarray(new_state_input),
np.asarray(new_action_seq), train_mask
],
axis=2)
pred_target_data = np.asarray(np.asarray(pred_target))
pred_trace_lengths = new_state_trace_length
pred_selection_matrix = generate_selection_matrix(
new_state_trace_length,
max_trace_length=config.Learn.max_seq_length)
if training_flag:
for i in range(len(new_state_input)):
cache_label = np.argmax(pred_target_data[i], axis=0)
Prediction_MemoryBuffer.push([
pred_input_data[i], pred_target_data[i],
pred_trace_lengths[i], pred_selection_matrix[i]
], cache_label)
# reward_count = sum(reward)
# print ("reward number" + str(reward_count))
if len(state_reward_input) != len(reward) or len(
state_trace_length) != len(reward):
raise Exception('state length does not equal to reward length')
kl_loss_game = []
ll_game = []
train_len = len(state_reward_input)
train_number = 0
s_t0 = state_reward_input[train_number]
train_number += 1
while True:
# try:
batch_return, \
train_number, \
s_tl, \
print_flag = get_together_training_batch(s_t0=s_t0,
state_input=state_reward_input,
reward=reward,
player_index=player_index_seq,
train_number=train_number,
train_len=train_len,
state_trace_length=state_trace_length,
action=action_seq,
team_id=team_id_seq,
win_info=score_diff,
score_info=score_difference_game,
config=config)
# get the batch variables
# s_t0, s_t1, r_t0_combine, s_length_t0, s_length_t1, action_id_t0, action_id_t1, team_id_t0,
# team_id_t1, 0, 0
s_t0_batch = [d[0] for d in batch_return]
s_t1_batch = [d[1] for d in batch_return]
r_t_batch = [d[2] for d in batch_return]
trace_t0_batch = [d[3] for d in batch_return]
trace_t1_batch = [d[4] for d in batch_return]
action_id_t0 = [d[5] for d in batch_return]
action_id_t1 = [d[6] for d in batch_return]
team_id_t0_batch = [d[7] for d in batch_return]
team_id_t1_batch = [d[8] for d in batch_return]
player_id_t0_batch = [d[9] for d in batch_return]
player_id_t1_batch = [d[10] for d in batch_return]
win_id_t_batch = [d[11] for d in batch_return]
terminal_batch = [d[-2] for d in batch_return]
cut_batch = [d[-1] for d in batch_return]
score_diff_t_batch = [d[11] for d in batch_return]
score_diff_base_t0_batch = [d[12] for d in batch_return]
outcome_data = score_diff_t_batch
score_diff_base_t0 = score_diff_base_t0_batch
if training_flag:
train_mask = np.asarray([[[1]] * config.Learn.max_seq_length] *
len(s_t0_batch))
else:
train_mask = np.asarray([[[0]] * config.Learn.max_seq_length] *
len(s_t0_batch))
# (player_id, state ,action flag)
for i in range(0, len(terminal_batch)):
terminal = terminal_batch[i]
cut = cut_batch[i]
input_data_t0 = np.concatenate([
np.asarray(player_id_t0_batch),
np.asarray(team_id_t0_batch),
np.asarray(s_t0_batch),
np.asarray(action_id_t0), train_mask
],
axis=2)
target_data_t0 = np.asarray(np.asarray(player_id_t0_batch))
trace_lengths_t0 = trace_t0_batch
selection_matrix_t0 = generate_selection_matrix(
trace_lengths_t0, max_trace_length=config.Learn.max_seq_length)
input_data_t1 = np.concatenate([
np.asarray(player_id_t1_batch),
np.asarray(team_id_t1_batch),
np.asarray(s_t1_batch),
np.asarray(action_id_t1), train_mask
],
axis=2)
target_data_t1 = np.asarray(np.asarray(player_id_t1_batch))
trace_lengths_t1 = trace_t1_batch
selection_matrix_t1 = generate_selection_matrix(
trace_t1_batch, max_trace_length=config.Learn.max_seq_length)
if training_flag:
if config.Learn.apply_stochastic:
for i in range(len(input_data_t0)):
General_MemoryBuffer.push([
input_data_t0[i], target_data_t0[i],
trace_lengths_t0[i], selection_matrix_t0[i],
input_data_t1[i], target_data_t1[i],
trace_lengths_t1[i], selection_matrix_t1[i],
r_t_batch[i], win_id_t_batch[i], terminal_batch[i],
cut_batch[i]
])
sampled_data = General_MemoryBuffer.sample(
batch_size=config.Learn.batch_size)
sample_input_data_t0 = np.asarray(
[sampled_data[j][0] for j in range(len(sampled_data))])
sample_target_data_t0 = np.asarray(
[sampled_data[j][1] for j in range(len(sampled_data))])
sample_trace_lengths_t0 = np.asarray(
[sampled_data[j][2] for j in range(len(sampled_data))])
sample_selection_matrix_t0 = np.asarray(
[sampled_data[j][3] for j in range(len(sampled_data))])
# sample_input_data_t1 = np.asarray([sampled_data[j][4] for j in range(len(sampled_data))])
# sample_target_data_t1 = np.asarray([sampled_data[j][5] for j in range(len(sampled_data))])
# sample_trace_lengths_t1 = np.asarray([sampled_data[j][6] for j in range(len(sampled_data))])
# sample_selection_matrix_t1 = np.asarray([sampled_data[j][7] for j in range(len(sampled_data))])
# sample_r_t_batch = np.asarray([sampled_data[j][8] for j in range(len(sampled_data))])
# sample_terminal_batch = np.asarray([sampled_data[j][10] for j in range(len(sampled_data))])
# sample_cut_batch = np.asarray([sampled_data[j][11] for j in range(len(sampled_data))])
sampled_outcome_t = np.asarray(
[sampled_data[j][9] for j in range(len(sampled_data))])
for i in range(0, len(terminal_batch)):
batch_terminal = terminal_batch[i]
batch_cut = cut_batch[i]
else:
sample_input_data_t0 = input_data_t0
sample_target_data_t0 = target_data_t0
sample_trace_lengths_t0 = trace_lengths_t0
sample_selection_matrix_t0 = selection_matrix_t0
sampled_outcome_t = win_id_t_batch
kld_all, ll_all = train_cvrnn_model(model, sess, config,
sample_input_data_t0,
sample_target_data_t0,
sample_trace_lengths_t0,
sample_selection_matrix_t0)
kl_loss_game += kld_all
ll_game += ll_all
# """we skip sampling for TD learning"""
# train_td_model(model, sess, config, input_data_t0, trace_lengths_t0, selection_matrix_t0,
# input_data_t1, trace_lengths_t1, selection_matrix_t1, r_t_batch, terminal, cut)
#
# train_score_diff(model, sess, config, input_data_t0, trace_lengths_t0, selection_matrix_t0,
# input_data_t1, trace_lengths_t1, selection_matrix_t1, r_t_batch, outcome_data,
# score_diff_base_t0, terminal, cut)
# if add_pred_flag:
# sampled_data = Prediction_MemoryBuffer.sample(batch_size=config.Learn.batch_size)
# sample_pred_input_data = np.asarray([sampled_data[j][0] for j in range(len(sampled_data))])
# sample_pred_target_data = np.asarray([sampled_data[j][1] for j in range(len(sampled_data))])
# sample_pred_trace_lengths = np.asarray([sampled_data[j][2] for j in range(len(sampled_data))])
# sample_pred_selection_matrix = np.asarray([sampled_data[j][3] for j in range(len(sampled_data))])
#
# train_prediction(model, sess, config,
# sample_pred_input_data,
# sample_pred_target_data,
# sample_pred_trace_lengths,
# sample_pred_selection_matrix)
else:
# for i in range(0, len(r_t_batch)):
# if i == len(r_t_batch) - 1:
# if terminal or cut:
# print(r_t_batch[i])
if validate_cvrnn_flag:
output_decoder = cvrnn_validation(sess, model, input_data_t0,
target_data_t0,
trace_lengths_t0,
selection_matrix_t0, config)
if output_decoder_all is None:
output_decoder_all = output_decoder
target_data_all = target_data_t0
selection_matrix_all = selection_matrix_t0
else:
# try:
output_decoder_all = np.concatenate(
[output_decoder_all, output_decoder], axis=0)
# except:
# print output_decoder_all.shape
# print output_decoder.shape
target_data_all = np.concatenate(
[target_data_all, target_data_t0], axis=0)
selection_matrix_all = np.concatenate(
[selection_matrix_all, selection_matrix_t0], axis=0)
if validate_td_flag:
# validate_variance_flag = validate_variance_flag if train_number <= 500 else False
q_values, match_q_values_players_dict = \
td_validation(sess, model, trace_lengths_t0, selection_matrix_t0,
player_id_t0_batch, s_t0_batch, action_id_t0, input_data_t0,
train_mask, config, match_q_values_players_dict,
r_t_batch, terminal, cut, train_number, validate_variance_flag=False)
if q_values_all is None:
q_values_all = q_values
else:
q_values_all = np.concatenate([q_values_all, q_values],
axis=0)
if validate_diff_flag:
output_label, real_label = diff_validation(
sess, model, input_data_t0, trace_lengths_t0,
selection_matrix_t0, score_diff_base_t0, config,
outcome_data)
if real_label_all is None:
real_label_all = real_label
else:
real_label_all = np.concatenate(
[real_label_all, real_label], axis=0)
if output_label_all is None:
output_label_all = output_label
else:
output_label_all = np.concatenate(
[output_label_all, output_label], axis=0)
s_t0 = s_tl
if terminal:
break
if training_file is not None:
training_file.write("Mean kl:{0}, Mean ll:{1}\n".format(
np.mean(kl_loss_game), np.mean(ll_game)))
training_file.flush()
if validate_predict_flag and add_pred_flag:
input_data, pred_output_prob = prediction_validation(
model, sess, config, pred_input_data, pred_target_data,
pred_trace_lengths, pred_selection_matrix)
if pred_target_data_all is None:
pred_target_data_all = pred_target_data
else:
pred_target_data_all = np.concatenate(
[pred_target_data_all, pred_target_data], axis=0)
if pred_output_prob_all is None:
pred_output_prob_all = pred_output_prob
else:
pred_output_prob_all = np.concatenate(
[pred_output_prob_all, pred_output_prob], axis=0)
return [
output_decoder_all, target_data_all, selection_matrix_all,
q_values_all, real_label_all, output_label_all, pred_target_data_all,
pred_output_prob_all, match_q_values_players_dict
]
def validation_model(testing_dir_games_all, data_store, config, sess, model,
predicted_target, embedding2validate):
model_output_all = None
target_data_all = None
selection_matrix_all = None
print('validating model')
game_number = 0
for dir_game in testing_dir_games_all:
if dir_game == '.DS_Store':
continue
game_number += 1
state_trace_length, state_input, reward, action, team_id, player_index = get_icehockey_game_data(
data_store=data_store, dir_game=dir_game, config=config)
# state_trace_length = np.asarray([10] * len(state_trace_length))
action_seq = transfer2seq(data=action,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
team_id_seq = transfer2seq(data=team_id,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
player_id_seq = transfer2seq(data=player_index,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
# print ("\n training file" + str(dir_game))
# reward_count = sum(reward)
# print ("reward number" + str(reward_count))
if len(state_input) != len(reward) or len(state_trace_length) != len(
reward):
raise Exception('state length does not equal to reward length')
train_len = len(state_input)
train_number = 0
s_t0 = state_input[train_number]
train_number += 1
while True:
# try:
batch_return, \
train_number, \
s_tl, \
print_flag = get_together_training_batch(s_t0=s_t0,
state_input=state_input,
reward=reward,
player_index=player_index,
train_number=train_number,
train_len=train_len,
state_trace_length=state_trace_length,
action=action_seq,
team_id=team_id_seq,
config=config)
# get the batch variables
# s_t0, s_t1, r_t0_combine, s_length_t0, s_length_t1, action_id_t0, action_id_t1, team_id_t0,
# team_id_t1, 0, 0
s_t0_batch = [d[0] for d in batch_return]
s_t1_batch = [d[1] for d in batch_return]
r_t_batch = [d[2] for d in batch_return]
trace_t0_batch = [d[3] for d in batch_return]
trace_t1_batch = [d[4] for d in batch_return]
action_id_t0_batch = [d[5] for d in batch_return]
action_id_t1_batch = [d[6] for d in batch_return]
team_id_t0_batch = [d[7] for d in batch_return]
team_id_t1_batch = [d[8] for d in batch_return]
player_id_t0_batch = [d[9] for d in batch_return]
player_id_t1_batch = [d[10] for d in batch_return]
r_t_seq_batch = transfer2seq(
data=np.asarray(r_t_batch),
trace_length=trace_t0_batch,
max_length=config.Learn.max_seq_length)
current_state, history_state = handle_de_history(
data_seq_all=s_t0_batch, trace_lengths=trace_t0_batch)
current_action, history_action = handle_de_history(
data_seq_all=action_id_t0_batch, trace_lengths=trace_t0_batch)
current_reward, history_reward = handle_de_history(
data_seq_all=r_t_seq_batch, trace_lengths=trace_t0_batch)
if embedding2validate:
embed_data = []
for player_id in player_id_t0_batch:
player_index_scalar = np.where(player_id == 1)[0][0]
embed_data.append(embedding2validate[player_index_scalar])
embed_data = np.asarray(embed_data)
else:
embed_data = np.asarray(player_id_t0_batch)
if predicted_target == 'action':
input_seq_data = np.concatenate([
np.asarray(history_state),
np.asarray(history_action),
np.asarray(history_reward)
],
axis=2)
input_obs_data = np.concatenate(
[np.asarray(current_state),
np.asarray(current_reward)],
axis=1)
target_data = np.asarray(current_action)
trace_lengths = [tl - 1 for tl in trace_t0_batch
] # reduce 1 from trace length
elif predicted_target == 'state':
input_seq_data = np.concatenate([
np.asarray(history_state),
np.asarray(history_action),
np.asarray(history_reward)
],
axis=2)
input_obs_data = np.concatenate(
[np.asarray(current_action),
np.asarray(current_reward)],
axis=1)
target_data = np.asarray(current_state)
trace_lengths = [tl - 1 for tl in trace_t0_batch
] # reduce 1 from trace length
elif predicted_target == 'reward':
input_seq_data = np.concatenate([
np.asarray(history_state),
np.asarray(history_action),
np.asarray(history_reward)
],
axis=2)
input_obs_data = np.concatenate(
[np.asarray(current_state),
np.asarray(current_action)],
axis=1)
target_data = np.asarray(current_reward)
trace_lengths = [tl - 1 for tl in trace_t0_batch
] # reduce 1 from trace length
else:
raise ValueError('undefined predicted target')
for i in range(0, len(batch_return)):
terminal = batch_return[i][-2]
# cut = batch_return[i][8]
[output_prob, _] = sess.run(
[model.read_out, model.train_op],
feed_dict={
model.rnn_input_ph: input_seq_data,
model.feature_input_ph: input_obs_data,
model.y_ph: target_data,
model.embed_label_ph: embed_data,
model.trace_lengths_ph: trace_lengths
})
if model_output_all is None:
model_output_all = output_prob
target_data_all = target_data
else:
model_output_all = np.concatenate(
[model_output_all, output_prob], axis=0)
target_data_all = np.concatenate(
[target_data_all, target_data], axis=0)
s_t0 = s_tl
if terminal:
# save progress after a game
# model.saver.save(sess, saved_network + '/' + config.learn.sport + '-game-',
# global_step=game_number)
# v_diff_record_average = sum(v_diff_record) / len(v_diff_record)
# game_diff_record_dict.update({dir_game: v_diff_record_average})
break
if predicted_target == 'action':
acc = compute_acc(output_prob=model_output_all,
target_label=target_data_all,
if_print=True)
print("testing acc is {0}".format(str(acc)))
else:
mae = compute_mae(output_actions_prob=model_output_all,
target_actions_prob=target_data_all,
if_print=True)
print("mae is {0}".format(str(mae)))
def run_network(
sess,
model,
config,
training_dir_games_all,
testing_dir_games_all,
data_store,
predicted_target,
save_network_dir,
validate_embedding_tag,
):
if validate_embedding_tag is not None:
validate_msg = save_network_dir.split('/')[-1].split('_')[0] + '_'
save_embed_dir = save_network_dir.replace('de_embed_saved_networks', 'store_embedding'). \
replace('de_model_saved_NN', 'de_model_save_embedding').replace(validate_msg, '')
print('Applying embedding_matrix_game{0}.csv'.format(
str(validate_embedding_tag)))
with open(
save_embed_dir + '/embedding_matrix_game{0}.csv'.format(
str(validate_embedding_tag)), 'r') as csv_file:
csv_reader = csv.reader(csv_file, delimiter=',')
embedding2validate = []
for row in csv_reader:
embedding2validate.append(row)
else:
embedding2validate = None
game_number = 0
converge_flag = False
saver = tf.train.Saver(max_to_keep=300)
while True:
game_diff_record_dict = {}
iteration_now = game_number / config.Learn.number_of_total_game + 1
game_diff_record_dict.update({"Iteration": iteration_now})
if converge_flag:
break
elif game_number >= len(
training_dir_games_all) * config.Learn.iterate_num:
break
# else:
# converge_flag = True
for dir_game in training_dir_games_all:
if dir_game == '.DS_Store':
continue
print("\n training file" + str(dir_game))
game_number += 1
game_cost_record = []
state_trace_length, state_input, reward, action, team_id, player_index = get_icehockey_game_data(
data_store=data_store, dir_game=dir_game, config=config)
# state_trace_length = np.asarray([10] * len(state_trace_length))
action_seq = transfer2seq(data=action,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
team_id_seq = transfer2seq(data=team_id,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
player_id_seq = transfer2seq(
data=player_index,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
# reward_count = sum(reward)
# print ("reward number" + str(reward_count))
if len(state_input) != len(reward) or len(
state_trace_length) != len(reward):
raise Exception('state length does not equal to reward length')
train_len = len(state_input)
train_number = 0
s_t0 = state_input[train_number]
train_number += 1
while True:
# try:
batch_return, \
train_number, \
s_tl, \
print_flag = get_together_training_batch(s_t0=s_t0,
state_input=state_input,
reward=reward,
player_index=player_index,
train_number=train_number,
train_len=train_len,
state_trace_length=state_trace_length,
action=action_seq,
team_id=team_id_seq,
config=config)
# get the batch variables
# s_t0, s_t1, r_t0_combine, s_length_t0, s_length_t1, action_id_t0, action_id_t1, team_id_t0,
# team_id_t1, 0, 0
s_t0_batch = [d[0] for d in batch_return]
s_t1_batch = [d[1] for d in batch_return]
r_t_batch = [d[2] for d in batch_return]
trace_t0_batch = [d[3] for d in batch_return]
trace_t1_batch = [d[4] for d in batch_return]
action_id_t0_batch = [d[5] for d in batch_return]
action_id_t1_batch = [d[6] for d in batch_return]
team_id_t0_batch = [d[7] for d in batch_return]
team_id_t1_batch = [d[8] for d in batch_return]
player_id_t0_batch = [d[9] for d in batch_return]
player_id_t1_batch = [d[10] for d in batch_return]
r_t_seq_batch = transfer2seq(
data=np.asarray(r_t_batch),
trace_length=trace_t0_batch,
max_length=config.Learn.max_seq_length)
current_state, history_state = handle_de_history(
data_seq_all=s_t0_batch, trace_lengths=trace_t0_batch)
current_action, history_action = handle_de_history(
data_seq_all=action_id_t0_batch,
trace_lengths=trace_t0_batch)
current_reward, history_reward = handle_de_history(
data_seq_all=r_t_seq_batch, trace_lengths=trace_t0_batch)
if embedding2validate:
embed_data = []
for player_id in player_id_t0_batch:
player_index_scalar = np.where(player_id == 1)[0][0]
embed_data.append(
embedding2validate[player_index_scalar])
embed_data = np.asarray(embed_data)
else:
embed_data = np.asarray(player_id_t0_batch)
if predicted_target == 'action':
input_seq_data = np.concatenate([
np.asarray(history_state),
np.asarray(history_action),
np.asarray(history_reward)
],
axis=2)
input_obs_data = np.concatenate([
np.asarray(current_state),
np.asarray(current_reward)
],
axis=1)
target_data = np.asarray(current_action)
trace_lengths = [tl - 1 for tl in trace_t0_batch
] # reduce 1 from trace length
elif predicted_target == 'state':
input_seq_data = np.concatenate([
np.asarray(history_state),
np.asarray(history_action),
np.asarray(history_reward)
],
axis=2)
input_obs_data = np.concatenate([
np.asarray(current_action),
np.asarray(current_reward)
],
axis=1)
target_data = np.asarray(current_state)
trace_lengths = [tl - 1 for tl in trace_t0_batch
] # reduce 1 from trace length
elif predicted_target == 'reward':
input_seq_data = np.concatenate([
np.asarray(history_state),
np.asarray(history_action),
np.asarray(history_reward)
],
axis=2)
input_obs_data = np.concatenate([
np.asarray(current_state),
np.asarray(current_action)
],
axis=1)
target_data = np.asarray(current_reward)
trace_lengths = [tl - 1 for tl in trace_t0_batch
] # reduce 1 from trace length
else:
raise ValueError('undefined predicted target')
for i in range(0, len(batch_return)):
terminal = batch_return[i][-2]
# cut = batch_return[i][8]
# pretrain_flag = True if game_number <= 10 else False
train_model(model, sess, config, input_seq_data,
input_obs_data, target_data, trace_lengths,
embed_data, terminal)
s_t0 = s_tl
if terminal:
# save progress after a game
# model.saver.save(sess, saved_network + '/' + config.learn.sport + '-game-',
# global_step=game_number)
# v_diff_record_average = sum(v_diff_record) / len(v_diff_record)
# game_diff_record_dict.update({dir_game: v_diff_record_average})
break
if game_number % 100 == 1:
if validate_embedding_tag is None:
collect_de_player_embeddings(
sess=sess,
model=model,
save_network_dir=save_network_dir,
game_number=game_number)
if save_network_dir is not None:
print 'saving game {0} in {1}'.format(
str(game_number), save_network_dir)
saver.save(sess,
save_network_dir + '/' + config.Learn.sport +
'-game-',
global_step=game_number)
if game_number % 10 == 1:
validation_model(testing_dir_games_all, data_store, config,
sess, model, predicted_target,
embedding2validate)
def gathering_running_and_run(dir_game,
config,
player_id_cluster_dir,
data_store,
model,
sess,
training_flag,
game_number,
validate_cvrnn_flag=False,
validate_td_flag=False,
validate_variance_flag=False,
output_decoder_all=None,
target_data_all=None,
selection_matrix_all=None,
q_values_all=None):
if validate_variance_flag:
match_q_values_players_dict = {}
for i in range(config.Learn.player_cluster_number):
match_q_values_players_dict.update({i: []})
else:
match_q_values_players_dict = None
state_trace_length, state_input, reward, action, team_id, player_index = get_icehockey_game_data(
data_store=data_store,
dir_game=dir_game,
config=config,
player_id_cluster_dir=player_id_cluster_dir)
action_seq = transfer2seq(data=action,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
team_id_seq = transfer2seq(data=team_id,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
player_index_seq = transfer2seq(data=player_index,
trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
# reward_count = sum(reward)
# print ("reward number" + str(reward_count))
if len(state_input) != len(reward) or len(state_trace_length) != len(
reward):
raise Exception('state length does not equal to reward length')
initial_prior_mu_train = np.zeros(
shape=[config.Learn.batch_size, config.Arch.CVRNN.latent_dim])
previous_prior_mu_train = initial_prior_mu_train
initial_prior_sigma_train = np.zeros(
shape=[config.Learn.batch_size, config.Arch.CVRNN.latent_dim])
previous_prior_sigma_train = initial_prior_sigma_train
train_len = len(state_input)
train_number = 0
s_t0 = state_input[train_number]
train_number += 1
while True:
# try:
batch_return, \
train_number, \
s_t1, \
print_flag = get_together_training_batch(s_t0=s_t0,
state_input=state_input,
reward=reward,
player_index=player_index_seq,
train_number=train_number,
train_len=train_len,
state_trace_length=state_trace_length,
action=action_seq,
team_id=team_id_seq,
config=config)
# get the batch variables
# s_t0, s_t1, r_t0_combine, s_length_t0, s_length_t1, action_id_t0, action_id_t1, team_id_t0,
# team_id_t1, 0, 0
s_t0_batch = [d[0] for d in batch_return]
s_t1_batch = [d[1] for d in batch_return]
r_t_batch = [d[2] for d in batch_return]
trace_t0_batch = [d[3] for d in batch_return]
trace_t1_batch = [d[4] for d in batch_return]
action_id_t0 = [d[5] for d in batch_return]
action_id_t1 = [d[6] for d in batch_return]
team_id_t0_batch = [d[7] for d in batch_return]
team_id_t1_batch = [d[8] for d in batch_return]
player_id_t0_batch = [d[9] for d in batch_return]
player_id_t1_batch = [d[10] for d in batch_return]
terminal_batch = [d[-2] for d in batch_return]
cut_batch = [d[-1] for d in batch_return]
if training_flag:
train_mask = np.asarray([[[1]] * config.Learn.max_seq_length] *
len(s_t0_batch))
else:
train_mask = np.asarray([[[0]] * config.Learn.max_seq_length] *
len(s_t0_batch))
# (player_id, state ,action flag)
for i in range(0, len(terminal_batch)):
terminal = terminal_batch[i]
cut = cut_batch[i]
previous_prior_mu_train_seq = np.zeros(shape=[
len(batch_return), config.Learn.max_seq_length,
config.Arch.CVRNN.latent_dim
])
previous_prior_mu_train_seq[:, 0, :] = initial_prior_mu_train
previous_prior_sigma_train_seq = np.zeros(shape=[
len(batch_return), config.Learn.max_seq_length,
config.Arch.CVRNN.latent_dim
])
previous_prior_sigma_train_seq[:, 0, :] = initial_prior_sigma_train
if config.Learn.predict_target == 'PlayerLocalId':
input_data_t0 = np.concatenate([
np.asarray(player_id_t0_batch),
np.asarray(team_id_t0_batch),
np.asarray(s_t0_batch),
np.asarray(action_id_t0), previous_prior_mu_train_seq,
previous_prior_sigma_train_seq, train_mask
],
axis=2)
target_data_t0 = np.asarray(np.asarray(player_id_t0_batch))
trace_lengths_t0 = trace_t0_batch
selection_matrix_t0 = generate_selection_matrix(
trace_lengths_t0, max_trace_length=config.Learn.max_seq_length)
input_data_t1 = np.concatenate([
np.asarray(player_id_t1_batch),
np.asarray(team_id_t1_batch),
np.asarray(s_t1_batch),
np.asarray(action_id_t1), previous_prior_mu_train_seq,
previous_prior_sigma_train_seq, train_mask
],
axis=2)
target_data_t1 = np.asarray(np.asarray(player_id_t1_batch))
trace_lengths_t1 = trace_t1_batch
selection_matrix_t1 = generate_selection_matrix(
trace_t1_batch, max_trace_length=config.Learn.max_seq_length)
else:
input_data_t0 = np.concatenate([
np.asarray(player_id_t0_batch),
np.asarray(s_t0_batch),
np.asarray(action_id_t0), previous_prior_mu_train_seq,
previous_prior_sigma_train_seq, train_mask
],
axis=2)
target_data_t0 = np.asarray(np.asarray(player_id_t0_batch))
trace_lengths_t0 = trace_t0_batch
selection_matrix_t0 = generate_selection_matrix(
trace_lengths_t0, max_trace_length=config.Learn.max_seq_length)
input_data_t1 = np.concatenate([
np.asarray(player_id_t1_batch),
np.asarray(s_t1_batch),
np.asarray(action_id_t1), previous_prior_mu_train_seq,
previous_prior_sigma_train_seq, train_mask
],
axis=2)
target_data_t1 = np.asarray(np.asarray(player_id_t1_batch))
trace_lengths_t1 = trace_t1_batch
selection_matrix_t1 = generate_selection_matrix(
trace_t1_batch, max_trace_length=config.Learn.max_seq_length)
if training_flag:
if config.Learn.apply_stochastic:
for i in range(len(input_data_t0)):
MemoryBuffer.push([
input_data_t0[i], target_data_t0[i],
trace_lengths_t0[i], selection_matrix_t0[i],
input_data_t1[i], target_data_t1[i],
trace_lengths_t1[i], selection_matrix_t1[i],
r_t_batch[i], terminal_batch[i], cut_batch[i]
])
sampled_data = MemoryBuffer.sample(
batch_size=config.Learn.batch_size)
sample_input_data_t0 = np.asarray(
[sampled_data[j][0] for j in range(len(sampled_data))])
sample_target_data_t0 = np.asarray(
[sampled_data[j][1] for j in range(len(sampled_data))])
sample_trace_lengths_t0 = np.asarray(
[sampled_data[j][2] for j in range(len(sampled_data))])
sample_selection_matrix_t0 = np.asarray(
[sampled_data[j][3] for j in range(len(sampled_data))])
# sample_input_data_t1 = np.asarray([sampled_data[j][4] for j in range(len(sampled_data))])
# sample_target_data_t1 = np.asarray([sampled_data[j][5] for j in range(len(sampled_data))])
# sample_trace_lengths_t1 = np.asarray([sampled_data[j][6] for j in range(len(sampled_data))])
# sample_selection_matrix_t1 = np.asarray([sampled_data[j][7] for j in range(len(sampled_data))])
# sample_r_t_batch = np.asarray([sampled_data[j][8] for j in range(len(sampled_data))])
# sample_terminal_batch = np.asarray([sampled_data[j][9] for j in range(len(sampled_data))])
# sample_cut_batch = np.asarray([sampled_data[j][10] for j in range(len(sampled_data))])
pretrain_flag = False
for i in range(0, len(terminal_batch)):
batch_terminal = terminal_batch[i]
batch_cut = cut_batch[i]
else:
sample_input_data_t0 = input_data_t0
sample_target_data_t0 = target_data_t0
sample_trace_lengths_t0 = trace_lengths_t0
sample_selection_matrix_t0 = selection_matrix_t0
train_cvrnn_model(model, sess, config, sample_input_data_t0,
sample_target_data_t0, sample_trace_lengths_t0,
sample_selection_matrix_t0, pretrain_flag)
# train_td_model(model, sess, config, input_data_t0, trace_lengths_t0, selection_matrix_t0,
# input_data_t1, trace_lengths_t1, selection_matrix_t1, r_t_batch, terminal, cut)
else:
if validate_cvrnn_flag:
output_decoder = cvrnn_validation(sess, model, input_data_t0,
target_data_t0,
trace_lengths_t0,
selection_matrix_t0, config)
if output_decoder_all is None:
output_decoder_all = output_decoder
target_data_all = target_data_t0
selection_matrix_all = selection_matrix_t0
else:
# try:
output_decoder_all = np.concatenate(
[output_decoder_all, output_decoder], axis=0)
# except:
# print output_decoder_all.shape
# print output_decoder.shape
target_data_all = np.concatenate(
[target_data_all, target_data_t0], axis=0)
selection_matrix_all = np.concatenate(
[selection_matrix_all, selection_matrix_t0], axis=0)
# if validate_td_flag:
# # validate_variance_flag = validate_variance_flag if train_number <= 500 else False
# q_values, match_q_values_players_dict = \
# td_validation(sess, model, trace_lengths_t0, selection_matrix_t0,
# player_id_t0_batch, s_t0_batch, action_id_t0, input_data_t0,
# train_mask, config, match_q_values_players_dict,
# r_t_batch, terminal, cut, train_number, validate_variance_flag=False)
#
# if q_values_all is None:
# q_values_all = q_values
# else:
# q_values_all = np.concatenate([q_values_all, q_values], axis=0)
s_t0 = s_t1
if terminal:
break
return [
output_decoder_all, target_data_all, selection_matrix_all,
q_values_all, match_q_values_players_dict
]
def gathering_running_and_run(dir_game, config, player_id_cluster_dir, data_store, source_data_dir,
model, sess, training_flag, game_number,
validate_cvrnn_flag=False,
validate_td_flag=False,
validate_diff_flag=False,
validate_variance_flag=False,
validate_predict_flag=False,
output_decoder_all=None,
target_data_all=None,
q_values_all=None,
output_label_all=None,
real_label_all=None,
pred_target_data_all=None,
pred_output_prob_all=None
):
if validate_variance_flag:
match_q_values_players_dict = {}
for i in range(config.Learn.player_cluster_number):
match_q_values_players_dict.update({i: []})
else:
match_q_values_players_dict = None
state_trace_length, state_input, reward, actions, team_id, player_index = get_icehockey_game_data(
data_store=data_store, dir_game=dir_game, config=config, player_id_cluster_dir=player_id_cluster_dir)
if config.Learn.apply_lstm:
# raise ValueError('no! do not use LSTM!')
encoder_trace = state_trace_length
encoder_state_input = state_input
player_index_seq = transfer2seq(data=player_index, trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
encoder_player_index = player_index_seq
action_seq = transfer2seq(data=actions, trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
encoder_action = action_seq
team_id_seq = transfer2seq(data=team_id, trace_length=state_trace_length,
max_length=config.Learn.max_seq_length)
encoder_team_id = team_id_seq
axis_concat = 2
else:
state_zero_trace = [1] * len(state_trace_length)
state_zero_input = []
for trace_index in range(0, len(state_trace_length)):
trace_length = state_trace_length[trace_index]
trace_length = trace_length - 1
if trace_length > 9:
trace_length = 9
state_zero_input.append(state_input[trace_index, trace_length, :])
state_zero_input = np.asarray(state_zero_input)
encoder_trace = state_zero_trace
encoder_state_input = state_zero_input
encoder_player_index = player_index
encoder_action = actions
encoder_team_id = team_id
axis_concat = 1
if config.Arch.Predict.predict_target == 'ActionGoal':
add_pred_flag = True
actions_all = read_feature_within_events(directory=dir_game,
data_path=source_data_dir,
feature_name='name')
next_goal_label = []
data_length = state_trace_length.shape[0]
new_reward = []
new_action = []
new_state_input = []
new_state_trace = []
new_team_id = []
new_player_index = []
for action_index in range(0, data_length):
action = actions_all[action_index]
if 'shot' in action:
if action_index + 1 == data_length:
continue
if config.Learn.apply_lstm:
new_reward.append(reward[action_index])
new_action.append(encoder_action[action_index])
new_state_input.append(encoder_state_input[action_index])
new_state_trace.append(encoder_trace[action_index])
new_team_id.append(encoder_team_id[action_index])
new_player_index.append(encoder_player_index[action_index])
else:
new_reward.append(reward[action_index])
new_action.append(encoder_action[action_index])
new_state_input.append(encoder_state_input[action_index])
new_state_trace.append(encoder_trace[action_index])
new_team_id.append(encoder_team_id[action_index])
new_player_index.append(encoder_player_index[action_index])
if actions_all[action_index + 1] == 'goal':
# print(actions_all[action_index+1])
next_goal_label.append([1, 0])
else:
# print(actions_all[action_index + 1])
next_goal_label.append([0, 1])
pred_target = next_goal_label
# elif config.Arch.Predict.predict_target == 'Action':
# add_pred_flag = True
# pred_target = actions[1:, :]
# new_reward = reward[:-1]
# new_action = actions[:-1, :, :]
# new_state_zero_input = state_zero_input[:-1, :, :]
# new_state_zero_trace = state_zero_trace[:-1]
# new_team_id = team_id[:-1, :, :]
# new_player_index = player_index[:-1, :, :]
else:
# raise ValueError()
add_pred_flag = False
if add_pred_flag:
# if training_flag:
# pred_train_mask = np.asarray([1] * len(new_state_zero_input))
# else:
# pred_train_mask = np.asarray([1] * len(new_state_zero_input)) # we should apply encoder output
if config.Learn.predict_target == 'PlayerLocalId':
pred_input_data = np.concatenate([np.asarray(new_player_index),
np.asarray(new_team_id),
np.asarray(new_state_input),
np.asarray(new_action), ], axis=axis_concat)
pred_target_data = np.asarray(np.asarray(pred_target))
pred_trace_lengths = new_state_trace
else:
pred_input_data = np.concatenate([np.asarray(new_player_index),
np.asarray(new_state_input),
np.asarray(new_action)], axis=axis_concat)
pred_target_data = np.asarray(np.asarray(pred_target))
pred_trace_lengths = new_state_trace
if training_flag:
for i in range(len(new_state_input)):
cache_label = np.argmax(pred_target_data[i], axis=0)
Prediction_MemoryBuffer.push([pred_input_data[i],
pred_target_data[i],
pred_trace_lengths[i],
# pred_train_mask[i]
], cache_label)
score_diff = compute_game_score_diff_vec(rewards=reward)
score_difference_game = read_feature_within_events(dir_game,
source_data_dir,
'scoreDifferential',
transfer_home_number=True,
data_store=data_store)
# reward_count = sum(reward)
# print ("reward number" + str(reward_count))
if len(state_input) != len(reward) or len(state_trace_length) != len(reward):
raise Exception('state length does not equal to reward length')
train_len = len(state_input)
train_number = 0
s_t0 = encoder_state_input[train_number]
train_number += 1
while True:
# try:
batch_return, \
train_number, \
s_tl, \
print_flag = get_together_training_batch(s_t0=s_t0,
state_input=encoder_state_input,
reward=reward,
player_index=encoder_player_index,
train_number=train_number,
train_len=train_len,
state_trace_length=encoder_trace,
action=encoder_action,
team_id=encoder_team_id,
win_info=score_diff,
score_info=score_difference_game,
config=config)
# get the batch variables
# s_t0, s_t1, r_t0_combine, s_length_t0, s_length_t1, action_id_t0, action_id_t1, team_id_t0,
# team_id_t1, 0, 0
s_t0_batch = [d[0] for d in batch_return]
s_t1_batch = [d[1] for d in batch_return]
r_t_batch = [d[2] for d in batch_return]
trace_t0_batch = [d[3] for d in batch_return]
trace_t1_batch = [d[4] for d in batch_return]
action_id_t0 = [d[5] for d in batch_return]
action_id_t1 = [d[6] for d in batch_return]
team_id_t0_batch = [d[7] for d in batch_return]
team_id_t1_batch = [d[8] for d in batch_return]
player_id_t0_batch = [d[9] for d in batch_return]
player_id_t1_batch = [d[10] for d in batch_return]
win_id_t_batch = [d[11] for d in batch_return]
terminal_batch = [d[-2] for d in batch_return]
cut_batch = [d[-1] for d in batch_return]
score_diff_t_batch = [d[11] for d in batch_return]
score_diff_base_t0_batch = [d[12] for d in batch_return]
outcome_data = score_diff_t_batch
score_diff_base_t0 = score_diff_base_t0_batch
# (player_id, state ,action flag)
if training_flag:
train_mask = np.asarray([1] * len(s_t0_batch))
else:
train_mask = np.asarray([0] * len(s_t0_batch))
for i in range(0, len(terminal_batch)):
terminal = terminal_batch[i]
cut = cut_batch[i]
if config.Learn.apply_lstm:
target_data_t0 = []
for trace_index in range(0, len(trace_t0_batch)):
trace_length = trace_t0_batch[trace_index]
trace_length = trace_length - 1
target_data_t0.append(player_id_t0_batch[trace_index][trace_length, :])
target_data_t0 = np.asarray(target_data_t0)
target_data_t1 = []
for trace_index in range(0, len(trace_t1_batch)):
trace_length = trace_t1_batch[trace_index]
trace_length = trace_length - 1
target_data_t1.append(player_id_t1_batch[trace_index][trace_length, :])
target_data_t1 = np.asarray(target_data_t1)
else:
target_data_t0 = np.asarray(player_id_t0_batch)
target_data_t1 = np.asarray(player_id_t1_batch)
if config.Learn.predict_target == 'PlayerLocalId':
input_data_t0 = np.concatenate([np.asarray(player_id_t0_batch),
np.asarray(team_id_t0_batch),
np.asarray(s_t0_batch),
np.asarray(action_id_t0)], axis=axis_concat)
# target_data_t0 = np.asarray(np.asarray(player_id_t0_batch))
input_data_t1 = np.concatenate([np.asarray(player_id_t1_batch),
np.asarray(team_id_t1_batch),
np.asarray(s_t1_batch),
np.asarray(action_id_t1)], axis=axis_concat)
# target_data_t1 = np.asarray(np.asarray(player_id_t1_batch))
trace_length_t0 = np.asarray(trace_t0_batch)
trace_length_t1 = np.asarray(trace_t1_batch)
else:
input_data_t0 = np.concatenate([np.asarray(player_id_t0_batch), np.asarray(s_t0_batch),
np.asarray(action_id_t0)], axis=axis_concat)
# target_data_t0 = np.asarray(np.asarray(player_id_t0_batch))
input_data_t1 = np.concatenate([np.asarray(player_id_t1_batch), np.asarray(s_t1_batch),
np.asarray(action_id_t1)], axis=axis_concat)
# target_data_t1 = np.asarray(np.asarray(player_id_t1_batch))
trace_length_t0 = np.asarray(trace_t0_batch)
trace_length_t1 = np.asarray(trace_t1_batch)
if training_flag:
if config.Learn.apply_stochastic:
for i in range(len(input_data_t0)):
MemoryBuffer.push([input_data_t0[i], target_data_t0[i], trace_length_t0[i],
input_data_t1[i], target_data_t1[i], trace_length_t1[i],
r_t_batch[i], win_id_t_batch[i],
terminal_batch[i], cut_batch[i]
])
sampled_data = MemoryBuffer.sample(batch_size=config.Learn.batch_size)
sample_input_data_t0 = np.asarray([sampled_data[j][0] for j in range(len(sampled_data))])
sample_target_data_t0 = np.asarray([sampled_data[j][1] for j in range(len(sampled_data))])
sample_trace_length_t0 = np.asarray([sampled_data[j][2] for j in range(len(sampled_data))])
if training_flag:
train_mask = np.asarray([1] * len(sample_input_data_t0))
else:
train_mask = np.asarray([0] * len(sample_input_data_t0))
for i in range(0, len(terminal_batch)):
batch_terminal = terminal_batch[i]
batch_cut = cut_batch[i]
else:
sample_input_data_t0 = input_data_t0
sample_target_data_t0 = target_data_t0
sample_trace_length_t0 = trace_length_t0
likelihood_loss = train_encoder_model(model, sess, config, sample_input_data_t0,
sample_target_data_t0, sample_trace_length_t0,
train_mask)
if terminal or cut:
print('\n ll lost is {0}'.format(str(likelihood_loss)))
"""we skip sampling for TD learning"""
train_td_model(model, sess, config,
input_data_t0, target_data_t0, trace_length_t0,
input_data_t1, target_data_t1, trace_length_t1,
r_t_batch, train_mask, terminal, cut)
train_score_diff(model, sess, config,
input_data_t0, target_data_t0, trace_length_t0,
input_data_t1, target_data_t1, trace_length_t1,
r_t_batch, score_diff_base_t0, outcome_data,
train_mask, terminal, cut)
if add_pred_flag:
sampled_data = Prediction_MemoryBuffer.sample(batch_size=config.Learn.batch_size)
sample_pred_input_data = np.asarray([sampled_data[j][0] for j in range(len(sampled_data))])
sample_pred_target_data = np.asarray([sampled_data[j][1] for j in range(len(sampled_data))])
sample_pred_trace_lengths = np.asarray([sampled_data[j][2] for j in range(len(sampled_data))])
# sample_pred_train_mask = np.asarray([sampled_data[j][2] for j in range(len(sampled_data))])
train_prediction(model, sess, config,
sample_pred_input_data,
sample_pred_target_data,
sample_pred_trace_lengths)
else:
# for i in range(0, len(r_t_batch)):
# if i == len(r_t_batch) - 1:
# if terminal or cut:
# print(r_t_batch[i])
if validate_cvrnn_flag:
output_decoder = encoder_validation(sess, model, input_data_t0, target_data_t0,
trace_length_t0, train_mask, config)
if output_decoder_all is None:
output_decoder_all = output_decoder
target_data_all = target_data_t0
else:
# try:
output_decoder_all = np.concatenate([output_decoder_all, output_decoder], axis=0)
target_data_all = np.concatenate([target_data_all, target_data_t0], axis=0)
if validate_td_flag:
# validate_variance_flag = validate_variance_flag if train_number <= 500 else False
q_values = td_validation(sess, model, input_data_t0, target_data_t0, trace_length_t0, train_mask,
config)
if q_values_all is None:
q_values_all = q_values
else:
q_values_all = np.concatenate([q_values_all, q_values], axis=0)
if validate_diff_flag:
output_label, real_label = diff_validation(sess, model, input_data_t0, target_data_t0,
trace_length_t0, train_mask,
score_diff_base_t0,
config, outcome_data)
if real_label_all is None:
real_label_all = real_label
else:
real_label_all = np.concatenate([real_label_all, real_label], axis=0)
if output_label_all is None:
output_label_all = output_label
else:
output_label_all = np.concatenate([output_label_all, output_label, ], axis=0)
s_t0 = s_tl
if terminal:
break
if validate_predict_flag and add_pred_flag:
pred_target_data, pred_output_prob = validate_prediction(model, sess, config,
pred_input_data,
pred_target_data,
pred_trace_lengths)
if pred_target_data_all is None:
pred_target_data_all = pred_target_data
else:
pred_target_data_all = np.concatenate([pred_target_data_all, pred_target_data], axis=0)
if pred_output_prob_all is None:
pred_output_prob_all = pred_output_prob
else:
pred_output_prob_all = np.concatenate([pred_output_prob_all, pred_output_prob], axis=0)
return [output_decoder_all, target_data_all,
q_values_all, real_label_all, output_label_all, pred_target_data_all, pred_output_prob_all]