def create(self, weights=None, submodel=None):
"""Load a model
Args:
weights: a saved model weights from previous run
name: a model name from DeepTreeAttention.models
"""
self.HSI_model, self.HSI_spatial, self.HSI_spectral = Hang.create_models(self.HSI_size, self.HSI_size, self.HSI_channels, self.classes, self.config["train"]["learning_rate"])
self.RGB_model, self.RGB_spatial, self.RGB_spectral = Hang.create_models(self.RGB_size, self.RGB_size, self.RGB_channels, self.classes, self.config["train"]["learning_rate"])
def test_ensemble(HSI_image, metadata_data):
batch, height, width, channels = HSI_image.shape
sensor_inputs, sensor_outputs, spatial, spectral = Hang.define_model(classes=2, height=height, width=width, channels=channels)
model1 = tf.keras.Model(inputs=sensor_inputs, outputs=sensor_outputs)
metadata_model = metadata.create(classes=2, sites=10, domains =10, learning_rate=0.001)
ensemble = Hang.learned_ensemble(HSI_model=model1, metadata_model=metadata_model, classes=2)
prediction = ensemble.predict([HSI_image] + metadata_data)
assert prediction.shape == (1, 2)
def test_ensemble(RGB_image, HSI_image):
batch, height, width, channels = HSI_image[0].shape
sensor_inputs, sensor_outputs, spatial, spectral = Hang.define_model(
classes=2, height=height, width=width, channels=channels)
model1 = tf.keras.Model(inputs=sensor_inputs, outputs=sensor_outputs)
batch, height, width, channels = RGB_image[0].shape
sensor_inputs, sensor_outputs, spatial, spectral = Hang.define_model(
classes=2, height=height, width=width, channels=channels)
model2 = tf.keras.Model(inputs=sensor_inputs, outputs=sensor_outputs)
ensemble = Hang.ensemble(models=[model1, model2], classes=2)
prediction = ensemble.predict([HSI_image[0], RGB_image[0]])
assert prediction.shape == (1, 2)
def test_model(RGB_image, classes):
batch, height, width, channels = RGB_image[0].shape
sensor_inputs, sensor_outputs, spatial, spectral = Hang.define_model(
classes=classes, height=height, width=width, channels=channels)
model = tf.keras.Model(inputs=sensor_inputs, outputs=sensor_outputs)
prediction = model.predict(RGB_image)
assert prediction.shape == (1, classes)
def ensemble_model():
sensor_inputs, sensor_outputs, spatial, spectral = Hang2020_geographic.define_model(
classes=2, height=20, width=20, channels=369)
model1 = tf.keras.Model(inputs=sensor_inputs, outputs=sensor_outputs)
metadata_model = metadata.create(classes=2,
sites=10,
domains=10,
learning_rate=0.001)
ensemble_model = Hang2020_geographic.learned_ensemble(
HSI_model=model1, metadata_model=metadata_model, classes=2)
ensemble_model = tf.keras.Model(
ensemble_model.inputs,
ensemble_model.get_layer("submodel_concat").output)
return ensemble_model
def create(self, weights=None, submodel=None):
"""Load a model
Args:
weights: a saved model weights from previous run
name: a model name from DeepTreeAttention.models
"""
self.classes = pd.read_csv(self.classes_file).shape[0]
if self.config["train"]["gpus"] > 1:
self.strategy = tf.distribute.MirroredStrategy()
print("Running in parallel on {} GPUs".format(
self.strategy.num_replicas_in_sync))
self.config["train"]["batch_size"] = self.config["train"][
"batch_size"] * self.strategy.num_replicas_in_sync
with self.strategy.scope():
self.HSI_model, self.HSI_spatial, self.HSI_spectral = Hang.create_models(
self.HSI_size, self.HSI_size, self.HSI_channels,
self.classes, self.config["train"]["learning_rate"])
self.RGB_model, self.RGB_spatial, self.RGB_spectral = Hang.create_models(
self.RGB_size, self.RGB_size, self.RGB_channels,
self.classes, self.config["train"]["learning_rate"])
#create a metadata model
self.metadata_model = metadata.create(
classes=self.classes,
sites=self.sites,
domains=self.domains,
learning_rate=self.config["train"]["learning_rate"])
else:
self.HSI_model, self.HSI_spatial, self.HSI_spectral = Hang.create_models(
self.HSI_size, self.HSI_size, self.HSI_channels, self.classes,
self.config["train"]["learning_rate"])
self.RGB_model, self.RGB_spatial, self.RGB_spectral = Hang.create_models(
self.RGB_size, self.RGB_size, self.RGB_channels, self.classes,
self.config["train"]["learning_rate"])
#create a metadata model
self.metadata_model = metadata.create(
classes=self.classes,
sites=self.sites,
domains=self.domains,
learning_rate=self.config["train"]["learning_rate"])
def test_define(HSI_image, metadata_data):
batch, height, width, channels = HSI_image.shape
sensor_inputs, sensor_outputs, spatial, spectral = Hang.define_model(
classes=2, height=height, width=width, channels=channels)
model1 = tf.keras.Model(inputs=sensor_inputs, outputs=sensor_outputs)
metadata_model = metadata.create(classes=2,
sites=10,
domains=10,
learning_rate=0.001)
ensemble = Hang.learned_ensemble(HSI_model=model1,
metadata_model=metadata_model,
classes=2)
extractor = tf.keras.Model(ensemble.inputs, ensemble.output)
neighbor_array = []
neighbor_distance = []
for x in np.arange(5):
prediction = extractor.predict([HSI_image] + metadata_data)
neighbor_array.append(prediction)
neighbor_distance.append(np.random.rand())
#stack and batch
neighbor_array = np.vstack(neighbor_array)
neighbor_array = np.expand_dims(neighbor_array, axis=0)
neighbor_distance = np.expand_dims(neighbor_distance, axis=0)
neighbor_model = neighbors_model.create(ensemble_model=ensemble,
freeze=False,
k_neighbors=5,
classes=2)
prediction = neighbor_model.predict([HSI_image] + metadata_data +
[neighbor_array] + [neighbor_distance])
assert prediction.shape == (1, 2)
def ensemble(self, experiment, class_weight=None, freeze = True, train=True):
#Manually override batch size
self.config["train"]["batch_size"] = self.config["train"]["ensemble"]["batch_size"]
self.read_data(mode="ensemble")
self.ensemble_model = Hang.ensemble([self.HSI_model, self.RGB_model], freeze=freeze, classes=self.classes)
if train:
self.ensemble_model.compile(
loss="categorical_crossentropy",
optimizer=tf.keras.optimizers.Adam(
lr=float(self.config["train"]["learning_rate"])),
metrics=[tf.keras.metrics.CategoricalAccuracy(
name='acc')])
if self.val_split is None:
print("Cannot run callbacks without validation data, skipping...")
callback_list = None
label_names = None
elif experiment is None:
print("Cannot run callbacks without comet experiment, skipping...")
callback_list = None
label_names = None
else:
if self.classes_file is not None:
labeldf = pd.read_csv(self.classes_file)
label_names = list(labeldf.taxonID.values)
else:
label_names = None
callback_list = callbacks.create(log_dir=self.log_dir,
experiment=experiment,
validation_data=self.val_split,
train_data=self.train_split,
label_names=label_names,
submodel="ensemble")
#Train ensemble layer
self.ensemble_model.fit(
self.train_split,
epochs=self.config["train"]["epochs"],
validation_data=self.val_split,
callbacks=callback_list,
class_weight=class_weight)
def create(self, weights=None, submodel=None):
"""Load a model
Args:
weights: a saved model weights from previous run
name: a model name from DeepTreeAttention.models
"""
self.classes = pd.read_csv(self.classes_file).shape[0]
if self.config["train"]["gpus"] > 1:
self.strategy = tf.distribute.MirroredStrategy()
print("Running in parallel on {} GPUs".format(
self.strategy.num_replicas_in_sync))
self.config["train"]["batch_size"] = self.config["train"][
"batch_size"] * self.strategy.num_replicas_in_sync
with self.strategy.scope():
self.HSI_model, self.HSI_spatial, self.HSI_spectral = Hang.create_models(
self.HSI_size, self.HSI_size, self.HSI_channels,
self.classes, self.config["train"]["learning_rate"])
self.RGB_model, self.RGB_spatial, self.RGB_spectral = Hang.create_models(
self.RGB_size, self.RGB_size, self.RGB_channels,
self.classes, self.config["train"]["learning_rate"])
#create a metadata model
self.metadata_model = metadata.create(
classes=self.classes,
sites=self.sites,
domains=self.domains,
learning_rate=self.config["train"]["learning_rate"])
self.ensemble_model = Hang.learned_ensemble(
HSI_model=self.HSI_model,
metadata_model=self.metadata_model,
freeze=self.config["ensemble"]["freeze"],
classes=self.classes)
self.ensemble_model.compile(
loss="categorical_crossentropy",
optimizer=tf.keras.optimizers.Adam(
lr=float(self.config["train"]["learning_rate"])),
metrics=[tf.keras.metrics.CategoricalAccuracy(name='acc')])
else:
self.HSI_model, self.HSI_spatial, self.HSI_spectral = Hang.create_models(
self.HSI_size, self.HSI_size, self.HSI_channels, self.classes,
self.config["train"]["learning_rate"])
self.RGB_model, self.RGB_spatial, self.RGB_spectral = Hang.create_models(
self.RGB_size, self.RGB_size, self.RGB_channels, self.classes,
self.config["train"]["learning_rate"])
#create a metadata model
self.metadata_model = metadata.create(
classes=self.classes,
sites=self.sites,
domains=self.domains,
learning_rate=self.config["train"]["learning_rate"])
#create an ensemble model
self.ensemble_model = Hang.learned_ensemble(
HSI_model=self.HSI_model,
metadata_model=self.metadata_model,
freeze=self.config["train"]["ensemble"]["freeze"],
classes=self.classes)
#Compile ensemble
self.ensemble_model.compile(
loss="categorical_crossentropy",
optimizer=tf.keras.optimizers.Adam(
lr=float(self.config["train"]["learning_rate"])),
metrics=[tf.keras.metrics.CategoricalAccuracy(name='acc')])
def ensemble(self, experiment, class_weight=None, freeze=True, train=True):
self.classes = pd.read_csv(self.classes_file).shape[0]
self.read_data(mode="ensemble")
if self.val_split is None:
print("Cannot run callbacks without validation data, skipping...")
callback_list = None
label_names = None
elif experiment is None:
print("Cannot run callbacks without comet experiment, skipping...")
callback_list = None
label_names = None
else:
if self.classes_file is not None:
labeldf = pd.read_csv(self.classes_file)
label_names = list(labeldf.taxonID.values)
else:
label_names = None
callback_list = callbacks.create(log_dir=self.log_dir,
experiment=experiment,
validation_data=self.val_split,
train_data=self.train_split,
label_names=label_names,
train_shp=self.train_shp,
submodel="ensemble")
print("callback list is {}".format(callback_list))
if self.config["train"]["gpus"] > 1:
with self.strategy.scope():
self.ensemble_model = Hang.learned_ensemble(
HSI_model=self.HSI_model,
metadata_model=self.metadata_model,
freeze=freeze,
classes=self.classes)
if train:
self.ensemble_model.compile(
loss="categorical_crossentropy",
optimizer=tf.keras.optimizers.Adam(
lr=float(self.config["train"]["learning_rate"])),
metrics=[
tf.keras.metrics.CategoricalAccuracy(name='acc')
])
#Train ensemble layer
self.ensemble_model.fit(
self.train_split,
epochs=self.config["train"]["ensemble"]["epochs"],
validation_data=self.val_split,
callbacks=callback_list,
class_weight=class_weight)
else:
self.ensemble_model = Hang.learned_ensemble(
HSI_model=self.HSI_model,
metadata_model=self.metadata_model,
freeze=freeze,
classes=self.classes)
if train:
self.ensemble_model.compile(
loss="categorical_crossentropy",
optimizer=tf.keras.optimizers.Adam(
lr=float(self.config["train"]["learning_rate"])),
metrics=[tf.keras.metrics.CategoricalAccuracy(name='acc')])
#Train ensemble layer
self.ensemble_model.fit(
self.train_split,
epochs=self.config["train"]["ensemble"]["epochs"],
validation_data=self.val_split,
callbacks=callback_list,
class_weight=class_weight)