First execute compute_score.py file, get different scores with different number of extractions, then execute score_analysis.py file to save best suitable score. Next use score to run several tests with MLP or KNN.
score_analysis.py: The score curve will be saved first, and then the console will ask for the appropriate extractions (only the extractions saved in check points can be entered).
Object view dataset Parameters: Classes: 7 Objects: 3 Views: 37
Component view dataset Parameters: Classes: 7 Objects: 1 Views: 48
Python: 3.7.3 Pytorch: 1.9.1 CUDA: V11.4.120
Use following command to create environment:
conda env create -f env.yamlAll configurations are in folder config.
compute_score.json: Parameters used to calculate scores and analyze them (compute_score.py and score_analysis.py).
"data_set": contain dataset path, number of classes, number of objects per class and number of views per object.
"nb_tirages": maximum number of extraction.
"check_points": set of check points, when number of extraction hit check point, program will save the score.
"nb_sample_set_min & nb_sample_set_max": minimum and maximum number of sample classes. (ex: for object dataset, set to 3 and 5, for component dataset, set to 3 and 7)
"nb_sample_min & nb_sample_max": minimum and maximum number of sample views in one class. (ex: for object dataset, set to 5 and 12, for component dataset, set to 3 and 7)
"score_save_path": path to store the scores saved by each check points. (need to create the folder manually)
"score_analysis_results_save_path": path to store the result of analyzing scores (get best round of extraction(optimal global)). (need to create the folder manually)
train_mlp.json: Parameters used to train model (train_mlp.py).
"train_set & eval_set": contain dataset path, number of classes, number of objects per class and number of views per object.
"score_file": score record file from the result of analyzing scores.
"training_type": two options: "top" use optimal views, "random" use random views.
"num_execution": number of execution, if "training_type" set to "top", by default will be 1, if "training_type" set to "random", each execution will select different set of random views.
"train_size_per_class & eval_size_per_class": number of views used for each class in training or evaluation. Note: "eval_size_per_class" will not be used if "cross_view" set to "union", "inter" and "complet".
"cross_view": Four options: "union": use whole evalset, "inter": use same angles of views in training set, "complet": use views with different angles of views in training set, "other": can manually set the number of "eval_size_per_class".
"model_save_dir": path to save trained model weights. (need to create the folder manually)
"hyperparameters": Note: "nb_layer_fix": set first number of layers fix(frozen parameters). "cosine_annealing": false or true, use cosine annealing algorithm on learning rate or not.
test_mlp.json: Parameters used to test model (test_mlp.py).
"test_set": contain dataset path, number of classes, number of objects per class and number of views per object.
"score_file": score record file from the result of analyzing scores.
"test_size_per_class": number of views used for each class in testing. Note: will not be used if "cross_view" set to "union", "inter" and "complet".
"cross_view": Four options: "union": use whole testset, "inter": use same angles of views in training set, "complet": use views with different angles of views in training set, "other": can manually set the number of "test_size_per_class".
"weights_files_dir": folder where store model weights in training process.
"nb_layer_fix": need to be set to the same value in training process.
"device": two options: "cuda": use cuda device, "cpu": use cpu device.
train_test_knn.json: Parameters used to train and test on knn (train_test_knn.py).
"train_set & test_set": contain dataset path, number of classes, number of objects per class and number of views per object.
"score_file": score record file from the result of analyzing scores.
"training_type": two options: "top" use optimal views, "random" use random views.
"num_execution": number of execution, if "training_type" set to "top", by default will be 1, if "training_type" set to "random", each execution will select different set of random views.
"train_size_per_class & test_size_per_class": number of views used for each class in training or testing. Note: "test_size_per_class" will not be used if "cross_view" set to "union", "inter" and "complet".
"cross_view": Four options: "union": use whole testset, "inter": use same angles of views in training set, "complet": use views with different angles of views in training set, "other": can manually set the number of "test_size_per_class".
"weights_files_dir": folder where store model weights in training process or set to "None" if only used as a pre-trained feature extractor.
"nb_layer_fix": need to be set to the same value in training process, will not be used if "weights_files_dir" set to "None".
Direct use python3 + filename, ex:
python3 compute_score.py