Merge pull request #47 from hpi-epic/add_normalization_after_sample

Add normalization after sampling

README.md

@@ -101,6 +101,7 @@ python3 -m twine upload dist/*
 | beta_lower_limit | (0, Inf) | 0.5 | Lower limit for beta values for influence of continuous parents. Betas are sampled uniform from the union of [-upper,-lower] and [lower,upper]. Upper limit see below. |
 | beta_upper_limit | (0, Inf) | 1 | Upper limit for beta values for influence of continuous parents. Betas are sampled uniform from the union of [-upper,-lower] and [lower,upper]. Lower limit see above. |
 | graph_structure_file |  | None | Defines a path to a .gml file for a fixed DAG structure (ignoring node and edge characteristics) used during manm_cs graph building. Note graph_structure_file is mutually exclusive to num_nodes and edge_density. |
+| normalized | 0 or 1 | 0 | Defines if the generated samples of continuous variables are normalized after sampling. |
 
 ## License
 

manm_cs/__main__.py

@@ -124,6 +124,8 @@ def parse_args():
     parser.add_argument('--output_samples_file', type=str, required=False, default=SAMPLES_FILE,
                     help='Output file (path) for the generated samples csv. Relative to the directory from which the library is executed.'
                     'Specify without file extension.')
+    parser.add_argument('--normalize', type=to_bool, required=False, default=False,
+                    help='Normalize the continuous variables in the dataset once all samples are generated.')
     args = parser.parse_args()
 
     assert args.min_discrete_value_classes <= args.max_discrete_value_classes, \
@@ -171,6 +173,8 @@ def graph_from_args(args) -> Graph:
     graph = graph_from_args(args)
 
     dfs = graph.sample(num_observations=args.num_samples, num_processes=args.num_processes)
+    if args.normalize:
+        dfs = graph.normalize_continous_columns(dataframes=dfs)
     write_single_csv(dataframes=dfs, target_path=f"{args.output_samples_file}.csv")
 
     nx_graph = graph.to_networkx_graph()

manm_cs/graph/graph.py

@@ -6,7 +6,7 @@
 import numpy as np
 from pathos.multiprocessing import ProcessingPool
 
-from manm_cs.variables.variable import Variable
+from manm_cs.variables.variable import Variable, VariableType
 
 
 class Graph:
@@ -50,3 +50,14 @@ def to_networkx_graph(self) -> nx.DiGraph:
             for parent in var.parents:
                 nx_graph.add_edge(parent.idx, var.idx)
         return nx_graph
+
+    def normalize_continous_columns(self, dataframes: List[pd.DataFrame]) -> List[pd.DataFrame]:
+
+        # merge df in dataframes
+        merged_df = pd.concat(dataframes, axis=1)
+        for variable in self.variables:
+            if variable.type == VariableType.CONTINUOUS:
+                print(variable.idx, merged_df[variable.idx].mean(), merged_df[variable.idx].std())
+                merged_df[variable.idx] =  (merged_df[variable.idx] - merged_df[variable.idx].mean())/ merged_df[variable.idx].std()
+
+        return [merged_df]

setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = manm_cs
-version = 0.1.1
+version = 0.1.2
 author = Johannes Huegle, Christopher Hagedorn, Lukas Boehme, Mats Poerschke, Jonas Umland
 author_email = [email protected]
 description = Data generation for causal structure learning based on mixed additive noise model (MANM)