Function SHAP

pairwise/dataloader.py:1083–1361

(model, model_weights,train_val_dataset, test_loader,args)

    return network_weights

def SHAP(model, model_weights,train_val_dataset, test_loader,args):
    ####################
    # calcuate shapley
    ####################
    print('calculate shapely values')
    model.eval()
    model.load_state_dict(torch.load(model_weights))

    #这里需要修改
    # refer to "k_fold_trainer_graph_TGSynergy" to load dataset
    if args.model in ['TGSynergy', 'deepdds_wang']:
        # explainer = SubgraphX(model, num_classes=4, device="cpu", explain_graph=False,
        #                 reward_method='nc_mc_l_shapley')

        train_val_dataset_drug = train_val_dataset[0]
        train_val_dataset_drug2 = train_val_dataset[1]
        train_val_dataset_cell = train_val_dataset[2]
        train_val_dataset_target = train_val_dataset[3].tolist()
        trainval_df = [train_val_dataset_drug,train_val_dataset_drug2,train_val_dataset_cell,train_val_dataset_target]
        trainval_df = pd.DataFrame(trainval_df).T  # shape (138878, 4)
        train_dataset = MyDataset(trainval_df)
        train_val_loader = torch_geometric.data.DataLoader(train_dataset, batch_size=256,shuffle = False)
        #batch = next(iter(train_val_loader))
        #background, _ = batch[:-1], batch[-1]

        # a dict to store both activations
        activation = {}
        def getActivation(name):
            def hook(model, input, output):
                if name in activation:
                    activation[name].append(output.detach())
                else:
                    activation[name] = [output.detach()]
            return hook

        if args.model == 'TGSynergy':
            # register forward hooks on the layers of choice
            h1 = model.drug_emb.register_forward_hook(getActivation('drug_emb'))
            h2 = model.cell_emb.register_forward_hook(getActivation('cell_emb'))
        elif args.model == 'deepdds_wang':
            h1 = model.drug_emb.register_forward_hook(getActivation('drug_emb'))
            h2 = model.cell_emb.register_forward_hook(getActivation('cell_emb'))

        batch = next(iter(train_val_loader))
        background, _ = batch[:-1], batch[-1]
        # forward pass -- getting the outputs
        out = model(background)
            
            # detach the hooks
        #h1.remove()
        #h2.remove()

        #x_drug, x_drug2,x_cell
        bg_x = torch.cat([
            activation['drug_emb'][0], activation['drug_emb'][1],activation['cell_emb'][0]
        ], -1)