kaist-amsg/composition-conditioned-crystal-gan

Composition-Conditioned Crystal GAN pytorch code

composition-conditioned-crystal-gan's Issues

Data visualisation

Facing a problem while training

Dear Developers
@syaym
I am trying to train using mgmno_2000,pickle. I am running it on Google Colab, and in the last line of the train.py, I am getting a Runtime Error: Mismatch in shape: grad_output[0] has a shape of torch.Size([1]) and output[0] has a shape of torch.Size([]).
I am attaching a screenshot of the same.

The line where I am facing the error:
if name == 'main':
print("not import")
main()
else:
print("import")
pass

On the problems for the training data.

Dear Composition-Conditoned-Crystal Gan developers.

I'm now trying to run your github code of your paper
'Generative Adversarial Networks for Crystal Structure Prediction'

In order to run the train.py , it needs the training data mgmno_2000.pickle
and I cannot found the training data on the github.
Then, I had make the training data by run

1. 5.make_comp_dict.py and 2) 6.data_augmentation_mgmno.py from
   'unique_sc_mnmno.npy' and 'unique_sc_mgmno_name_list' on the github.

However, the train.py code did not work with the generated training data.

I see the code of train.py then, found the problems below,

The training data are assumed to be packed with crystal images and labels in the code
like,

for j, (imgs, label) in enumerate(dataloader):
batch_size = img.shape[0]
real_imgs = img.view(batch_size,1,30,3)

I think it assumes images (denoted as C in your paper : the representation of the crystal structure)
and labels (denoted as A in your paper, the atomic status ) for the training data.

However, in the codes of data preparation (6.data_augmentation_mgmno.py)
the only images data is dumped. Atom status is not generated from the scripts for data preparation on the github.

On the S.I of your paper, the loss function of the classifier are written as

L_class_comp = CE(C_real,\hat(C_real) + lamba1 CE(C_gen, \hat(C_gen))
L_class_atm = CE(A_real,\hat(A_real) + lamba1 CE(A_gen, \hat(A_gen))
L_class = L_class_atom + lambdaC * L_class_comp

On the other hand, in the code (train.py)

cat_loss_real = 0.3*(cat_loss_mg_real + cat_loss_mn_real + cat_loss_o_real) + cat_loss_mg_real2+cat_loss_mn_real2 + cat_loss_o_real

it seems L_class is given like

L_class = L_class_comp + 0.3 * L_class_atom

Furthermore, in the code (train.py), it does not considered the loss function term

CE(Agen,\hat(A_gen) . In the code of train.,py

fake_mg_label fake_mn_label, fale_o_label, fake_mg_cat, ,,,,, = net_Q(fake)

fake_mg_label, fake_mn_label, fake_o_label are not used in the code, and not implemented the

loss function term CE(A_gen, \hat(A_gen)) .

If I can get your reply, I'm very happy.

Sincerely,

Yukihiro Okuno.