def start_experiment(self, engine):
# log the initial artifacts in the dir
mlflow.log_artifacts(args.logdir, "training")
mlflow.log_metric("finished", False)
def on_train_end(self, args, state, control, **kwargs):
if self._initialized and state.is_world_process_zero:
if self._log_artifacts:
logger.info("Logging artifacts. This may take time.")
mlflow.log_artifacts(args.output_dir)
mlflow.end_run()
log_scalar('test_loss', test_loss, step)
log_scalar('test_accuracy', test_accuracy, step)
def log_scalar(name, value, step):
"""Log a scalar value to both MLflow and TensorBoard"""
writer.add_scalar(name, value, step)
mlflow.log_metric(name, value)
with mlflow.start_run():
# Log our parameters into mlflow
for key, value in vars(args).items():
mlflow.log_param(key, value)
# Create a SummaryWriter to write TensorBoard events locally
output_dir = dirpath = tempfile.mkdtemp()
writer = SummaryWriter(output_dir)
print("Writing TensorBoard events locally to %s\n" % output_dir)
# Perform the training
for epoch in range(1, args.epochs + 1):
train(epoch)
test(epoch)
# Upload the TensorBoard event logs as a run artifact
print("Uploading TensorBoard events as a run artifact...")
mlflow.log_artifacts(output_dir, artifact_path="events")
print("\nLaunch TensorBoard with:\n\ntensorboard --logdir=%s" %
os.path.join(mlflow.get_artifact_uri(), "events"))
def run_evaluation(
*,
model_hash: str,
thr: float = 0.5,
dpath: str = "datasets/clouds/38-Cloud/38-Cloud_test",
gtpath: str = "datasets/clouds/38-Cloud/38-Cloud_test/Entire_scene_gts",
vpath: str = "datasets/clouds/38-Cloud/38-Cloud_test/Natural_False_Color",
rpath: str = "artifacts/",
vids: ("v", multi(min=1)),
batch_size: int = 32,
img_ids: ("iid", multi(min=0)),
mlflow: bool = False,
run_name: str = None,
):
"""
Load model given model hash and get evaluation metrics on 38-Cloud testset.
:param model_hash: MLFlow hash of the model to load.
:param thr: threshold to be used during evaluation.
:param dpath: path to dataset.
:param gtpath: path to dataset ground truths.
:param vpath: path to dataset (false color) visualisation images.
:param rpath: path to directory where results
and artifacts should be logged.
:param vids: tuple of ids of images which should be used to create
visualisations. If contains '*' visualisations will be
created for all images in the dataset.
:type vids: tuple[str]
:param batch_size: size of generated batches, only one batch is loaded
to memory at a time.
:param img_ids: if given, process only these images.
:type img_ids: list[int]
:param mlflow: whether to use MLFlow.
:param run_name: name of the run.
"""
snow_imgs = [
"LC08_L1TP_064015_20160420_20170223_01_T1",
"LC08_L1TP_035035_20160120_20170224_01_T1",
"LC08_L1TP_050024_20160520_20170324_01_T1",
]
if img_ids == []:
img_ids = None
else:
snow_imgs = list(set(snow_imgs) & set(img_ids))
dpath, gtpath, vpath, rpath = make_paths(dpath, gtpath, vpath, rpath)
rpath = rpath / uuid.uuid4().hex
print(f"Working dir: {os.getcwd()}, "
+ f"artifacts dir: {rpath}",
flush=True)
mpath = Path(
f"/media/ML/mlflow/beetle/artifacts/34/{model_hash}/"
# Change init_model to model for old models
+ "artifacts/init_model/data/model.h5"
)
# WARNING: If ever upgraded to tf2.x, this way of using metrics will not work
# because loaded metrics will become MeanMetricWrapper objects in tf2.x and
# this script isn't prepared for such objects (because MeanMetricWrapper has state,
# as opposed to present stateless metric functions).
model = keras.models.load_model(
mpath,
custom_objects={
"jaccard_index_loss": losses.JaccardIndexLoss(),
"jaccard_index_metric": losses.JaccardIndexMetric(),
"dice_coeff_metric": losses.DiceCoefMetric(),
"recall": losses.recall,
"precision": losses.precision,
"specificity": losses.specificity,
# F1 score is needed for old models
# "f1_score": losses.f1_score,
"tf": tf,
},
)
model.load_weights(
f"/media/ML/mlflow/beetle/artifacts/34/{model_hash}/"
+ "artifacts/best_weights/best_weights"
)
metrics, _ = evaluate_model(
model=model,
thr=thr,
dpath=dpath,
gtpath=gtpath,
vpath=vpath,
rpath=rpath,
vids=vids,
batch_size=batch_size,
img_ids=img_ids,
mlflow=mlflow,
run_name=run_name,
)
mean_metrics = {}
mean_metrics_snow = {}
for key, value in metrics.items():
mean_metrics[key] = np.mean(list(value.values()))
mean_metrics_snow[f"snow_{key}"] = np.mean(
[value[x] for x in snow_imgs])
print(mean_metrics, mean_metrics_snow)
if mlflow:
log_metrics(mean_metrics)
log_metrics(mean_metrics_snow)
log_artifacts(rpath)
def log_artifacts(cls, dir_path, artifact_path=None):
mlflow.log_artifacts(dir_path, artifact_path)
def main():
with mlflow.start_run(run_name="Gen Expert Data"):
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('expert_policy_file', type=str)
parser.add_argument('envname', type=str)
parser.add_argument('--render',
type=str2bool,
nargs='?',
const=True,
default=False)
parser.add_argument("--max_timesteps", type=int)
parser.add_argument('--num_rollouts',
type=int,
default=20,
help='Number of expert roll outs')
args = parser.parse_args()
for k, v in vars(args).items():
mlflow.log_param(k, v)
print('loading and building expert policy')
policy_fn = load_policy.load_policy(args.expert_policy_file)
print('loaded and built')
with tf.Session():
tf_util.initialize()
returns = []
observations = []
actions = []
steps = []
for i in range(args.num_rollouts):
env = gym.make(args.envname)
if args.render:
video_dir = "./videos/{0}/".format(time())
env = wrappers.Monitor(env, video_dir, force=True)
max_steps = args.max_timesteps or env.spec.max_episode_steps
print("max_steps set to {0}".format(max_steps))
print('iter', i)
obs = env.reset()
done = False
totalr = 0.
trial_steps = 0
while not done:
action = policy_fn(obs[None, :])
#action = env.action_space.sample()
observations.append(obs)
actions.append(action)
obs, r, done, _ = env.step(action)
totalr += r
trial_steps += 1
if args.render:
env.render(mode='rgb_array')
if trial_steps % 100 == 0:
print("%i/%i" % (trial_steps, max_steps))
if trial_steps >= max_steps:
print("hit max_steps")
break
returns.append(totalr)
steps.append(trial_steps)
env.close()
if args.render:
mlflow.log_artifacts(video_dir)
for s in steps:
mlflow.log_metric('steps', s)
for r in returns:
mlflow.log_metric('returns', r)
mlflow.log_metric('mean return', np.mean(returns))
mlflow.log_metric('std of return', np.std(returns))
expert_data = {
'observations': np.array(observations),
'actions': np.array(actions)
}
if not os.path.exists('expert_data'):
os.makedirs('expert_data')
filename = os.path.join('expert_data', args.envname + '.pkl')
with open((filename), 'wb') as f:
pickle.dump(expert_data, f, pickle.HIGHEST_PROTOCOL)
mlflow.log_artifact(filename, artifact_path="expert_data_file")
def train_models(self, args, base_line=True):
"""
Train the model and log all the MLflow Metrics
:param args: command line arguments. If no arguments then use default
:param base_line: Default flag. Create Baseline model
"""
# Create TensorFlow Session
sess = tf.InteractiveSession()
# Configure output_dir
output_dir = tempfile.mkdtemp()
#
# initialize some classes
#
kdata_cls = KIMDB_Data_Utils()
ktrain_cls = KTrain()
kplot_cls = KPlot()
#
# get IMDB Data
#
(train_data,
train_labels), (test_data, test_labels) = kdata_cls.fetch_imdb_data()
#
# prepare and vectorize data
#
x_train = kdata_cls.prepare_vectorized_sequences(train_data)
x_test = kdata_cls.prepare_vectorized_sequences(test_data)
y_train = kdata_cls.prepare_vectorized_labels(train_labels)
y_test = kdata_cls.prepare_vectorized_labels(test_labels)
image_dir = ktrain_cls.get_directory_path("images")
model_dir = ktrain_cls.get_directory_path("models")
graph_label_loss = 'Baseline Model: Training and Validation Loss'
graph_label_acc = 'Baseline Model: Training and Validation Accuracy'
graph_image_loss_png = os.path.join(image_dir, 'baseline_loss.png')
graph_image_acc_png = os.path.join(image_dir, 'baseline_accuracy.png')
if not base_line:
graph_label_loss = 'Experimental: Training and Validation Loss'
graph_label_acc = 'Experimental Model: Training and Validation Accuracy'
graph_image_loss_png = os.path.join(image_dir,
'experimental_loss.png')
graph_image_acc_png = os.path.join(image_dir,
'experimental_accuracy.png')
kmodel = KModel()
if base_line:
print("Baseline Model:")
model = kmodel.build_basic_model()
else:
print("Experiment Model:")
model = kmodel.build_experimental_model(args.hidden_layers,
args.output)
history = ktrain_cls.compile_and_fit_model(model,
x_train,
y_train,
epochs=args.epochs,
loss=args.loss,
output_dir=output_dir)
model.summary()
ktrain_cls.print_metrics(history)
figure_loss = kplot_cls.plot_loss_graph(history, graph_label_loss)
figure_loss.savefig(graph_image_loss_png)
figure_acc = kplot_cls.plot_accuracy_graph(history, graph_label_acc)
figure_acc.savefig(graph_image_acc_png)
results = ktrain_cls.evaluate_model(model, x_test, y_test)
print("Average Probability Results:")
print(results)
print()
print("Predictions Results:")
predictions = model.predict(x_test)
print(predictions)
mlflow_server = args.tracking_server
#
# We don't want to force people to have tracking server
# running on localhost as it tracks in mlruns directory
if mlflow_server:
# Tracking URI
if not mlflow_server.startswith("http"):
mlflow_tracking_uri = 'http://' + mlflow_server + ':5000'
else:
mlflow_tracking_uri = mlflow_server
# Set the Tracking URI
mlflow.set_tracking_uri(mlflow_tracking_uri)
print("MLflow Tracking URI: %s" % mlflow_tracking_uri)
else:
print("MLflow Tracking URI: %s" % "local directory 'mlruns'")
with mlflow.start_run():
# print out current run_uuid
run_uuid = mlflow.active_run().info.run_uuid
print("MLflow Run ID: %s" % run_uuid)
# log parameters
mlflow.log_param("hidden_layers", args.hidden_layers)
mlflow.log_param("output", args.output)
mlflow.log_param("epochs", args.epochs)
mlflow.log_param("loss_function", args.loss)
# calculate metrics
binary_loss = ktrain_cls.get_binary_loss(history)
binary_acc = ktrain_cls.get_binary_acc(history)
validation_loss = ktrain_cls.get_validation_loss(history)
validation_acc = ktrain_cls.get_validation_acc(history)
average_loss = results[0]
average_acc = results[1]
# log metrics
mlflow.log_metric("binary_loss", binary_loss)
mlflow.log_metric("binary_acc", binary_acc)
mlflow.log_metric("validation_loss", validation_loss)
mlflow.log_metric("validation_acc", validation_acc)
mlflow.log_metric("average_loss", average_loss)
mlflow.log_metric("average_acc", average_acc)
# log artifacts
mlflow.log_artifacts(image_dir, "images")
# log model
mlflow.keras.log_model(model, "models")
# save model locally
pathdir = "keras_models/" + run_uuid
model_dir = self.get_directory_path(pathdir, False)
ktrain_cls.keras_save_model(model, model_dir)
# Write out TensorFlow events as a run artifact
print("Uploading TensorFlow events as a run artifact.")
mlflow.log_artifacts(output_dir, artifact_path="events")
print("loss function use", args.loss)
numpy.ones((points.shape[0], 1),
dtype=numpy.float64))))
for t, points in zip(timepoints, samples)]
ret = list(
scopyon.generate_images(inputs,
num_frames=num_frames,
config=config,
rng=rng,
full_output=True))
inputs_ = []
for t, data in inputs:
inputs_.extend(([t] + list(row) for row in data))
inputs_ = numpy.array(inputs_)
numpy.save(artifacts / f"inputs{i:03d}.npy", inputs_)
numpy.save(artifacts / f"images{i:03d}.npy",
numpy.array([img.as_array() for img, infodict in ret]))
true_data = []
for t, (_, infodict) in zip(timepoints, ret):
true_data.extend([t, key] + list(value)
for key, value in infodict['true_data'].items())
true_data = numpy.array(true_data)
numpy.save(artifacts / f"true_data{i:03d}.npy", true_data)
#!ls ./artifacts
log_artifacts(artifactsPath)
mlflow.end_run()
def train_models(self, args, base_line=True):
"""
Train the model and log all the MLflow Metrics
:param args: command line arguments. If no arguments then use default
:param base_line: Default flag. Create Baseline model
"""
#
# initialize some classes
#
kdata_cls = KIMDB_Data_Utils()
ktrain_cls = KTrain()
kplot_cls = KPlot()
start_time = time()
#
# get IMDB Data
#
(train_data, train_labels), (test_data, test_labels) = kdata_cls.fetch_imdb_data()
#
# prepare and vectorize data
#
x_train = kdata_cls.prepare_vectorized_sequences(train_data)
x_test = kdata_cls.prepare_vectorized_sequences(test_data)
y_train = kdata_cls.prepare_vectorized_labels(train_labels)
y_test = kdata_cls.prepare_vectorized_labels(test_labels)
image_dir = ktrain_cls.get_directory_path("images")
model_dir = ktrain_cls.get_directory_path("models")
graph_label_loss = 'Baseline Model: Training and Validation Loss'
graph_label_acc = 'Baseline Model: Training and Validation Accuracy'
graph_image_loss_png = os.path.join(image_dir,'baseline_loss.png')
graph_image_acc_png = os.path.join(image_dir, 'baseline_accuracy.png')
if not base_line:
graph_label_loss = 'Experimental: Training and Validation Loss'
graph_label_acc = 'Experimental Model: Training and Validation Accuracy'
graph_image_loss_png = os.path.join(image_dir, 'experimental_loss.png')
graph_image_acc_png = os.path.join(image_dir,'experimental_accuracy.png')
kmodel = KModel()
if base_line:
print("Baseline Model:")
model = kmodel.build_basic_model()
else:
print("Experiment Model:")
model = kmodel.build_experimental_model(args.hidden_layers, args.output)
history = ktrain_cls.compile_and_fit_model(model, x_train, y_train, epochs=args.epochs, loss=args.loss)
model.summary()
ktrain_cls.print_metrics(history)
figure_loss = kplot_cls.plot_loss_graph(history, graph_label_loss)
figure_loss.savefig(graph_image_loss_png )
figure_acc = kplot_cls.plot_accuracy_graph(history, graph_label_acc)
figure_acc.savefig(graph_image_acc_png)
results = ktrain_cls.evaulate_model(model, x_test, y_test)
print("Average Probability Results:")
print(results)
print()
print("Predictions Results:")
predictions = model.predict(x_test)
print(predictions)
timed = time() - start_time
with mlflow.start_run():
# log parameters
mlflow.log_param("hidden_layers", args.hidden_layers)
mlflow.log_param("output", args.output)
mlflow.log_param("epochs", args.epochs)
mlflow.log_param("loss_function", args.loss)
# log metrics
mlflow.log_metric("binary_loss", ktrain_cls.get_binary_loss(history))
mlflow.log_metric("binary_acc", ktrain_cls.get_binary_acc(history))
mlflow.log_metric("validation_loss", ktrain_cls.get_binary_loss(history))
mlflow.log_metric("validation_acc", ktrain_cls.get_validation_acc(history))
mlflow.log_metric("average_loss", results[0])
mlflow.log_metric("average_acc", results[1])
# log artifacts
mlflow.log_artifacts(image_dir, "images")
# log model
mlflow.keras.log_model(model, "models")
print("This model took", timed, " seconds to train and test.")
print("loss function use", args.loss)
def train_model(args, base_line=True):
'''
Train model function
'''
graph_label_loss = 'Baseline Model: Training and Validation Loss'
graph_label_acc = 'Baseline Model: Training and Validation Accuracy'
graph_image_loss_png = os.path.join(image_dir, 'baseline_loss.png')
graph_image_acc_png = os.path.join(image_dir, 'baseline_accuracy.png')
if not base_line:
graph_label_loss = 'Experimental: Training and Validation Loss'
graph_label_acc = 'Experimental Model: Training and Validation Accuracy'
graph_image_loss_png = os.path.join(image_dir, 'experimental_loss.png')
graph_image_acc_png = os.path.join(image_dir,
'experimental_accuracy.png')
image_data_generator = tf.keras.preprocessing.image.ImageDataGenerator(
validation_split=validation_split)
train_generator = image_data_generator.flow_from_directory(
TRAIN_DATA_DIR,
target_size=(TRAIN_IMAGE_SIZE, TRAIN_IMAGE_SIZE),
batch_size=TRAIN_BATCH_SIZE,
class_mode='categorical',
subset='training')
validation_generator = image_data_generator.flow_from_directory(
TRAIN_DATA_DIR,
target_size=(TRAIN_IMAGE_SIZE, TRAIN_IMAGE_SIZE),
batch_size=TRAIN_BATCH_SIZE,
class_mode='categorical',
subset='validation')
# Create the model
model = Sequential()
model.add(
Conv2D(args.filters,
kernel_size=args.kernel_size,
activation='relu',
padding='same',
input_shape=(img_width, img_height, img_num_channels)))
model.add(Flatten())
model.add(Dense(args.output, activation='softmax'))
# Compile the model
model.compile(loss=args.loss,
optimizer=args.optimizer,
metrics=['accuracy'])
history = model.fit_generator(train_generator,
epochs=args.epochs,
validation_data=validation_generator)
model.summary()
print_metrics(history)
figure_loss = plot_loss_graph(history, graph_label_loss)
figure_loss.savefig(graph_image_loss_png)
figure_acc = plot_accuracy_graph(history, graph_label_acc)
figure_acc.savefig(graph_image_acc_png)
# print('==================================================')
# predictions = model.predict(TEST_DATA_DIR)
# print(predictions)
# print('==================================================')
#mlflow.set_experiment(args.experiment_name)
with mlflow.start_run():
# print out current run_uuid
run_uuid = mlflow.active_run().info.run_uuid
print("MLflow Run ID: %s" % run_uuid)
# mlflow.create_experiment("Training CNN Model", artifact_location=None)
# log parameters
mlflow.log_param("Filters", args.filters)
mlflow.log_param("Kernel Size", args.kernel_size)
mlflow.log_param("Output", args.output)
mlflow.log_param("Epochs", args.epochs)
mlflow.log_param("Loss", args.loss)
mlflow.log_param("Optimize", args.optimizer)
# calculate metrics
binary_loss = get_binary_loss(history)
binary_acc = get_binary_acc(history)
validation_loss = get_validation_loss(history)
validation_acc = get_validation_acc(history)
# log metrics
mlflow.log_metric("binary_loss", binary_loss)
mlflow.log_metric("binary_acc", binary_acc)
mlflow.log_metric("validation_loss", validation_loss)
mlflow.log_metric("validation_acc", validation_acc)
# log artifacts
mlflow.log_artifacts(image_dir, "images")
# log model
mlflow.keras.log_model(model, "models")
# save model locally
pathdir = "../data/out/keras_models/" + run_uuid
# keras_save_model(model, pathdir)
# Write out TensorFlow events as a run artifact
print("Uploading TensorFlow events as a run artifact.")
mlflow.log_artifacts(output_dir, artifact_path="events")
mlflow.end_run()
def run(self, args: argparse.Namespace) -> None:
logger.info("Load config from %s", args.config)
config = load_yaml(minato.cached_path(args.config), args.overrides)
logger.info("Configuration: %s", str(config))
builder = ConfigBuilder.build(config)
model = builder.model
train_file = args.train or builder.train_file
validation_file = args.validation or builder.validation_file
if not train_file:
raise ConfigurationError("train file is required.")
logger.info("Start training...")
logger.info("Training data: %s", str(train_file))
logger.info("Validation data: %s", str(validation_file))
params = {
"command": " ".join(sys.argv),
"config_file": args.config,
"train_file": train_file,
"validation_file": validation_file,
"serialization_dir": args.serialization_dir,
"config": config,
}
with _mlflow_start_run():
serialization_dir = args.serialization_dir
if args.serialization_dir is None and mlflow is None:
serialization_dir = "./output"
with create_workdir(
serialization_dir,
exist_ok=args.force,
) as workdir:
workdir = workdir.absolute()
try:
with open(workdir / "config.yaml", "w") as f:
yaml.dump(config, f)
with open(workdir / "params.json", "w") as f:
json.dump(params, f, indent=2)
if mlflow is not None:
logger.info("Log params to mlflow")
mlflow.log_params(params)
metrics = model.train(train_file, validation_file, workdir)
if mlflow is not None:
logger.info("Log metrics to mlflow")
mlflow.log_metrics(metrics)
logger.info("Training completed")
logger.info("Training metrics: %s",
json.dumps(metrics, indent=2))
with open(workdir / "metrics.json", "w") as metrics_file:
json.dump(metrics, metrics_file)
with open(workdir / "model.pkl", "wb") as model_file:
pickle.dump(model, model_file)
finally:
if mlflow is not None:
logger.info("Log metrics to mlflow")
mlflow.log_artifacts(str(workdir))
logger.info("Done!")
def main(learning_rate, batch_size, checkpoint_base_path, data_path,
tracking_url):
checkpoint_path = os.path.join(checkpoint_base_path, current_timestamp())
os.makedirs(checkpoint_path, exist_ok=True)
params = {
'hidden_size': 512,
'keep_rate': 0.8,
'learning_rate': learning_rate,
'nb_epochs': 1,
'batch_size': batch_size,
'checkpoint_path': checkpoint_path
}
# Configure the location where tracking data will be written to. In real-life
# this would be a remote MLFlow Tracking Servinc (using HTTP) or something like
# S3, HDFS etc.
mlflow.set_tracking_uri(tracking_url)
# Set name of experiment
mlflow.set_experiment('MNIST_TF_Estimator')
with mlflow.start_run() as run:
# Log parameters in MLFlow
for name, value in params.items():
mlflow.log_param(name, value)
def train_input_fn():
ds = dataset.train(data_path)
ds = ds.shuffle(buffer_size=50000)
ds = ds.take(5000) # just to speed up training
ds = ds.batch(params['batch_size'])
ds = ds.repeat(params['nb_epochs'])
return ds
def eval_input_fn():
ds = dataset.test(data_path)
ds = ds.batch(params['batch_size'])
return ds
run_config = tf.estimator.RunConfig(log_step_count_steps=50)
estimator = tf.estimator.Estimator(model_fn=model_fn,
model_dir=checkpoint_path,
params=params,
config=run_config)
estimator.train(input_fn=train_input_fn)
eval_results = estimator.evaluate(input_fn=eval_input_fn)
tf.logging.info('Eval loss: %s' % eval_results['loss'])
tf.logging.info('Eval accuracy: %s' % eval_results['accuracy'])
# Log results in MLFlow
mlflow.log_metric("eval_loss", eval_results['loss'])
mlflow.log_metric("eval_acc", eval_results['accuracy'])
# Send checkpoint and event files to MLFlow
mlflow.log_artifacts(checkpoint_path)
# Export the latest checkpoint as SavedModel
feat_spec = {
"images": tf.placeholder("float", name="images", shape=[None, 784])
}
receiver_fn = tf.estimator.export.build_raw_serving_input_receiver_fn(
feat_spec)
export_dir_base = os.path.join(checkpoint_path, 'saved_models')
saved_estimator_path = estimator.export_saved_model(
export_dir_base, receiver_fn).decode("utf-8")
tf.logging.info('SavedModel has been exported to %s' %
saved_estimator_path)
# Log the SavedModel as MLFlow model
mlflow.tensorflow.log_model(tf_saved_model_dir=saved_estimator_path,
tf_meta_graph_tags=[tag_constants.SERVING],
tf_signature_def_key="serving_default",
artifact_path="exported_model")
#
# Code snippet for https://mlflow.org/docs/latest/python_api/mlflow.html#mlflow.log_artifacts
#
import warnings
import os
import json
import mlflow
if __name__ == "__main__":
warnings.filterwarnings("ignore")
print(mlflow.__version__)
# Create some artifacts data to preserve
features = "rooms, zipcode, median_price, school_rating, transport"
data = {"state": "TX", "Available": 25, "Type": "Detached"}
# Create couple of artifact files under the directory "data"
os.makedirs("data", exist_ok=True)
with open("data/data.json", 'w', encoding='utf-8') as f:
json.dump(data, f, indent=2)
with open("data/features.txt", 'w') as f:
f.write(features)
# Write all files in "data" to root artifact_uri/states
with mlflow.start_run():
mlflow.log_artifacts("data", artifact_path="states")
def log_artifact(source_dir, artifact_path=None):
print(f"[INFO] Logging artifacts in {source_dir}...")
mlflow.log_artifacts(local_dir, artifact_path=None)
# Log a metric; metrics can be updated throughout the run
mlflow.log_metric("foo", random())
mlflow.log_metric("foo", random()+1)
mlflow.log_metric("foo", random()+2)
mlflow.log_metric("foo", random()+3)
mlflow.log_metric("bar", random())
mlflow.log_metric("bar", random()+1)
mlflow.log_metric("bar", random()+2)
mlflow.log_metric("bar", random()+3)
# Log an artifact (output file)
os.makedirs("outputs", exist_ok=True)
with open("outputs/test.txt", "w") as f:
f.write("hello world!")
mlflow.log_artifacts("outputs")
mlflow.end_run()
# Step 1: Set your Environment Variables
get_ipython().run_line_magic('env', 'NEPTUNE_API_TOKEN=ANONYMOUS')
get_ipython().run_line_magic('env', 'NEPTUNE_PROJECT=shared/mlflow-integration')
# Step 2: Sync your MLruns with Neptune
get_ipython().system(' neptune mlflow')
# **Note:**
# You can specify the path to the directory where the 'mlruns' directory is.
random_state=123)
rf.fit(train_resampled.drop(outcome, axis=1), train_resampled[outcome])
# evaluate model
train_predictions = rf.predict(train_resampled.drop(outcome, axis=1))
test_predictions = rf.predict(test.drop(outcome, axis=1))
train_auc = roc_auc_score(train_resampled[outcome], train_predictions)
test_auc = roc_auc_score(test[outcome], test_predictions)
# log data
import mlflow
import tempfile
mlflow.set_tracking_uri('http://localhost:5000')
mlflow.set_experiment('case-study-one')
mlflow.start_run()
mlflow.log_param('n_estimators', n_estimators)
mlflow.log_param('max_features', max_features)
mlflow.log_metric('train_auc', train_auc)
mlflow.log_metric('test_auc', test_auc)
with tempfile.TemporaryDirectory() as tmp:
path = os.path.join(tmp, 'train.csv')
train.to_csv(path)
mlflow.log_artifacts(tmp)
mlflow.sklearn.log_model(rf, 'model')
mlflow.end_run()
def train_elasticnet(datasets,
in_alpha,
in_l1_ratio,
trial=None,
verbose=True):
from sklearn.linear_model import ElasticNet
train_x = datasets['train_x']
train_y = datasets['train_y']
test_x = datasets['test_x']
test_y = datasets['test_y']
alpha, l1_ratio = eval_parameters(in_alpha, in_l1_ratio)
print("Parameters (alpha=%f, l1_ratio=%f):" % (alpha, l1_ratio))
server_uri = "http://localhost:5000"
#mlflow.set_tracking_uri(server_uri)
#mlflow.set_experiment("wine6")
with mlflow.start_run():
# train with ElasticNet
lr = ElasticNet(alpha=alpha, l1_ratio=l1_ratio, random_state=42)
lr.fit(train_x, train_y)
# Evaluate Metrics
predicted_qualities = lr.predict(test_x)
(rmse, mae, r2) = eval_metrics(test_y, predicted_qualities)
# Print out metrics
print("Elasticnet model (alpha=%f, l1_ratio=%f):" % (alpha, l1_ratio))
print(" RMSE: %s" % rmse)
print(" MAE: %s" % mae)
print(" R2: %s" % r2)
# Log parameter, metrics, and model to MLflow
mlflow.log_param("alpha", alpha)
mlflow.log_param("l1_ratio", l1_ratio)
mlflow.log_metric("rmse", rmse)
mlflow.log_metric("mae", mae)
mlflow.log_metric("r2", r2)
mlflow.set_tag("algo", "ElastiNet")
if trial != None:
mlflow.set_tag("trial", trial)
# store info
if verbose:
workdir = tempfile.mkdtemp()
with tempfile.TemporaryDirectory() as tmpdirname:
output_train_data_summary(tmpdirname, train_x, train_y)
output_enet_coefs(tmpdirname, train_x.columns, lr)
# plots
plot_enet_feature_importance(tmpdirname, train_x.columns,
lr.coef_)
# Call plot_enet_descent_path
#image = plot_enet_descent_path(tmpdirname, train_x, train_y, l1_ratio)
# Log artifacts (output files)
mlflow.log_artifacts(tmpdirname, artifact_path="artifacts")
# store model
mlflow.sklearn.log_model(lr, "model")
return rmse
def log_artifacts(cls, dir_path, artifact_path=None):
try:
mlflow.log_artifacts(dir_path, artifact_path)
except ConnectionError:
logger.warning(f"ConnectionError in logging artifacts to MLFlow")
def train_and_evaluate(args):
"""Trains and evaluates the Keras model.
Uses the Keras model defined in model.py and trains on data loaded and
preprocessed in utils.py. Saves the trained model in TensorFlow SavedModel
format to the path defined in part by the --job-dir argument.
History objects returns:
{'loss': [0.5699903990809373, 0.3629718415849791],
'acc': [0.78604823, 0.8331693],
'val_loss': [0.3966572880744934, 0.3477487564086914],
'val_acc': [0.8278044, 0.8281116],
'lr': [0.02, 0.015]}
Args:
args: dictionary of arguments - see get_args() for details
"""
logging.info('Resume training: {}'.format(args.reuse_job_dir))
if not args.reuse_job_dir:
if tf.io.gfile.exists(args.job_dir):
tf.io.gfile.rmtree(args.job_dir)
logging.info('Deleted job_dir {} to avoid re-use'.format(
args.job_dir))
else:
logging.info('Reusing job_dir {} if it exists'.format(args.job_dir))
train_x, train_y, eval_x, eval_y = utils.load_data(args.train_files,
args.eval_files)
# dimensions
num_train_examples, input_dim = train_x.shape
num_eval_examples = eval_x.shape[0]
# Create the Keras Model
keras_model = model.create_keras_model(input_dim=input_dim,
learning_rate=args.learning_rate)
# Pass a numpy array by passing DataFrame.values
training_dataset = model.input_fn(features=train_x.values,
labels=train_y,
shuffle=True,
num_epochs=args.num_epochs,
batch_size=args.batch_size)
# Pass a numpy array by passing DataFrame.values
validation_dataset = model.input_fn(features=eval_x.values,
labels=eval_y,
shuffle=False,
num_epochs=args.num_epochs,
batch_size=num_eval_examples)
start_time = time()
# Set MLflow tracking URI
if args.mlflow_tracking_uri:
mlflow.set_tracking_uri(args.mlflow_tracking_uri)
# Train model
with mlflow.start_run() as active_run:
run_id = active_run.info.run_id
# Callbacks
class MlflowCallback(tf.keras.callbacks.Callback):
# This function will be called after training completes.
def on_train_end(self, logs=None):
mlflow.log_param('num_layers', len(self.model.layers))
mlflow.log_param('optimizer_name',
type(self.model.optimizer).__name__)
# MLflow callback
mlflow_callback = MlflowCallback()
# Setup Learning Rate decay callback.
lr_decay_callback = tf.keras.callbacks.LearningRateScheduler(
lambda epoch: args.learning_rate + 0.02 * (0.5**(1 + epoch)),
verbose=False)
# Setup TensorBoard callback.
tensorboard_path = os.path.join(args.job_dir, run_id, 'tensorboard')
tensorboard_callback = tf.keras.callbacks.TensorBoard(tensorboard_path,
histogram_freq=1)
history = keras_model.fit(
training_dataset,
steps_per_epoch=int(num_train_examples / args.batch_size),
epochs=args.num_epochs,
validation_data=validation_dataset,
validation_steps=args.eval_steps,
verbose=1,
callbacks=[
lr_decay_callback, tensorboard_callback, mlflow_callback
])
metrics = history.history
logging.info(metrics)
keras_model.summary()
mlflow.log_param('train_files', args.train_files)
mlflow.log_param('eval_files', args.eval_files)
mlflow.log_param('num_epochs', args.num_epochs)
mlflow.log_param('batch_size', args.batch_size)
mlflow.log_param('learning_rate', args.learning_rate)
mlflow.log_param('train_samples', num_train_examples)
mlflow.log_param('eval_samples', num_eval_examples)
mlflow.log_param('eval_steps', args.eval_steps)
mlflow.log_param('steps_per_epoch',
int(num_train_examples / args.batch_size))
# Add metrics
_mlflow_log_metrics(metrics, 'loss')
_mlflow_log_metrics(metrics, 'acc')
_mlflow_log_metrics(metrics, 'val_loss')
_mlflow_log_metrics(metrics, 'val_acc')
_mlflow_log_metrics(metrics, 'lr')
# Export SavedModel
model_local_path = os.path.join(args.job_dir, run_id, 'model')
tf.keras.experimental.export_saved_model(keras_model, model_local_path)
# Define artifacts.
logging.info('Model exported to: {}'.format(model_local_path))
# MLflow workaround since is unable to read GCS path.
# https://github.com/mlflow/mlflow/issues/1765
if model_local_path.startswith('gs://'):
logging.info('Creating temp folder')
temp = tempfile.mkdtemp()
model_deployment.copy_artifacts(model_local_path, temp)
model_local_path = os.path.join(temp, 'model')
if tensorboard_path.startswith('gs://'):
logging.info('Creating temp folder')
temp = tempfile.mkdtemp()
model_deployment.copy_artifacts(tensorboard_path, temp)
tensorboard_path = temp
mlflow.tensorflow.log_model(tf_saved_model_dir=model_local_path,
tf_meta_graph_tags=[tag_constants.SERVING],
tf_signature_def_key='serving_default',
artifact_path='model')
# Reloading the model
pyfunc_model = mlflow.pyfunc.load_model(
mlflow.get_artifact_uri('model'))
logging.info('Uploading TensorFlow events as a run artifact.')
mlflow.log_artifacts(tensorboard_path)
logging.info('Launch TensorBoard with:\n\ntensorboard --logdir=%s' %
tensorboard_path)
duration = time() - start_time
mlflow.log_metric('duration', duration)
mlflow.end_run()
if model_local_path.startswith(
'gs://') and tensorboard_path.startswith('gs://'):
shutil.rmtree(model_local_path)
shutil.rmtree(tensorboard_path)
# Deploy model to AI Platform.
if args.deploy_gcp:
# Create AI Platform helper instance.
if not args.project_id:
raise ValueError('No Project is defined')
if not args.gcs_bucket:
raise ValueError('No GCS bucket')
model_helper = model_deployment.AIPlatformModel(
project_id=args.project_id)
# Copy local model to GCS for deployment.
if not model_local_path.startswith('gs://'):
model_gcs_path = os.path.join('gs://', args.gcs_bucket, run_id,
'model')
model_deployment.copy_artifacts(model_local_path, model_gcs_path)
# Create model
model_helper.create_model(args.model_name)
# Create model version
model_helper.deploy_model(model_gcs_path, args.model_name, run_id,
args.run_time_version)
logging.info('Model deployment in GCP completed')
logging.info(
'This model took: {} seconds to train and test.'.format(duration))
def run_evaluation(
*,
model_hash: str,
thr: float = 0.5,
dpath: str = "datasets/clouds/Landsat-Cloud-Cover-Assessment-" +
"Validation-Data-Partial",
rpath: str = "artifacts/",
vids: ("v", multi(min=1)),
batch_size: int = 32,
bands: ("b", multi(min=1)),
bands_names: ("bn", multi(min=1)),
img_ids: ("iid", multi(min=0)),
resize: bool = False,
normalize: bool = False,
standardize: bool = False,
mlflow: bool = False,
run_name: str = None,
):
"""
Load model given model hash and get evaluation metrics on L8CCA testset.
:param model_hash: MLFlow hash of the model to load.
:param thr: threshold to be used during evaluation.
:param dpath: path to dataset.
:param rpath: path to directory where results
and artifacts should be logged.
:param vids: tuple of ids of images which should be used to create
visualisations. If contains '*' visualisations will be
created for all images in the dataset.
:type vids: tuple[str]
:param batch_size: size of generated batches, only one batch is loaded
to memory at a time.
:param bands: band numbers to load
:type bands: list[int]
:param bands_names: names of the bands to load. Should have the same number
of elements as bands.
:type bands_names: list[str]
:param img_ids: if given, process only these images.
:type img_ids: list[int]
:param resize: whether to resize loaded img to gt.
:param normalize: whether to normalize the image.
:param standardize: whether to standardize the image.
:param mlflow: whether to use MLFlow.
:param run_name: name of the run.
"""
snow_imgs = ["LC82271192014287LGN00", "LC81321192014054LGN00"]
if img_ids == []:
img_ids = None
else:
snow_imgs = list(set(snow_imgs) & set(img_ids))
dpath, rpath = make_paths(dpath, rpath)
rpath = rpath / uuid.uuid4().hex
print(f"Working dir: {os.getcwd()}, " + f"artifacts dir: {rpath}",
flush=True)
mpath = Path(f"/media/ML/mlflow/beetle/artifacts/34/{model_hash}/"
# Change init_model to model for old models
+ "artifacts/init_model/data/model.h5")
# WARNING: If ever upgraded to tf2.x, this way of using metrics will not work
# because loaded metrics will become MeanMetricWrapper objects in tf2.x and
# this script isn't prepared for such objects (because MeanMetricWrapper has state,
# as opposed to present stateless metric functions).
model = keras.models.load_model(
mpath,
custom_objects={
"jaccard_index_loss": losses.JaccardIndexLoss(),
"jaccard_index_metric": losses.JaccardIndexMetric(),
"dice_coeff_metric": losses.DiceCoefMetric(),
"recall": losses.recall,
"precision": losses.precision,
"specificity": losses.specificity,
# F1 score is needed for old models
# "f1_score": losses.f1_score,
"tf": tf,
},
)
model.load_weights(f"/media/ML/mlflow/beetle/artifacts/34/{model_hash}/" +
"artifacts/best_weights/best_weights")
metrics, _ = evaluate_model(
model=model,
thr=thr,
dpath=dpath,
rpath=rpath,
vids=vids,
batch_size=batch_size,
bands=bands,
bands_names=bands_names,
img_ids=img_ids,
resize=resize,
normalize=normalize,
standardize=standardize,
mlflow=mlflow,
run_name=run_name,
)
mean_metrics = {}
mean_metrics_snow = {}
for key, value in metrics.items():
mean_metrics[key] = np.mean(list(value.values()))
mean_metrics_snow[f"snow_{key}"] = np.mean(
[value[x] for x in snow_imgs])
print(mean_metrics, mean_metrics_snow)
if mlflow:
log_metrics(mean_metrics)
log_metrics(mean_metrics_snow)
log_artifacts(rpath)
if gen_method in ratio_estimator_methods:
llr_means = np.mean(llr, axis=0)
mlflow.log_metrics({
"theta 0 LLR": llr_means[0],
"theta 1 LLR": llr_means[1],
})
elif gen_method in score_estimator_methods:
score_means = np.mean(scores, axis=0)
mlflow.log_metrics({
"theta 0 score": score_means[0],
"theta 1 score": score_means[1],
})
mlflow.log_artifacts(f"{results_dir}/{gen_method}")
#################################
## Calculating expected limits ##
#################################
for flag in include_xsec:
_, p_values, best_fit_index, _, _, _ = limits.expected_limits(
mode="histo",
theta_true=theta_true,
grid_ranges=theta_ranges,
grid_resolutions=resolutions,
include_xsec=flag,
luminosity=luminosity,
hist_vars=[histogram_var],
def train(logger):
"""
perform the training routine for a given fold. saves plots and selected parameters to the experiment dir
specified in the configs.
"""
logger.info(
'performing training in {}D over fold {} on experiment {} with model {}'
.format(cf.dim, cf.fold, cf.exp_dir, cf.model))
net = model.net(cf, logger).cuda()
if hasattr(cf, "optimizer") and cf.optimizer.lower() == "adam":
logger.info("Using Adam optimizer.")
optimizer = torch.optim.Adam(utils.parse_params_for_optim(
net,
weight_decay=cf.weight_decay,
exclude_from_wd=cf.exclude_from_wd),
lr=cf.learning_rate[0])
else:
logger.info("Using AdamW optimizer.")
optimizer = torch.optim.AdamW(utils.parse_params_for_optim(
net,
weight_decay=cf.weight_decay,
exclude_from_wd=cf.exclude_from_wd),
lr=cf.learning_rate[0])
if cf.dynamic_lr_scheduling:
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
mode=cf.scheduling_mode,
factor=cf.lr_decay_factor,
patience=cf.scheduling_patience)
model_selector = utils.ModelSelector(cf, logger)
train_evaluator = Evaluator(cf, logger, mode='train')
val_evaluator = Evaluator(cf, logger, mode=cf.val_mode)
starting_epoch = 1
# prepare monitoring
monitor_metrics = utils.prepare_monitoring(cf)
if cf.resume:
checkpoint_path = os.path.join(cf.fold_dir, "last_checkpoint")
starting_epoch, net, optimizer, monitor_metrics = \
utils.load_checkpoint(checkpoint_path, net, optimizer)
logger.info('resumed from checkpoint {} to epoch {}'.format(
checkpoint_path, starting_epoch))
logger.info('loading dataset and initializing batch generators...')
batch_gen = data_loader.get_train_generators(cf, logger)
# Prepare MLFlow
best_loss = 1e3
step = 1
mlflow.log_artifacts(cf.exp_dir, "exp")
for epoch in range(starting_epoch, cf.num_epochs + 1):
logger.info('starting training epoch {}'.format(epoch))
start_time = time.time()
net.train()
train_results_list = []
bix = 0
seen_pids = []
while True:
bix = bix + 1
try:
batch = next(batch_gen['train'])
except StopIteration:
break
for pid in batch['pid']:
seen_pids.append(pid)
# print(f'\rtr. batch {bix}: {batch["pid"]}')
tic_fw = time.time()
results_dict = net.train_forward(batch)
tic_bw = time.time()
optimizer.zero_grad()
results_dict['torch_loss'].backward()
optimizer.step()
print(
'\rtr. batch {0} (ep. {1}) fw {2:.2f}s / bw {3:.2f} s / total {4:.2f} s || '
.format(bix + 1, epoch, tic_bw - tic_fw,
time.time() - tic_bw,
time.time() - tic_fw) + results_dict['logger_string'],
flush=True,
end="")
train_results_list.append(
({k: v
for k, v in results_dict.items()
if k != "seg_preds"}, batch["pid"]))
print(f"Seen pids (unique): {len(np.unique(seen_pids))}")
print()
_, monitor_metrics['train'] = train_evaluator.evaluate_predictions(
train_results_list, monitor_metrics['train'])
logger.info('generating training example plot.')
utils.split_off_process(
plot_batch_prediction,
batch,
results_dict,
cf,
outfile=os.path.join(cf.plot_dir,
'pred_example_{}_train.png'.format(cf.fold)))
train_time = time.time() - start_time
logger.info('starting validation in mode {}.'.format(cf.val_mode))
with torch.no_grad():
net.eval()
if cf.do_validation:
val_results_list = []
val_predictor = Predictor(cf, net, logger, mode='val')
while True:
try:
batch = next(batch_gen[cf.val_mode])
except StopIteration:
break
if cf.val_mode == 'val_patient':
results_dict = val_predictor.predict_patient(batch)
elif cf.val_mode == 'val_sampling':
results_dict = net.train_forward(batch,
is_validation=True)
val_results_list.append(({
k: v
for k, v in results_dict.items() if k != "seg_preds"
}, batch["pid"]))
_, monitor_metrics['val'] = val_evaluator.evaluate_predictions(
val_results_list, monitor_metrics['val'])
best_model_path = model_selector.run_model_selection(
net, optimizer, monitor_metrics, epoch)
# Save best model
mlflow.log_artifacts(
best_model_path,
os.path.join("exp", os.path.basename(cf.fold_dir),
'best_checkpoint'))
# Save logs and plots
mlflow.log_artifacts(os.path.join(cf.exp_dir, "logs"),
os.path.join("exp", 'logs'))
mlflow.log_artifacts(
cf.plot_dir, os.path.join("exp",
os.path.basename(cf.plot_dir)))
# update monitoring and prediction plots
monitor_metrics.update({
"lr": {
str(g): group['lr']
for (g, group) in enumerate(optimizer.param_groups)
}
})
# replace tboard metrics with MLFlow
#logger.metrics2tboard(monitor_metrics, global_step=epoch)
mlflow.log_metric('learning rate', optimizer.param_groups[0]['lr'],
cf.num_epochs * cf.fold + epoch)
for key in ['train', 'val']:
for tag, val in monitor_metrics[key].items():
val = val[
-1] # maybe remove list wrapping, recording in evaluator?
if 'loss' in tag.lower() and not np.isnan(val):
mlflow.log_metric(f'{key}_{tag}', val,
cf.num_epochs * cf.fold + epoch)
elif not np.isnan(val):
mlflow.log_metric(f'{key}_{tag}', val,
cf.num_epochs * cf.fold + epoch)
epoch_time = time.time() - start_time
logger.info('trained epoch {}: took {} ({} train / {} val)'.format(
epoch, utils.get_formatted_duration(epoch_time, "ms"),
utils.get_formatted_duration(train_time, "ms"),
utils.get_formatted_duration(epoch_time - train_time, "ms")))
batch = next(batch_gen['val_sampling'])
results_dict = net.train_forward(batch, is_validation=True)
logger.info('generating validation-sampling example plot.')
utils.split_off_process(plot_batch_prediction,
batch,
results_dict,
cf,
outfile=os.path.join(
cf.plot_dir,
'pred_example_{}_val.png'.format(
cf.fold)))
# -------------- scheduling -----------------
if cf.dynamic_lr_scheduling:
scheduler.step(monitor_metrics["val"][cf.scheduling_criterion][-1])
else:
for param_group in optimizer.param_groups:
param_group['lr'] = cf.learning_rate[epoch - 1]
# Save whole experiment to MLFlow
mlflow.log_artifacts(cf.exp_dir, "exp")
import tempfile
import mlflow
from mlflow import log_metric, log_param, log_artifacts, get_artifact_uri, active_run,\
get_tracking_uri, log_artifact
if __name__ == "__main__":
print("Running {} with tracking URI {}".format(sys.argv[0],
get_tracking_uri()))
log_param("param1", 5)
log_metric("foo", 5)
log_metric("foo", 6)
log_metric("foo", 7)
log_metric("random_int", random.randint(0, 100))
run_id = active_run().info.run_id
# Get run metadata & data from the tracking server
service = mlflow.tracking.MlflowClient()
run = service.get_run(run_id)
print("Metadata & data for run with UUID %s: %s" % (run_id, run))
local_dir = tempfile.mkdtemp()
message = "test artifact written during run %s within artifact URI %s\n" \
% (active_run().info.run_id, get_artifact_uri())
try:
file_path = os.path.join(local_dir, "some_output_file.txt")
with open(file_path, "w") as handle:
handle.write(message)
log_artifacts(local_dir, "some_subdir")
log_artifact(file_path, "another_dir")
finally:
shutil.rmtree(local_dir)
for i in range(num_samples):
imgs = [
scopyon.Image(data)
for data in numpy.load(inputpath / f"images{i:03d}.npy")
]
spots = [
scopyon.analysis.spot_detection(img.as_array(),
processes=nproc,
min_sigma=min_sigma,
max_sigma=max_sigma,
threshold=threshold,
overlap=overlap) for img in imgs
]
spots_ = []
for t, data in zip(timepoints, spots):
spots_.extend(([t] + list(row) for row in data))
spots_ = numpy.array(spots_)
numpy.save(artifacts / f"spots{i:03d}.npy", spots_)
print("{} spots are detected in {} frames.".format(len(spots_), len(imgs)))
warnings.resetwarnings()
#!ls ./artifacts
#log_artifacts("./artifacts")
log_artifacts(generated_data)
mlflow.end_run()
train_time = timeit.default_timer()
os.system('python {} --mode {} --train-csv {} --model-dir {}'.format(
train_file,
'regression' if dataset[-1] == 'r' else 'classification',
'{}/{}/train.csv'.format(data_dir, dataset),
'{}/{}/'.format(result_dir, dataset)))
train_time = timeit.default_timer() - train_time
pred_time = timeit.default_timer()
os.system('python {} --prediction-csv {} --test-csv {} --model-dir {}'.
format(predict_file,
'{}/{}/pred.csv'.format(result_dir, dataset),
'{}/{}/test.csv'.format(data_dir, dataset),
'{}/{}/'.format(result_dir, dataset)))
pred_time = timeit.default_timer() - pred_time
df = pd.read_csv('{}/{}/test-target.csv'.format(data_dir, dataset))
df_pred = pd.read_csv('{}/{}/pred.csv'.format(result_dir, dataset))
df = pd.merge(df, df_pred, on='line_id', left_index=True)
score = roc_auc_score(df.target.values, df.prediction.values) if dataset[-1] == 'c' else \
np.sqrt(mean_squared_error(df.target.values, df.prediction.values))
print('Score {:0.5f}'.format(score))
n = dataset.split('_')[1]
mlflow.log_metric('score_{}'.format(n), score)
mlflow.log_metric('train_time_{}'.format(n), train_time)
mlflow.log_metric('test_time_{}'.format(n), pred_time)
mlflow.log_artifacts('./')
def rbf_svr_tuning(c = [0.001, 0.01, 0.1, 1, 10], gamma = [0.001, 0.01, 0.1, 1, 10], k = 5,
train_data_path = '../data/training_data.csv', save_model = False, tracking_uri = "http://0.0.0.0:5000"):
# Log the parameters with mlflow
mlflow.log_param("c", c)
mlflow.set_tag("k", k)
# Set random seed for reproducibility
np.random.seed(RANDOM_SEED)
random.seed(RANDOM_SEED)
# Get data shuffled and split into training and test sets
mdr = MiningDataReader(path = train_data_path)
(variable_names, X_train, X_test, y_train, y_test) = mdr.get_splitted_data()
pipeline = Pipeline(steps = [('scaling', StandardScaler()),
('regression', SVR(kernel = 'rbf'))])
### TRAINING ###
################
# Generate grid search for hyperparam tuning
hyperparams = {}
hyperparams['regression__C'] = c
hyperparams['regression__gamma'] = gamma
print("Training started...\n")
# Create an instance of Random Forest Regressor and fit the data for the grid parameters using all processors
modelCV = GridSearchCV(estimator = pipeline,
param_grid = hyperparams,
cv = k,
scoring = 'neg_mean_squared_error',
n_jobs = -1)
with ProgressBar():
modelCV.fit(X_train, y_train)
# Iterate over the results storing training error for each hyperparameter combination
results = modelCV.cv_results_
param_list, training_err_list, training_dev_list = [], [], []
for i in range(len(results['params'])):
param = results['params'][i]
score = (-1) * results['mean_test_score'][i] # NEGATIVE MSE
std = results['std_test_score'][i]
param_list.append(param)
training_err_list.append(score)
training_dev_list.append(std)
print(f"\nBest parameter set found for the training set:\n{modelCV.best_params_}")
# Store the index of the best combination
best_index = param_list.index(modelCV.best_params_)
# Get the best values for hyperparams
best_c = modelCV.best_params_['regression__C']
best_gamma = modelCV.best_params_['regression__gamma']
print("\nTraining finished. Evaluating model...\n")
### EVALUATION ###
##################
# Criteria is C
criteria = 'c'
mlflow.set_tag("criteria", criteria)
param_values = c
# Predict test data variying criteria param and evaluate the models
training_err_by_criteria, training_dev_by_criteria, test_err_list = [], [], []
rmse_score, mae_score, r2_score = -1, -1, -1
feature_names, feature_importances = [], []
for param_value in tqdm(param_values):
model = Pipeline(steps = [('scaler', StandardScaler()),
('regression', SVR(
C = param_value,
gamma = best_gamma,
kernel = 'rbf'))])
param = {'regression__C': param_value, 'regression__gamma': best_gamma}
# Fit model and evaluate results
model.fit(X_train, y_train)
prediction = model.predict(X_test)
index = param_list.index(param)
training_err = training_err_list[index]
training_dev = training_dev_list[index]
(training_mse, test_mse, rmse, mae, r2) = get_test_metrics(training_err, y_test, prediction)
# Store metrics
training_err_by_criteria.append(training_mse)
training_dev_by_criteria.append(training_dev)
test_err_list.append(test_mse)
# Set aditional metrics for the best combination
if index == best_index:
rmse_score = rmse
mae_score = mae
r2_score = r2
# Generate the plots
empty_img_folder()
plot_errors(criteria, param_values, training_err_by_criteria, training_dev_by_criteria, test_err_list)
# Once hyperparameters are selected, train and save the best model
if save_model:
print("\nEvaluation finished. Training final model with train + test data with the best hyperparameters...")
final_model = Pipeline(steps = [('scaler', StandardScaler()),
('regression', SVR(
C = param_list[best_index]['regression__C'],
gamma = best_gamma,
kernel = 'rbf'))])
# Train the best model with all the data (training + test)
full_X = np.vstack((X_train, X_test))
full_y = np.concatenate((y_train, y_test))
final_model.fit(full_X, full_y)
# Log plots and model with mlflow
mlflow.log_artifacts('./img')
mlflow.sklearn.log_model(final_model, 'model')
# Log results with mlflow
mlflow.log_metric("train_mse", training_err_list[best_index])
mlflow.log_metric("test_mse", min(test_err_list))
mlflow.log_metric("rmse", rmse_score)
mlflow.log_metric("mae", mae_score)
mlflow.log_metric("r2", r2_score)
mlflow.set_tag("best_params", param_list[best_index])
# Output the results
print(f'''
-----------------------------------------------------------------------------------------------------------------------
RESULTS
-----------------------------------------------------------------------------------------------------------------------
Best params: {param_list[best_index]}
Training MSE: {training_err_list[best_index]}
Test MSE: {min(test_err_list)}
RMSE: {rmse_score}
MAE: {mae_score}
R2: {r2_score}
-----------------------------------------------------------------------------------------------------------------------
''')
plt.title('XGBoost Classification ' + metric_name)
plt.savefig("artifacts/" + metric_name + ".png")
plt.show()
##################################
# Plot decision tree, feature importance stack and precision-recall curve
# ................................
precision, recall, _ = precision_recall_curve(test.get_label(),
predictions)
plt.step(recall, precision, color='b', alpha=0.2, where='post')
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.title('Precision-Recall curve')
plt.savefig("artifacts/PR_curve.png")
plt.show()
plt.rcParams["figure.figsize"] = (14, 7)
xgb.plot_importance(clf, grid=False)
plt.savefig("artifacts/importance.png")
plt.show()
plt.rcParams["figure.figsize"] = (14, 3)
xgb.plot_tree(clf)
plt.title('Decision Tree')
plt.savefig("artifacts/tree.png")
plt.show()
# persist the model and save artficats to mlflow
clf.save_model('artifacts/xgb.model')
mlflow.log_artifacts("artifacts")
callback=callback,
log_interval=args.log_interval)
model.save(env.simulator._env_working_dir_parent + '/' + name)
# If the algorithm doesn't reset or close the environment, this script will do it in
# order to correctly log all the simulation data (Energyplus + Sinergym
# logs)
if env.simulator._episode_existed:
env.close()
# ---------------------------------------------------------------------------- #
# Mlflow artifacts storege #
# ---------------------------------------------------------------------------- #
if args.mlflow_store:
# Code for send output and tensorboard to mlflow artifacts.
mlflow.log_artifacts(local_dir=env.simulator._env_working_dir_parent,
artifact_path=name)
if args.evaluation:
mlflow.log_artifacts(local_dir='best_model/' + name,
artifact_path='best_model/' + name)
# If tensorboard is active (in local) we should send to mlflow
if args.tensorboard and 'gs://' + args.bucket_name not in args.tensorboard:
mlflow.log_artifacts(local_dir=args.tensorboard + '/' + name,
artifact_path=os.path.abspath(
args.tensorboard).split('/')[-1] + '/' +
name)
# ---------------------------------------------------------------------------- #
# Google Cloud Bucket Storage #
# ---------------------------------------------------------------------------- #
if args.remote_store:
# Initiate Google Cloud client
import os
from random import random, randint
from mlflow import log_metric, log_param, log_artifacts
if __name__ == "__main__":
# Log a parameter (key-value pair)
log_param("param1", randint(0, 100))
# Log a metric; metrics can be updated throughout the run
log_metric("foo", random())
log_metric("foo", random() + 1)
log_metric("foo", random() + 2)
# Log an artifact (output file)
if not os.path.exists("outputs"):
os.makedirs("outputs")
with open("outputs/test.txt", "w") as f:
f.write("hello world!")
log_artifacts("outputs")
if not is_nni_run_standalone():
# Report final training results to NNI (NNI HP or NNI Classic NAS APIs)
# TODO: make sure `valid_state.metrics` is ordered so that reported default metric to NNI is always the same
nni.report_final_result({'default': valid_evaluator.state.metrics.values()[0], **train_evaluator.state.metrics, **valid_evaluator.state.metrics})
return (valid_evaluator.state.metrics, state)
except Exception as e:
logging.error(
f'Ignite training loop of "{type(model).__name__}" model failed, exception "{e}" raised{deepcv.utils.NL}### Traceback ###{deepcv.utils.NL}{traceback.format_exc()}')
raise RuntimeError(f'Error: `{e}` exception raised during ignite training loop of "{type(model).__name__}" model...') from e
finally:
if backend_conf.rank == 0:
tb_logger.close()
if hp['log_output_dir_to_mlflow'] and mlflow.active_run():
logging.info('Logging training output directory as mlfow artifacts...')
mlflow.log_artifacts(str(output_path))
# TODO: log and replace artifacts to mlflow at every epochs?
# TODO: make sure all artifacts are loaded synchronously here
# shutil.rmtree(output_path)
def _setup_distributed_training(device, backend_conf: BackendConfig, model: torch.nn.Module, batch_shape: torch.Size, use_sync_batch_norm: bool = False) -> torch.nn.Module:
if backend_conf.distributed:
# Setup distributed training with `torch.distributed`
dist.init_process_group(backend_conf.dist_backend, init_method=backend_conf.dist_url)
assert backend_conf.is_cuda, 'Error: Distributed training must be run on GPU(s).'
torch.cuda.set_device(backend_conf.device)
# TODO: make sure we dont want to add more device IDs here (see distributed examples in Ignite or PyTorch)
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[backend_conf.local_rank, ], output_device=backend_conf.local_rank)
if use_sync_batch_norm and any(map(model.modules(), lambda m: isinstance(m, (torch.nn.BatchNorm1d, torch.nn.BatchNorm2d, torch.nn.BatchNorm3d)))):