示例#1
0
n_seeds = 5
results_list = []
explanations = {i: [] for i in range(n_classes)}
for seed in range(n_seeds):
    seed_everything(seed)
    print(f'Seed [{seed + 1}/{n_seeds}]')
    train_loader = DataLoader(train_data, batch_size=len(train_data))
    val_loader = DataLoader(val_data, batch_size=len(val_data))
    test_loader = DataLoader(test_data, batch_size=len(test_data))

    checkpoint_callback = ModelCheckpoint(dirpath=base_dir, monitor='val_loss', save_top_k=1)
    trainer = Trainer(max_epochs=500, gpus=1, auto_lr_find=True, deterministic=True,
                      check_val_every_n_epoch=1, default_root_dir=base_dir,
                      weights_save_path=base_dir, callbacks=[checkpoint_callback])
    model = Explainer(n_concepts=n_concepts, n_classes=n_classes, l1=0, lr=0.01, explainer_hidden=[10])

    trainer.fit(model, train_loader, val_loader)
    print(f"Concept mask: {model.model[0].concept_mask}")
    model.freeze()
    model_results = trainer.test(model, test_dataloaders=test_loader)
    for j in range(n_classes):
        n_used_concepts = sum(model.model[0].concept_mask[j] > 0.5)
        print(f"Extracted concepts: {n_used_concepts}")
    results = {}
    results['model_accuracy'] = model_results[0]['test_acc']

    results_list.append(results)

    results_df = pd.DataFrame(results_list)
    results_df.to_csv(os.path.join(base_dir, 'results_aware_cub.csv'))
示例#2
0
            checkpoint_callback_xc = ModelCheckpoint(dirpath=base_dir,
                                                     monitor='val_loss',
                                                     save_top_k=1)
            trainer_xc = Trainer(max_epochs=200,
                                 gpus=1,
                                 auto_lr_find=True,
                                 deterministic=True,
                                 check_val_every_n_epoch=1,
                                 default_root_dir=base_dir + '_xc',
                                 weights_save_path=base_dir,
                                 callbacks=[checkpoint_callback_xc])
            model_xc = Explainer(n_concepts=x.shape[1],
                                 n_classes=c.shape[1],
                                 l1=0,
                                 lr=0.01,
                                 explainer_hidden=[100, 50],
                                 temperature=5000,
                                 loss=torch.nn.BCEWithLogitsLoss())
            trainer_xc.fit(model_xc, train_loader_xc, val_loader_xc)
            model_xc.freeze()
            c_train_pred = model_xc.model(x_train)
            c_val_pred = model_xc.model(x_val)
            c_test_pred = model_xc.model(x_test)

            # train C->Y
            train_data = TensorDataset(c_train_pred.squeeze(), y_train)
            val_data = TensorDataset(c_val_pred.squeeze(), y_val)
            test_data = TensorDataset(c_test_pred.squeeze(), y_test)
            train_loader = DataLoader(train_data, batch_size=train_size)
            val_loader = DataLoader(val_data, batch_size=val_size)
示例#3
0
    y_trainval, y_test = torch.FloatTensor(y[trainval_index]), torch.FloatTensor(y[test_index])
    x_train, x_val, y_train, y_val = train_test_split(x_trainval, y_trainval, test_size=0.2, random_state=42)
    print(f'{len(y_train)}/{len(y_val)}/{len(y_test)}')

    train_data = TensorDataset(x_train, y_train)
    val_data = TensorDataset(x_val, y_val)
    test_data = TensorDataset(x_test, y_test)
    train_loader = DataLoader(train_data, batch_size=train_size)
    val_loader = DataLoader(val_data, batch_size=val_size)
    test_loader = DataLoader(test_data, batch_size=test_size)

    checkpoint_callback = ModelCheckpoint(dirpath=base_dir, monitor='val_loss', save_top_k=1)
    trainer = Trainer(max_epochs=200, gpus=1, auto_lr_find=True, deterministic=True,
                      check_val_every_n_epoch=1, default_root_dir=base_dir,
                      weights_save_path=base_dir, callbacks=[checkpoint_callback])
    model = Explainer(n_concepts=n_concepts, n_classes=n_classes, l1=1e-3, lr=0.01,
                      explainer_hidden=[20], temperature=0.7)

    start = time.time()
    trainer.fit(model, train_loader, val_loader)
    print(f"Gamma: {model.model[0].concept_mask}")
    model.freeze()
    model_results = trainer.test(model, test_dataloaders=test_loader)
    for j in range(n_classes):
        n_used_concepts = sum(model.model[0].concept_mask[j] > 0.5)
        print(f"Extracted concepts: {n_used_concepts}")
    results, f = model.explain_class(val_loader, train_loader, test_loader,
                                     topk_explanations=10,
                                     concept_names=concept_names)
    end = time.time() - start
    results['model_accuracy'] = model_results[0]['test_acc']
    results['extraction_time'] = end
示例#4
0
    checkpoint_callback = ModelCheckpoint(dirpath=base_dir,
                                          monitor='val_loss',
                                          save_top_k=1)
    trainer = Trainer(max_epochs=10,
                      gpus=1,
                      auto_lr_find=True,
                      deterministic=True,
                      check_val_every_n_epoch=1,
                      default_root_dir=base_dir,
                      weights_save_path=base_dir,
                      callbacks=[checkpoint_callback])
    model = Explainer(n_concepts=n_concepts,
                      n_classes=n_classes,
                      l1=0.0000001,
                      temperature=5,
                      lr=0.01,
                      explainer_hidden=[10],
                      conceptizator='identity_bool')

    start = time.time()
    trainer.fit(model, train_loader, val_loader)
    print(f"Concept mask: {model.model[0].concept_mask}")
    model.freeze()
    model_results = trainer.test(model, test_dataloaders=test_loader)
    for j in range(n_classes):
        n_used_concepts = sum(model.model[0].concept_mask[j] > 0.5)
        print(f"Extracted concepts: {n_used_concepts}")
    results, f = model.explain_class(val_loader,
                                     val_loader,
                                     test_loader,
示例#5
0
    test_loader = DataLoader(test_data, batch_size=len(test_data))

    checkpoint_callback = ModelCheckpoint(dirpath=base_dir,
                                          monitor='val_loss',
                                          save_top_k=1)
    trainer = Trainer(max_epochs=500,
                      gpus=1,
                      auto_lr_find=True,
                      deterministic=True,
                      check_val_every_n_epoch=1,
                      default_root_dir=base_dir,
                      weights_save_path=base_dir,
                      callbacks=[checkpoint_callback])
    model = Explainer(n_concepts=n_concepts,
                      n_classes=n_classes,
                      l1=0.0001,
                      temperature=0.7,
                      lr=0.01,
                      explainer_hidden=[10])

    start = time.time()
    trainer.fit(model, train_loader, val_loader)
    print(f"Concept mask: {model.model[0].concept_mask}")
    model.freeze()
    model_results = trainer.test(model, test_dataloaders=test_loader)
    for j in range(n_classes):
        n_used_concepts = sum(model.model[0].concept_mask[j] > 0.5)
        print(f"Extracted concepts: {n_used_concepts}")
    results, f = model.explain_class(val_loader,
                                     train_loader,
                                     test_loader,
                                     topk_explanations=50,
示例#6
0
    checkpoint_callback_xc = ModelCheckpoint(dirpath=base_dir,
                                             monitor='val_loss',
                                             save_top_k=1)
    trainer_xc = Trainer(max_epochs=200,
                         gpus=1,
                         auto_lr_find=True,
                         deterministic=True,
                         check_val_every_n_epoch=1,
                         default_root_dir=base_dir + '_xc',
                         weights_save_path=base_dir,
                         callbacks=[checkpoint_callback_xc])
    model_xc = Explainer(n_concepts=x.shape[1],
                         n_classes=c.shape[1],
                         l1=0,
                         lr=0.01,
                         explainer_hidden=[100, 50],
                         temperature=5000,
                         loss=torch.nn.BCEWithLogitsLoss())
    trainer_xc.fit(model_xc, train_loader_xc, val_loader_xc)
    model_xc.freeze()
    c_train_pred = model_xc.model(x_train)
    c_val_pred = model_xc.model(x_val)
    c_test_pred = model_xc.model(x_test)

    # train C->Y
    train_data = TensorDataset(c_train_pred.squeeze(), y_train)
    val_data = TensorDataset(c_val_pred.squeeze(), y_val)
    test_data = TensorDataset(c_test_pred.squeeze(), y_test)
    train_loader = DataLoader(train_data, batch_size=train_size)
    val_loader = DataLoader(val_data, batch_size=val_size)
示例#7
0
n_seeds = 5
results_list = []
explanations = {i: [] for i in range(n_classes)}
for seed in range(n_seeds):
    seed_everything(seed)
    print(f'Seed [{seed + 1}/{n_seeds}]')
    train_loader = DataLoader(train_data, batch_size=len(train_data))
    val_loader = DataLoader(val_data, batch_size=len(val_data))
    test_loader = DataLoader(test_data, batch_size=len(test_data))

    checkpoint_callback = ModelCheckpoint(dirpath=base_dir, monitor='val_loss', save_top_k=1)
    trainer = Trainer(max_epochs=10, gpus=1, auto_lr_find=True, deterministic=True,
                      check_val_every_n_epoch=1, default_root_dir=base_dir,
                      weights_save_path=base_dir, callbacks=[checkpoint_callback])
    model = Explainer(n_concepts=n_concepts, n_classes=n_classes, l1=0, lr=0.01,
                        explainer_hidden=[10], conceptizator='identity_bool')

    trainer.fit(model, train_loader, val_loader)
    print(f"Concept mask: {model.model[0].concept_mask}")
    model.freeze()
    model_results = trainer.test(model, test_dataloaders=test_loader)
    for j in range(n_classes):
        n_used_concepts = sum(model.model[0].concept_mask[j] > 0.5)
        print(f"Extracted concepts: {n_used_concepts}")
    results = {}
    results['model_accuracy'] = model_results[0]['test_acc']

    results_list.append(results)

results_df = pd.DataFrame(results_list)
results_df.to_csv(os.path.join(base_dir, 'results_aware_mnist.csv'))