def summarize_dirs(self):
"""
Get the best and worst epochs of all given folders as a dict.
Returns
-------
minima : dict
Keys : Name of folder.
Values : [Epoch, metric] of where the metric is lowest.
maxima : dict
As above, but for where the metric is highest.
"""
minima, maxima = {}, {}
for folder_no, folder in enumerate(self.folders):
hist = HistoryHandler(folder)
smry_met_name = self._full_metrics[folder_no]
try:
max_line = hist.get_best_epoch_info(metric=smry_met_name,
mini=False)
min_line = hist.get_best_epoch_info(metric=smry_met_name,
mini=True)
except OSError as e:
warnings.warn(str(e))
continue
minima[folder] = [min_line["Epoch"], min_line[smry_met_name]]
maxima[folder] = [max_line["Epoch"], max_line[smry_met_name]]
return minima, maxima
def __init__(self,
output_folder,
list_file=None,
config_file=None,
tf_log_level=None,
discover_tomls=True):
"""
Set the attributes of the Configuration object.
Instead of using a config_file, the attributes of orga.cfg can
also be changed directly, e.g. by calling orga.cfg.batchsize.
Parameters
----------
output_folder : str
Name of the folder of this model in which everything will be saved,
e.g., the summary.txt log file is located in here.
Will be used to load saved files or to save new ones.
list_file : str, optional
Path to a toml list file with pathes to all the h5 files that should
be used for training and validation.
Will be used to extract samples and labels.
Default: Look for a file called 'list.toml' in the given output_folder.
config_file : str, optional
Path to a toml config file with settings that are used instead of
the default ones.
Default: Look for a file called 'config.toml' in the given output_folder.
tf_log_level : int/str
Sets the TensorFlow CPP_MIN_LOG_LEVEL environment variable.
0 = all messages are logged (default behavior).
1 = INFO messages are not printed.
2 = INFO and WARNING messages are not printed.
3 = INFO, WARNING, and ERROR messages are not printed.
discover_tomls : bool
If False, do not try to look for toml files in the given
output_folder if list_file or config_file is None [Default: True].
"""
if tf_log_level is not None:
os.environ['TF_CPP_MIN_LOG_LEVEL'] = str(tf_log_level)
if discover_tomls and list_file is None:
list_file = orcanet.misc.find_file(output_folder, "list.toml")
if discover_tomls and config_file is None:
config_file = orcanet.misc.find_file(output_folder, "config.toml")
self.cfg = Configuration(output_folder, list_file, config_file)
self.io = IOHandler(self.cfg)
self.history = HistoryHandler(output_folder)
self.xs_mean = None
self._auto_label_modifier = None
self._stored_model = None
self._strategy = None
def _summarize_folder(self, folder_no):
label = self.labels[folder_no]
folder = self.folders[folder_no]
hist = HistoryHandler(folder)
val_data, min_stat, max_stat = None, None, None
# read data from summary file
try:
smry_met_name = self._full_metrics[folder_no]
max_line = hist.get_best_epoch_info(metric=smry_met_name,
mini=False)
min_line = hist.get_best_epoch_info(metric=smry_met_name,
mini=True)
min_stat = [min_line[smry_met_name], label, min_line["Epoch"]]
max_stat = [max_line[smry_met_name], label, max_line["Epoch"]]
summary_data = hist.get_summary_data()
val_data = [
summary_data["Epoch"],
summary_data[self._full_metrics[folder_no]]
]
except OSError:
if self.verbose:
print(f"Warning: No summary file found for {folder}")
except ValueError as e:
if self.verbose:
print(f"Error reading summary file {hist.summary_file} ({e})")
# read data from training files
full_train_data = hist.get_train_data()
train_data = [
full_train_data["Batch_float"],
full_train_data[self._metric_names[folder_no]]
]
if not self.noplot:
if len(self.labels) == 1:
train_label, val_label = "training", "validation"
elif val_data is None:
train_label, val_label = label, None
else:
train_label, val_label = None, label
self._tvp.plot_curves(train_data=train_data,
val_data=val_data,
train_label=train_label,
val_label=val_label,
smooth_sigma=self.smooth,
tlw=0.5 * self.width,
vlw=0.5 * self.width,
vms=3 * self.width**0.5)
return min_stat, max_stat
class TestHistoryHandler(TestCase):
"""
Test the in out Handler on 2 dummy summary files
(summary.txt, summary_2.txt).
"""
def setUp(self):
self.output_folder = os.path.join(os.path.dirname(__file__), "data",
"dummy_model")
self.history = HistoryHandler(self.output_folder)
self.summary_filename_2 = "summary_2.txt"
def test_get_metrics(self):
metrics = self.history.get_metrics()
target = ["loss", "acc"]
np.testing.assert_array_equal(metrics, target)
def test_get_summary_data(self):
summary_data = self.history.get_summary_data()
target = np.array(
[(0.03488, 0.005, 0.07776, 0.05562, 0.971, 0.9808),
(0.06971, 0.00465, 0.05034, np.nan, 0.9822, np.nan)],
dtype=[('Epoch', '<f8'), ('LR', '<f8'), ('train_loss', '<f8'),
('val_loss', '<f8'), ('train_acc', '<f8'),
('val_acc', '<f8')])
assert_equal_struc_array(summary_data, target)
def test_get_train_data(self):
train_data = self.history.get_train_data()
print(train_data)
target = np.array([(250., 0.00018755, 0.487258, 0.756687),
(500., 0.00056265, 0.291661, 0.874188),
(250., 0.0350691, 0.0589843, 0.977938),
(500., 0.0354442, 0.0554589, 0.980375)],
dtype=[('Batch', '<f8'), ('Batch_float', '<f8'),
('loss', '<f8'), ('acc', '<f8')])
assert_equal_struc_array(train_data, target)
def test_get_column_names(self):
column_names = self.history.get_column_names()
target = ('Epoch', 'LR', 'train_loss', 'val_loss', 'train_acc',
'val_acc')
self.assertSequenceEqual(column_names, target)
def test_get_state(self):
self.history.summary_filename = self.summary_filename_2
state = self.history.get_state()
print(state)
target = [
{
'epoch': 0.1,
'is_trained': True,
'is_validated': True
},
{
'epoch': 0.2,
'is_trained': False,
'is_validated': True
},
{
'epoch': 0.3,
'is_trained': True,
'is_validated': False
},
{
'epoch': 0.4,
'is_trained': False,
'is_validated': False
},
]
self.assertSequenceEqual(state, target)
def test_plot_metric_unknown_metric(self):
with self.assertRaises(ValueError):
self.history.plot_metric("test")
"""
@patch('orcanet.history.plot_history')
def test_plot_metric_loss(self, mock_plot_history):
def plot_history(train_data, val_data, **kwargs):
return train_data, val_data, kwargs
mock_plot_history.side_effect = plot_history
value_train, value_val, value_kwargs = self.history.plot_metric("loss")
target_train = [
np.array([0.000187551, 0.000562653, 0.0350691, 0.0354442]),
np.array([0.487258, 0.291661, 0.0589843, 0.0554589]),
]
target_val = [
np.array([0.03488, 0.06971]),
np.array([0.05562, np.nan]),
]
target_kwargs = {'y_label': 'loss'}
np.testing.assert_array_almost_equal(target_train, value_train)
np.testing.assert_array_almost_equal(target_val, value_val)
self.assertDictEqual(target_kwargs, value_kwargs)
@patch('orcanet.history.plot_history')
def test_plot_metric_acc(self, mock_plot_history):
def plot_history(train_data, val_data, **kwargs):
return train_data, val_data, kwargs
mock_plot_history.side_effect = plot_history
value_train, value_val, value_kwargs = self.history.plot_metric("acc")
target_train = [
np.array([0.000187551, 0.000562653, 0.0350691, 0.0354442]),
np.array([0.756687, 0.874188, 0.977938, 0.980375]),
]
target_val = [
np.array([0.03488, 0.06971]),
np.array([0.9808, np.nan]),
]
target_kwargs = {'y_label': 'acc'}
np.testing.assert_array_almost_equal(target_train, value_train)
np.testing.assert_array_almost_equal(target_val, value_val)
self.assertDictEqual(target_kwargs, value_kwargs)
@patch('orcanet.history.plot_history')
def test_plot_lr(self, mock_plot_history):
def plot_history(train_data, val_data, **kwargs):
return train_data, val_data, kwargs
mock_plot_history.side_effect = plot_history
value_train, value_val, value_kwargs = self.history.plot_lr()
target_train = None
target_val = [
np.array([0.03488, 0.06971]),
np.array([0.005, 0.00465])
]
target_kwargs = {
'y_label': 'Learning rate',
'legend': False
}
self.assertEqual(target_train, value_train)
np.testing.assert_array_almost_equal(target_val, value_val)
self.assertDictEqual(target_kwargs, value_kwargs)
"""
def test_get_best_epoch_info(self):
self.history.summary_filename = self.summary_filename_2
value = self.history.get_best_epoch_info()
target = np.array([(0.1, 0.11, 0.12, 0.13, 0.14, 0.15)],
dtype=[('Epoch', '<f8'), ('LR', '<f8'),
('train_loss', '<f8'), ('val_loss', '<f8'),
('train_acc', '<f8'), ('val_acc', '<f8')])
np.testing.assert_array_equal(target, value)
def test_get_best_epoch_fileno(self):
self.history.summary_filename = self.summary_filename_2
value = self.history.get_best_epoch_fileno()
target = (1, 1)
self.assertEqual(target, value)
def setUp(self):
self.output_folder = os.path.join(os.path.dirname(__file__), "data",
"dummy_model")
self.history = HistoryHandler(self.output_folder)
self.summary_filename_2 = "summary_2.txt"
class Organizer:
"""
Core class for working with networks in OrcaNet.
Attributes
----------
cfg : orcanet.core.Configuration
Contains all configurable options.
io : orcanet.in_out.IOHandler
Utility functions for accessing the info in cfg.
history : orcanet.in_out.HistoryHandler
For reading and plotting data from the log files created
during training.
"""
def __init__(self, output_folder,
list_file=None,
config_file=None,
tf_log_level=None,
discover_tomls=True):
"""
Set the attributes of the Configuration object.
Instead of using a config_file, the attributes of orga.cfg can
also be changed directly, e.g. by calling orga.cfg.batchsize.
Parameters
----------
output_folder : str
Name of the folder of this model in which everything will be saved,
e.g., the summary.txt log file is located in here.
Will be used to load saved files or to save new ones.
list_file : str, optional
Path to a toml list file with pathes to all the h5 files that should
be used for training and validation.
Will be used to extract samples and labels.
Default: Look for a file called 'list.toml' in the given output_folder.
config_file : str, optional
Path to a toml config file with settings that are used instead of
the default ones.
Default: Look for a file called 'config.toml' in the given output_folder.
tf_log_level : int/str
Sets the TensorFlow CPP_MIN_LOG_LEVEL environment variable.
0 = all messages are logged (default behavior).
1 = INFO messages are not printed.
2 = INFO and WARNING messages are not printed.
3 = INFO, WARNING, and ERROR messages are not printed.
discover_tomls : bool
If False, do not try to look for toml files in the given
output_folder if list_file or config_file is None [Default: True].
"""
if tf_log_level is not None:
os.environ['TF_CPP_MIN_LOG_LEVEL'] = str(tf_log_level)
if discover_tomls and list_file is None:
list_file = orcanet.misc.find_file(output_folder, "list.toml")
if discover_tomls and config_file is None:
config_file = orcanet.misc.find_file(output_folder, "config.toml")
self.cfg = Configuration(output_folder, list_file, config_file)
self.io = IOHandler(self.cfg)
self.history = HistoryHandler(output_folder)
self.xs_mean = None
self._auto_label_modifier = None
self._stored_model = None
self._strategy = None
def train_and_validate(self, model=None, epochs=None, to_epoch=None):
"""
Train a model and validate according to schedule.
The various settings of this process can be controlled with the
attributes of orca.cfg.
The model will be trained on the given data, saved and validated.
Logfiles of the training are saved in the output folder.
Plots showing the training and validation history, as well as
the weights and activations of the network are generated in
the plots subfolder after every validation.
The training can be resumed by executing this function again.
Parameters
----------
model : ks.models.Model or str, optional
Compiled keras model to use for training. Required for the first
epoch (the start of training).
Can also be the path to a saved keras model, which will be laoded.
If model is None, the most recent saved model will be
loaded automatically to continue the training.
epochs : int, optional
How many epochs should be trained by running this function.
None for infinite. This includes the current epoch in case it
is not finished yet, i.e. 1 means complete the epoch if there
are files left, otherwise do the next epoch.
to_epoch : int, optional
Train up to and including this epoch. Can not be used together with
epochs.
Returns
-------
model : ks.models.Model
The trained keras model.
"""
latest_epoch = self.io.get_latest_epoch()
model = self._get_model(model, logging=False)
self._stored_model = model
# check if the validation is missing for the latest fileno
if latest_epoch is not None:
state = self.history.get_state()[-1]
if state["is_validated"] is False and self.val_is_due(latest_epoch):
self.validate()
next_epoch = self.io.get_next_epoch(latest_epoch)
n_train_files = self.io.get_no_of_files("train")
if to_epoch is None:
epochs_left = epochs
else:
if epochs is not None:
raise ValueError("Can not give both 'epochs' and 'to_epoch'")
if latest_epoch is None:
epochs_left = to_epoch
else:
epochs_left = max(
0, to_epoch - self.io.get_next_epoch(latest_epoch)[0] + 1)
trained_epochs = 0
while epochs_left is None or trained_epochs < epochs_left:
# Train on remaining files
for file_no in range(next_epoch[1], n_train_files + 1):
curr_epoch = (next_epoch[0], file_no)
self.train(model)
if self.val_is_due(curr_epoch):
self.validate()
next_epoch = (next_epoch[0] + 1, 1)
trained_epochs += 1
self._stored_model = None
return model
def train(self, model=None):
"""
Trains a model on the next file.
The progress of the training is also logged and plotted.
Parameters
----------
model : ks.models.Model or str, optional
Compiled keras model to use for training. Required for the first
epoch (the start of training).
Can also be the path to a saved keras model, which will be laoded.
If model is None, the most recent saved model will be
loaded automatically to continue the training.
Returns
-------
history : dict
The history of the training on this file. A record of training
loss values and metrics values.
"""
# Create folder structure
self.io.get_subfolder(create=True)
latest_epoch = self.io.get_latest_epoch()
model = self._get_model(model, logging=True)
self._set_up(model, logging=True)
# epoch about to be trained
next_epoch = self.io.get_next_epoch(latest_epoch)
next_epoch_float = self.io.get_epoch_float(*next_epoch)
if latest_epoch is None:
self.io.check_connections(model)
logging.log_start_training(self)
model_path = self.io.get_model_path(*next_epoch)
model_path_local = self.io.get_model_path(*next_epoch, local=True)
if os.path.isfile(model_path):
raise FileExistsError(
"Can not train model in epoch {} file {}, this model has "
"already been saved!".format(*next_epoch))
smry_logger = logging.SummaryLogger(self, model)
if self.cfg.learning_rate is not None:
tf.keras.backend.set_value(
model.optimizer.lr, self.io.get_learning_rate(next_epoch)
)
files_dict = self.io.get_file("train", next_epoch[1])
line = "Training in epoch {} on file {}/{}".format(
next_epoch[0], next_epoch[1], self.io.get_no_of_files("train"))
self.io.print_log(line)
self.io.print_log("-" * len(line))
self.io.print_log("Learning rate is at {}".format(
tf.keras.backend.get_value(model.optimizer.lr)))
self.io.print_log('Inputs and files:')
for input_name, input_file in files_dict.items():
self.io.print_log(" {}: \t{}".format(input_name,
os.path.basename(
input_file)))
start_time = time.time()
history = backend.train_model(self, model, next_epoch, batch_logger=True)
elapsed_s = int(time.time() - start_time)
model.save(model_path)
smry_logger.write_line(
next_epoch_float,
tf.keras.backend.get_value(model.optimizer.lr),
history_train=history,
)
self.io.print_log('Training results:')
for metric_name, loss in history.items():
self.io.print_log(f" {metric_name}: \t{loss}")
self.io.print_log(f"Elapsed time: {timedelta(seconds=elapsed_s)}")
self.io.print_log(f"Saved model to: {model_path_local}\n")
update_summary_plot(self)
if self.cfg.cleanup_models:
self.cleanup_models()
return history
def validate(self):
"""
Validate the most recent saved model on all validation files.
Will also log the progress, as well as update the summary plot and
plot weights and activations of the model.
Returns
-------
history : dict
The history of the validation on all files. A record of validation
loss values and metrics values.
"""
latest_epoch = self.io.get_latest_epoch()
if latest_epoch is None:
raise ValueError("Can not validate: No saved model found")
if self.history.get_state()[-1]["is_validated"] is True:
raise ValueError("Can not validate in epoch {} file {}: "
"Has already been validated".format(*latest_epoch))
if self._stored_model is None:
model = self.load_saved_model(*latest_epoch)
else:
model = self._stored_model
self._set_up(model, logging=True)
epoch_float = self.io.get_epoch_float(*latest_epoch)
smry_logger = logging.SummaryLogger(self, model)
logging.log_start_validation(self)
start_time = time.time()
history = backend.validate_model(self, model)
elapsed_s = int(time.time() - start_time)
self.io.print_log('Validation results:')
for metric_name, loss in history.items():
self.io.print_log(f" {metric_name}: \t{loss}")
self.io.print_log(f"Elapsed time: {timedelta(seconds=elapsed_s)}\n")
smry_logger.write_line(epoch_float, "n/a", history_val=history)
update_summary_plot(self)
if self.cfg.cleanup_models:
self.cleanup_models()
return history
def predict(self, epoch=None, fileno=None, samples=None):
"""
Make a prediction if it does not exist yet, and return its filepath.
Load the model with the lowest validation loss, let it predict on
all samples of the validation set
in the toml list, and save this prediction together with all the
y_values as h5 file(s) in the predictions subfolder.
Parameters
----------
epoch : int, optional
Epoch of a model to load. Default: lowest val loss.
fileno : int, optional
File number of a model to load. Default: lowest val loss.
samples : int, optional
Don't use the full validation files, but just the given number
of samples.
Returns
-------
pred_filename : List
List to the paths of all the prediction files.
"""
if fileno is None and epoch is None:
epoch, fileno = self.history.get_best_epoch_fileno()
print(f"Automatically set epoch to epoch {epoch} file {fileno}.")
elif fileno is None or epoch is None:
raise ValueError(
"Either both or none of epoch and fileno must be None")
if self._check_if_pred_already_done(epoch, fileno):
print("Prediction has already been done.")
pred_filepaths = self.io.get_pred_files_list(epoch, fileno)
else:
if self._stored_model is None:
model = self.load_saved_model(epoch, fileno, logging=False)
else:
model = self._stored_model
self._set_up(model)
start_time = time.time()
backend.make_model_prediction(
self, model, epoch, fileno, samples=samples)
elapsed_s = int(time.time() - start_time)
print('Finished predicting on all validation files.')
print("Elapsed time: {}\n".format(timedelta(seconds=elapsed_s)))
pred_filepaths = self.io.get_pred_files_list(epoch, fileno)
return pred_filepaths
def inference(self, epoch=None, fileno=None, as_generator=False):
"""
Make an inference and return the filepaths.
Load the model with the lowest validation loss, let
it predict on all samples of all inference files
in the toml list, and save these predictions as h5 files in the
predictions subfolder. y values will only be added if they are in
the input file, so this can be used on un-labeled data as well.
Parameters
----------
epoch : int, optional
Epoch of a model to load [default: lowest val loss].
fileno : int, optional
File number of a model to load [default: lowest val loss].
as_generator : bool
If true, return a generator, which yields the output filename
after the inference of each file.
If false (default), do all files back to back.
Returns
-------
filenames : list
List to the paths of all created output files.
"""
if fileno is None and epoch is None:
epoch, fileno = self.history.get_best_epoch_fileno()
print("Automatically set epoch to epoch {} file {}.".format(epoch, fileno))
elif fileno is None or epoch is None:
raise ValueError(
"Either both or none of epoch and fileno must be None")
if self._stored_model is None:
model = self.load_saved_model(epoch, fileno, logging=False)
else:
model = self._stored_model
self._set_up(model)
filenames = []
for files_dict in self.io.yield_files("inference"):
# output filename is based on name of file in first input
first_filename = os.path.basename(list(files_dict.values())[0])
output_filename = "model_epoch_{}_file_{}_on_{}".format(
epoch, fileno, first_filename)
output_path = os.path.join(self.io.get_subfolder("inference"),
output_filename)
filenames.append(output_path)
if os.path.exists(output_path):
warnings.warn("Warning: {} exists already, skipping "
"file".format(output_filename))
continue
print(f'Working on file {first_filename}')
start_time = time.time()
backend.h5_inference(
self, model, files_dict, output_path, use_def_label=False)
elapsed_s = int(time.time() - start_time)
print(f'Finished on file {first_filename} in {elapsed_s/60} min')
if as_generator:
yield output_path
return filenames
def cleanup_models(self):
"""
Delete all models except for the the most recent one (to continue
training), and the ones with the highest and lowest loss/metrics.
"""
all_epochs = self.io.get_all_epochs()
epochs_to_keep = {self.io.get_latest_epoch(), }
try:
for metric in self.history.get_metrics():
epochs_to_keep.add(
self.history.get_best_epoch_fileno(
metric=f"val_{metric}", mini=True))
epochs_to_keep.add(
self.history.get_best_epoch_fileno(
metric=f"val_{metric}", mini=False))
except ValueError:
# no best epoch exists
pass
for epoch in epochs_to_keep:
if epoch not in all_epochs:
warnings.warn(
f"ERROR: keeping_epoch {epoch} not in available epochs {all_epochs}, "
f"skipping clean-up of models!")
return
print("\nClean-up saved models:")
for epoch in all_epochs:
model_path = self.io.get_model_path(epoch[0], epoch[1])
model_name = os.path.basename(model_path)
if epoch in epochs_to_keep:
print("Keeping model {}".format(model_name))
else:
print("Deleting model {}".format(model_name))
os.remove(model_path)
def _check_if_pred_already_done(self, epoch, fileno):
"""
Checks if the prediction has already been done before.
(-> predicted on all validation files)
Returns
-------
pred_done : bool
Boolean flag to specify if the prediction has
already been fully done or not.
"""
latest_pred_file_no = self.io.get_latest_prediction_file_no(epoch, fileno)
total_no_of_val_files = self.io.get_no_of_files('val')
if latest_pred_file_no is None:
pred_done = False
elif latest_pred_file_no == total_no_of_val_files:
return True
else:
pred_done = False
return pred_done
def get_xs_mean(self, logging=False):
"""
Set and return the zero center image for each list input.
Requires the cfg.zero_center_folder to be set. If no existing
image for the given input files is found in the folder, it will
be calculated and saved by averaging over all samples in the
train dataset.
Parameters
----------
logging : bool
If true, the execution of this function will be logged into the
full summary in the output folder if called for the first time.
Returns
-------
dict
Dict of numpy arrays that contains the mean_image of the x dataset
(1 array per list input).
Example format:
{ "input_A" : ndarray, "input_B" : ndarray }
"""
if self.xs_mean is None:
if self.cfg.zero_center_folder is None:
raise ValueError("Can not calculate zero center: "
"No zero center folder given")
self.xs_mean = load_zero_center_data(self, logging=logging)
return self.xs_mean
def load_saved_model(self, epoch, fileno, logging=False):
"""
Load a saved model.
Parameters
----------
epoch : int
Epoch of the saved model. If both this and fileno are -1,
load the most recent model.
fileno : int
Fileno of the saved model.
logging : bool
If True, will log this function call into the log.txt file.
Returns
-------
model : keras model
"""
path_of_model = self.io.get_model_path(epoch, fileno)
path_loc = self.io.get_model_path(epoch, fileno, local=True)
self.io.print_log("Loading saved model: " + path_loc, logging=logging)
return self._load_model(path_of_model)
def _get_model(self, model, logging=False):
""" Load most recent saved model or use user model. """
latest_epoch = self.io.get_latest_epoch()
if latest_epoch is None:
# new training, log info about model
if model is None:
raise ValueError("You need to provide a compiled keras model "
"for the start of the training! (You gave None)")
elif isinstance(model, str):
# path to a saved model
self.io.print_log("Loading model from " + model, logging=logging)
model = self._load_model(model)
if logging:
self._save_as_json(model)
model.summary(print_fn=self.io.print_log)
try:
plots_folder = self.io.get_subfolder("plots", create=True)
tf.keras.utils.plot_model(
model, plots_folder + "/model_plot.png", show_shapes=True)
except (ImportError, AttributeError) as e:
# TODO remove AttributeError once https://github.com/tensorflow/tensorflow/issues/38988 is fixed
warnings.warn("Can not plot model: " + str(e))
else:
# resuming training, load model if it is not given
if model is None:
model = self.load_saved_model(*latest_epoch, logging=logging)
elif isinstance(model, str):
# path to a saved model
self.io.print_log("Loading model from " + model, logging=logging)
model = self._load_model(model)
return model
def _load_model(self, filepath):
""" Load from path, with custom objects and parallized. """
with self.get_strategy().scope():
model = tf.keras.models.load_model(
filepath, custom_objects=self.cfg.get_custom_objects())
return model
def _save_as_json(self, model):
""" Save the architecture of a model as json to fixed path. """
json_filename = "model_arch.json"
json_string = model.to_json(indent=1)
model_folder = self.io.get_subfolder("saved_models", create=True)
with open(os.path.join(model_folder, json_filename), "w") as f:
f.write(json_string)
def _set_up(self, model, logging=False):
""" Necessary setup for training, validating and predicting. """
if self.cfg.get_list_file() is None:
raise ValueError("No files specified. Need to load a toml "
"list file with pathes to h5 files first.")
if self.cfg.label_modifier is None:
self._auto_label_modifier = lib.label_modifiers.ColumnLabels(model)
if self.cfg.zero_center_folder is not None:
self.get_xs_mean(logging)
def val_is_due(self, epoch=None):
"""
True if validation is due on given epoch according to schedule.
Does not check if it has been done already.
"""
if epoch is None:
epoch = self.io.get_latest_epoch()
n_train_files = self.io.get_no_of_files("train")
val_sched = (epoch[1] == n_train_files) or \
(self.cfg.validate_interval is not None and
epoch[1] % self.cfg.validate_interval == 0)
return val_sched
def get_strategy(self):
""" Get the strategy for distributed training. """
if self._strategy is None:
if self.cfg.multi_gpu and len(
tf.config.list_physical_devices('GPU')) > 1:
self._strategy = tf.distribute.MirroredStrategy()
print(f'Number of GPUs: {self._strategy.num_replicas_in_sync}')
else:
self._strategy = tf.distribute.get_strategy()
return self._strategy