def test_run_with_artifact_path(tmpdir):
artifact_file = tmpdir.join("model.pkl")
artifact_file.write("Hello world")
with kiwi.start_run() as run:
kiwi.log_artifact(artifact_file)
submitted_run = kiwi.projects.run(
TEST_PROJECT_DIR,
entry_point="test_artifact_path",
parameters={"model": "runs:/%s/model.pkl" % run.info.run_id},
use_conda=False,
experiment_id=FileStore.DEFAULT_EXPERIMENT_ID)
validate_exit_status(submitted_run.get_status(), RunStatus.FINISHED)
def objective(args):
# Define model
model = RankedNetworkCNNModule(args['learning_rate'],
dataset.get_embeddings(),
hidden_dim=args['hidden'],
output_labels=2)
# Evaluation on held-out test-set
with torch.no_grad():
model.eval()
results = pd.DataFrame(columns=['labels', 'predictions'])
for batch_idx, batch in enumerate(test_loader):
y_hat = model(batch['a'], batch['b'])
results: pd.DataFrame = results.append(pd.DataFrame({
'labels':
batch['label'].flatten(),
'predictions':
y_hat.detach().argmax(axis=1)
}),
ignore_index=True)
results.to_csv()
# With a nice confusion matrix
confusion_matrix(y_pred=results['predictions'].values,
y_true=results['labels'].values,
classes=[0, 1])
cm = ConfusionMatrix(actual_vector=results['labels'].values,
predict_vector=results['predictions'].values)
output_test_results = "cm.txt"
cm.save_stat(output_test_results)
output_test_predictions_file = "test_predictions.txt"
np.savetxt(output_test_predictions_file,
results['predictions'].values,
delimiter=",")
kiwi.log_metric(key="test_acc", value=cm.Overall_ACC)
kiwi.log_metric(key="test_f1_micro", value=cm.F1_Micro)
kiwi.log_metric(key="test_f1_macro", value=cm.F1_Macro)
kiwi.log_metric(key="test_ci_pm",
value=cm.CI95[1] - cm.Overall_ACC)
kiwi.log_metric(key="test_ci_pm",
value=cm.CI95[1] - cm.Overall_ACC)
kiwi.log_artifact(output_test_predictions_file)
kiwi.log_artifact(output_test_results + ".pycm")
return cm.Overall_ACC
def test_artifact_can_be_downloaded_from_absolute_uri_successfully(tmpdir):
artifact_file_name = "artifact.txt"
artifact_text = "Sample artifact text"
local_artifact_path = tmpdir.join(artifact_file_name).strpath
with open(local_artifact_path, "w") as out:
out.write(artifact_text)
logged_artifact_path = "artifact"
with kiwi.start_run():
kiwi.log_artifact(local_path=local_artifact_path, artifact_path=logged_artifact_path)
artifact_uri = kiwi.get_artifact_uri(artifact_path=logged_artifact_path)
downloaded_artifact_path = os.path.join(
_download_artifact_from_uri(artifact_uri), artifact_file_name)
assert downloaded_artifact_path != local_artifact_path
assert downloaded_artifact_path != logged_artifact_path
with open(downloaded_artifact_path, "r") as f:
assert f.read() == artifact_text
def test_download_artifacts_from_uri():
with kiwi.start_run() as run:
with TempDir() as tmp:
local_path = tmp.path("test")
with open(local_path, "w") as f:
f.write("test")
kiwi.log_artifact(local_path, "test")
command = ["mlflow", "artifacts", "download", "-u"]
# Test with run uri
run_uri = "runs:/{run_id}/test".format(run_id=run.info.run_id)
actual_uri = posixpath.join(run.info.artifact_uri, "test")
for uri in (run_uri, actual_uri):
p = Popen(command + [uri], stdout=PIPE,
stderr=STDOUT)
output = p.stdout.readlines()
downloaded_file_path = output[-1].strip()
downloaded_file = os.listdir(downloaded_file_path)[0]
with open(os.path.join(downloaded_file_path, downloaded_file), "r") as f:
assert f.read() == "test"
def load_raw_data(url):
with kiwi.start_run() as mlrun:
local_dir = tempfile.mkdtemp()
local_filename = os.path.join(local_dir, "ml-20m.zip")
print("Downloading %s to %s" % (url, local_filename))
r = requests.get(url, stream=True)
with open(local_filename, 'wb') as f:
for chunk in r.iter_content(chunk_size=1024):
if chunk: # filter out keep-alive new chunks
f.write(chunk)
extracted_dir = os.path.join(local_dir, 'ml-20m')
print("Extracting %s into %s" % (local_filename, extracted_dir))
with zipfile.ZipFile(local_filename, 'r') as zip_ref:
zip_ref.extractall(local_dir)
ratings_file = os.path.join(extracted_dir, 'ratings.csv')
print("Uploading ratings: %s" % ratings_file)
kiwi.log_artifact(ratings_file, "ratings-csv-dir")
def test_download_artifact_from_absolute_uri_persists_data_to_specified_output_directory(tmpdir):
artifact_file_name = "artifact.txt"
artifact_text = "Sample artifact text"
local_artifact_path = tmpdir.join(artifact_file_name).strpath
with open(local_artifact_path, "w") as out:
out.write(artifact_text)
logged_artifact_subdir = "logged_artifact"
with kiwi.start_run():
kiwi.log_artifact(local_path=local_artifact_path, artifact_path=logged_artifact_subdir)
artifact_uri = kiwi.get_artifact_uri(artifact_path=logged_artifact_subdir)
artifact_output_path = tmpdir.join("artifact_output").strpath
os.makedirs(artifact_output_path)
_download_artifact_from_uri(artifact_uri=artifact_uri, output_path=artifact_output_path)
assert logged_artifact_subdir in os.listdir(artifact_output_path)
assert artifact_file_name in os.listdir(
os.path.join(artifact_output_path, logged_artifact_subdir))
with open(os.path.join(
artifact_output_path, logged_artifact_subdir, artifact_file_name), "r") as f:
assert f.read() == artifact_text
def test_log_artifact():
artifact_src_dir = tempfile.mkdtemp()
# Create artifacts
_, path0 = tempfile.mkstemp(dir=artifact_src_dir)
_, path1 = tempfile.mkstemp(dir=artifact_src_dir)
for i, path in enumerate([path0, path1]):
with open(path, "w") as handle:
handle.write("%s" % str(i))
# Log an artifact, verify it exists in the directory returned by get_artifact_uri
# after the run finishes
artifact_parent_dirs = ["some_parent_dir", None]
for parent_dir in artifact_parent_dirs:
with start_run():
artifact_uri = kiwi.get_artifact_uri()
run_artifact_dir = local_file_uri_to_path(artifact_uri)
kiwi.log_artifact(path0, parent_dir)
expected_dir = os.path.join(run_artifact_dir, parent_dir) \
if parent_dir is not None else run_artifact_dir
assert os.listdir(expected_dir) == [os.path.basename(path0)]
logged_artifact_path = os.path.join(expected_dir, path0)
assert filecmp.cmp(logged_artifact_path, path0, shallow=False)
# Log multiple artifacts, verify they exist in the directory returned by get_artifact_uri
for parent_dir in artifact_parent_dirs:
with start_run():
artifact_uri = kiwi.get_artifact_uri()
run_artifact_dir = local_file_uri_to_path(artifact_uri)
kiwi.log_artifacts(artifact_src_dir, parent_dir)
# Check that the logged artifacts match
expected_artifact_output_dir = os.path.join(run_artifact_dir, parent_dir) \
if parent_dir is not None else run_artifact_dir
dir_comparison = filecmp.dircmp(artifact_src_dir,
expected_artifact_output_dir)
assert len(dir_comparison.left_only) == 0
assert len(dir_comparison.right_only) == 0
assert len(dir_comparison.diff_files) == 0
assert len(dir_comparison.funny_files) == 0
def test_log_artifact_with_dirs(tmpdir):
# Test log artifact with a directory
art_dir = tmpdir.mkdir("parent")
file0 = art_dir.join("file0")
file0.write("something")
file1 = art_dir.join("file1")
file1.write("something")
sub_dir = art_dir.mkdir("child")
with start_run():
artifact_uri = kiwi.get_artifact_uri()
run_artifact_dir = local_file_uri_to_path(artifact_uri)
kiwi.log_artifact(str(art_dir))
base = os.path.basename(str(art_dir))
assert os.listdir(run_artifact_dir) == [base]
assert set(os.listdir(os.path.join(run_artifact_dir, base))) == \
{'child', 'file0', 'file1'}
with open(os.path.join(run_artifact_dir, base, "file0")) as f:
assert f.read() == "something"
# Test log artifact with directory and specified parent folder
art_dir = tmpdir.mkdir("dir")
with start_run():
artifact_uri = kiwi.get_artifact_uri()
run_artifact_dir = local_file_uri_to_path(artifact_uri)
kiwi.log_artifact(str(art_dir), "some_parent")
assert os.listdir(run_artifact_dir) == [
os.path.basename("some_parent")
]
assert os.listdir(os.path.join(run_artifact_dir, "some_parent")) == \
[os.path.basename(str(art_dir))]
sub_dir = art_dir.mkdir("another_dir")
with start_run():
artifact_uri = kiwi.get_artifact_uri()
run_artifact_dir = local_file_uri_to_path(artifact_uri)
kiwi.log_artifact(str(art_dir), "parent/and_child")
assert os.listdir(os.path.join(run_artifact_dir, "parent", "and_child")) == \
[os.path.basename(str(art_dir))]
assert os.listdir(os.path.join(run_artifact_dir,
"parent", "and_child",
os.path.basename(str(art_dir)))) == \
[os.path.basename(str(sub_dir))]
import tempfile
import kiwi
from kiwi 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 = kiwi.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)
print("Computing regularization path using the elastic net.")
alphas_enet, coefs_enet, _ = enet_path(X,
y,
eps=eps,
l1_ratio=l1_ratio,
fit_intercept=False)
# Display results
fig = plt.figure(1)
ax = plt.gca()
colors = cycle(['b', 'r', 'g', 'c', 'k'])
neg_log_alphas_enet = -np.log10(alphas_enet)
for coef_e, c in zip(coefs_enet, colors):
l2 = plt.plot(neg_log_alphas_enet, coef_e, linestyle='--', c=c)
plt.xlabel('-Log(alpha)')
plt.ylabel('coefficients')
title = 'ElasticNet Path by alpha for l1_ratio = ' + str(l1_ratio)
plt.title(title)
plt.axis('tight')
# Save figures
fig.savefig("ElasticNet-paths.png")
# Close plot
plt.close(fig)
# Log artifacts (output files)
kiwi.log_artifact("ElasticNet-paths.png")
def run(training_data, max_runs, batch_size, max_p, epochs, metric, gpy_model,
gpy_acquisition, initial_design, seed):
bounds = [
{
'name': 'lr',
'type': 'continuous',
'domain': (1e-5, 1e-1)
},
{
'name': 'momentum',
'type': 'continuous',
'domain': (0.0, 1.0)
},
]
# create random file to store run ids of the training tasks
tracking_client = kiwi.tracking.MlflowClient()
def new_eval(nepochs,
experiment_id,
null_train_loss,
null_valid_loss,
null_test_loss,
return_all=False):
"""
Create a new eval function
:param nepochs: Number of epochs to train the model.
:experiment_id: Experiment id for the training run
:valid_null_loss: Loss of a null model on the validation dataset
:test_null_loss: Loss of a null model on the test dataset.
:return_test_loss: Return both validation and test loss if set.
:return: new eval function.
"""
def eval(params):
"""
Train Keras model with given parameters by invoking MLflow run.
Notice we store runUuid and resulting metric in a file. We will later use these to pick
the best run and to log the runUuids of the child runs as an artifact. This is a
temporary workaround until MLflow offers better mechanism of linking runs together.
:param params: Parameters to the train_keras script we optimize over:
learning_rate, drop_out_1
:return: The metric value evaluated on the validation data.
"""
lr, momentum = params[0]
with kiwi.start_run(nested=True) as child_run:
p = kiwi.projects.run(run_id=child_run.info.run_id,
uri=".",
entry_point="train",
parameters={
"training_data": training_data,
"epochs": str(nepochs),
"learning_rate": str(lr),
"momentum": str(momentum),
"seed": str(seed)
},
experiment_id=experiment_id,
synchronous=False)
succeeded = p.wait()
if succeeded:
training_run = tracking_client.get_run(p.run_id)
metrics = training_run.data.metrics
# cap the loss at the loss of the null model
train_loss = min(null_valid_loss,
metrics["train_{}".format(metric)])
valid_loss = min(null_valid_loss,
metrics["val_{}".format(metric)])
test_loss = min(null_test_loss,
metrics["test_{}".format(metric)])
else:
# run failed => return null loss
tracking_client.set_terminated(p.run_id, "FAILED")
train_loss = null_train_loss
valid_loss = null_valid_loss
test_loss = null_test_loss
kiwi.log_metrics({
"train_{}".format(metric): train_loss,
"val_{}".format(metric): valid_loss,
"test_{}".format(metric): test_loss
})
if return_all:
return train_loss, valid_loss, test_loss
else:
return valid_loss
return eval
with kiwi.start_run() as run:
experiment_id = run.info.experiment_id
# Evaluate null model first.
# We use null model (predict everything to the mean) as a reasonable upper bound on loss.
# We need an upper bound to handle the failed runs (e.g. return NaNs) because GPyOpt can not
# handle Infs.
# Always including a null model in our results is also a good ML practice.
train_null_loss, valid_null_loss, test_null_loss = new_eval(
0, experiment_id, _inf, _inf, _inf, True)(params=[[0, 0]])
myProblem = GPyOpt.methods.BayesianOptimization(
new_eval(epochs, experiment_id, train_null_loss, valid_null_loss,
test_null_loss),
bounds,
evaluator_type="local_penalization"
if min(batch_size, max_p) > 1 else "sequential",
batch_size=batch_size,
num_cores=max_p,
model_type=gpy_model,
acquisition_type=gpy_acquisition,
initial_design_type=initial_design,
initial_design_numdata=max_runs >> 2,
exact_feval=False)
myProblem.run_optimization(max_runs)
matplotlib.use('agg')
plt.switch_backend('agg')
with TempDir() as tmp:
acquisition_plot = tmp.path("acquisition_plot.png")
convergence_plot = tmp.path("convergence_plot.png")
myProblem.plot_acquisition(filename=acquisition_plot)
myProblem.plot_convergence(filename=convergence_plot)
if os.path.exists(convergence_plot):
kiwi.log_artifact(convergence_plot, "converegence_plot")
if os.path.exists(acquisition_plot):
kiwi.log_artifact(acquisition_plot, "acquisition_plot")
# find the best run, log its metrics as the final metrics of this run.
client = MlflowClient()
runs = client.search_runs(
[experiment_id], "tags.mlflow.parentRunId = '{run_id}' ".format(
run_id=run.info.run_id))
best_val_train = _inf
best_val_valid = _inf
best_val_test = _inf
best_run = None
for r in runs:
if r.data.metrics["val_rmse"] < best_val_valid:
best_run = r
best_val_train = r.data.metrics["train_rmse"]
best_val_valid = r.data.metrics["val_rmse"]
best_val_test = r.data.metrics["test_rmse"]
kiwi.set_tag("best_run", best_run.info.run_id)
kiwi.log_metrics({
"train_{}".format(metric): best_val_train,
"val_{}".format(metric): best_val_valid,
"test_{}".format(metric): best_val_test
})
def call_tracking_apis():
kiwi.log_metric("some_key", 3)
with tempfile.NamedTemporaryFile("w") as temp_file:
temp_file.write("Temporary content.")
kiwi.log_artifact(temp_file.name)
