Worse accuracy while continuing training due to a possible mistake in initializing `setup`

[Kumoi0728]

I trained a MVTN model with 100 epochs with the following command, and stopped training after 57 epochs.

python run_mvtn.py --data_dir data/ModelNet40/ --run_mode train --mvnetwork mvcnn --epochs 100 --nb_views 1 --views_config learned_circular

And the output of the 57th epoch is like this,

Epoch: [57/100]
Iter [50/492] Loss: 0.7633
Iter [100/492] Loss: 0.7892
Iter [150/492] Loss: 0.3939
Iter [200/492] Loss: 0.1820
Iter [250/492] Loss: 0.2282
Iter [300/492] Loss: 0.6939
Iter [350/492] Loss: 0.4468
Iter [400/492] Loss: 0.2383
Iter [450/492] Loss: 0.5454
Evaluation:
train acc: 82.03 - train Loss: 0.6457
Val Acc: 71.31 - val Loss: 1.0960
Current best val acc: 72.61

When I load the trained model to continue training, although it started training from the 58th epoch correctly, the accuracies got lower,

Epoch: [58/100]
Iter [50/492] Loss: 1.2060
Iter [100/492] Loss: 0.6699
Iter [150/492] Loss: 0.5014
Iter [200/492] Loss: 0.4189
Iter [250/492] Loss: 0.2721
Iter [300/492] Loss: 0.3099
Iter [350/492] Loss: 1.0518
Iter [400/492] Loss: 1.1512
Iter [450/492] Loss: 0.2506
Evaluation:
train acc: 55.48 - train Loss: 1.6519
Val Acc: 60.13 - val Loss: 1.4470
Current best val acc: 72.61

I found that in ops.py line 260-264, only when is_learning_views = True , the trained MVTN model will be loaded,

if setup["is_learning_views"]:
        models_bag["mvtn"].load_state_dict(
            checkpoint['mvtn'])
        models_bag["mvtn_optimizer"].load_state_dict(
            checkpoint['mvtn_optimizer'])

and in line 55-56, is_learning_views in setup is initialized like this,

setup["is_learning_views"] = setup["views_config"] in ["learned_offset",
                                                       "learned_direct", "learned_spherical", "learned_random", "learned_transfer"]

should the learned_offset in line 55 be repalced by learned_circular?
Becaues the choices of learned views_config must be learned_circular, learned_spherical, learned_direct, learned_random or learned_transfer.

I am sorry if the reason is not here. I would appreciate it if you could tell me the correct way. :) @ajhamdi

[Kumoi0728]

I checked the results of the continued training. With epoch=57 as the cut-off point, the position of camera also changed a lot.
When epoch=57, camera 0 was like:

However, when epoch=60, camera 0 was like:

According to other epochs, the position of the camera should not have changed so much. I think this is due to the fact that the trained MVTN model was not loaded correctly when the training continued.

[ajhamdi]

yes @Kumoi0728 you are right . I think this is a bug in the cod. I will look into it

[auniquesun]

I trained a MVTN model with 100 epochs with the following command, and stopped training after 57 epochs.

python run_mvtn.py --data_dir data/ModelNet40/ --run_mode train --mvnetwork mvcnn --epochs 100 --nb_views 1 --views_config learned_circular

And the output of the 57th epoch is like this,

Epoch: [57/100]
Iter [50/492] Loss: 0.7633
Iter [100/492] Loss: 0.7892
Iter [150/492] Loss: 0.3939
Iter [200/492] Loss: 0.1820
Iter [250/492] Loss: 0.2282
Iter [300/492] Loss: 0.6939
Iter [350/492] Loss: 0.4468
Iter [400/492] Loss: 0.2383
Iter [450/492] Loss: 0.5454
Evaluation:
train acc: 82.03 - train Loss: 0.6457
Val Acc: 71.31 - val Loss: 1.0960
Current best val acc: 72.61

When I load the trained model to continue training, although it started training from the 58th epoch correctly, the accuracies got lower,

Epoch: [58/100]
Iter [50/492] Loss: 1.2060
Iter [100/492] Loss: 0.6699
Iter [150/492] Loss: 0.5014
Iter [200/492] Loss: 0.4189
Iter [250/492] Loss: 0.2721
Iter [300/492] Loss: 0.3099
Iter [350/492] Loss: 1.0518
Iter [400/492] Loss: 1.1512
Iter [450/492] Loss: 0.2506
Evaluation:
train acc: 55.48 - train Loss: 1.6519
Val Acc: 60.13 - val Loss: 1.4470
Current best val acc: 72.61

I found that in ops.py line 260-264, only when is_learning_views = True , the trained MVTN model will be loaded,

if setup["is_learning_views"]:
        models_bag["mvtn"].load_state_dict(
            checkpoint['mvtn'])
        models_bag["mvtn_optimizer"].load_state_dict(
            checkpoint['mvtn_optimizer'])

and in line 55-56, is_learning_views in setup is initialized like this,

setup["is_learning_views"] = setup["views_config"] in ["learned_offset",
                                                       "learned_direct", "learned_spherical", "learned_random", "learned_transfer"]

should the learned_offset in line 55 be repalced by learned_circular? Becaues the choices of learned views_config must be learned_circular, learned_spherical, learned_direct, learned_random or learned_transfer.

I am sorry if the reason is not here. I would appreciate it if you could tell me the correct way. :) @ajhamdi

Recently, I have experimented with the code in this repo.
I agree with you that in line 55 in ops.py, learn_offset should be replaced by learned_circular.

I found even though training from scratch, the resutls are not unsatisfactory, shown in the following figure.

In my case, I set views_config=learned_spherical and test on ScanObjectNN. According to the code, the model will adjust the scene parameters for choosing a better position to render the point clouds into images, then classify the images using MVCNN.
However, after 21 epochs, I only get 18.7% accuracy. I think the score is too low and the process is abnormal.

I have read the code and understood their working process.
There is no bug during training and evaluation, but I am not sure whether I used the proper settings.
The running command is shown in the following figure.

Do you have any insight or advice on the poor performance? Thanks.
@ajhamdi @Kumoi0728

[meifangyan05-ux]

我检查了继续训练的结果。以 epoch=57 为截止点，摄像头的位置也发生了很大变化。epoch =57 时，摄像头 0 的位置如下： 然而，epoch=60 时，摄像头 0 的位置如下：

根据其他训练轮次的数据，摄像头的位置不应该发生如此大的变化。我认为这是因为在训练继续进行时，已训练的MVTN模型没有正确加载。

This is how it runs and can be seen.I want to know if my training has any effective perspective changes.