def train_model(orga, model, epoch, batch_logger=False):
"""
Train a model on one file and return the history.
Parameters
----------
orga : orcanet.core.Organizer
Contains all the configurable options in the OrcaNet scripts.
model : keras.Model
A compiled keras model.
epoch : tuple
Current epoch and the no of the file to train on.
batch_logger : bool
Use the orcanet batchlogger to log the training.
Returns
-------
history : dict
The history of the training on this file. A record of training
loss values and metrics values.
"""
callbacks = [
nn_utilities.RaiseOnNaN(),
nn_utilities.TimeModel(print_func=orga.io.print_log),
]
if batch_logger:
callbacks.append(BatchLogger(orga, epoch))
if orga.cfg.callback_train is not None:
try:
callbacks.extend(orga.cfg.callback_train)
except TypeError:
callbacks.append(orga.cfg.callback_train)
training_generator = h5_generator.get_h5_generator(
orga,
files_dict=orga.io.get_file("train", epoch[1]),
f_size=orga.cfg.n_events,
phase="training",
zero_center=orga.cfg.zero_center_folder is not None,
shuffle=orga.cfg.shuffle_train,
)
# status tf.2.5: In order to use ragged Tensors as input to fit,
# we have to use a tf dataset and not a generator
dataset = h5_generator.make_dataset(training_generator)
history = model.fit(
dataset,
steps_per_epoch=len(training_generator),
verbose=orga.cfg.verbose_train,
max_queue_size=orga.cfg.max_queue_size,
callbacks=callbacks,
initial_epoch=epoch[0] - 1,
epochs=epoch[0],
)
training_generator.print_timestats(print_func=orga.io.print_log)
# get a dict with losses and metrics
# only trained for one epoch, so value is list of len 1
history = {key: value[0] for key, value in history.history.items()}
return history
def test_batch(self):
filepaths = self.filepaths_file_1
gene = iter(get_h5_generator(self.orga, filepaths))
target_xs_batch_1 = {
"input_A": self.train_A_file_1_ctnt[0][:2],
"input_B": self.train_B_file_1_ctnt[0][:2],
}
target_ys_batch_1 = label_modifier(
{"y_values": self.train_A_file_1_ctnt[1][:2]})
target_xs_batch_2 = {
"input_A": self.train_A_file_1_ctnt[0][2:4],
"input_B": self.train_B_file_1_ctnt[0][2:4],
}
target_ys_batch_2 = label_modifier(
{"y_values": self.train_A_file_1_ctnt[1][2:4]})
xs, ys = next(gene)
assert_dict_arrays_equal(xs, target_xs_batch_1)
assert_dict_arrays_equal(ys, target_ys_batch_1)
xs, ys = next(gene)
assert_dict_arrays_equal(xs, target_xs_batch_2)
assert_dict_arrays_equal(ys, target_ys_batch_2)
with self.assertRaises(StopIteration):
next(gene)
def test_batch_mc_infos(self):
filepaths = self.filepaths_file_1
gene = iter(get_h5_generator(self.orga, filepaths, keras_mode=False))
target_xs_batch_1 = {
"input_A": self.train_A_file_1_ctnt[0][:2],
"input_B": self.train_B_file_1_ctnt[0][:2],
}
target_ys_batch_1 = label_modifier(
{"y_values": self.train_A_file_1_ctnt[1][:2]})
target_mc_info_batch_1 = self.train_A_file_1_ctnt[1][:2]
target_xs_batch_2 = {
"input_A": self.train_A_file_1_ctnt[0][2:4],
"input_B": self.train_B_file_1_ctnt[0][2:4],
}
target_ys_batch_2 = label_modifier(
{"y_values": self.train_A_file_1_ctnt[1][2:4]})
target_mc_info_batch_2 = self.train_A_file_1_ctnt[1][2:4]
info_blob = next(gene)
assert_dict_arrays_equal(info_blob["xs"], target_xs_batch_1)
assert_dict_arrays_equal(info_blob["ys"], target_ys_batch_1)
assert_equal_struc_array(info_blob["y_values"], target_mc_info_batch_1)
info_blob = next(gene)
assert_dict_arrays_equal(info_blob["xs"], target_xs_batch_2)
assert_dict_arrays_equal(info_blob["ys"], target_ys_batch_2)
assert_equal_struc_array(info_blob["y_values"], target_mc_info_batch_2)
def test_batch_sample_modifier(self):
filepaths = self.filepaths_file_1
def sample_modifier(info_blob):
xs_in = info_blob["x_values"]
mod = {name: val * 2 for name, val in xs_in.items()}
return mod
self.orga.cfg.sample_modifier = sample_modifier
gene = iter(get_h5_generator(self.orga, filepaths))
target_xs_batch_1 = {
"input_A": self.train_A_file_1_ctnt[0][:2] * 2,
"input_B": self.train_B_file_1_ctnt[0][:2] * 2,
}
target_ys_batch_1 = label_modifier(
{"y_values": self.train_A_file_1_ctnt[1][:2]})
target_xs_batch_2 = {
"input_A": self.train_A_file_1_ctnt[0][2:4] * 2,
"input_B": self.train_B_file_1_ctnt[0][2:4] * 2,
}
target_ys_batch_2 = label_modifier(
{"y_values": self.train_A_file_1_ctnt[1][2:4]})
xs, ys = next(gene)
assert_dict_arrays_equal(xs, target_xs_batch_1)
assert_dict_arrays_equal(ys, target_ys_batch_1)
xs, ys = next(gene)
assert_dict_arrays_equal(xs, target_xs_batch_2)
assert_dict_arrays_equal(ys, target_ys_batch_2)
def test_batch_zero_center(self):
filepaths = self.filepaths_file_1
xs_mean = {name: np.ones(shape) * 0.5 for name, shape in self.n_bins.items()}
self.orga.get_xs_mean = MagicMock(return_value=xs_mean)
gene = iter(get_h5_generator(self.orga, filepaths, zero_center=True))
target_xs_batch_1 = {
"input_A": np.subtract(self.train_A_file_1_ctnt[0][:2], xs_mean["input_A"]),
"input_B": np.subtract(self.train_B_file_1_ctnt[0][:2], xs_mean["input_B"]),
}
target_ys_batch_1 = label_modifier({"y_values": self.train_A_file_1_ctnt[1][:2]})
target_xs_batch_2 = {
"input_A": np.subtract(self.train_A_file_1_ctnt[0][2:], xs_mean["input_A"]),
"input_B": np.subtract(self.train_B_file_1_ctnt[0][2:], xs_mean["input_B"]),
}
target_ys_batch_2 = label_modifier({"y_values": self.train_A_file_1_ctnt[1][2:]})
xs, ys = next(gene)
assert_dict_arrays_equal(xs, target_xs_batch_1)
assert_dict_arrays_equal(ys, target_ys_batch_1)
xs, ys = next(gene)
assert_dict_arrays_equal(xs, target_xs_batch_2)
assert_dict_arrays_equal(ys, target_ys_batch_2)
def validate_model(orga, model):
"""
Validates a model on all validation files and return the history.
Parameters
----------
orga : orcanet.core.Organizer
Contains all the configurable options in the OrcaNet scripts.
model : keras.Model
A compiled keras model.
Returns
-------
history : dict
The history of the validation on all files. A record of validation
loss values and metrics values.
"""
# One history for each val file
histories = []
f_sizes = orga.io.get_file_sizes("val")
for i, files_dict in enumerate(orga.io.yield_files("val")):
f_size = f_sizes[i]
if orga.cfg.n_events is not None:
f_size = orga.cfg.n_events # for testing purposes
val_generator = get_h5_generator(
orga,
files_dict,
f_size=f_size,
phase="validation",
zero_center=orga.cfg.zero_center_folder is not None)
history_file = model.evaluate(val_generator,
steps=int(f_size / orga.cfg.batchsize),
max_queue_size=orga.cfg.max_queue_size,
verbose=orga.cfg.verbose_val)
if not isinstance(history_file, list):
history_file = [
history_file,
]
histories.append(history_file)
# average over all val files
history = weighted_average(histories, f_sizes)
# This history is just a list, not a dict like with fit_generator
# so transform to dict
history = dict(zip(model.metrics_names, history))
return history
def validate_model(orga, model):
"""
Validates a model on all validation files and return the history.
Parameters
----------
orga : orcanet.core.Organizer
Contains all the configurable options in the OrcaNet scripts.
model : keras.Model
A compiled keras model.
Returns
-------
history : dict
The history of the validation on all files. A record of validation
loss values and metrics values.
"""
# One history for each val file
histories = []
f_sizes = orga.io.get_file_sizes("val")
for i, files_dict in enumerate(orga.io.yield_files("val")):
val_generator = h5_generator.get_h5_generator(
orga,
files_dict,
f_size=orga.cfg.n_events,
phase="validation",
zero_center=orga.cfg.zero_center_folder is not None,
)
# status tf.2.5: In order to use ragged Tensors as input to fit,
# we have to use a tf dataset and not a generator
dataset = h5_generator.make_dataset(val_generator)
history_file = model.evaluate(
dataset,
steps=len(val_generator),
max_queue_size=orga.cfg.max_queue_size,
verbose=orga.cfg.verbose_val)
if not isinstance(history_file, list):
history_file = [history_file, ]
histories.append(history_file)
# average over all val files
history = weighted_average(histories, f_sizes)
# This history is just a list, not a dict like with fit_generator
# so transform to dict
history = dict(zip(model.metrics_names, history))
return history
def train_model(orga, model, epoch, batch_logger=False):
"""
Train a model on one file and return the history.
Parameters
----------
orga : orcanet.core.Organizer
Contains all the configurable options in the OrcaNet scripts.
model : keras.Model
A compiled keras model.
epoch : tuple
Current epoch and the no of the file to train on.
batch_logger : bool
Use the orcanet batchlogger to log the training.
Returns
-------
history : dict
The history of the training on this file. A record of training
loss values and metrics values.
"""
files_dict = orga.io.get_file("train", epoch[1])
if orga.cfg.n_events is not None:
# TODO Can throw an error if n_events is larger than the file
f_size = orga.cfg.n_events # for testing purposes
else:
f_size = orga.io.get_file_sizes("train")[epoch[1] - 1]
callbacks = [
nn_utilities.RaiseOnNaN(),
nn_utilities.TimeModel(print_func=orga.io.print_log),
]
if batch_logger:
callbacks.append(BatchLogger(orga, epoch))
if orga.cfg.callback_train is not None:
try:
callbacks.extend(orga.cfg.callback_train)
except TypeError:
callbacks.append(orga.cfg.callback_train)
training_generator = get_h5_generator(
orga,
files_dict,
f_size=f_size,
phase="training",
zero_center=orga.cfg.zero_center_folder is not None,
shuffle=orga.cfg.shuffle_train)
history = model.fit(
training_generator,
steps_per_epoch=int(f_size / orga.cfg.batchsize),
verbose=orga.cfg.verbose_train,
max_queue_size=orga.cfg.max_queue_size,
callbacks=callbacks,
initial_epoch=epoch[0] - 1,
epochs=epoch[0],
)
training_generator.print_timestats(print_func=orga.io.print_log)
# get a dict with losses and metrics
# only trained for one epoch, so value is list of len 1
history = {key: value[0] for key, value in history.history.items()}
return history
def h5_inference(orga,
model,
files_dict,
output_path,
samples=None,
use_def_label=True):
"""
Let a model predict on all samples in a h5 file, and save it as a h5 file.
Per default, the h5 file will contain a datagroup y_values straight from
the given files, as well as two datagroups per output layer of the network,
which have the labels and the predicted values in them as numpy arrays,
respectively.
Parameters
----------
orga : orcanet.core.Organizer
Contains all the configurable options in the OrcaNet scripts.
model : keras.Model
Trained Keras model of a neural network.
files_dict : dict
Dict mapping model input names to h5 file paths.
output_path : str
Name of the output h5 file containing the predictions.
samples : int, optional
Dont use all events in the file, but instead only the given number.
use_def_label : bool
If True and no label modifier is given by user, use the default
label modifier instead of none.
"""
file_size = h5_get_number_of_rows(list(files_dict.values())[0],
datasets=[orga.cfg.key_x_values])
generator = get_h5_generator(
orga,
files_dict,
zero_center=orga.cfg.zero_center_folder is not None,
keras_mode=False,
use_def_label=use_def_label,
phase="inference",
)
itergen = iter(generator)
if samples is None:
steps = len(generator)
else:
steps = int(samples / orga.cfg.batchsize)
print_every = max(100, min(int(round(steps / 10, -2)), 1000))
model_time_total = 0.
temp_output_path = os.path.join(
os.path.dirname(output_path), "temp_" + os.path.basename(output_path) +
"_" + time.strftime("%d-%m-%Y-%H-%M-%S", time.gmtime()))
print(f"Creating temporary file {temp_output_path}")
with h5py.File(temp_output_path, 'x') as h5_file:
# add version and paths of h5files
h5_file.attrs.create("orcanet", orcanet.__version__)
for input_key, file in files_dict.items():
h5_file.attrs.create(f"orcanet_inp_{input_key}", file)
for s in range(steps):
if s % print_every == 0:
print('Predicting in step {}/{} ({:0.2%})'.format(
s, steps, s / steps))
info_blob = next(itergen)
start_time = time.time()
y_pred = model.predict_on_batch(info_blob["xs"])
model_time_total += time.time() - start_time
if not isinstance(y_pred, list):
# if only one output, transform to a list
y_pred = [y_pred]
# transform y_pred to dict
y_pred = {
out: y_pred[i]
for i, out in enumerate(model.output_names)
}
info_blob["y_pred"] = y_pred
if info_blob.get("org_batchsize") is not None:
_slice_to_size(info_blob)
if orga.cfg.dataset_modifier is None:
datasets = dataset_modifiers.as_array(info_blob)
else:
datasets = orga.cfg.dataset_modifier(info_blob)
if s == 0: # create datasets in the first step
for dataset_name, data in datasets.items():
h5_file.create_dataset(
dataset_name,
data=data,
maxshape=(file_size, ) + data.shape[1:],
chunks=True, # (batchsize,) + data.shape[1:]
compression="gzip",
compression_opts=1,
)
else:
for dataset_name, data in datasets.items():
# append data at the end of the dataset
h5_file[dataset_name].resize(
h5_file[dataset_name].shape[0] + data.shape[0], axis=0)
h5_file[dataset_name][-data.shape[0]:] = data
if os.path.exists(output_path):
raise FileExistsError(
f"{output_path} exists already! But file {temp_output_path} "
f"is finished and can be safely used.")
os.rename(temp_output_path, output_path)
generator.print_timestats()
print("Statistics of model prediction:")
print(f"\tTotal time:\t{model_time_total / 60:.2f} min")
print(f"\tPer batch:\t{1000 * model_time_total / steps:.5} ms")
def test_y_field_names(self):
y_field_names = ("mc_A", )
filepaths = self.filepaths_file_1
self.orga.cfg.y_field_names = y_field_names
gene = get_h5_generator(self.orga, filepaths, keras_mode=False)
self.assertTupleEqual(gene[0]["y_values"].dtype.names, y_field_names)
