Clean UP & Dockerization eos92sw

[HellenNamulinda]

Hello @GemmaTuron, This model couldn't work from the go. I tested it on Colab as seen in this notebook and also from CLI using the cloned repo. All were raising an EmptyOutput error. eos92sw_fetch.log

EmptyOutputError
Detailed error:
Model API eos92sw:predict did not produce an output/root/eos/repository/eos92sw/20230614075837_7D4403/eos92sw/artifacts/framework/fpscores.pkl.gz: /root/eos/repository/eos92sw/20230614075837_7D4403/eos92sw/artifacts/framework/fpscores.pkl.gz: cannot execute binary file

Specifically, cannot execute binary file. The error was related to the line trying to load a pickled file in sascore.py using pickle.load(gzip.open("%s.pkl.gz" % name))

On installing the exact packages in an environment with python=3.7 and running the code in model/framework, I got an import error related to libboost

from .rdBase import rdkitVersion as __version__
ImportError: libboost_python37.so.1.68.0: cannot open shared object file: No such file or directory

After searching for solutions, I changed python=3.7 to python=3.8 and also fixed the incompatibility versions, like scikit-learn==0.23.2 instead of scikit-learn==0.23 which installs scikit-learn==0.23.0, and also numpy, matplotlib and pandas. The new commands are;

conda install python=3.8
conda install -c conda-forge rdkit
pip install scikit-learn==0.23.2
pip install matplotlib==3.7
pip install numpy==1.20.3
pip install pandas==2.0.2
pip install joblib

With the new installs, I was able to run the code using python etoxpred_predict.py --datafile {0} --modelfile {1} --outputfile {2} But fetching on ersilia CLI is still failing. It's what I will fix next now that I know the code works.

[GemmaTuron]

Ok thanks @HellenNamulinda , very helpful explanation Do you have the log file of the CLI failing so I can have a look and help you?

[HellenNamulinda]

Hi @GemmaTuron, Sure I attached the log file before the update, eos92sw_fetch.log The model is still failing with the same EmptyOutputError, resulting from ../fpscores.pkl.gz: cannot execute binary file Since the model env is cloned from the base env, I think there might be some package conflicts. So, I'm fetching the model again, but I added --vv(verbose) on conda commands so I could debug. I will share a new logfile file, or if I figure it out.

[HellenNamulinda]

Hello @GemmaTuron, After tracing the conflict in this log file, eos92sw_updated_fetch.log, I realized it was originating from installing rdkit using conda-forge RUN conda install -c conda-forge rdkit.

After changing it to be installed using pip RUN pip install rdkit-pypi, the model fetched successfully eos92sw_updated1_fetch.log And I was also able to run inferences on the model

{
    "input": {
        "key": "RZVAJINKPMORJF-UHFFFAOYSA-N",
        "input": "CC(=O)NC1=CC=C(O)C=C1",
        "text": "CC(=O)NC1=CC=C(O)C=C1"
    },
    "output": {
        "Tox-score": 0.7498964851421736,
        "SAscore": 0.6654453269695645
    }
}

Plus the file; eml_etoxpred_eos92sw.csv

[HellenNamulinda]

Now that the model works well, I will go ahead and refactor it.

But I'm going to change the model code a bit so that run.sh won't need three parameters as we see in service.py(python etoxpred_predict.py --datafile {0} --modelfile {1} --outputfile {2}) I'm going to make the model file static and remove it from the arguments needed in etoxpred_predict.py,

def myargs():
    parser = argparse.ArgumentParser()
    parser.add_argument("--datafile", required=True, help="training data filename")
    parser.add_argument("--modelfile", required=True, help="path to the model to load")
    parser.add_argument(
        "--outputfile",
        required=False,
        default="./results.csv",
        help="output file to save the result",
    )
    args = parser.parse_args()
    return args

[HellenNamulinda]

Hi @GemmaTuron, I made the necessary changes and I created a pull request here

After the changes, I tested the model locally and it worked well. 👍 Model eos92sw fetched successfully! eos92sw_fetch1.log

The model was able to predict when given string inputs

10:21:19 | DEBUG    | Done with unique posting
{
    "input": {
        "key": "RZVAJINKPMORJF-UHFFFAOYSA-N",
        "input": "CC(=O)NC1=CC=C(O)C=C1",
        "text": "CC(=O)NC1=CC=C(O)C=C1"
    },
    "output": {
        "Tox-score": 0.7498964851421736,
        "SAscore": 0.6654453269695645
    }
}

And also for file inputs, I used the eml dataset.
eml_etoxpred_eos92sw.csv eos92sw_predict_file.log

[miquelduranfrigola]

Thanks @HellenNamulinda - I have merged the PR, let's see if it works.

In any case, since we have made changes to the workflows, I will update them afterwards after the new workflows are validated

[GemmaTuron]

It seems that we cannot open the pickle file. I suggest trying to load it directly (from the CLI, do pickle.load from the same repo)and see if it is a problem with the versions of the packages we have in the conda environment

[GemmaTuron]

@HellenNamulinda

Did you manage to test the pickle load?

[HellenNamulinda]

Hello @GemmaTuron, I tested and was able to load the pickle file, just like before.

I cloned this repo and tested it locally using the below commands(cloned it inside a folder, not the root folder). It fetched successfully, eos92sw_repo_fetch.log 👍 Model eos92sw fetched successfully!

hellenah@hellenah-elitebook:~$ cd Outreachy
hellenah@hellenah-elitebook:~/Outreachy$ conda activate ersilia
(ersilia) hellenah@hellenah-elitebook:~/Outreachy$ git clone https://github.com/ersilia-os/eos92sw.git
Cloning into 'eos92sw'...
remote: Enumerating objects: 98, done.
remote: Counting objects: 100% (98/98), done.
remote: Compressing objects: 100% (75/75), done.
remote: Total 98 (delta 36), reused 65 (delta 19), pack-reused 0
Receiving objects: 100% (98/98), 308.07 KiB | 215.00 KiB/s, done.
Resolving deltas: 100% (36/36), done.
Filtering content: 100% (2/2), 57.35 MiB | 1.26 MiB/s, done.
(ersilia) hellenah@hellenah-elitebook:~/Outreachy$ ersilia -v fetch eos92sw --repo_path eos92sw > eos92sw_repo_fetch.log 2>&1

When I cloned the repo into the root folder, this model failed to fetch giving the error we saw; FileNotFoundError: [Errno 2] No such file or directory: './fpscores.pkl.gz'. eos92sw_repo_fetch_root.log I cloned to the root folder and fetched like below;

hellenah@hellenah-elitebook:~$ conda activate ersilia
(ersilia) hellenah@hellenah-elitebook:~$ git clone https://github.com/ersilia-os/eos92sw.git
Cloning into 'eos92sw'...
remote: Enumerating objects: 98, done.
remote: Counting objects: 100% (98/98), done.
remote: Compressing objects: 100% (75/75), done.
remote: Total 98 (delta 36), reused 65 (delta 19), pack-reused 0
Receiving objects: 100% (98/98), 308.07 KiB | 418.00 KiB/s, done.
Resolving deltas: 100% (36/36), done.
Filtering content: 100% (2/2), 57.35 MiB | 1.28 MiB/s, done.
(ersilia) hellenah@hellenah-elitebook:~$ ersilia -v fetch eos92sw --repo_path eos92sw > eos92sw_repo_fetch_root.log 2>&1

I did test the repo in the root directory by running the run.sh file, which raised the same error.

(eos92sw) hellenah@hellenah-elitebook:~/eos92sw$ bash model/framework/run.sh model/framework ~/input_15.csv output.csv
...loading models
...starts prediction
Traceback (most recent call last):
  File "model/framework/etoxpred_predict.py", line 86, in <module>
    outputs = predict(smiles_list)
  File "model/framework/etoxpred_predict.py", line 68, in predict
    sa_score = reg(smiles_list[i]) 
  File "/home/hellenah/eos92sw/model/framework/sascore.py", line 29, in __call__
    self.readFragmentScores()
  File "/home/hellenah/eos92sw/model/framework/sascore.py", line 112, in readFragmentScores
    _fscores = pickle.load(gzip.open("%s.pkl.gz" % name))
  File "/home/hellenah/anaconda3/envs/eos92sw/lib/python3.8/gzip.py", line 58, in open
    binary_file = GzipFile(filename, gz_mode, compresslevel)
  File "/home/hellenah/anaconda3/envs/eos92sw/lib/python3.8/gzip.py", line 173, in __init__
    fileobj = self.myfileobj = builtins.open(filename, mode or 'rb')
FileNotFoundError: [Errno 2] No such file or directory: './fpscores.pkl.gz'
(eos92sw) hellenah@hellenah-elitebook:~/eos92sw$ 

To me, I knew the issue was because of the relative path instead of an absolute path. So, this code in sascore.py file under frameworks subfolder

        global _fscores
        # generate the full path filename:
        name = os.path.join("./", name)
        _fscores = pickle.load(gzip.open("%s.pkl.gz" % name))

I changed it to

        global _fscores
        # generate the full path filename:
        file_path = os.path.abspath(os.path.join(__file__,".."))  
        name = os.path.join(file_path, name)
        _fscores = pickle.load(gzip.open("%s.pkl.gz" % name))

And run.sh worked well, output.csv

(eos92sw) hellenah@hellenah-elitebook:~/eos92sw$ bash model/framework/run.sh model/framework ~/input_15.csv output.csv
...loading models
...starts prediction
...prediction done!
returning output
saving output
saving output done
(eos92sw) hellenah@hellenah-elitebook:~/eos92sw$ 

Likewise, fetching this model from its repo in the root directory also worked successfully. eos92sw_repo_fetch_root_updated.log

01:13:58 | DEBUG    | Repo path specified: eos92sw
01:13:58 | DEBUG    | Absolute path: /home/hellenah/eos92sw
⬇️  Fetching model eos92sw: etoxpred
.
.
.
01:18:32 | INFO     | Fetching eos92sw done successfully: 0:04:24.263940
Checking setup: 1.290s
Preparing model: 18.917778253555298s
Getting model: 5.072348117828369s
Packing model: 205.6379256248474s
Checking if model needs to be integrated to a tool: 0.0014548301696777344s
Getting model card: 1.755021572113037s
[]
Checking that autoservice works: 6.63561749458313s
Sniffing model: 20.31470251083374s
👍 Model eos92sw fetched successfully!

The model served model runs inferences successfully

01:23:29 | DEBUG    | Done with unique posting
{
    "input": {
        "key": "QTBSBXVTEAMEQO-UHFFFAOYSA-N",
        "input": "CC(=O)O",
        "text": "CC(=O)O"
    },
    "output": {
        "Tox-score": 0.8160144342015792,
        "SAscore": 0.597541979674131
    }
}

In conclusion, paths need to be provided as absolute and not relative. This ensures that the file path is consistent regardless of the current working directory. I made a pull request for these changes here

[HellenNamulinda]

@GemmaTuron, The Model Test on PR failed when installing ersilia, an error related to metadata-generation and Bentmol

AttributeError: 'BentoMLRequirement' object has no attribute '_is_installed'. Did you mean: 'is_installed'?
      [end of output]

  note: This error originates from a subprocess, and is likely not a problem with pip.
error: metadata-generation-failed

Although metadata-generation failed, no change was made to the metadata file.

This is an issue with ersilia and not this model repository code.

[GemmaTuron]

@HellenNamulinda

You are right this was a temporary bug int he code which should now be fixed. I am rerunning the Model Test action

[HellenNamulinda]

Hi @GemmaTuron, the build succeeded for both amd64, at 20306

#9 [linux/amd64 2/2] RUN ersilia -v fetch eos92sw --from_github
...
#9 192.1 👍 Model eos92sw fetched successfully!
#9 DONE 193.6s

and arm64 at 22993

#8 [linux/arm64 2/2] RUN ersilia -v fetch eos92sw --from_github
...
#8 8206.1 👍 Model eos92sw fetched successfully!
#8 DONE 8207.6s

I've tested the model on Colab and it works well