def test_logger():
logger = Logger()
logger('iteration', 1)
logger('learning_rate', 1e-3)
logger('train_loss', 0.12)
logger('eval_loss', 0.14)
logger('iteration', 2)
logger('learning_rate', 5e-4)
logger('train_loss', 0.11)
logger('eval_loss', 0.13)
logger('iteration', 3)
logger('learning_rate', 1e-4)
logger('train_loss', 0.09)
logger('eval_loss', 0.10)
def check(logs):
assert len(logs) == 4
assert list(logs.keys()) == [
'iteration', 'learning_rate', 'train_loss', 'eval_loss'
]
assert logs['iteration'] == [1, 2, 3]
assert np.allclose(logs['learning_rate'], [1e-3, 5e-4, 1e-4])
assert np.allclose(logs['train_loss'], [0.12, 0.11, 0.09])
assert np.allclose(logs['eval_loss'], [0.14, 0.13, 0.10])
check(logger.logs)
logger.dump()
logger.dump(border='-' * 50)
logger.dump(keys=['iteration'])
logger.dump(keys=['iteration', 'train_loss'])
logger.dump(index=0)
logger.dump(index=[1, 2])
logger.dump(index=0)
logger.dump(keys=['iteration', 'eval_loss'], index=1)
logger.dump(keys=['iteration', 'learning_rate'], indent=1)
logger.dump(keys=['iteration', 'train_loss'],
index=[0, 2],
indent=1,
border='#' * 50)
f = Path('./logger_file')
logger.save(f)
f = f.with_suffix('.pkl')
assert f.exists()
logs = pickle_load(f)
check(logs)
f.unlink()
assert not f.exists()
logger.clear()
assert len(logger.logs) == 0
def test_pickle_yaml():
a = {'one': 1, 'two': [2, 3]}
b = {'three': 3, 'four': [4, 5]}
c = [a, b]
def check(x):
assert isinstance(x, list)
assert len(x) == 2
assert all([isinstance(i, dict) for i in x])
assert list(x[0].keys()) == ['one', 'two']
assert list(x[1].keys()) == ['three', 'four']
assert list(x[0].values()) == [1, [2, 3]]
assert list(x[1].values()) == [3, [4, 5]]
pickle_dump(c, '.tmp_pickle')
check(pickle_load('.tmp_pickle.pkl'))
os.unlink('.tmp_pickle.pkl')
yaml_dump(c, '.tmp_yaml')
check(yaml_load('.tmp_yaml.yml'))
os.unlink('.tmp_yaml.yml')
def read_xy(log_folder,
file_name,
get_x,
get_y,
smooth_out=False,
smooth_kws=None,
point_step=1):
glob_dir = lambda x: [
p for p in x.glob('*/') if p.is_dir() and str(p.name).isdigit()
]
dfs = []
for id_folder in glob_dir(Path(log_folder)):
x = []
y = []
for seed_folder in glob_dir(id_folder):
logs = pickle_load(seed_folder / file_name)
x.append([get_x(log) for log in logs])
y.append([get_y(log) for log in logs])
new_x, ys = interp_curves(x, y) # all seeds share same x values
if smooth_out:
if smooth_kws is None:
smooth_kws = {'window_length': 51, 'polyorder': 3}
ys = [smooth_filter(y, **smooth_kws) for y in ys]
if point_step > 1:
idx = np.arange(0, new_x.size, step=point_step)
new_x = new_x[idx, ...]
ys = [y[idx, ...] for y in ys]
df = pd.DataFrame({'x': np.tile(new_x, len(ys)), 'y': np.hstack(ys)})
config = yaml_load(id_folder / 'config.yml')
config = pd.DataFrame([config.values()], columns=config.keys())
config = config.applymap(lambda x: tuple(x)
if isinstance(x, list) else x)
df = pd.concat([df, config], axis=1, ignore_index=False)
df = df.fillna(method='pad') # padding all NaN configs
dfs.append(df)
dfs = pd.concat(dfs, axis=0, ignore_index=True)
return dfs
