 TensorFlow.org で表示 |
 Google Colab で実行 |
 GitHub でソースを表示 |
 ノートブックをダウンロード |
概要
トレーニング中にモデルを保存して読み込むことは一般的な作業です。Keras モデルの保存と読み込みに使用する API には、高レベル API と低レベル API の 2 つがあります。このチュートリアルでは、tf.distribute.Strategy を使用してる場合に SavedModel API を使用する方法を実演しています。SavedModel とシリアル化に関する一般的な情報は、SavedModel ガイドとKeras モデルのシリアル化ガイドをご覧ください。では、簡単な例から始めましょう。
依存関係をインポートします。
import tensorflow_datasets as tfds
import tensorflow as tf
tfds.disable_progress_bar()
tf.distribute.Strategy を使ってデータとモデルを準備します。
mirrored_strategy = tf.distribute.MirroredStrategy()
def get_data():
datasets, ds_info = tfds.load(name='mnist', with_info=True, as_supervised=True)
mnist_train, mnist_test = datasets['train'], datasets['test']
BUFFER_SIZE = 10000
BATCH_SIZE_PER_REPLICA = 64
BATCH_SIZE = BATCH_SIZE_PER_REPLICA * mirrored_strategy.num_replicas_in_sync
def scale(image, label):
image = tf.cast(image, tf.float32)
image /= 255
return image, label
train_dataset = mnist_train.map(scale).cache().shuffle(BUFFER_SIZE).batch(BATCH_SIZE)
eval_dataset = mnist_test.map(scale).batch(BATCH_SIZE)
return train_dataset, eval_dataset
def get_model():
with mirrored_strategy.scope():
model = tf.keras.Sequential([
tf.keras.layers.Conv2D(32, 3, activation='relu', input_shape=(28, 28, 1)),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(64, activation='relu'),
tf.keras.layers.Dense(10)
])
model.compile(loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer=tf.keras.optimizers.Adam(),
metrics=['accuracy'])
return model
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
モデルをトレーニングします。
model = get_model()
train_dataset, eval_dataset = get_data()
model.fit(train_dataset, epochs=2)
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
Epoch 1/2
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
938/938 [==============================] - 10s 4ms/step - loss: 0.4205 - accuracy: 0.8744
Epoch 2/2
938/938 [==============================] - 3s 3ms/step - loss: 0.0789 - accuracy: 0.9769
<tensorflow.python.keras.callbacks.History at 0x7fa340042828>
モデルを保存して読み込む
作業に使用する単純なモデルの準備ができたので、保存と読み込みの API を見てみましょう。利用できる API には次の 2 種類があります。
- 高レベルの Keras
model.saveとtf.keras.models.load_model - 低レベルの
tf.saved_model.saveとtf.saved_model.load
Keras API
次に、Keras API を使ってモデルを保存して読み込む例を示します。
keras_model_path = "/tmp/keras_save"
model.save(keras_model_path)
INFO:tensorflow:Assets written to: /tmp/keras_save/assets INFO:tensorflow:Assets written to: /tmp/keras_save/assets
tf.distribute.Strategy を使用せずにモデルを復元します。
restored_keras_model = tf.keras.models.load_model(keras_model_path)
restored_keras_model.fit(train_dataset, epochs=2)
Epoch 1/2 938/938 [==============================] - 3s 2ms/step - loss: 0.0510 - accuracy: 0.9843 Epoch 2/2 938/938 [==============================] - 2s 2ms/step - loss: 0.0355 - accuracy: 0.9893 <tensorflow.python.keras.callbacks.History at 0x7fa3c3374fd0>
モデルを復元したら、compile() をもう一度呼び出すことなくそのトレーニングを続行できます。保存前にコンパイル済みであるからです。モデルは、TensorFlow の標準的な SavedModel プロと形式で保存されています。その他の詳細は、saved_model 形式ガイドをご覧ください。
tf.distribute.Strategy を使用して、モデルを読み込んでトレーニングします。
another_strategy = tf.distribute.OneDeviceStrategy("/cpu:0")
with another_strategy.scope():
restored_keras_model_ds = tf.keras.models.load_model(keras_model_path)
restored_keras_model_ds.fit(train_dataset, epochs=2)
Epoch 1/2 938/938 [==============================] - 9s 9ms/step - loss: 0.0503 - accuracy: 0.9845 Epoch 2/2 938/938 [==============================] - 8s 9ms/step - loss: 0.0355 - accuracy: 0.9892
ご覧の通り、tf.distribute.Strategy を使って期待通りに読み込まれました。ここで使用されるストラテジーは、保存前に使用したストラテジーと同じものである必要はありません。
tf.saved_model API
では、低レベル API を見てみましょう。モデルの保存は Keras API に類似しています。
model = get_model() # get a fresh model
saved_model_path = "/tmp/tf_save"
tf.saved_model.save(model, saved_model_path)
INFO:tensorflow:Assets written to: /tmp/tf_save/assets INFO:tensorflow:Assets written to: /tmp/tf_save/assets
読み込みは tf.saved_model.load() で行えますが、より低いレベルにある API（したがって広範なユースケースのある API）であるため、Keras モデルを返しません。代わりに、推論を行うために使用できる関数を含むオブジェクトを返します。次に例を示します。
DEFAULT_FUNCTION_KEY = "serving_default"
loaded = tf.saved_model.load(saved_model_path)
inference_func = loaded.signatures[DEFAULT_FUNCTION_KEY]
読み込まれたオブジェクトには複数の関数が含まれ、それぞれにキーが関連付けられている可能性があります。"serving_default" は、保存された Keras モデルを使用した推論関数のデフォルトのキーです。この関数で推論を実行するには、次のようにします。
predict_dataset = eval_dataset.map(lambda image, label: image)
for batch in predict_dataset.take(1):
print(inference_func(batch))
{'dense_3': <tf.Tensor: shape=(64, 10), dtype=float32, numpy=
array([[-2.80148745e-01, 1.33514032e-01, -2.87131630e-02,
7.16523156e-02, 5.96208051e-02, -1.41107485e-01,
-1.91312388e-01, -2.37503886e-01, -3.35927419e-02,
-5.19809276e-02],
[-2.36288786e-01, 5.03980108e-02, -1.62936170e-02,
-4.73961532e-02, 6.72176480e-02, 1.00629888e-01,
-7.40617514e-02, -8.68952796e-02, -4.73142453e-02,
-2.62654185e-01],
[-1.34671092e-01, 2.75589246e-02, -3.82595509e-02,
8.22241753e-02, 2.00248569e-01, 5.53470105e-04,
-6.37755543e-02, -1.81547310e-02, 5.63463122e-02,
-2.53453195e-01],
[-3.00725941e-02, 1.14580847e-01, -1.36942551e-01,
1.96914107e-01, -7.57494345e-02, 3.88330817e-02,
-4.73932028e-02, -9.81834531e-02, 5.53093925e-02,
6.25574738e-02],
[-6.90577328e-02, 4.01481241e-02, 7.58580863e-05,
8.74225870e-02, 1.08554110e-01, -4.93706316e-02,
-7.11562335e-02, -1.20140880e-01, 1.04657030e-02,
-8.07161778e-02],
[-1.52959123e-01, 2.04001382e-01, -6.20923229e-02,
1.99777156e-01, 1.34099647e-01, 6.88168854e-02,
-3.72684970e-02, -1.52993947e-01, -1.06225178e-01,
-6.87185004e-02],
[-9.58189666e-02, 1.03578016e-01, 5.63422069e-02,
1.02696434e-01, 7.79873282e-02, 4.63175029e-02,
-2.90759206e-02, -2.39548281e-01, -6.80998862e-02,
-1.75888419e-01],
[-1.15482561e-01, 1.87766463e-01, -1.92467883e-01,
1.34843200e-01, 2.81673819e-02, 1.05498143e-01,
4.87741642e-02, -1.11555547e-01, -3.73479575e-02,
9.21436101e-02],
[-2.25425810e-01, 1.47895426e-01, 1.94478929e-02,
1.77668944e-01, 1.70605719e-01, 1.46347091e-01,
-5.64843491e-02, -2.27678254e-01, -1.48098990e-01,
-2.97955386e-02],
[-1.05673417e-01, -8.63498822e-03, -1.94058135e-01,
9.95621681e-02, -3.76955494e-02, 1.36648178e-01,
-2.47661006e-02, 1.94414444e-02, -5.16721681e-02,
2.19522677e-02],
[-2.61032432e-02, 1.65352836e-01, -1.02065675e-01,
2.87719220e-01, -5.69451880e-03, 8.70772228e-02,
-2.04757065e-01, -1.68713421e-01, 4.19157147e-02,
-1.29969329e-01],
[-1.95437104e-01, 1.99485153e-01, -3.33056971e-02,
2.20946252e-01, 1.19004212e-01, 1.11005664e-01,
-9.24758017e-02, -8.89133066e-02, -8.90364498e-02,
-8.05964768e-02],
[-4.46876436e-02, 6.00830466e-02, -1.53938048e-02,
1.52039722e-01, 9.05125514e-02, 4.67565246e-02,
-1.09522790e-03, -2.59150207e-01, -5.84345236e-02,
6.93883970e-02],
[-8.20917487e-02, 4.75821830e-02, -6.70246705e-02,
2.01006785e-01, 2.47687325e-02, 1.15976043e-01,
-1.12001128e-01, -1.99499398e-01, -2.36950591e-02,
-5.70925698e-02],
[-1.84315890e-01, 1.87751845e-01, -1.78156570e-02,
1.75649762e-01, 3.24658900e-02, 2.98027880e-03,
-1.06299669e-01, -5.05354106e-02, 1.73205137e-03,
-2.45052539e-02],
[-1.45228893e-01, 1.31983459e-01, -2.78562456e-02,
1.78783461e-01, 1.19794048e-02, -6.04133755e-02,
-2.23527014e-01, -4.63448241e-02, 6.95200786e-02,
-1.87706128e-01],
[-1.04228586e-01, 2.13943012e-02, -1.34355389e-02,
1.70563966e-01, -8.66526365e-03, -1.23366416e-02,
-1.44198984e-01, -6.58028498e-02, 5.76171651e-02,
6.96113054e-03],
[-1.25256106e-02, 1.78468786e-02, -6.78840131e-02,
1.37728944e-01, 5.45562617e-02, 1.26407653e-01,
-4.08719778e-02, -1.13486752e-01, -1.54493749e-03,
-5.20698875e-02],
[ 4.62122858e-02, 1.31210506e-01, -2.38493439e-02,
1.17228568e-01, -2.61356831e-02, 1.13679111e-01,
-1.19934157e-01, -3.48669030e-02, -3.59272584e-02,
-2.71111056e-02],
[-2.17378423e-01, 1.83873564e-01, -5.18436730e-02,
2.37078965e-01, 4.16856520e-02, 4.85838056e-02,
-1.38593808e-01, -6.06887136e-03, -1.43550113e-02,
-7.06137270e-02],
[-1.57423168e-01, 9.36834812e-02, 8.97897035e-02,
1.21637769e-01, 1.31796181e-01, -1.65549647e-02,
-6.42024204e-02, -5.14324754e-03, -4.08546552e-02,
-1.05977021e-01],
[-1.48703009e-02, -2.58602649e-02, -1.69074256e-02,
3.98873910e-02, 9.76919457e-02, 5.71911037e-03,
-3.73653397e-02, -8.49235132e-02, 6.72312453e-02,
-1.33198313e-02],
[-6.34206086e-02, 1.51487142e-02, -1.02290526e-01,
1.40237004e-01, 6.86008409e-02, 1.94813907e-01,
8.79306868e-02, -2.19192445e-01, -9.22468454e-02,
-4.28461432e-02],
[-6.74271733e-02, 1.97617367e-01, 1.98064953e-01,
2.82037742e-02, 1.44770145e-02, -6.68774731e-03,
-9.42453146e-02, -1.13627508e-01, -9.45969895e-02,
-1.99111566e-01],
[-1.30129820e-02, -8.39958489e-02, -4.42412421e-02,
1.34765178e-01, 6.02031462e-02, 1.96596861e-01,
-8.18365067e-02, -1.45619735e-01, -1.62957266e-01,
1.95069984e-03],
[-8.67712125e-02, 3.25389206e-03, -1.14180908e-01,
1.31985873e-01, 2.01569423e-01, 1.31158561e-01,
2.54417323e-02, -2.45438382e-01, -4.70624864e-02,
-1.40900053e-02],
[-8.00902992e-02, 3.45779657e-02, -1.50647402e-01,
1.35409802e-01, -6.10385723e-02, 9.71267074e-02,
-5.65597489e-02, -1.04390785e-01, 2.37799883e-02,
-1.82536673e-02],
[-9.35142711e-02, 2.52702534e-01, -3.18659022e-02,
1.71243101e-01, 9.54027772e-02, -3.82800400e-03,
-1.23053603e-01, 1.48397125e-02, -1.73588656e-02,
-1.66330785e-01],
[-1.02943242e-01, 7.12655038e-02, 3.56826186e-03,
5.46290874e-02, -6.33837134e-02, 3.07019185e-02,
-2.03796327e-02, 1.32264495e-02, -5.46035208e-02,
-3.71890403e-02],
[-2.06419140e-01, 7.59046897e-02, 6.42012581e-02,
1.59068286e-01, 1.42954856e-01, 8.00145715e-02,
-7.57730007e-02, -1.48105875e-01, 6.09865040e-02,
-9.88312662e-02],
[-8.47662836e-02, 1.88201651e-01, -1.71128549e-02,
2.62053698e-01, 1.46339133e-01, -1.07536577e-01,
-1.78948760e-01, -1.54565811e-01, 4.36789021e-02,
-1.95843250e-01],
[ 1.27585530e-02, 8.68005827e-02, -1.02946900e-01,
1.94241941e-01, 1.64123014e-01, 1.97950214e-01,
4.28899303e-02, -1.68846846e-01, -9.03922990e-02,
3.18114907e-02],
[-1.08418033e-01, 8.62024128e-02, 6.11752942e-02,
1.10205211e-01, 1.96092814e-01, 1.94776863e-01,
-4.78504933e-02, -1.81526337e-02, -1.78298637e-01,
1.71261281e-03],
[ 2.66609285e-02, -7.09919445e-03, -5.94112761e-02,
6.90286160e-02, -3.69608551e-02, 1.24270767e-02,
-9.61745158e-02, -3.92623395e-02, 1.74935497e-02,
-6.76374137e-03],
[-6.61516711e-02, 1.35748640e-01, -4.90215495e-02,
1.68750688e-01, 5.45104742e-02, 8.80523920e-02,
-9.94434953e-02, -2.45334193e-01, -1.74186528e-01,
-1.34910107e-01],
[-1.28005132e-01, 1.31362900e-01, 7.90357366e-02,
1.23638391e-01, 5.19953631e-02, -2.46395007e-01,
-2.11327657e-01, 2.35061012e-02, 1.04565367e-01,
-1.46918595e-01],
[-2.01322585e-01, 7.69232735e-02, -9.20683444e-02,
1.00531250e-01, 2.19418138e-01, 1.92265585e-02,
-1.49796382e-01, -1.71407014e-01, -1.24774791e-01,
-1.32502884e-01],
[-1.08242273e-01, -3.59819531e-02, 7.11274445e-02,
9.92041901e-02, -1.46615505e-02, 5.15304357e-02,
-8.56937319e-02, -6.29515573e-02, 5.88932820e-02,
-1.80848707e-02],
[-1.14341632e-01, 5.85459657e-02, -4.80359755e-02,
7.93043151e-02, 2.74222828e-02, 2.41690613e-02,
-3.36848386e-02, 1.59112103e-02, 4.32283431e-03,
-8.20446610e-02],
[-2.34901775e-02, 1.00804538e-01, 2.91342884e-02,
1.78647369e-01, -1.14274524e-01, 3.73394713e-02,
-8.32463875e-02, -8.76808316e-02, 7.43689612e-02,
-3.84302288e-02],
[-9.46902260e-02, 1.83316410e-01, -5.19321933e-02,
1.19176000e-01, 9.62891579e-02, 1.64513245e-01,
-1.04709119e-02, -5.51389456e-02, -1.17363684e-01,
-7.77934268e-02],
[-3.82615477e-02, -3.67629528e-03, -4.03916314e-02,
1.12081997e-01, 1.19140416e-01, 5.62198013e-02,
-4.06569764e-02, -1.84003294e-01, -2.34463289e-02,
-3.09241600e-02],
[-1.21338770e-01, 9.93697941e-02, -2.27549188e-02,
1.63415670e-01, 9.50213596e-02, 8.96587819e-02,
-3.41728330e-02, -5.34512773e-02, -4.43711840e-02,
-8.76995735e-03],
[-1.16379865e-01, 5.37211001e-02, -4.90345992e-02,
1.44882053e-01, 2.42444009e-01, -3.14327031e-02,
-7.67233521e-02, -6.67005181e-02, -2.68945061e-02,
-5.56297600e-02],
[-9.97139513e-02, 7.09423870e-02, 1.01991177e-01,
1.43535435e-01, 1.15161538e-01, -1.60139859e-01,
-1.66834265e-01, -3.67481560e-02, 5.15153185e-02,
-1.02712616e-01],
[-8.57860744e-02, -1.12610072e-01, -7.90568590e-02,
1.05949417e-01, 2.50540107e-01, 1.68823555e-01,
-2.17120834e-02, -1.22933879e-01, -7.00456277e-02,
-5.86306602e-02],
[-1.47669703e-01, -3.68478559e-02, -5.91939762e-02,
1.18016057e-01, 1.15264602e-01, 8.11087936e-02,
-8.63386020e-02, -2.58504689e-01, 8.07600096e-03,
-7.26428553e-02],
[-1.49649456e-01, 5.63042201e-02, 3.82883176e-02,
1.14912771e-01, 9.47139189e-02, 4.54579964e-02,
-6.99455440e-02, -1.71200782e-01, -2.88073197e-02,
-2.11609229e-01],
[-5.54483011e-02, 1.00146934e-01, -9.60576907e-02,
1.83109671e-01, 7.37194493e-02, 1.91168159e-01,
-6.74408302e-02, -1.50882870e-01, -6.12113550e-02,
-1.33752763e-01],
[-9.04634148e-02, -3.53388935e-02, -1.22361317e-01,
8.95344317e-02, 4.54211123e-02, 1.52289450e-01,
-1.46961687e-02, 3.73999104e-02, -1.19142778e-01,
9.15640667e-02],
[-1.52703285e-01, 3.30357015e-01, 3.76335606e-02,
2.96404436e-02, 1.62578374e-01, 7.83460289e-02,
5.31926751e-04, -5.24859875e-02, -1.86474130e-01,
-2.14560941e-01],
[-1.26740202e-01, -6.25290424e-02, -5.49526513e-02,
1.13773897e-01, 7.84891397e-02, 5.27034029e-02,
-8.14578682e-03, -2.37367094e-01, -3.00243702e-02,
2.55652480e-02],
[-1.03566192e-01, 2.99373325e-02, -1.74053665e-02,
1.47834927e-01, 5.51711097e-02, 1.39649704e-01,
-8.15002024e-02, -1.37584135e-01, 1.32908411e-02,
-5.03157005e-02],
[-2.28102490e-01, 1.54016480e-01, 9.73745361e-02,
9.99387503e-02, -3.97506319e-02, -2.74454225e-02,
-6.78737313e-02, -6.45283610e-02, -1.02310404e-02,
-4.17116582e-02],
[-2.68130243e-01, 1.38102412e-01, 6.27536625e-02,
1.75527334e-01, 7.54192397e-02, -7.78935030e-02,
-1.61849275e-01, -1.18933164e-01, 1.40640512e-03,
-8.56635794e-02],
[-9.10969675e-02, 1.76204905e-01, -3.20403352e-02,
1.16631843e-01, 1.41420662e-01, 8.71218741e-04,
-2.10343331e-01, -4.64888737e-02, -3.82080674e-03,
-2.35398933e-01],
[-8.60667378e-02, 1.90748304e-01, -1.00961223e-01,
7.83518553e-02, 1.05100065e-01, 9.62026268e-02,
-3.96616012e-02, -1.07762277e-01, -8.43076408e-02,
-1.62416585e-02],
[-1.39651194e-01, 8.80849212e-02, -9.33513194e-02,
9.95803922e-02, 1.43332034e-03, 2.32762739e-01,
3.05913612e-02, -5.54927550e-02, -8.69316310e-02,
3.88408825e-02],
[-3.37045580e-01, 1.41151592e-01, -4.73942608e-03,
1.36001334e-01, 8.12572762e-02, 5.91648370e-02,
-7.05533326e-02, -2.73337178e-02, -2.79111490e-02,
-2.35716682e-02],
[-1.60612002e-01, 8.30742866e-02, -8.68474022e-02,
-4.64118645e-02, -2.08184980e-02, -3.72342840e-02,
-1.16365984e-01, 1.41748518e-01, 1.16192520e-01,
-1.86156020e-01],
[-5.42840548e-02, 2.05266312e-01, -4.47835587e-02,
1.47917837e-01, 7.08672255e-02, -5.21331169e-02,
-1.61334664e-01, -1.19976789e-01, 1.44696319e-02,
-1.21052071e-01],
[-1.74211279e-01, 1.90779641e-01, 8.71634856e-03,
1.31276980e-01, 4.35346104e-02, -8.93380493e-02,
-1.15072154e-01, -9.72218290e-02, 3.98243703e-02,
-1.02335595e-01],
[-7.44505227e-02, -8.93713087e-02, 6.19246177e-02,
1.36901751e-01, 4.21357229e-02, 2.36863568e-02,
-6.85738474e-02, -1.94917679e-01, 2.22067088e-02,
-7.60869086e-02],
[-2.96473689e-02, 6.17601499e-02, -5.22884279e-02,
1.47381663e-01, 1.22329667e-01, 6.64027259e-02,
-1.27000526e-01, 7.84885287e-02, 3.67109478e-03,
1.11098588e-03]], dtype=float32)>}
また、分散方法で読み込んで推論を実行することもできます。
another_strategy = tf.distribute.MirroredStrategy()
with another_strategy.scope():
loaded = tf.saved_model.load(saved_model_path)
inference_func = loaded.signatures[DEFAULT_FUNCTION_KEY]
dist_predict_dataset = another_strategy.experimental_distribute_dataset(
predict_dataset)
# Calling the function in a distributed manner
for batch in dist_predict_dataset:
another_strategy.run(inference_func,args=(batch,))
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
WARNING:tensorflow:Using MirroredStrategy eagerly has significant overhead currently. We will be working on improving this in the future, but for now please wrap `call_for_each_replica` or `experimental_run` or `run` inside a tf.function to get the best performance.
WARNING:tensorflow:Using MirroredStrategy eagerly has significant overhead currently. We will be working on improving this in the future, but for now please wrap `call_for_each_replica` or `experimental_run` or `run` inside a tf.function to get the best performance.
WARNING:tensorflow:Using MirroredStrategy eagerly has significant overhead currently. We will be working on improving this in the future, but for now please wrap `call_for_each_replica` or `experimental_run` or `run` inside a tf.function to get the best performance.
WARNING:tensorflow:Using MirroredStrategy eagerly has significant overhead currently. We will be working on improving this in the future, but for now please wrap `call_for_each_replica` or `experimental_run` or `run` inside a tf.function to get the best performance.
WARNING:tensorflow:Using MirroredStrategy eagerly has significant overhead currently. We will be working on improving this in the future, but for now please wrap `call_for_each_replica` or `experimental_run` or `run` inside a tf.function to get the best performance.
WARNING:tensorflow:Using MirroredStrategy eagerly has significant overhead currently. We will be working on improving this in the future, but for now please wrap `call_for_each_replica` or `experimental_run` or `run` inside a tf.function to get the best performance.
WARNING:tensorflow:Using MirroredStrategy eagerly has significant overhead currently. We will be working on improving this in the future, but for now please wrap `call_for_each_replica` or `experimental_run` or `run` inside a tf.function to get the best performance.
WARNING:tensorflow:Using MirroredStrategy eagerly has significant overhead currently. We will be working on improving this in the future, but for now please wrap `call_for_each_replica` or `experimental_run` or `run` inside a tf.function to get the best performance.
WARNING:tensorflow:Using MirroredStrategy eagerly has significant overhead currently. We will be working on improving this in the future, but for now please wrap `call_for_each_replica` or `experimental_run` or `run` inside a tf.function to get the best performance.
WARNING:tensorflow:Using MirroredStrategy eagerly has significant overhead currently. We will be working on improving this in the future, but for now please wrap `call_for_each_replica` or `experimental_run` or `run` inside a tf.function to get the best performance.
復元された関数の呼び出しは、保存されたモデル（predict）に対するフォワードパスです。読み込まれた関数をトレーニングし続ける場合はどうでしょうか。または読み込まれた関数をより大きなモデルに埋め込むには？一般的には、この読み込まれたオブジェクトを Keras レイヤーにラップして行うことができます。幸いにも、TF Hub には、次に示すとおり、この目的に使用できる hub.KerasLayer が用意されています。
import tensorflow_hub as hub
def build_model(loaded):
x = tf.keras.layers.Input(shape=(28, 28, 1), name='input_x')
# Wrap what's loaded to a KerasLayer
keras_layer = hub.KerasLayer(loaded, trainable=True)(x)
model = tf.keras.Model(x, keras_layer)
return model
another_strategy = tf.distribute.MirroredStrategy()
with another_strategy.scope():
loaded = tf.saved_model.load(saved_model_path)
model = build_model(loaded)
model.compile(loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer=tf.keras.optimizers.Adam(),
metrics=['accuracy'])
model.fit(train_dataset, epochs=2)
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
Epoch 1/2
938/938 [==============================] - 6s 3ms/step - loss: 0.4257 - accuracy: 0.8715
Epoch 2/2
938/938 [==============================] - 3s 3ms/step - loss: 0.0818 - accuracy: 0.9762
ご覧の通り、hub.KerasLayer は tf.saved_model.load() から読み込まれた結果を、別のモデルの構築に使用できる Keras レイヤーにラップしています。学習を転送する場合に非常に便利な手法です。
どの API を使用すべきですか？
保存の場合は、Keras モデルを使用しているのであれば、Keras の model.save() API をほぼ必ず使用することが推奨されます。保存しているものが Keras モデルでなければ、低レベル API しか使用できません。
読み込みの場合は、読み込み API から得ようとしているものによって選択肢がきまs理ます。Keras モデルを使用できない（または使用を希望しない）のであれば、tf.saved_model.load() を使用し、そうでなければ、tf.keras.models.load_model() を使用します。Keras モデルを保存した場合にのみ Keras モデルを読み込めることに注意してください。
API を混在させることも可能です。model.save で Keras モデルを保存し、低レベルの tf.saved_model.load API を使用して、非 Keras モデルを読み込むことができます。
model = get_model()
# Saving the model using Keras's save() API
model.save(keras_model_path)
another_strategy = tf.distribute.MirroredStrategy()
# Loading the model using lower level API
with another_strategy.scope():
loaded = tf.saved_model.load(keras_model_path)
INFO:tensorflow:Assets written to: /tmp/keras_save/assets
INFO:tensorflow:Assets written to: /tmp/keras_save/assets
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
ローカルデバイスで保存または読み込む
クラウド TPU を使用するなど、リモートで実行中にローカルの IO デバイスに保存したり、そこから読み込んだりする場合、experimental_io_device オプションを使用して、IO デバイスを localhost に設定する必要があります。
model = get_model()
# Saving the model to a path on localhost.
saved_model_path = "/tmp/tf_save"
save_options = tf.saved_model.SaveOptions(experimental_io_device='/job:localhost')
model.save(saved_model_path, options=save_options)
# Loading the model from a path on localhost.
another_strategy = tf.distribute.MirroredStrategy()
with another_strategy.scope():
load_options = tf.saved_model.LoadOptions(experimental_io_device='/job:localhost')
loaded = tf.keras.models.load_model(saved_model_path, options=load_options)
INFO:tensorflow:Assets written to: /tmp/tf_save/assets
INFO:tensorflow:Assets written to: /tmp/tf_save/assets
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0',)
警告
特殊なケースは、入力が十分に定義されていない Keras モデルがある場合です。たとえば、Seeuqntial モデルは、入力形状（Sequential([Dense(3), ...]）を使用せずに作成できます。Subclassed モデルにも、初期化後は十分に定義された入力がありません。この場合、保存と読み込みの両方に低レベル API を使用する必要があります。そうしない場合はエラーが発生します。
モデルの入力が十分に定義されたものであるかを確認するには、model.inputs が None であるかどうかを確認します。None でなければ問題ありません。入力形状は、モデルが .fit、.evaluate、.predict で使用されている場合、またはモデルを呼び出す場合（model(inputs)）に自動的に定義されます。
次に例を示します。
class SubclassedModel(tf.keras.Model):
output_name = 'output_layer'
def __init__(self):
super(SubclassedModel, self).__init__()
self._dense_layer = tf.keras.layers.Dense(
5, dtype=tf.dtypes.float32, name=self.output_name)
def call(self, inputs):
return self._dense_layer(inputs)
my_model = SubclassedModel()
# my_model.save(keras_model_path) # ERROR!
tf.saved_model.save(my_model, saved_model_path)
WARNING:tensorflow:Skipping full serialization of Keras layer <__main__.SubclassedModel object at 0x7fa3c2b3ea58>, because it is not built. WARNING:tensorflow:Skipping full serialization of Keras layer <__main__.SubclassedModel object at 0x7fa3c2b3ea58>, because it is not built. WARNING:tensorflow:Skipping full serialization of Keras layer <tensorflow.python.keras.layers.core.Dense object at 0x7fa3b1b572e8>, because it is not built. WARNING:tensorflow:Skipping full serialization of Keras layer <tensorflow.python.keras.layers.core.Dense object at 0x7fa3b1b572e8>, because it is not built. INFO:tensorflow:Assets written to: /tmp/tf_save/assets INFO:tensorflow:Assets written to: /tmp/tf_save/assets