Following up from 4 days ago (Dec 14, 2019) issue...No luck for me. Instructions starting "Benchmarking your models" and below are actually not clear to me. Exactly what needs to be done step by step to run each of the 5 models? I would appreciate knowing the steps. For example, what exactly are the path --gen_path for the aae, charRNN, organ, latentgan and vae models under Training, Generation and Evaluation?. What special needs to be done for latentgan, e.g. re. ddc_pub v3, molvengen, ...? How exactly to execute sh scripts/run_all_models.sh? I am using Python 3.7.3, Jupyter Notebook 6.0.2, and have installed rdkit 2019.03.4.0, and molsets 0.2 and many other packages? Is there an exact requirements list of packages required to run MOSES?
Will appreciate your clarifications. Thanks,

When I trained char-rnn model on my pc and got 30,000 samples from this generative model. After I evalutated the results between MOSES and my own , something weird happened, my FCD Test value and FCD TestSF value is much smaller than your results. So , why ?

|char-rnn | MOSES| MY OWN RESULT| |----------|:-------------:|------:| |FCD (↓) Test | 0.355 | 02616| |FCD (↓) TestSF | 0.8995 | 0.7881|

I run the code use cuda, but I face this problem, so how can I solvie the problem?

Traceback (most recent call last): File "scripts/run.py", line 219, in main(config) File "scripts/run.py", line 199, in main train_model(config, model, train_path, test_path) File "scripts/run.py", line 127, in train_model trainer_script.main(model, trainer_config) File "/home/cyy/Code/moses/scripts/train.py", line 62, in main trainer.fit(model, train_data, val_data) File "/home/cyy/anaconda3/envs/moses/lib/python3.7/site-packages/moses/aae/trainer.py", line 284, in fit self._train(model, train_loader, val_loader, logger) File "/home/cyy/anaconda3/envs/moses/lib/python3.7/site-packages/moses/aae/trainer.py", line 219, in _train tqdm_data, criterions, optimizers) File "/home/cyy/anaconda3/envs/moses/lib/python3.7/site-packages/moses/aae/trainer.py", line 116, in _train_epoch latent_codes = model.encoder_forward(*encoder_inputs) File "/home/cyy/anaconda3/envs/moses/lib/python3.7/site-packages/moses/aae/model.py", line 109, in encoder_forward return self.encoder(*args, **kwargs) File "/home/cyy/anaconda3/envs/moses/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(*input, **kwargs) File "/home/cyy/anaconda3/envs/moses/lib/python3.7/site-packages/moses/aae/model.py", line 26, in forward x = pack_padded_sequence(x, lengths, batch_first=True) File "/home/cyy/anaconda3/envs/moses/lib/python3.7/site-packages/torch/nn/utils/rnn.py", line 249, in pack_padded_sequence _VF._pack_padded_sequence(input, lengths, batch_first) RuntimeError: 'lengths' argument should be a 1D CPU int64 tensor, but got 1D cuda:1 Long tensor

Following the instructions on the readme, I did an End-to-End launch, by running the run.py in the scripts folder.

Traceback (most recent call last): File "run.py", line 208, in main(config) File "run.py", line 190, in main sample_from_model(config, model) File "run.py", line 127, in sample_from_model sampler_script.main(model, sampler_config) File "/content/moses/scripts/moses/scripts/sample.py", line 47, in main min(n, config.n_batch), config.max_len File "/usr/local/lib/python3.7/site-packages/molsets-0.1.4-py3.7.egg/moses/vae/model.py", line 218, in sample i_eos_mask = ~eos_mask & (w == self.eos) RuntimeError: Expected object of scalar type Byte but got scalar type Bool for argument #2 'other'

How could I resolve this?

This is the error message:

(moses-env) [trial0@xps8500 moses]$ python scripts/train.py organ --help Traceback (most recent call last): File "scripts/train.py", line 5, in import rdkit File "/home/trial0/anaconda3/envs/moses-env/lib/python3.6/site-packages/rdkit/init.py", line 2, in from .rdBase import rdkitVersion as version ImportError: /lib64/libstdc++.so.6: version `GLIBCXX_3.4.21' not found (required by /home/trial0/anaconda3/envs/moses-env/lib/python3.6/site-packages/rdkit/rdBase.so)

However, rdkit itself is working:

(moses-env) [trial0@xps8500 ~]$ python Python 3.6.8 |Anaconda, Inc.| (default, Dec 30 2018, 01:22:34) [GCC 7.3.0] on linux Type "help", "copyright", "credits" or "license" for more information.

from future import print_function from rdkit import Chem m = Chem.MolFromSmiles('Cc1ccccc1') m <rdkit.Chem.rdchem.Mol object at 0x7f1e30b8bf80>

This under CentOS 7.6 after introducing the FCD module with pytorch 1.01. It looks this is not rdkit problem because ORGAN is still running under this env. without any problem. Any idea? Thanks!

Hi, Did you forget to tell us how to specify "ddc_pub"? in from .model import LatentGAN moses/latentgan/model.py", line 4 from ddc_pub import ddc_v3 as ddc

Hello,

In most implementations of VAE for molecular generation on the web there seems to be a trend to downweight the KL penalty/suppress the reparametrization in the VAE training, basically degenerating the model into a standard AE. The reason being that a vanilla VAE seems to be unable to learn from a SMILES dataset otherwise.

In Moses VAE's training scheme it seems to me that the KL penalty has a weighting factor that starts at 0 and grows linearly towards 1 with the number of epochs increasing.

Do you find that this annealing of the KL term allows to get the best of both worlds, i.e not train just a plain AE while still being able to acheive low reconstruction error and high log-likelihood? Any insights on the impact of this scheme and of the KL penalty in general on the training of a VAE on a SMILES dataset?

Thanks for your amazing repo!

Hi,

Have there been any effort in scaling the models to multi-GPU systems for training?. I am not a molecular domain expert. I am curious if there are situations where using multi-GPU would help for the models used in the benchmark suite.

Best, Trinayan

January 15, 2020... Simon, Thanks for the help so far. I have been able to download and install all packages from the new fork: https://github.com/SeemonJ/moses which you shared on Dec 21, 2019. Thanks for that. I am now facing errors listed in attached 3 txt files. Request your help to fix these errors. Errro Running Moses_15Jan2020.txt Errro Running sh scripts_run_all_models.txt List of installed packages.txt

any modification from last update? I got an error message from: ########### python scripts/eval.py --n_jobs 4 --device cuda:0 --test_path data/test.csv --ptest_path data/test_stats.npz --test_scaffolds_path data/test_scaffolds.csv --ptest_scaffolds_path data/test_scaffolds_stats.npz --gen_path checkpoints/organ_generated.csv Traceback (most recent call last): File "scripts/eval.py", line 87, in main(config) File "scripts/eval.py", line 42, in main train=train) TypeError: get_all_metrics() got an unexpected keyword argument 'train' ############ It was OK before. Oh Happy 1024!

In accordance with mail conversation, here is the part of the LatentGAN that has so far been made public. The heteroencoder will be published soon, after which the code refactoring will be performed.

pip install molsets on Ubuntu 20.04.4 LTS with Python 3.8.13 and GCC 7.5.0 fails due to an error in installing a dependency.

Building wheel for pomegranate (setup.py) resulted in the following error:

building 'pomegranate.distributions.NeuralNetworkWrapper' extension

  gcc -pthread -B /home/dhanajayb/anaconda3/envs/DeepChem/compiler_compat -Wl,--sysroot=/ -Wsign-compare 
  -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC 
  -I/home/dhanajayb/anaconda3/envs/DeepChem/include/python3.8 
  -I/home/dhanajayb/anaconda3/envs/DeepChem/lib/python3.8/site-packages/numpy/core/include 
  -c pomegranate/distributions/NeuralNetworkWrapper.c 
  -o build/temp.linux-x86_643.8/pomegranate/distributions/NeuralNetworkWrapper.o
  
  gcc: error: pomegranate/distributions/NeuralNetworkWrapper.c: No such file or directory
  error: command '/usr/bin/gcc' failed with exit code 1
  ----------------------------------------
  ERROR: Failed building wheel for pomegranate

pip install pomegranate successfully installs pomegranate-0.14.8 on this machine. Has anyone else experienced this issue?

Hi! Thanks a lot for your efforts in integrating these models. In the README, you show the result of JTN-VAE. However, it seems that there has no real implementation for JTN-VAE. Could you show me a way to use JTN-VAE? Thanks!

Hi,

Thanks for your wonderful software. I was just wondering since you have cited ChemVAE in your supported models whether you have implemented it. In VAE class I didn't see any GRU layer followed by conv1d layers as described in ChemVAE paper.

Thanks, Mohsen

Firstly, whether the variable 'vocab' was generated by CharVocab.from_data function as mentioned in README.md. Additionally, if so, I couldn't find a property called 'vectors'. What the exactly type of the variable 'vocab' there? Whether I missed other key points on 'vocab'?

I have this function to check the validity of smiles that is based on RDKit :

from rdkit import Chem
def checksmi(smi):
    m = Chem.MolFromSmiles(smi,sanitize=False)
    if m is None:
        #print('invalid SMILES')
        v = 0
    else:
        #print("valid smiles.")
        v = 1
        try:
            Chem.SanitizeMol(m)
        except:
            #print('invalid chemistry')
            v = 0
     return v

is this the same as the validity given by moses when we type the below code, does moses validity checks for valid smiles (grammar) or also valid chemistry:

import moses
metrics = moses.get_all_metrics(list_to_evaluate)
print(metrics)
{'valid': 0.8571428571428572,
 'unique@1000': 1.0,
 'unique@10000': 1.0,
 'FCD/Test': 52.710485508527654,
 'SNN/Test': 0.2737954681118329,
 'Frag/Test': 0.26661035762724716,
 'Scaf/Test': nan,
 'FCD/TestSF': 54.30007202575979,
 'SNN/TestSF': 0.23380156854788461,
 'Frag/TestSF': 0.3777161249748274,
 'Scaf/TestSF': nan,
 'IntDiv': 0.7468284898334079,
 'IntDiv2': 0.578859979666881,
 'Filters': 0.6666666666666666,
 'logP': 3.048971913797608,
 'SA': 0.8860967344024802,
 'QED': 0.3780715536953819,
 'weight': 184.10358258270205,
 'Novelty': 1.0}