Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    9H-Fluorene (Kasitinard M.)
    @ninehfluorene:matrix.org
    [m]
    This is interesting i just found paper that working on QM9 prediction and just find out that order to improve on loss (mae) i should do data clean up (decrease Noise on data) and add more features (for example group up some of molecules that similar etc.)
    https://pubs.acs.org/doi/10.1021/acs.jpca.0c05969
    9H-Fluorene (Kasitinard M.)
    @ninehfluorene:matrix.org
    [m]
    And it bag a question in the future of deepchem molnet is there are gonna be improved version of any database (for example qm9_v2 or something like that) what do you think if this concept ?
    AshW360
    @AshW360
    Hi, how do we separate compounds from .csv file with water solubility?
    Omid Tarkhaneh
    @OmidTarkhaneh
    I am looking for some papers with their code using deepchem for potential energy prediction, if any I wonder someone send me. Thanks in advance.
    Omid Tarkhaneh
    @OmidTarkhaneh
    In the past we used import deepchem.models.tensorgraph.layers as layers for defining our arbitrary architecture in deepchem, now when I am going to use this I can not. Is there any sugesstion, thank you.
    9H-Fluorene (Kasitinard M.)
    @ninehfluorene:matrix.org
    [m]
    @AshW360: if you mean delaney dataset you can use pandas 'drop' , here the example from my code
    Atreya Majumdar
    @atreyamaj
    @OmidTarkhaneh something like this might work:
    layer = dc.models.layers.GraphConv(64)
    Omid Tarkhaneh
    @OmidTarkhaneh
    @atreyamaj Thanks a lot for your help.
    Atreya Majumdar
    @atreyamaj
    :D
    9H-Fluorene (Kasitinard M.)
    @ninehfluorene:matrix.org
    [m]
    after i drop label (properties) from 15 to just 3 label my loss is increse from 0.6XXX to 0.80XX this is really nice
    i reference activation function vs property tabel is really gone up
    https://www.sciencedirect.com/science/article/abs/pii/S0255270120306358
    (LUMO, energy gap (E gap) and dipole moment (u) in QM9)
    Omid Tarkhaneh
    @OmidTarkhaneh
    @ninehfluorene: Can you share the source code of the paper??
    3 replies
    9H-Fluorene (Kasitinard M.)
    @ninehfluorene:matrix.org
    [m]
    by used tensorflow
    Bharath Ramsundar
    @rbharath
    Hey folks, I'm mostly offline this week (my wedding is at the end of the week) so won't be able to answer many questions here. I'll be back online as usual mid next week
    2 replies
    Arthur Funnell
    @elemets
    Best wishes!
    Omid Tarkhaneh
    @OmidTarkhaneh
    hello. Is anyone help how can I work with CML and XML datasets. I have some xml datasets but do not know how to feed network with them
    Vignesh Venkataraman
    @VIGNESHinZONE
    Hi @OmidTarkhaneh
    At present, I don't think deepchem provides any featurization support for XML datasets but I was curious, Are you by any chance working on OPSIN datasets(They also happen to be XML representations of Molecules)?
    Omid Tarkhaneh
    @OmidTarkhaneh
    @VIGNESHinZONE Hi. Thanks for your response. No my datasets are related to DUDE
    Masun Nabhan Homsi
    @MasunNabhanHoms_twitter
    Hello!, I am rerunning in Google-colab a script that I programmed it with Deepchem two months ago, but now when I run it the colab's session is crashing.
    I can run another script that uses a different deep neural network library without any problem. Please, give me some hints to solve this this problem. Thank you.
    James Y
    @yuanjames
    Does anyone have the idea for improving GPU utility?
    ignaczgerg
    @ignaczgerg
    Hi all, is there any news regarding the migration to python=3.8/3.9?
    Arthur Funnell
    @elemets
    Hey does anyone know a way of featurizing molecules quickly? e.g. vectorized featurization, specifically with the WeaveFeaturizer? I have ~500k molecules I'd like to train on...
    Bharath Ramsundar
    @rbharath
    Sorry just starting to come back online
    I'll try to work backwords and answer questions
    @elemets You should definitely be able to featurize 500K compounds on a decent CPU. I've featurized 1M+ datasets within an hour or two. What issues are you seeing?
    @ignaczgerg I'm currently working on this! I was offline most of the last week but just starting to come back online and get to work on the migration. I'll post more information soon
    @MasunNabhanHoms_twitter Can you report more details about the error that you're seeing? I'm not sure what the issue is
    @yuanjames I think GPU utilization should be pretty good for most models, from 50-90%. Which models are you seeing poor utilization with?
    Arthur Funnell
    @elemets
    @rbharath Thanks, I guess I'm using an inefficient method then? It's taken over 8 hours to featurize 600k SMILES. Given they are peptides so are quite long if that makes a difference?
    The code I am using to featurize is:
    def featurizing_smis(smi):
        featurized = featurizer(smi)
        return featurized
    Should I be using a pandas loop instead?
    vectorized_featurizing = np.vectorize(featurizing_smis)
    array_of_feats = vectorized_featurizing(array_of_smis)
    ignaczgerg
    @ignaczgerg

    @ignaczgerg I'm currently working on this! I was offline most of the last week but just starting to come back online and get to work on the migration. I'll post more information soon

    @rbharath That is awesome, thank you! If I could help with anything regarding this, I am more than happy to help.

    Bharath Ramsundar
    @rbharath
    @elemets Ah, being peptides would explain the slowdown! The numbers I mentioned above are for small molecules. Featurization time is either linear or quadratic in molecular length (depending on featurization) so there will definitely be a slowdown. Weave is quadratic, but convmol is linear so perhaps it might be useful to experiment with different featurizations?
    @ignaczgerg Will do! I'll be posting issues and updates at deepchem/deepchem#2512 as I work on the upgrades
    James Y
    @yuanjames
    @rbharath Thanks, from deepchem.models import GCNModel, I am using RTX 3090, it only gives me 2% utility. and deepchem.models.GraphConvModel is around 10%, but MPNN from pytorch could reach 70% all on delaney datasets.
    1 reply
    Akhil Pandey
    @akhilpandey95
    @rbharath what is a good place to start for implementing bayesian GCN using deepchem ?
    Masun Nabhan Homsi
    @MasunNabhanHoms_twitter
    @MasunNabhanHoms_twitter Can you report more details about the error that you're seeing? I'm not sure what the issue is
    @rbharath
    When it crashes, it say "Session crashes with no reason". The log messages are :
    WARNING:root:kernel dab80b04-c3b5-45ae-827c-33975f71d502 restarted
    KernelRestarter: restarting kernel (1/5), keep random ports 2021-06-04 08:15:54.625105: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0
    Akhil Pandey
    @akhilpandey95
    Can anyone help me on this ?
    :point_up:
    su-chao
    @su-chao
    NotImplementedError: Cannot convert a symbolic Tensor (gradient_tape/private__graph_conv_keras_model/graph_gather/sub:0) to a numpy array. This error may indicate that you're trying to pass a Tensor to a NumPy call, which is not supported
    how to deal this Error?
    ignaczgerg
    @ignaczgerg

    NotImplementedError: Cannot convert a symbolic Tensor (gradient_tape/private__graph_conv_keras_model/graph_gather/sub:0) to a numpy array. This error may indicate that you're trying to pass a Tensor to a NumPy call, which is not supported

    Hi, which python version are you using?
    I got this error when I accidentally used python=3.9 with deepchem. Try to downgrade to python=3.7 and you will be fine.

    su-chao
    @su-chao
    My version : Python 3.7.10 ; deepchem 2.5 ;tensorflow-gpu 2.41 @ignaczgerg
    su-chao
    @su-chao
    Where can I see the update logs about deepchem-version? I found some "import * "code can't work when run the examples files
    kingscolour
    @kingscolour
    Anyone know if dc=2.5.0 is incompatible with tf=2.5.0? On linux, dc would not import with tf-gpu=2.5.0: AttributeError: module 'deepchem' has no attribute 'data'
    I had to uninstall tf-gpu=2.5.0 and install tf=2.4.0 to resolve the error.
    Bharath Ramsundar
    @rbharath
    Sorry for the lack of responses folks
    I've been travellign and just starting to catch up