class Trainer(object):
def __init__(self, config, debug=True):
self.config = config
self.debug = debug
self.logger = LogAPI.create_logger(self.__class__.__name__, self.debug)
self.data_processor = DataProcessor(config)
self.data_processor.prepare()
n_vocab = len(self.data_processor.vocab)
self.model = L.Classifier(Img2Seq(n_vocab, config))
self.model.compute_accuracy = False # I want loss, not accuracy
self.optimizer = chainer.optimizers.Adam()
self.optimizer.setup(self.model)
if self.config['use_gpu'] >= 0:
chainer.cuda.get_device(self.config['use_gpu']).use()
self.model.to_gpu()
self.xp = cuda.cupy
else:
self.xp = np
# パラメータ・実験結果を保存しておくdict
self.result_storage = {}
self.result_storage["result"] = {
"total_loss": [], "average_loss": [], "time_taken": [], "hyper_params": config}
def _calc_loss(self, batch):
boke, img = batch
boke = self.xp.asarray(boke, dtype=np.int32)
img = self.xp.asarray(img, dtype=np.float32)
# 1. ベクトル化してある画像をCNNに入れて特徴ベクトルにする
img_vec = self.model.predictor.encode_image(img)
# この時点で画像は入力(word embedding)と同じ次元
# もしかして: これLinkの外に出したほうがいいか？
# 3. bokeをデコードするように学習
accum_loss = 0
n = 0
for curr_words, next_words in zip(boke.T, boke[:, 1:].T):
if n == 0:
accum_loss += self.model(curr_words, img_vec, next_words)
else:
accum_loss += self.model(curr_words, next_words)
n += 1
return accum_loss
def run(self):
for epoch in xrange(self.config["total_epoch"]):
self.logger.info("Currently at Epoch #{}".format(epoch + 1))
self.epoch_start_time = datetime.now()
for batch in self.data_processor.batch_iter():
loss = self._calc_loss(batch)
self.optimizer.target.zerograds()
loss.backward()
self.optimizer.update()
self.epoch_end_time = datetime.now()
if (epoch + 1) % 3 == 0:
self._take_snapshot(epoch)
self.evaluate(epoch)
def evaluate(self, epoch):
self.logger.info("Processing Validation...")
accum_loss = 0
n_samples = 0
for batch in self.data_processor.batch_iter(kind="dev"):
accum_loss += cuda.to_cpu(self._calc_loss(batch).data)
n_samples += len(batch[0]) # batchsize
# epoch単位の実行時間の計算
td = self.epoch_end_time - self.epoch_start_time # td: timedelta
hours, remainder = divmod(td.seconds, 3600)
minutes, seconds = divmod(remainder, 60)
time_taken = "{hours:02d}:{minutes:02d}".format(
hours=hours, minutes=minutes)
total_loss = float(accum_loss)
average_loss = total_loss / n_samples
self.logger.info("Epoch: #{epoch}\tTotal Loss:{total:.4f}\tAverage Loss:{average:.4f}\tTime Taken:{time_taken}\n".format(
epoch=epoch,
total=total_loss,
average=average_loss,
time_taken=time_taken
))
# accuracy/total loss/average lossをそれぞれ格納
self.result_storage["result"]["total_loss"].append(total_loss)
self.result_storage["result"]["average_loss"].append(average_loss)
self.result_storage["result"]["time_taken"].append(time_taken)
# 格納したstatsをファイルに保存
self._save_stats()
def _take_snapshot(self, epoch):
"""
訓練したモデルをepochごとに保存していくメソッド
"""
path = "{}/{:02d}.npz".format(self.config["model_path"], epoch)
S.save_npz(path, self.model)
def _save_stats(self):
"""
stats.json (accuracy/average loss/total lossを保存するjson) を
epochごとに書き換えるメソッド
"""
with open(self.config["log_path"], "w") as result:
result.write(json.dumps(self.result_storage))
result.flush()
