133 lines
4.4 KiB
Python
133 lines
4.4 KiB
Python
# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# ==============================================================================
|
|
"""Public APIs for the HParams plugin.
|
|
|
|
This module supports a spectrum of use cases, depending on how much
|
|
structure you want. In the simplest case, you can simply collect your
|
|
hparams into a dict, and use a Keras callback to record them:
|
|
|
|
>>> from tensorboard.plugins.hparams import api as hp
|
|
>>> hparams = {
|
|
... "optimizer": "adam",
|
|
... "fc_dropout": 0.2,
|
|
... "neurons": 128,
|
|
... # ...
|
|
... }
|
|
>>>
|
|
>>> model = model_fn(hparams)
|
|
>>> callbacks = [
|
|
>>> tf.keras.callbacks.TensorBoard(logdir),
|
|
>>> hp.KerasCallback(logdir, hparams),
|
|
>>> ]
|
|
>>> model.fit(..., callbacks=callbacks)
|
|
|
|
The Keras callback requires that TensorFlow eager execution be enabled.
|
|
|
|
If not using Keras, use the `hparams` function to write the values
|
|
directly:
|
|
|
|
>>> # In eager mode:
|
|
>>> with tf.create_file_writer(logdir).as_default():
|
|
... hp.hparams(hparams)
|
|
>>>
|
|
>>> # In legacy graph mode:
|
|
>>> with tf.compat.v2.create_file_writer(logdir).as_default() as w:
|
|
... sess.run(w.init())
|
|
... sess.run(hp.hparams(hparams))
|
|
... sess.run(w.flush())
|
|
|
|
To control how hyperparameters and metrics appear in the TensorBoard UI,
|
|
you can define `HParam` and `Metric` objects, and write an experiment
|
|
summary to the top-level log directory:
|
|
|
|
>>> HP_OPTIMIZER = hp.HParam("optimizer")
|
|
>>> HP_FC_DROPOUT = hp.HParam(
|
|
... "fc_dropout",
|
|
... display_name="f.c. dropout",
|
|
... description="Dropout rate for fully connected subnet.",
|
|
... )
|
|
>>> HP_NEURONS = hp.HParam("neurons", description="Neurons per dense layer")
|
|
>>>
|
|
>>> with tf.summary.create_file_writer(base_logdir).as_default():
|
|
... hp.hparams_config(
|
|
... hparams=[
|
|
... HP_OPTIMIZER,
|
|
... HP_FC_DROPOUT,
|
|
... HP_NEURONS,
|
|
... ],
|
|
... metrics=[
|
|
... hp.Metric("xent", group="validation", display_name="cross-entropy"),
|
|
... hp.Metric("f1", group="validation", display_name="F₁ score"),
|
|
... hp.Metric("loss", group="train", display_name="training loss"),
|
|
... ],
|
|
... )
|
|
|
|
You can continue to pass a string-keyed dict to the Keras callback or
|
|
the `hparams` function, or you can use `HParam` objects as the keys. The
|
|
latter approach enables better static analysis: your favorite Python
|
|
linter can tell you if you misspell a hyperparameter name, your IDE can
|
|
help you find all the places where a hyperparameter is used, etc:
|
|
|
|
>>> hparams = {
|
|
... HP_OPTIMIZER: "adam",
|
|
... HP_FC_DROPOUT: 0.2,
|
|
... HP_NEURONS: 128,
|
|
... # ...
|
|
... }
|
|
>>>
|
|
>>> model = model_fn(hparams)
|
|
>>> callbacks = [
|
|
>>> tf.keras.callbacks.TensorBoard(logdir),
|
|
>>> hp.KerasCallback(logdir, hparams),
|
|
>>> ]
|
|
|
|
Finally, you can choose to annotate your hparam definitions with domain
|
|
information:
|
|
|
|
>>> HP_OPTIMIZER = hp.HParam("optimizer", hp.Discrete(["adam", "sgd"]))
|
|
>>> HP_FC_DROPOUT = hp.HParam("fc_dropout", hp.RealInterval(0.1, 0.4))
|
|
>>> HP_NEURONS = hp.HParam("neurons", hp.IntInterval(64, 256))
|
|
|
|
The TensorBoard HParams plugin does not provide tuners, but you can
|
|
integrate these domains into your preferred tuning framework if you so
|
|
desire. The domains will also be reflected in the TensorBoard UI.
|
|
|
|
See the `Experiment`, `HParam`, `Metric`, and `KerasCallback` classes
|
|
for API specifications. Consult the `hparams_demo.py` script in the
|
|
TensorBoard repository for an end-to-end MNIST example.
|
|
"""
|
|
|
|
|
|
from tensorboard.plugins.hparams import _keras
|
|
from tensorboard.plugins.hparams import summary_v2
|
|
|
|
|
|
Discrete = summary_v2.Discrete
|
|
Domain = summary_v2.Domain
|
|
HParam = summary_v2.HParam
|
|
IntInterval = summary_v2.IntInterval
|
|
Metric = summary_v2.Metric
|
|
RealInterval = summary_v2.RealInterval
|
|
hparams = summary_v2.hparams
|
|
hparams_pb = summary_v2.hparams_pb
|
|
hparams_config = summary_v2.hparams_config
|
|
hparams_config_pb = summary_v2.hparams_config_pb
|
|
|
|
KerasCallback = _keras.Callback
|
|
|
|
|
|
del _keras
|
|
del summary_v2
|