Hello,

I've been trying to train a model with the same instructions you provide in the Readme of molvae, but I have not been able to get the same reconstruct as you. First I used your code, and had some problems with this part of the code in vaetrain.py:

if (it + 1) % 1500 == 0: #Fast annealing
            scheduler.step()
            print "learning rate: %.6f" % scheduler.get_lr()[0]
            torch.save(model.state_dict(), opts.save_path + "/model.iter-%d-%d" % (epoch, it + 1))
            beta = max(1.0, beta + anneal) 
 

When the code first enters into this if, the LOGs of the reconstruction goes crazy and the final model I obtained had 0 reconstruction score. After reading and understanding more of your code I thought that the max in beta = max(1.0, beta + anneal) may be a min, so I changed that line of code to beta = min(1.0, beta + anneal).

With this new change, I trained the whole model and now the LOGs seemed fine, but after I had the whole model, I tried to test this new model with the reconstruct.py and the model is giving me about a 0.52 reconstruct score, way too far from the 0.77 reconstruction score that the model MPNVAE-h450-L56-d3-beta0.005 has.

I would like to know if the change I made to the beta = max(1.0, beta + anneal) it's correct, also why with or without this change, I'm not being able to train a new model following your instructions? Is there something I would have to do that I'm not doing to train a model? Did you trained your model with the same instructions that the Readme says?

Thank you, I'm going crazy because this doesn't make sense at all, I would be very grateful if you had any idea of what's happening.

Hey, when running your model with my own dataset I'm getting the following error:

Traceback (most recent call last):
  File "pretrain.py", line 69, in <module>
    loss, kl_div, wacc, tacc, sacc, dacc, pacc = model(batch, beta=0)
  File "/home/.conda/envs/jtnn_pytorch/lib/python2.7/site-packages/torch/nn/modules/module.py", line 224, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/molopt/jtnn/jtprop_vae.py", line 76, in forward
    tree_mess, tree_vec, mol_vec = self.encode(mol_batch)
  File "/home/molopt/jtnn/jtprop_vae.py", line 60, in encode
    mol_vec = self.mpn(mol2graph(smiles_batch))
  File "/home/molopt/jtnn/mpn.py", line 75, in mol2graph
    agraph[a,i] = b
IndexError: index 6 is out of range for dimension 0 (of size 6)

--> PS: I changed MAXNB in jtnn_dec and jtnn_enc to 32.

Hi, I'm using my own train data to run Bayesian Optimization, but I'm not sure how to generate my Vocab file using my train file. Can somebody help me with this?

Thanks

Dear Wengong: You use the following code to run gen_latent.py in the bo folder.: python gen_latent.py --data ../data/train.txt --vocab ../data/vocab.txt
--hidden 450 --depth 3 --latent 56
--model ../molvae/MPNVAE-h450-L56-d3-beta0.005/model.4 But the system throws the following error: No such file or directory: '../molvae/MPNVAE-h450-L56-d3-beta0.005/model.4 as the file in molvae is named model.iter-4 instead of model.4

I tried passing model-iter.4 instead of model.4 but it still shows an error. Could you help resolve this? Thanks.

Hi, I want to run the sample file on the provided model under the fast_molvae directory. But I found that the size of vocabulary in the mose directory may not match with the provided model. And I got the error below. The size of the learned embedding in the mode is 800 which indicates that that model is built on a vocab of size 800.

I'm not sure whether the problem is because of vocabulary, but if it is, can you provide that vocab file?

Thanks.

le.py\", line 845, in load_state_dict
    self.__class__.__name__, \"\n\t\".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for JTNNVAE:
        size mismatch for jtnn.embedding.weight: copying a param with shape torch.Size([531, 450]) from checkpoint, the shape in current model is torch.Size([800, 450]).
        size mismatch for decoder.embedding.weight: copying a param with shape torch.Size([531, 450]) from checkpoint, the shape in current model is torch.Size([800, 450]).
        size mismatch for decoder.W_o.bias: copying a param with shape torch.Size([531]) from checkpoint, the shape in current model is torch.Size([800]).
        size mismatch for decoder.W_o.weight: copying a param with shape torch.Size([531, 450]) from checkpoint, the shape in current model is torch.Size([800, 450]).

Hi Wengong,

Thank you for sharing this wonderful project.

I have issues on the second step of training in fast_molvae. Due to the MOSES dataset is too large for my hardware to training, I chose to train and test on ZINC dataset. The step of Deriving Vocabulary and Step 1 of Training works perfectly. But there are some "CUDA out of memory" issues happened at Step 2 of Training. One major problem it caused is there is no model output in vae_model folder (actually it is empty after my Step 2). It makes impossible for me to do the testing.

The testing platform on my side is: Ubuntu 18.04/python 2.7/ cuda v9.1.85 / pytorch 1.0.1.post2 /rdkit 2018.09.2 GPU: NVIDIA Quadro P620

Thanks!

Hi! This error occurs when I ran sample.py in fast_molvae without any changes.

Traceback (most recent call last): File "/fast_molvae/sample.py", line 39, in print model.sample_prior() File "/jtnn_vae.py", line 119, in sample_prior return self.decode(z_tree, z_mol, prob_decode) File "/fast_jtnn/jtnn_vae.py", line 200, in decode jtenc_holder, mess_dict = JTNNEncoder.tensorize_nodes(pred_nodes, scope) File "/fast_jtnn/jtnn_enc.py", line 73, in tensorize_nodes cid = y.cid[1] if y.cid[0] == x.idx else 0 AttributeError: 'MolTreeNode' object has no attribute 'cid'

Process finished with exit code 1

Maybe something wrong with your codes?

Thank you!

I was trying to get the fast_molvae code to run on my own dataset and I changed the preprocess.py function code a little bit to run on my parallel worker setup. In the process I ended up removing the rdkit logger that's set to critical and I noticed a bunch of these exceptions:

[23:37:03] Explicit valence for atom # 1 C, 5, is greater than permitted [23:37:03] Explicit valence for atom # 3 C, 5, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 6, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 5, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 6, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 6, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 5, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 6, is greater than permitted [23:37:03] Explicit valence for atom # 3 C, 5, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 5, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 7, is greater than permitted [23:37:03] Explicit valence for atom # 3 C, 5, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 5, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 7, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 5, is greater than permitted [23:37:03] Explicit valence for atom # 1 C, 6, is greater than permitted [[2323::3737::0303] ] Explicit valence for atom # 1 C, 5, is greater than permitted Explicit valence for atom # 4 C, 5, is greater

I think they are coming from get_clique_mol-> sanitize() in the chemutils.py. I am not sure if I am missing something but is this supposed to happen? I tried to repeat the same thing with your training set and same set of logs appear. And next to the return statement in the sanitize method in utils I notice that there is a comment - #We assume this is not None . So are we supposed to "clean" our data to not include the smiles for which this happens ?

Dear Wengong Jin

I'd like to ask your help about molopt during running pretrain.py I have successfully done all of example in molopt with data/train.txt , data/vocab.txt, data/train.logP-SA.

However RuntimeError has occurred with my own training dataset , vocab generated with python ../jtnn/mol_tree.py < my_dataset.txt and my own logP property file.

It seems to be wrong dimension during node aggregation. What's your opinion about this issue ?

Best Regards, Minkyu Ha

( environment is same with you. python 2.7, cuda 8.0, pytorch 0.3.1)

Model #Params: 4271K Traceback (most recent call last): File "pretrain.py", line 69, in loss, kl_div, wacc, tacc, sacc, dacc, pacc = model(batch, beta=0) File "/home/minkyuha/anaconda2/lib/python2.7/site-packages/torch/nn/modules/module.py", line 357, in call result = self.forward(*input, **kwargs) File "/home/minkyuha/new-jtnn/icml18-jtnn/jtnn/jtprop_vae.py", line 76, in forward tree_mess, tree_vec, mol_vec = self.encode(mol_batch) File "/home/minkyuha/new-jtnn/icml18-jtnn/jtnn/jtprop_vae.py", line 57, in encode tree_mess,tree_vec = self.jtnn(root_batch) File "/home/minkyuha/anaconda2/lib/python2.7/site-packages/torch/nn/modules/module.py", line 357, in call result = self.forward(*input, **kwargs) File "/home/minkyuha/new-jtnn/icml18-jtnn/jtnn/jtnn_enc.py", line 62, in forward cur_h_nei = torch.cat(cur_h_nei, dim=0).view(-1,MAX_NB,self.hidden_size) RuntimeError: invalid argument 2: size '[-1 x 8 x 420]' is invalid for input with 144900 elements at /opt/conda/conda-bld/pytorch_1523240155148/work/torch/lib/TH/THStorage.c:37

While running the run_bo.py file in Bayesian Optimization, I get the following error: File "run_bo.py", line 87, in sgp.train_via_ADAM(X_train, 0 * X_train, y_train, X_test, X_test * 0, y_test, minibatch_size = 10 * M, max_iterations = 100, learning_rate = 0.001) File "/home/naman_churiwala_quantiphi_com/icml18-jtnn/bo/sparse_gp.py", line 213, in train_via_ADAM e = self.getEnergy() File "/home/naman_churiwala_quantiphi_com/icml18-jtnn/bo/sparse_gp.py", line 102, in getEnergy self.sparse_gp.compute_output() File "/home/naman_churiwala_quantiphi_com/icml18-jtnn/bo/sparse_gp_theano_internal.py", line 77, in compute_output self.KzzInv = T.nlinalg.MatrixInversePSD()(self.Kzz) AttributeError: 'module' object has no attribute 'MatrixInversePSD'

Is it possible that the attribute is MatrixInverse instead of MatrixInversePSD?

I have run run_bo.py previously without encountering the above error, but this seems to be a problem now. Can you help me with this? Thanks

Hi there,

When I have a molecule in training with a single character SMILES (such as "C" for methane), the encoder is failing with IndexError: list index out of range on the following:

def encode(self, mol_batch):
        set_batch_nodeID(mol_batch, self.vocab)
        root_batch = [mol_tree.nodes[0] for mol_tree in mol_batch]

because mol_tree.nodes is empty. From my understanding, I thought methane should be a single node graph. "C" is present in my vocabulary, and this also happens for other single character SMILES.

I removed these molecules from my training since they aren't helpful for my purposes anyway, but wasn't sure if this was a bug.

Hi there! I have a question regarding the logp-sa score computation. Using the simple script below, I can reproduce the log-sa scores in the test data data/zinc/opt.test.logP-SA BUT NOT in the training data data/zinc/train.logP-SA. Suggestions, advice, and explanations appreciated regarding this mismatch. Thanks!

from rdkit import Chem
from rdkit.Chem import Descriptors
from molopt import sascorer

# the smiles below is the first one in `data/zinc/opt.test.logP-SA`
# the score computed below (-2.5248038322) matches that in the file (-2.5248038322)
smiles = 'CC(C)OC(=O)c1cccc(-c2ccc([C@H]3[NH2+][C@H](C(=O)[O-])C(C)(C)S3)o2)c1'
score = Descriptors.MolLogP(Chem.MolFromSmiles(smiles)) - sascorer.calculateScore(Chem.MolFromSmiles(smiles))

# the smiles below is the first one in `data/zinc/train.logP-SA`
# the score computed below (3.412092566642019) DOES NOT matches that in the file (2.878620321486616174)
smiles = 'CCCCCCC1=NN2C(=N)/C(=C\c3cc(C)n(-c4ccc(C)cc4C)c3C)C(=O)N=C2S1'
score = Descriptors.MolLogP(Chem.MolFromSmiles(smiles)) - sascorer.calculateScore(Chem.MolFromSmiles(smiles))

While running preprocess.py - python3 preprocess.py --train ../data/moses/train.txt --split 100 --jobs 16 Traceback (most recent call last): File "preprocess.py", line 10, in from fast_jtnn import * File "/home/shreya/Desktop/jtnn/icml18-jtnn-master/fast_molvae/fast_jtnn/init.py", line 2, in from jtnn_vae import JTNNVAE File "/home/shreya/Desktop/jtnn/icml18-jtnn-master/fast_molvae/jtnn_vae.py", line 5, in from nnutils import create_var, flatten_tensor, avg_pool ImportError: cannot import name 'flatten_tensor' from 'nnutils'

Can anyone help with this?

While generating latent representations for BO, there are a few molecules that are perhaps not in the vocabulary, I get the following error: File "gen_latent.py", line 84, in <module> mol_vec = model.encode_latent_mean(batch) File "JT-VAE/icml18-jtnn/jtnn/jtnn_vae.py", line 61, in encode_latent_mean _, tree_vec, mol_vec = self.encode(mol_batch) File "JT-VAE/icml18-jtnn/jtnn/jtnn_vae.py", line 48, in encode set_batch_nodeID(mol_batch, self.vocab) File "JT-VAE/icml18-jtnn/jtnn/jtnn_vae.py", line 22, in set_batch_nodeID node.wid = vocab.get_index(node.smiles) File "JT-VAE/icml18-jtnn/jtnn/mol_tree.py", line 18, in get_index return self.vmap[smiles] KeyError: 'C1=CN=NC=C1'

Is there any way that we could measure variable MAX_NB? While pretraining the code with our new dataset, we get below scores during training, [2100] Beta: 0.000, KL: 548.74, Word: 101.77, Topo: 106.81, Assm: 98.06, PNorm: 160.23, GNorm: 37.14 We thought that MAX_NB should be set correctly by each dataset. Thus, is there any way that we could measure MAX_NB variable?

Hi, I want to know how to print

valid = 1.0
unique@1000 = 1.0
unique@10000 = 0.9992
FCD/Test = 0.42235413520261034
SNN/Test = 0.5560595345050097
Frag/Test = 0.996223352989786
Scaf/Test = 0.8924981494347503
FCD/TestSF = 0.9962165008703465
SNN/TestSF = 0.5272934146558245
Frag/TestSF = 0.9947901514732745
Scaf/TestSF = 0.10049873444911761
IntDiv = 0.8511712225340441
IntDiv2 = 0.8453088593783662
Filters = 0.9778
logP = 0.0054694810121243
SA = 0.015992957588069068
QED = 1.15692473423544e-05
NP = 0.021087573878091237
weight = 0.5403194879856983

this result. So, at MOSES I know that if I print that result, I need test, test_scaffold, train, precalculated test npz, precalculated test_scaffold npz and generated(sampled) molecules.

At /data/moses and moses-h450z56, I can get the test, test_scaffold and train molecules and generated molecules. The questions are here,

1. How can I get the scaffold molecules? for custom dataset?
2. What is the precalculated test( or test_scaffold) npz?
3. Do you use .txt not .csv for printing that result? ( at MOSES, they use .csv)

Thank you :)

when I run preprocess.py, I meet a problem. Traceback (most recent call last): File "preprocess.py", line 9, in from fast_jtnn import * File "/home/zhai/cxh/icml18-jtnn/fast_jtnn/init.py", line 1, in from mol_tree import Vocab, MolTree ModuleNotFoundError: No module named 'mol_tree' But I add export PYTHONPATH=$PREFIX/icml18-jtnn to my environment variables.