def model_fn_for_npair_loss(features, labels, mode, params):
# TARGET_SIZE = 512 #
TARGET_SIZE = 128 #
learning_rate = 0.01
loss, train_op = None, None
predictions = None
training_hooks = []
if mode != tf.estimator.ModeKeys.PREDICT:
anc_xs = features['prod_img']
pos_xs = features['snap_img']
ys1 = create_network3(anc_xs, TARGET_SIZE)
ys2 = create_network3(pos_xs, TARGET_SIZE, reuse=True)
loss = metric_learning.npairs_loss(labels, ys1, ys2)
optimizer = tf.train.AdamOptimizer(learning_rate)
if params["use_tpu"]:
# tpu の場合、optimizer を以下のようにラップしてやる必要がある.
optimizer = tf.contrib.tpu.CrossShardOptimizer(optimizer)
train_op = optimizer.minimize(loss,
global_step=tf.train.get_global_step())
# tf.train.Saver() の生成は graph 生成後でないとエラーになるのでここで設定している.
training_hooks = create_hooks(loss, params)
spec = tf.contrib.tpu.TPUEstimatorSpec(mode=mode,
training_hooks=training_hooks,
predictions=predictions,
loss=loss,
train_op=train_op)
return spec
def model_fn_for_npair_loss(features, labels, mode, params):
TARGET_SIZE = 3 # 教師画像の種類数, 0 - 9 までの 10個.
learning_rate = 0.01
loss, train_op = None, None
predictions = None
training_hooks = []
if mode != tf.estimator.ModeKeys.PREDICT:
anc_xs = features['anc_image']
pos_xs = features['pos_image']
ys1 = create_network2(anc_xs, TARGET_SIZE)
ys2 = create_network2(pos_xs, TARGET_SIZE, reuse=True)
loss = metric_learning.npairs_loss(labels, ys1, ys2)
train_op = tf.train.AdamOptimizer(learning_rate).minimize(
loss, global_step=tf.train.get_global_step())
# tf.train.Saver() の生成は graph 生成後でないとエラーになるのでここで設定している.
training_hooks = create_hooks(loss, params)
spec = tf.estimator.EstimatorSpec(mode=mode,
training_hooks=training_hooks,
predictions=predictions,
loss=loss,
train_op=train_op)
return spec
def model_fn_by_resnet(features, labels, mode, params):
if mode != tf.estimator.ModeKeys.PREDICT:
learning_rate = 0.01
anc_xs = features['prod_img']
pos_xs = features['snap_img']
TARGET_SIZE = 128 #
ys1 = create_network_resnet(anc_xs, TARGET_SIZE)
ys2 = create_network_resnet(pos_xs, TARGET_SIZE, reuse=True)
loss = metric_learning.npairs_loss(labels, ys1, ys2)
optimizer = tf.train.AdamOptimizer(learning_rate)
train_op = optimizer.minimize(loss,
global_step=tf.train.get_global_step())
spec = tf.contrib.tpu.TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
return spec
def train():
random_seed_set(1) # seed=1 はたまたまうまくいったので使っている.
TARGET_SIZE = 3 # 教師画像の種類数, 0 - 9 までの 10個.
BATCH_SIZE = 20 # バッチサイズ
MAX_STEPS = 1000 # 学習回数
NUM_THREADS = 4 # スレッド
INPUT_TFRECORD_TRAIN = "npair_train.tfrecord" # TFRecordファイル名（学習用）
INPUT_TFRECORD_TEST = "npair_test.tfrecord" # TFRecordファイル名（評価用）
filenames = tf.placeholder(tf.string, shape=[None])
dataset = tf.data.TFRecordDataset(filenames) # ファイル名を遅延評価するパイプを作成.
dataset = dataset.map(read_function,
NUM_THREADS) # ファイル名からデータを作成する遅延評価するパイプを作成.
dataset = dataset.shuffle(60000)
dataset = dataset.batch(BATCH_SIZE) # 返すレコードの個数を指定.
dataset = dataset.repeat(-1) # 無限に繰り返す設定にする.
# 型だけを設定した Iterator を用意する.
iterator = tf.data.Iterator.from_structure(dataset.output_types,
dataset.output_shapes)
iter_init_op = iterator.make_initializer(dataset) # イテレータを初期化するオペレータ.
anc_xs, pos_xs, ys_ = iterator.get_next() # 遅延評価されるして返される要素.
ys1 = create_network2(anc_xs, TARGET_SIZE)
ys2 = create_network2(pos_xs, TARGET_SIZE, reuse=True)
loss = metric_learning.npairs_loss(ys_, ys1, ys2)
train_op = tf.train.AdamOptimizer(0.01).minimize(loss)
start_time = time.time()
with tf.Session() as sess:
sess.run(iter_init_op, feed_dict={filenames: [INPUT_TFRECORD_TRAIN]})
init = tf.global_variables_initializer()
sess.run(init) # 変数の初期化処理
for step in range(MAX_STEPS):
#print(sess.run(ys_))
_, loss_v = sess.run([train_op, loss]) # 最急勾配法でパラメータ更新
duration = time.time() - start_time
if (step + 1) % 100 == 0:
print("step={:4d}, loss={:5.2f}, time=({:.3f} sec)".format(
step + 1, loss_v, duration))
return