def reduce_mean_over_time(exps, expected_len=300):
new_data = {}
for data_key in exps[0].progress.keys():
agg_data = [exp.progress[data_key] for exp in exps]
filtered_agg_data = [
data for data in agg_data if len(data) == expected_len
]
new_data[data_key] = scipy.stats.trim_mean(filtered_agg_data,
0.1,
axis=0)
return log_processor.ExperimentLog({}, new_data, None)
def reduce_mean_partitions_over_time(partitions, split_key, expected_len=300):
for partition_key in partitions:
if isinstance(split_key, (list, tuple)):
new_params = dict(zip(split_key, partition_key))
else:
new_params = {split_key: partition_key}
exps = partitions[partition_key]
new_data = {}
for data_key in exps[0].progress.keys():
if data_key in ['validation_stop_step']:
continue
agg_data = [exp.progress[data_key] for exp in exps]
filtered_agg_data = [data for data in agg_data if len(data)==expected_len]
new_data[data_key] = scipy.stats.trim_mean(filtered_agg_data, 0.1, axis=0)
yield partition_key, log_processor.ExperimentLog(new_params, new_data, None)