julessanchez / x-nesyl

Main project of the internship done at ENSTA. FasterRCNN + shap backprop

architectural_style_classification

Usage

Provide path to the data inside config.py It is expected to follow the branching present originally in the dataset, otherwise perform adequat change to config.py and dataloader.py

Folder organisation

• Root
  • Data
    • Dataset-IGRB1092_14cls
      • original disposition of the dataset
      • csv for training are created here (provided in this project, feel free to move them at the right place)
    • pascalPartDataset
      • original disposition of the dataset
      • txt for training are created here (provided in this project, feel free to move them at the right place)
  • MonuMAI-AutomaticStyleClassification (my branch)
  • pytorch-retinanet (my branch)
  • architectural_style_classification (this project)

If the folder organisation if different, change the various import path + the config.py file.

How to run

First run 'python build-csv.py' Then 'train.py' default parameter should work

Models

Models can be found at : https://cloud.minesparis.psl.eu/index.php/s/TBKECKgGZoMp1aQ

x-nesyl's Issues

bug - compare_shap_and_KG

Values written to

X-NeSyL/utils/knowledge_graph.py

Line 13 in da333ce

archi_features = [el for sublist in list(MonMonumenAI.ELEMENT_DIC.values()) for el in sublist]

are overwritten in

X-NeSyL/utils/knowledge_graph.py

Line 23 in da333ce

archi_features = [el for sublist in list(MonPascal.ELEMENT_DIC.values()) for el in sublist]

Meaning that MonuMAI dataset can't be used without modifying the previous code and

X-NeSyL/utils/knowledge_graph.py

Lines 34 to 38 in da333ce

	if dataset == 'MonumenAI':
	knowledge_graph = knowledge_graph_monumenai
	if dataset == 'PascalPart':
	knowledge_graph = knowledge_graph_pascal
	contrib = np.zeros((len(true_labels),len(archi_features)))

The should have either unique identifiers or change

X-NeSyL/utils/knowledge_graph.py

Line 38 in da333ce

contrib = np.zeros((len(true_labels),len(archi_features)))

to contrib = np.zeros((len(true_labels), knowledge_graph.shape[0]))

bug - GED_metric

X-NeSyL/utils/knowledge_graph.py

Line 167 in da333ce

facade = np.copy(shap_array[i,:,k])*features[k]

presumably should be features[i], otherwise only the first 4 are used, due to

X-NeSyL/utils/knowledge_graph.py

Line 162 in da333ce

shap_array = np.dstack((shap_values[0],shap_values[1],shap_values[2],shap_values[3]))

and

X-NeSyL/utils/knowledge_graph.py

Line 166 in da333ce

for k in range(shap_array.shape[-1]):

and if the number of object classes is greater than the batch size the code would produce an Index Out of Bounds exception, if line 162 didn't limit it to 4.

X-NeSyL/utils/knowledge_graph.py

Lines 180 to 181 in da333ce

	facade[facade<threshold] = 1
	facade[facade>=threshold] = 0

makes all features 0 always, which mean it does no contribution to the facade graph.