Evaluationg other dataset

[watashihageteru]

Hi, I've been trying to reproduce the results on the dataset other than ShapeNet. I didn't face too much difficulty during the training phase, but I'm struggling with the evaluation phase.

It seems that we need to create "datasets/test_data/xxxxxx.pt" for evaluating other dataset. Would it be possible to provide the script for generating "datasets/test_data/xxxxxx.pt"?

[ZENGXH]

this is my old script used to generate the test data for the 55 classes (the dataset initialization argument may need to align with the current dataset). Is this what you are looking for?

if __name__ == '__main__':
    #te_dataset = ShapeNet15kPointClouds(
    #    categories=['chair'], #cfg.cates,
    #    split='train', #eval_split,
    #    tr_sample_size=2048, #cfg.tr_max_sample_points,
    #    te_sample_size=2048, #cfg.te_max_sample_points,
    #    # normalize_per_shape=True,
    #    normalize_global=True,
    #)
    for cats in ['all']: ##, 'car', 'airplane']: 
        ## for s in ['train', 'val', 'test']:
        train_data = ShapeNet15kPointClouds(categories=[cats], split='train',
            tr_sample_size=2048, te_sample_size=2048,
            normalize_global=True)  

        data = ShapeNet15kPointClouds(categories=[cats], split='val',
            tr_sample_size=2048, te_sample_size=2048,
            all_points_mean=train_data.all_points_mean,
            all_points_std=train_data.all_points_std,
            normalize_global=True)  

        ref, m, s, uid = [], [], [], []
        for i, d in enumerate(data):
            ref.append(d['tr_points'])
            m.append(d['mean'])
            s.append(d['std']) 
            uid.append(d['oid'])

        ref_pcs, m_pcs, s_pcs = torch.from_numpy(np.stack(ref)), torch.from_numpy(np.stack(m)), torch.from_numpy(np.stack(s)) 
        logger.info('ref_pcs: {}', ref_pcs.shape) 
        np.random.seed(0) 
        N = 1000
        xperm = np.random.permutation(np.arange(ref_pcs.shape[0]))[:N] 
        ref_pcs = ref_pcs[xperm]
        m_pcs = m_pcs[xperm] 
        s_pcs = s_pcs[xperm]
        ref_name = '/workspace/data_cache_local/test_data/ref_ns_val_%s.pt'%cats 
        print('save as: ', ref_name)
        torch.save({'ref': ref_pcs, 'mean': m_pcs, 'std': s_pcs, 'obj_id': uid}, ref_name)

[watashihageteru]

Thank you very much for providing your script. When I ran the script in combination with "LION/datasets/dataset.py", I encountered errors in the following three places.

① ・The location where the error occurred：normalize_global=True) ・Error Messsage: TypeError: init() got an unexpected keyword argument 'normalize_global' ⇒I was able to run it without errors after removing 'normalize_global', but is that okay?

② ・The location where the error occurred：ref.append(d['tr_points']) ・Error Messsage:KeyError: 'tr_points' ⇒When I checked the contents of 'd,' I found a similarly named 'train_points.' Is it okay to replace 'tr_points' with 'train_points'?"

③ ・The location where the error occurred：uid.append(d['oid'] ・Error Messsage:KeyError: 'oid' ⇒When I checked the contents of 'datasets/test_data/ref_val_airplane.pt', I found that only three variables - 'ref,' 'mean,' and 'std' - were registered as dictionaries. So, Is it okay to comment out and ignore the above?"

[ZENGXH]

for 2 and 3: yes your change makes sense. for 1: the argument is for the init function is originally defined here: https://github.com/nv-tlabs/LION/blob/main/datasets/pointflow_datasets.py#L100C18-L100C34. If your datasets does not have this argument, it's ok to remove it.

[noahcao]

@watashihageteru Can you provide more details about how to train the model on other datasets? I tried to fine-tune it while facing some issues: https://github.com/nv-tlabs/LION/issues/68