instructions for generating vocab.pkl?

[manavsingh415]

Hi. Would it be possible for the authors to either upload vocab.pkl (for the pretrained model), or give instructions and code about how to generate the vocab.pkl file from the CHEMBL24 dataset (or any other dataset used)? Thanks

[shionhonda]

Hi. I'm sorry but I lost access to the resources. Does this issue help? https://github.com/DSPsleeporg/smiles-transformer/issues/11

[manavsingh415]

Hi Shion

Thanks for getting back to me! I checked that issue and it points to the ChEMBL24 dataset. I am interested in how to generate vocab.pkl from this dataset.

Actually, I just wish to run the pretrained model on a set of molecules I have, to generate their vector representations. If this is possible without the vocab.pkl file, please let me know!

Thanks

Regards, Manav

---

From: Shion Honda @.> Sent: Monday, June 28, 2021 3:36 PM To: DSPsleeporg/smiles-transformer @.> Cc: manavsingh415 @.>; Author @.> Subject: Re: [DSPsleeporg/smiles-transformer] instructions for generating vocab.pkl? (#17)

Hi. I'm sorry but I lost access to the resources. Does this issue help? #11https://github.com/DSPsleeporg/smiles-transformer/issues/11

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHubhttps://github.com/DSPsleeporg/smiles-transformer/issues/17#issuecomment-869740150, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AMRX2BDUJKUPHPRPEPZEY43TVCCIDANCNFSM47N53UPA.

[shionhonda]

Well, I meant to mention this comment. I'm glad if it helps. https://github.com/DSPsleeporg/smiles-transformer/issues/11#issuecomment-775982033

[miquelduranfrigola]

Hi Shion,

I have tried to generate the vocab.pkl from ChEMBL 24. Using default parameters in the build_vocab.py file, I get a vocabulary size of 75.

If I am not mistaken, this is not compatible with the pretrained model provided:

size mismatch for embed.weight: copying a param with shape torch.Size([45, 256]) from checkpoint, the shape in current model is torch.Size([75, 256]). size mismatch for out.weight: copying a param with shape torch.Size([45, 256]) from checkpoint, the shape in current model is torch.Size([75, 256]). size mismatch for out.bias: copying a param with shape torch.Size([45]) from checkpoint, the shape in current model is torch.Size([75]).

Thanks! M

[shionhonda]

Thanks for reporting. That's strange. Then I might have used different parameters... I'm sorry that it's not set properly.

[shionhonda]

Does it help? https://github.com/DSPsleeporg/smiles-transformer/pull/19

[miquelduranfrigola]

Hi..! Unfortunately the vocab.pkl file from #19 does not help either...

size mismatch for embed.weight: copying a param with shape torch.Size([45, 256]) from checkpoint, the shape in current model is torch.Size([50, 256]). size mismatch for out.weight: copying a param with shape torch.Size([45, 256]) from checkpoint, the shape in current model is torch.Size([50, 256]). size mismatch for out.bias: copying a param with shape torch.Size([45]) from checkpoint, the shape in current model is torch.Size([50]).

[dinabandhu50]

I was able to reproduce the vocab.pkl with following steps

1. Download chemble_24 data form chemble_24_1 with name chembl_24_1_chemreps.txt.gz , this is the same data as mentioned by the author in this issue.

2. Then open the 01_data_prepare.ipynb file and start running from the following cell

3. Run till the following line (to obtain chembl_24.csv)

4. After obtaining the csv file run the build_corpus.py, I have only changed the file reading location and the pandas dataframe column to obtain SMILES. Running this file will take some time.

5. After obtaining the data/chembl24_corpus.txt by running above, run the build_vocab.py file

6. Now this vocab will have len(vocab)==45, I am attaching the obtained result below

PS: don't forget to change the the n_layers from 3 to 4 - trfm = TrfmSeq2seq(len(vocab), 256, len(vocab), 4)

Thanks

Regards, Dinabandhu

[shionhonda]

@dinabandhu50 Thank you so much!!

[sevencheung2021]

copying a param with shape torch.Size([45, 256]) from checkpoint, the shape in current model is torch.Size([75, 256]). size mismatch for out.weight: copying a param with shape torch.Size([45, 256]) from checkpoint, the

have you solved this mismatch problem ?

[VincentBt]

It indeed solves the mismatch problem

[AachenOcean]

where could I fin 01_data_prepare.ipynb ? Thanks.

[AachenOcean]

I find the data_prepare.ipynb, however, I still have a problem in step of runing the build_corpus.py. At the beginning It shows i don't have the utils module, then I install it with pip install utils. However, when I run it again, it shows the error "cannot import name 'split' from 'utils'". I use Python3 to run this command, do you have any suggestion on it? Thanks.

[madiha1ahmed]

Where did you get the "01_data_prepare.ipnyb'?

[GGuuu]

Where did you get the "01_data_prepare.ipnyb'?

I think that is 'prepare_data.ipynb' in 'experiments' folder I could make file from that file