COMET(Chemically Omnipotent Molecular Encoder from Transformer)
-
branch_v1_weighted_sampling_dataloader
- Data Loader : data loader sampling masking atom based on their inverse occurence probability
- Logging : record macro f1 score, confusion matrix, weight histogram
-
branch_v2_weighted_sampling_dataset_fixed_size_batch
- Data : dataset consist more balanced molecule sample with more abundant rare symbol.
- Data Set : Normalize molecular property with mean & std value. It is very fast with pandas operation.
- Data Set : It preprocess each molecule and hold their Adjacency Matrix and Feature Matrix. Also each molecule is parsed into fixed size vector.
- Inside Iteration : masking indices were selected based on the symbol distribution and return A, X, masked_A, masked_X, masked_idx, P
- Ground Truth : previous ground truth matrix is indexed inside the training iteration.
- Put Masking Task in Batch_process fn
- Normalize A matrix in dataloader or iteration (Note. masked A should be generate from the Normalized A)
- Crop A matrix with max-atom length in order to increase speed
- Auxilary Regression is done in one time
- Masking Rate is adjusted with given Masking Radius. Avg num_masking get close to Avg connected num_atom
-
branch_v3_radius_masking_dataloader
- Data Loader : Firstly sampling center atom with occurence distribution. Secondly, it find out adjacent atom by multiplying A matrix with r(radius) times. Construct index set and truncate with num_masking
-
branch_v4_adj_masking_dataloader
- Data Loader : Masked A would provide
- Model : A should be calculate from previous A.
Total Number of Molecules in Raw Zinc Dataset : 531,354,040
| Name | Train Size | Train Coverage | Valid Size | Valid Coverage | Sampling Rate |
|---|---|---|---|---|---|
| COMET_L | 19.9M(19,919,005) | 4.9M(4,980,881) | 4.7% | ||
| COMET_M | 5.9M (5,975,109) | 1.4M(1,494,480) | 1.4% | ||
| COMET_S | 1.9M (1,979,256) | 0.5M (495,380) | 0.47% | ||
| COMET_XXS | 197K (197,189) | 49K (49,514) | 0.047% |
Register conda environment to jupyter notebook : https://stackoverflow.com/questions/39604271/conda-environments-not-showing-up-in-jupyter-notebook
Install RDkit : https://anaconda.org/rdkit/rdkit
Handling Large Dataset : https://machinelearningmastery.com/large-data-files-machine-learning/
Neat Tutorial to use HDF5 with python : https://www.pythonforthelab.com/blog/how-to-use-hdf5-files-in-python/
Convert String into HDF5 encoding : https://stackoverflow.com/questions/23220513/storing-a-list-of-strings-to-a-hdf5-dataset-from-python
Loading List of HDF5 files with pytorch Dataset : https://discuss.pytorch.org/t/loading-huge-data-functionality/346/9
Installing TensorboardX : https://github.com/lanpa/tensorboardX
git clone https://github.com/lanpa/tensorboardX && cd tensorboardX && python setup.py install
Compress and Extract datasetfile : https://www.cyberciti.biz/faq/how-do-i-compress-a-whole-linux-or-unix-directory/
compress : tar -zcvf dataset.tar.gz dataset
extrace : tar -zxvf dataset.tar.gz