def get_batches(self, batch_size, current_epoch,
random_sample=True, is_training=True):
dataset_path = os.path.join(self.dataset_path, '%03d' % current_epoch)
data = self.read_data(dataset_path)
for i in range(self.iterations_per_epoch):
batch = recDotDefaultDict()
batch.is_training = is_training
if random_sample:
replays = random.sample(data, batch_size)
for x in replays:
batching_dicts(batch, x)
yield batch
def create_test_batch(roles, picks, bans, vocab):
batch = recDotDefaultDict()
example = create_example(roles, picks, bans, vocab)
batch = batching_dicts(
batch, example) # list of dictionaries to dictionary of lists.
batch = padding(batch)
return batch
def tensorize(self, data):
batch = recDotDefaultDict()
for d in data:
batch = batching_dicts(
batch, d) # list of dictionaries to dictionary of lists.
batch = self.padding(batch)
return batch
def add_replay(self, fpath):
log = self.read_log(fpath)
if not log:
return
fkey, p1log, p2log = log
p1log_tensors = recDotDefaultDict()
for d in p1log:
batching_dicts(p1log_tensors, d)
p2log_tensors = recDotDefaultDict()
for d in p2log:
batching_dicts(p2log_tensors, d)
# Propagate rewards from the last state for N-step TD.
# T = len(p1log)
# for t in range(T):
# p1log[t].reward = (self.td_gamma ** (T - t - 1)) * p1log[-1].reward
# p2log[t].reward = (self.td_gamma ** (T - t - 1)) * p2log[-1].reward
data = [p1log_tensors, p2log_tensors]
return data