def __init__(self, config, triage_model):
super(AstroCNNModelVetting, self).__init__()
hps = config.vetting_hparams
self.triage_model = astro_cnn_model.AstroCNNModel(
config, triage_model, embeds_only=not hps.use_preds_layer)
self.config = config
self.ts_blocks = self._create_ts_blocks(config)
self.final = [tf.keras.layers.Concatenate()]
for i in range(hps.num_pre_logits_hidden_layers):
self.final.append(
tf.keras.layers.Dense(units=hps.pre_logits_hidden_layer_size,
activation='relu'))
if hps.use_batch_norm:
self.final.append(tf.keras.layers.BatchNormalization())
self.final.append(
tf.keras.layers.Dropout(hps.pre_logits_dropout_rate))
if config.inputs.get('exclusive_labels', False):
self.final.append(
tf.keras.layers.Dense(units=len(config.inputs.label_columns),
activation=None))
self.final.append(tf.keras.layers.Softmax())
else:
self.final.append(
tf.keras.layers.Dense(units=len(config.inputs.label_columns),
activation='sigmoid'))
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 testTwoTimeSeriesFeatures(self):
# Build config.
feature_spec = {
"time_feature_1": {
"length": 20,
"is_time_series": True,
},
"time_feature_2": {
"length": 5,
"is_time_series": True,
},
"aux_feature_1": {
"length": 1,
"is_time_series": False,
},
}
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,
},
"time_feature_2": {
"cnn_num_blocks": 1,
"cnn_block_size": 1,
"cnn_initial_num_filters": 5,
"cnn_block_filter_factor": 1,
"cnn_kernel_size": 2,
"convolution_padding": "same",
"pool_size": 0,
"pool_strides": 0,
}
}
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()
# TODO(shallue): TensorFlow 2.0 doesn't have global variable collections.
# If we want to keep testing variable shapes in 2.0, we must keep track of
# the individual Keras Layer objects in the model class.
variables = {v.op.name: v for v in tf.global_variables()}
# Validate Tensor shapes.
feature_1_block_1_conv_1 = variables[
"time_feature_1_hidden/block_1/conv_1/kernel"]
self.assertShapeEquals((3, 1, 4), feature_1_block_1_conv_1)
feature_1_block_1_conv_2 = variables[
"time_feature_1_hidden/block_1/conv_2/kernel"]
self.assertShapeEquals((3, 4, 4), feature_1_block_1_conv_2)
feature_1_block_2_conv_1 = variables[
"time_feature_1_hidden/block_2/conv_1/kernel"]
self.assertShapeEquals((3, 4, 6), feature_1_block_2_conv_1)
feature_1_block_2_conv_2 = variables[
"time_feature_1_hidden/block_2/conv_2/kernel"]
self.assertShapeEquals((3, 6, 6), feature_1_block_2_conv_2)
feature_2_block_1_conv_1 = variables[
"time_feature_2_hidden/block_1/conv_1/kernel"]
self.assertShapeEquals((2, 1, 5), feature_2_block_1_conv_1)
self.assertItemsEqual(["time_feature_1", "time_feature_2"],
model.time_series_hidden_layers.keys())
self.assertShapeEquals((None, 30),
model.time_series_hidden_layers["time_feature_1"])
self.assertShapeEquals((None, 25),
model.time_series_hidden_layers["time_feature_2"])
self.assertItemsEqual(["aux_feature_1"], model.aux_hidden_layers.keys())
self.assertIs(model.aux_features["aux_feature_1"],
model.aux_hidden_layers["aux_feature_1"])
self.assertShapeEquals((None, 56), model.pre_logits_concat)
# Execute the TensorFlow graph.
scaffold = tf.train.Scaffold()
scaffold.finalize()
with self.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)