bug: state randomness (test time)

[zuoxingdong]

rlbidder/scripts/scale_transitions.py

Lines 68 to 93 in 1d2ee92

	('feature_jitter', FeatureWiseJitter( # tiny jitter to smooth quantile estimation
	scale_factor=0.1,
	distribution='uniform',
	random_state=DEFAULT_SEED,
	copy=True,
	)),
	('quantile', QuantileTransformer(
	output_distribution='normal',
	n_quantiles=100_000, # HACK: critical to be sufficiently large otherwise performance degrades
	subsample=None, # Use all data points
	random_state=DEFAULT_SEED,
	)),
	],
	verbose=True,
	)
	logger.info("Setting up action scaler for policy type '%s'.", cfg.policy_type)
	if cfg.policy_type == "tanh_normal":
	action_scaler = Pipeline(
	steps=[
	("clip", ClipTransformer(min_value=0.1, max_value=None)),
	("winsor", WinsorizerTransformer(quantile_range=(0.005, 0.995))),
	(
	"feature_jitter",
	FeatureWiseJitter(