Spaces:
Sleeping
Sleeping
| # this is a wrapper prior that samples hyperparameters which are set to be ConfigSpace parameters | |
| import torch | |
| from .prior import Batch | |
| from ConfigSpace import hyperparameters as CSH | |
| import ConfigSpace as CS | |
| from copy import deepcopy | |
| def list_all_hps_in_nested(config): | |
| if isinstance(config, CSH.Hyperparameter): | |
| return [config] | |
| elif isinstance(config, dict): | |
| result = [] | |
| for k, v in config.items(): | |
| result += list_all_hps_in_nested(v) | |
| return result | |
| else: | |
| return [] | |
| def create_configspace_from_hierarchical(config): | |
| cs = CS.ConfigurationSpace() | |
| for hp in list_all_hps_in_nested(config): | |
| cs.add_hyperparameter(hp) | |
| return cs | |
| def fill_in_configsample(config, configsample): | |
| # config is our dict that defines config distribution | |
| # configsample is a CS.Configuration | |
| hierarchical_configsample = deepcopy(config) | |
| for k, v in config.items(): | |
| if isinstance(v, CSH.Hyperparameter): | |
| hierarchical_configsample[k] = configsample[v.name] | |
| elif isinstance(v, dict): | |
| hierarchical_configsample[k] = fill_in_configsample(v, configsample) | |
| return hierarchical_configsample | |
| def sample_configspace_hyperparameters(hyperparameters): | |
| cs = create_configspace_from_hierarchical(hyperparameters) | |
| cs_sample = cs.sample_configuration() | |
| return fill_in_configsample(hyperparameters, cs_sample) | |
| def get_batch(batch_size, *args, hyperparameters, get_batch, **kwargs): | |
| num_models = min(hyperparameters.get('num_hyperparameter_samples_per_batch', 1), batch_size) | |
| if num_models == -1: | |
| num_models = batch_size | |
| assert batch_size % num_models == 0, 'batch_size must be a multiple of num_models' | |
| cs = create_configspace_from_hierarchical(hyperparameters) | |
| sub_batches = [] | |
| for i in range(num_models): | |
| cs_sample = cs.sample_configuration() | |
| hyperparameters_sample = fill_in_configsample(hyperparameters, cs_sample) | |
| sub_batch = get_batch(batch_size//num_models, *args, hyperparameters=hyperparameters_sample, **kwargs) | |
| sub_batches.append(sub_batch) | |
| # concat x, y, target (and maybe style) | |
| #assert 3 <= len(sub_batch) <= 4 | |
| #return tuple(torch.cat([sb[i] for sb in sub_batches], dim=(0 if i == 3 else 1)) for i in range(len(sub_batch))) | |
| assert all(not b.other_filled_attributes(set_of_attributes=('x', 'y', 'target_y')) for b in sub_batches) | |
| return Batch(x=torch.cat([b.x for b in sub_batches], dim=1), | |
| y=torch.cat([b.y for b in sub_batches], dim=1), | |
| target_y=torch.cat([b.target_y for b in sub_batches], dim=1)) | |