-
Notifications
You must be signed in to change notification settings - Fork 4k
/
test_fastai_autolog.py
248 lines (197 loc) · 8.55 KB
/
test_fastai_autolog.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
import pytest
import numpy as np
from tests.conftest import tracking_uri_mock # pylint: disable=W0611
import pandas as pd
import sklearn.datasets as datasets
from fastai.tabular import tabular_learner, TabularList
from fastai.metrics import accuracy
import mlflow
import mlflow.fastai
from fastai.callbacks import EarlyStoppingCallback
np.random.seed(1337)
NUM_EPOCHS = 3
MIN_DELTA = 99999999 # Forces earlystopping
@pytest.fixture(params=[True, False])
def manual_run(request, tracking_uri_mock):
if request.param:
mlflow.start_run()
yield
mlflow.end_run()
def iris_dataframe():
iris = datasets.load_iris()
return pd.DataFrame(iris.data[:, :2], columns=iris.feature_names[:2])
@pytest.fixture(scope="session")
def iris_data():
iris = datasets.load_iris()
X = pd.DataFrame(iris.data[:, :2], columns=iris.feature_names[:2])
y = pd.Series(iris.target, name='label')
return (TabularList.from_df(pd.concat([X, y], axis=1), cont_names=list(X.columns))
.split_by_rand_pct(valid_pct=0.1, seed=42)
.label_from_df(cols='label')
.databunch())
def fastai_model(data, **kwargs):
return tabular_learner(data, metrics=accuracy, layers=[5, 3, 2], **kwargs)
@pytest.mark.large
@pytest.mark.parametrize('fit_variant', ['fit', 'fit_one_cycle'])
def test_fastai_autolog_ends_auto_created_run(iris_data, fit_variant):
mlflow.fastai.autolog()
model = fastai_model(iris_data)
if fit_variant == 'fit_one_cycle':
model.fit_one_cycle(1)
else:
model.fit(1)
assert mlflow.active_run() is None
@pytest.mark.large
@pytest.mark.parametrize('fit_variant', ['fit', 'fit_one_cycle'])
def test_fastai_autolog_persists_manually_created_run(iris_data, fit_variant):
mlflow.fastai.autolog()
with mlflow.start_run() as run:
model = fastai_model(iris_data)
if fit_variant == 'fit_one_cycle':
model.fit_one_cycle(NUM_EPOCHS)
else:
model.fit(NUM_EPOCHS)
assert mlflow.active_run()
assert mlflow.active_run().info.run_id == run.info.run_id
@pytest.fixture
def fastai_random_data_run(iris_data, fit_variant, manual_run):
mlflow.fastai.autolog()
model = fastai_model(iris_data)
if fit_variant == 'fit_one_cycle':
model.fit_one_cycle(NUM_EPOCHS)
else:
model.fit(NUM_EPOCHS)
client = mlflow.tracking.MlflowClient()
return model, client.get_run(client.list_run_infos(experiment_id='0')[0].run_id)
@pytest.mark.large
@pytest.mark.parametrize('fit_variant', ['fit', 'fit_one_cycle'])
def test_fastai_autolog_logs_expected_data(fastai_random_data_run, fit_variant):
model, run = fastai_random_data_run
data = run.data
# Testing metrics are logged
assert 'train_loss' in data.metrics
assert 'valid_loss' in data.metrics
for o in model.metrics:
assert o.__name__ in data.metrics
client = mlflow.tracking.MlflowClient()
metric_history = client.get_metric_history(run.info.run_id, 'valid_loss')
assert np.array_equal([m.value for m in metric_history], model.recorder.val_losses)
# Testing explicitly passed parameters are logged correctly
assert 'epochs' in data.params
assert data.params['epochs'] == str(NUM_EPOCHS)
# Testing implicitly passed parameters are logged correctly
assert 'wd' in data.params
# Testing unwanted parameters are not logged
assert 'callbacks' not in data.params
assert 'learn' not in data.params
# Testing optimizer parameters are logged
assert 'opt_func' in data.params
assert data.params['opt_func'] == 'Adam'
assert 'model_summary' in data.tags
# Testing model_summary.txt is saved
client = mlflow.tracking.MlflowClient()
artifacts = client.list_artifacts(run.info.run_id)
artifacts = map(lambda x: x.path, artifacts)
assert 'model_summary.txt' in artifacts
@pytest.mark.large
@pytest.mark.parametrize('fit_variant', ['fit', 'fit_one_cycle'])
def test_fastai_autolog_logs_default_params(fastai_random_data_run, fit_variant):
_, run = fastai_random_data_run
if fit_variant == 'fit':
assert 'lr' in run.data.params
assert run.data.params['lr'] == 'slice(None, 0.003, None)'
else:
assert 'pct_start' in run.data.params
assert run.data.params['pct_start'] == '0.3'
@pytest.mark.large
@pytest.mark.parametrize('fit_variant', ['fit', 'fit_one_cycle'])
def test_fastai_autolog_model_can_load_from_artifact(fastai_random_data_run):
run_id = fastai_random_data_run[1].info.run_id
client = mlflow.tracking.MlflowClient()
artifacts = client.list_artifacts(run_id)
artifacts = map(lambda x: x.path, artifacts)
assert 'model' in artifacts
model = mlflow.fastai.load_model("runs:/" + run_id + "/model")
model_wrapper = mlflow.fastai._FastaiModelWrapper(model)
model_wrapper.predict(iris_dataframe())
@pytest.fixture
def fastai_random_data_run_with_callback(iris_data, fit_variant, manual_run, callback, patience):
mlflow.fastai.autolog()
callbacks = []
if callback == 'early':
# min_delta is set as such to guarantee early stopping
callbacks.append(lambda learn: EarlyStoppingCallback(
learn, patience=patience, min_delta=MIN_DELTA))
model = fastai_model(iris_data, callback_fns=callbacks)
if fit_variant == 'fit_one_cycle':
model.fit_one_cycle(NUM_EPOCHS)
else:
model.fit(NUM_EPOCHS)
client = mlflow.tracking.MlflowClient()
return model, client.get_run(client.list_run_infos(experiment_id='0')[0].run_id)
@pytest.mark.large
@pytest.mark.parametrize('fit_variant', ['fit', 'fit_one_cycle'])
@pytest.mark.parametrize('callback', ['early'])
@pytest.mark.parametrize('patience', [0, 1, 5])
def test_fastai_autolog_early_stop_logs(fastai_random_data_run_with_callback, patience):
model, run = fastai_random_data_run_with_callback
params = run.data.params
assert 'early_stop_patience' in params
assert params['early_stop_patience'] == str(patience)
assert 'early_stop_monitor' in params
assert params['early_stop_monitor'] == 'valid_loss'
assert 'early_stop_mode' in params
assert params['early_stop_mode'] == 'auto'
assert 'early_stop_min_delta' in params
assert params['early_stop_min_delta'] == '-{}'.format(MIN_DELTA)
client = mlflow.tracking.MlflowClient()
metric_history = client.get_metric_history(run.info.run_id, 'valid_loss')
num_of_epochs = len(model.recorder.val_losses)
assert len(metric_history) == num_of_epochs
@pytest.mark.large
@pytest.mark.parametrize('fit_variant', ['fit', 'fit_one_cycle'])
@pytest.mark.parametrize('callback', ['early'])
@pytest.mark.parametrize('patience', [11])
def test_fastai_autolog_early_stop_no_stop_does_not_log(
fastai_random_data_run_with_callback, patience):
model, run, = fastai_random_data_run_with_callback
metrics = run.data.metrics
params = run.data.params
assert 'early_stop_patience' in params
assert params['early_stop_patience'] == str(patience)
assert 'early_stop_monitor' in params
assert params['early_stop_monitor'] == 'valid_loss'
assert 'early_stop_mode' in params
assert 'early_stop_min_delta' in params
assert params['early_stop_min_delta'] == '-{}'.format(99999999)
"""
assert 'stopped_epoch' in metrics
assert metrics['stopped_epoch'] == 0
assert 'restored_epoch' not in metrics
"""
num_of_epochs = len(model.recorder.val_losses)
client = mlflow.tracking.MlflowClient()
metric_history = client.get_metric_history(run.info.run_id, 'valid_loss')
# Check the test epoch numbers are correct
assert num_of_epochs == NUM_EPOCHS
assert len(metric_history) == num_of_epochs
@pytest.mark.large
@pytest.mark.parametrize('fit_variant', ['fit', 'fit_one_cycle'])
@pytest.mark.parametrize('callback', ['not-early'])
@pytest.mark.parametrize('patience', [5])
def test_fastai_autolog_non_early_stop_callback_does_not_log(fastai_random_data_run_with_callback):
model, run, = fastai_random_data_run_with_callback
metrics = run.data.metrics
params = run.data.params
assert 'early_stop_patience' not in params
assert 'early_stop_monitor' not in params
assert 'early_stop_mode' not in params
assert 'stopped_epoch' not in metrics
assert 'restored_epoch' not in metrics
assert 'early_stop_min_delta' not in params
num_of_epochs = len(model.recorder.val_losses)
client = mlflow.tracking.MlflowClient()
metric_history = client.get_metric_history(run.info.run_id, 'valid_loss')
# Check the test epoch numbers are correct
assert num_of_epochs == NUM_EPOCHS
assert len(metric_history) == num_of_epochs