def main(_):
model_class = models.get_model_class(FLAGS.model)
# Look up the model configuration.
assert (FLAGS.config_name is None) != (FLAGS.config_json is None), (
"Exactly one of --config_name or --config_json is required.")
config = (models.get_model_config(FLAGS.model, FLAGS.config_name) if
FLAGS.config_name else config_util.parse_json(FLAGS.config_json))
config = configdict.ConfigDict(config)
# Create the estimator.
estimator = estimator_util.create_estimator(model_class,
config.hparams,
model_dir=FLAGS.model_dir)
# Read and process the input features.
features = _process_tce(config.inputs.features)
# Create an input function.
def input_fn():
return {
"time_series_features":
tf.estimator.inputs.numpy_input_fn(features,
batch_size=1,
shuffle=False,
queue_capacity=1)()
}
# Generate the predictions.
for predictions in estimator.predict(input_fn):
assert len(predictions) == 1
print("Prediction:", predictions[0])
def testOneTimeSeriesFeature(self):
# Build config.
feature_spec = {
"time_feature_1": {
"length": 10,
"is_time_series": True,
}
}
config = configurations.base()
config["inputs"]["features"] = feature_spec
config = configdict.ConfigDict(config)
# Build model.
features = input_ops.build_feature_placeholders(config.inputs.features)
labels = input_ops.build_labels_placeholder()
model = astro_model.AstroModel(features, labels, config.hparams,
tf.estimator.ModeKeys.TRAIN)
model.build()
# Validate hidden layers.
self.assertItemsEqual(["time_feature_1"],
model.time_series_hidden_layers.keys())
self.assertIs(model.time_series_features["time_feature_1"],
model.time_series_hidden_layers["time_feature_1"])
self.assertEqual(len(model.aux_hidden_layers), 0)
self.assertIs(model.time_series_features["time_feature_1"],
model.pre_logits_concat)
def setUp(self):
super(BuildDatasetTest, self).setUp()
# The test dataset contains 10 tensorflow.Example protocol buffers. The i-th
# Example contains the following features:
# global_view = [0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0]
# local_view = [0.0, 1.0, 2.0, 3.0]
# aux_feature = 100 + i
# label_str = "PC" if i % 3 == 0 else "AFP" if i % 3 == 1 else "NTP"
self._file_pattern = _TEST_TFRECORD_FILE
self._input_config = configdict.ConfigDict({
"features": {
"global_view": {
"is_time_series": True,
"length": 8
},
"local_view": {
"is_time_series": True,
"length": 4
},
"aux_feature": {
"is_time_series": False,
"length": 1
}
}
})
def testZeroHiddenLayers(self):
# Build config.
feature_spec = {
"time_feature_1": {
"length": 10,
"is_time_series": True,
},
"time_feature_2": {
"length": 10,
"is_time_series": True,
},
"aux_feature_1": {
"length": 1,
"is_time_series": False,
},
}
config = configurations.base()
config["inputs"]["features"] = feature_spec
config = configdict.ConfigDict(config)
config.hparams.output_dim = 1
config.hparams.num_pre_logits_hidden_layers = 0
# Build model.
features = input_ops.build_feature_placeholders(config.inputs.features)
labels = input_ops.build_labels_placeholder()
model = astro_model.AstroModel(features, labels, config.hparams,
tf.estimator.ModeKeys.TRAIN)
model.build()
# Validate Tensor shapes.
self.assertShapeEquals((None, 21), model.pre_logits_concat)
logits_w = testing.get_variable_by_name("logits/kernel")
self.assertShapeEquals((21, 1), logits_w)
def main(_):
model_class = models.get_model_class(FLAGS.model)
# Look up the model configuration.
assert (FLAGS.config_name is None) != (FLAGS.config_json is None), (
"Exactly one of --config_name or --config_json is required.")
config = (
models.get_model_config(FLAGS.model, FLAGS.config_name)
if FLAGS.config_name else config_util.parse_json(FLAGS.config_json))
config = configdict.ConfigDict(config)
# Create the estimator.
estimator = estimator_util.create_estimator(
model_class, config.hparams, model_dir=FLAGS.model_dir)
# Create an input function that reads the evaluation dataset.
input_fn = estimator_util.create_input_fn(
file_pattern=FLAGS.eval_files,
input_config=config.inputs,
mode=tf.estimator.ModeKeys.EVAL)
# Run evaluation. This will log the result to stderr and also write a summary
# file in the model_dir.
estimator_util.evaluate(estimator, input_fn, eval_name=FLAGS.eval_name)
def testInvalidModeRaisesError(self):
# Build config.
config = configdict.ConfigDict(configurations.base())
# Build model.
features = input_ops.build_feature_placeholders(config.inputs.features)
labels = input_ops.build_labels_placeholder()
with self.assertRaises(ValueError):
_ = astro_model.AstroModel(features, labels, config.hparams,
"training")
def testZeroFeaturesRaisesError(self):
# Build config.
config = configurations.base()
config["inputs"]["features"] = {}
config = configdict.ConfigDict(config)
# Build model.
features = input_ops.build_feature_placeholders(config.inputs.features)
labels = input_ops.build_labels_placeholder()
model = astro_model.AstroModel(features, labels, config.hparams,
tf.estimator.ModeKeys.TRAIN)
with self.assertRaises(ValueError):
# Raises ValueError because at least one feature is required.
model.build()
def testEvalMode(self):
# Build config.
feature_spec = {
"time_feature_1": {
"length": 10,
"is_time_series": True,
},
"time_feature_2": {
"length": 10,
"is_time_series": True,
},
"aux_feature_1": {
"length": 1,
"is_time_series": False,
},
}
config = configurations.base()
config["inputs"]["features"] = feature_spec
config = configdict.ConfigDict(config)
config.hparams.output_dim = 1
# Build model.
features = input_ops.build_feature_placeholders(config.inputs.features)
labels = input_ops.build_labels_placeholder()
model = astro_model.AstroModel(features, labels, config.hparams,
tf.estimator.ModeKeys.TRAIN)
model.build()
# Validate Tensor shapes.
self.assertShapeEquals((None, 21), model.pre_logits_concat)
self.assertShapeEquals((None, 1), model.logits)
self.assertShapeEquals((None, 1), model.predictions)
self.assertShapeEquals((None, ), model.batch_losses)
self.assertShapeEquals((), model.total_loss)
# Execute the TensorFlow graph.
scaffold = tf.train.Scaffold()
scaffold.finalize()
with self.test_session() as sess:
sess.run([scaffold.init_op, scaffold.local_init_op])
step = sess.run(model.global_step)
self.assertEqual(0, step)
# Fetch total loss.
features = testing.fake_features(feature_spec, batch_size=16)
labels = testing.fake_labels(config.hparams.output_dim,
batch_size=16)
feed_dict = input_ops.prepare_feed_dict(model, features, labels)
total_loss = sess.run(model.total_loss, feed_dict=feed_dict)
self.assertShapeEquals((), total_loss)
def main(_):
model_class = models.get_model_class(FLAGS.model)
# Look up the model configuration.
assert (FLAGS.config_name is None) != (FLAGS.config_json is None), (
"Exactly one of --config_name or --config_json is required.")
config = (
models.get_model_config(FLAGS.model, FLAGS.config_name)
if FLAGS.config_name else config_util.parse_json(FLAGS.config_json))
config = configdict.ConfigDict(config)
config_util.log_and_save_config(config, FLAGS.model_dir)
# Create the estimator.
run_config = tf.estimator.RunConfig(keep_checkpoint_max=1)
estimator = estimator_util.create_estimator(model_class, config.hparams,
run_config, FLAGS.model_dir)
# Create an input function that reads the training dataset. We iterate through
# the dataset once at a time if we are alternating with evaluation, otherwise
# we iterate infinitely.
train_input_fn = estimator_util.create_input_fn(
file_pattern=FLAGS.train_files,
input_config=config.inputs,
mode=tf.estimator.ModeKeys.TRAIN,
shuffle_values_buffer=FLAGS.shuffle_buffer_size,
repeat=1 if FLAGS.eval_files else None)
if not FLAGS.eval_files:
estimator.train(train_input_fn, max_steps=FLAGS.train_steps)
else:
eval_input_fn = estimator_util.create_input_fn(
file_pattern=FLAGS.eval_files,
input_config=config.inputs,
mode=tf.estimator.ModeKeys.EVAL)
for _ in estimator_util.continuous_train_and_eval(
estimator=estimator,
train_input_fn=train_input_fn,
eval_input_fn=eval_input_fn,
train_steps=FLAGS.train_steps):
# continuous_train_and_eval() yields evaluation metrics after each
# training epoch. We don't do anything here.
pass
def testOneTimeSeriesFeature(self):
# Build config.
feature_spec = {
"time_feature_1": {
"length": 20,
"is_time_series": True,
}
}
hidden_spec = {
"time_feature_1": {
"cnn_num_blocks": 2,
"cnn_block_size": 2,
"cnn_initial_num_filters": 4,
"cnn_block_filter_factor": 1.5,
"cnn_kernel_size": 3,
"convolution_padding": "same",
"pool_size": 2,
"pool_strides": 2,
}
}
config = configurations.base()
config["inputs"]["features"] = feature_spec
config["hparams"]["time_series_hidden"] = hidden_spec
config = configdict.ConfigDict(config)
# Build model.
features = input_ops.build_feature_placeholders(config.inputs.features)
labels = input_ops.build_labels_placeholder()
model = astro_cnn_model.AstroCNNModel(features, labels, config.hparams,
tf.estimator.ModeKeys.TRAIN)
model.build()
# Validate Tensor shapes.
block_1_conv_1 = testing.get_variable_by_name(
"time_feature_1_hidden/block_1/conv_1/kernel")
self.assertShapeEquals((3, 1, 4), block_1_conv_1)
block_1_conv_2 = testing.get_variable_by_name(
"time_feature_1_hidden/block_1/conv_2/kernel")
self.assertShapeEquals((3, 4, 4), block_1_conv_2)
block_2_conv_1 = testing.get_variable_by_name(
"time_feature_1_hidden/block_2/conv_1/kernel")
self.assertShapeEquals((3, 4, 6), block_2_conv_1)
block_2_conv_2 = testing.get_variable_by_name(
"time_feature_1_hidden/block_2/conv_2/kernel")
self.assertShapeEquals((3, 6, 6), block_2_conv_2)
self.assertItemsEqual(["time_feature_1"],
model.time_series_hidden_layers.keys())
self.assertShapeEquals(
(None, 30), model.time_series_hidden_layers["time_feature_1"])
self.assertEqual(len(model.aux_hidden_layers), 0)
self.assertIs(model.time_series_hidden_layers["time_feature_1"],
model.pre_logits_concat)
# Execute the TensorFlow graph.
scaffold = tf.train.Scaffold()
scaffold.finalize()
with self.test_session() as sess:
sess.run([scaffold.init_op, scaffold.local_init_op])
step = sess.run(model.global_step)
self.assertEqual(0, step)
# Fetch predictions.
features = testing.fake_features(feature_spec, batch_size=16)
labels = testing.fake_labels(config.hparams.output_dim,
batch_size=16)
feed_dict = input_ops.prepare_feed_dict(model, features, labels)
predictions = sess.run(model.predictions, feed_dict=feed_dict)
self.assertShapeEquals((16, 1), predictions)
def testBuildFeaturePlaceholders(self):
# One time series feature.
config = configdict.ConfigDict(
{"time_feature_1": {
"length": 14,
"is_time_series": True,
}})
expected_shapes = {
"time_series_features": {
"time_feature_1": [None, 14],
},
"aux_features": {}
}
features = input_ops.build_feature_placeholders(config)
self.assertFeatureShapesEqual(expected_shapes, features)
# Two time series features.
config = configdict.ConfigDict({
"time_feature_1": {
"length": 14,
"is_time_series": True,
},
"time_feature_2": {
"length": 5,
"is_time_series": True,
}
})
expected_shapes = {
"time_series_features": {
"time_feature_1": [None, 14],
"time_feature_2": [None, 5],
},
"aux_features": {}
}
features = input_ops.build_feature_placeholders(config)
self.assertFeatureShapesEqual(expected_shapes, features)
# One aux feature.
config = configdict.ConfigDict({
"time_feature_1": {
"length": 14,
"is_time_series": True,
},
"aux_feature_1": {
"length": 1,
"is_time_series": False,
}
})
expected_shapes = {
"time_series_features": {
"time_feature_1": [None, 14],
},
"aux_features": {
"aux_feature_1": [None, 1]
}
}
features = input_ops.build_feature_placeholders(config)
self.assertFeatureShapesEqual(expected_shapes, features)
# Two aux features.
config = configdict.ConfigDict({
"time_feature_1": {
"length": 14,
"is_time_series": True,
},
"aux_feature_1": {
"length": 1,
"is_time_series": False,
},
"aux_feature_2": {
"length": 6,
"is_time_series": False,
},
})
expected_shapes = {
"time_series_features": {
"time_feature_1": [None, 14],
},
"aux_features": {
"aux_feature_1": [None, 1],
"aux_feature_2": [None, 6]
}
}
features = input_ops.build_feature_placeholders(config)
self.assertFeatureShapesEqual(expected_shapes, features)
def testOneTimeSeriesFeature(self):
# Build config.
feature_spec = {
"time_feature_1": {
"length": 14,
"is_time_series": True,
}
}
hidden_spec = {
"time_feature_1": {
"num_local_layers": 2,
"local_layer_size": 20,
"translation_delta": 2,
"pooling_type": "max",
"dropout_rate": 0.5,
}
}
config = configurations.base()
config["inputs"]["features"] = feature_spec
config["hparams"]["time_series_hidden"] = hidden_spec
config = configdict.ConfigDict(config)
# Build model.
features = input_ops.build_feature_placeholders(config.inputs.features)
labels = input_ops.build_labels_placeholder()
model = astro_fc_model.AstroFCModel(features, labels, config.hparams,
tf.estimator.ModeKeys.TRAIN)
model.build()
# Validate Tensor shapes.
conv = testing.get_variable_by_name(
"time_feature_1_hidden/conv1d/kernel")
self.assertShapeEquals((10, 1, 20), conv)
fc_1 = testing.get_variable_by_name(
"time_feature_1_hidden/fully_connected_1/weights")
self.assertShapeEquals((20, 20), fc_1)
self.assertItemsEqual(["time_feature_1"],
model.time_series_hidden_layers.keys())
self.assertShapeEquals(
(None, 20), model.time_series_hidden_layers["time_feature_1"])
self.assertEqual(len(model.aux_hidden_layers), 0)
self.assertIs(model.time_series_hidden_layers["time_feature_1"],
model.pre_logits_concat)
# Execute the TensorFlow graph.
scaffold = tf.train.Scaffold()
scaffold.finalize()
with self.test_session() as sess:
sess.run([scaffold.init_op, scaffold.local_init_op])
step = sess.run(model.global_step)
self.assertEqual(0, step)
# Fetch predictions.
features = testing.fake_features(feature_spec, batch_size=16)
labels = testing.fake_labels(config.hparams.output_dim,
batch_size=16)
feed_dict = input_ops.prepare_feed_dict(model, features, labels)
predictions = sess.run(model.predictions, feed_dict=feed_dict)
self.assertShapeEquals((16, 1), predictions)
