-
Notifications
You must be signed in to change notification settings - Fork 1
6. Advanced usage
leADS can be trained by users in multiple different ways. The training data made available to users that consists of i)- the Enzyme Commission (EC) number indices with embedding (biocyc21_Xe.pkl) and ii)- the pathway indices (biocyc21_y.pkl). To train leADS as per user specifications, a preprocessing step has to be first executed.
--------------------------------------------------
| Train the model based on features using files |
| other than [DATANAME]_X.pkl and [DATANAME]_y.pkl |
--------------------------------------------------
│
│
[Yes]───────────────────│───────────────────[No]
│ │
│ │
---------- ------------------------------
|Preprocess| | Train using [DATANAME]_X.pkl |
---------- | and [DATANAME]_y.pkl |
| ------------------------------
│
---------------------------------
| Train using data of your choice |
---------------------------------
Note: As before make sure to put the source code leADS (Installing leADS) into the same directory as explained in Download files. Additionally, create log and result folders (if you have not already created one during pathway prediction) in the same leADS_materials/ directory. The final structure should look like this:
leADS_materials/
├── objectset/
│ └── ...
├── model/
│ └── ...
├── dataset/
│ └── ...
├── result/
│ └── ...
├── log/
│ └── ...
└── leADS/
└── ...
For all experiments, using a terminal navigate to the src folder in the leADS directory and then run the commands. To display leADS's running options use: python main.py --help. It should be self-contained.
This step is crucial if users wish to use a different input file (other than just Enzyme Commission (EC) numbers) for training followed by pathway prediction. Preprocessing is only done if the user wants to generate embeddings files such as "[DATANAME]_Xe.pkl", "[DATANAME]_Xa.pkl" etc., from a "[DATANAME]_X.pkl" file in order to use it for training.
The input file used for preprocessing is any matrix file containing Enzyme Commission (EC) numbers (e.g. biocyc21_X.pkl, cami_X.pkl)
- biocyc.pkl
- pathway2ec.pkl
- pathway2ec_idx.pkl
- pathway2vec_embeddings.npz
- hin.pkl
python main.py \
--preprocess-dataset \
--object-name "biocyc.pkl" \
--pathway2ec-name "pathway2ec.pkl" \
--pathway2ec-idx-name "pathway2ec_idx.pkl" \
--hin-name "hin.pkl" \
--features-name "pathway2vec_embeddings.npz" \
--X-name "[DATANAME]_Xe.pkl" \
--file-name "[input (or save) file name]" \
--ospath "[absolute path to the object files directory (e.g. objectset)]" \
--dspath "[absolute path to the dataset directory (e.g. dataset)]" \
--batch 50 \
--num-jobs 2
The table below summarizes all the command-line arguments that are specific to this framework:
| Argument name | Description | Value |
|---|---|---|
| --preprocess-dataset | Preprocess inputs based on Biocyc collection | True |
| --object-name | The preprocessed MetaCyc database file | biocyc.pkl |
| --pathway2ec-name | The matrix file representing Pathway-EC association | pathway2ec.pkl |
| --pathway2ec-idx-name | The pathway2ec association indices file | pathway2ec_idx.pkl |
| --hin-name | The heterogeneous information network file | hin.pkl |
| --features-name | The features corresponding ECs and pathways | pathway2vec_embeddings.npz |
| --X-name | The Input file name to be provided for preprocessing | [input file name]_X.pkl |
| --file-name | The names of input preprocessed files (without extension) | [input (or save) file name] |
| --ospath | The path to the data object that contains extracted information from the MetaCyc database (biocyc.pkl) | Outside source code |
| --dspath | The path to the datasets | Outside source code |
| --batch | Batch size | 50 |
| --num-jobs | The number of parallel workers | 2 |
The output files generated after running the command are:
| File | Description |
|---|---|
| [DATANAME]_Xa.pkl | A matrix file (stored in the "dspath" location) representing organisms ids as rows and the concatenated abundance features in the columns |
| [DATANAME]_Xc.pkl | A matrix file (stored in the "dspath" location) representing organisms ids as rows and concatenated coverage features in the columns |
| [DATANAME]_Xe.pkl | A matrix file (stored in the "dspath" location) representing organisms ids as rows and the concatenated EC features in the columns |
| [DATANAME]_Xea.pkl | A matrix file (stored in the "dspath" location) representing organisms ids as rows and concatenated EC and abundance features in the columns |
| [DATANAME]_Xec.pkl | A matrix file (stored in the "dspath" location) representing organisms ids as rows and concatenated EC and coverage features in the columns |
| [DATANAME]_Xm.pkl | A matrix file (stored in the "dspath" location) representing organisms ids as rows and concatenated EC, abundance, and coverage features in the columns |
| [DATANAME]_Xp.pkl | A matrix file (stored in the "dspath" location) representing organisms ids as rows and the transformed instances to EC in the columns |
Note: Each of these files differ in the total number of columns they contain, which is why the file used for training should also be used during prediction if one decides to train their own model based on certain specifications mentioned above.
Execute the following command to preprocess "cami" data (as an example):
python main.py --preprocess-dataset --object-name "biocyc.pkl" --pathway2ec-name "pathway2ec.pkl" --pathway2ec-idx-name "pathway2ec_idx.pkl" --hin-name "hin.pkl" --features-name "pathway2vec_embeddings.npz" --X-name "cami_X.pkl" --file-name "cami" --batch 50 --num-jobs 2
After running the command, the output will be saved to the dataset/ folder. All the feature files described in the output table above are generated. The tree structure for the folder with the outputs will look like this:
leADS_materials/
├── objectset/
│ └── ...
├── model/
│ └── ...
├── dataset/
│ ├── cami_Xa.pkl
│ ├── cami_Xc.pkl
│ ├── cami_Xe.pkl
│ ├── cami_Xea.pkl
│ ├── cami_Xec.pkl
│ ├── cami_Xm.pkl
│ ├── cami_Xp.pkl
│ └── ...
├── result/
│ └── ...
└── leADS/
└── ...
Training can be done using one of the output files from the preprocessing step and the [DATANAME]_y.pkl file.
That being said, one has to keep in mind to use the same file during pathway predictions to avoid errors since each output file from the preprocessing step contains a different number of columns. Here we show you the recommended command but you can also use other flags as described in the argument descriptions table to suit your requirements.
The input to the command is the output obtained from the preprocessing step above (any one of the [DATANAME]_X*.pkl and the [DATANAME]_y.pkl)
python main.py \
--train \
--train-labels \
--calc-ads \
--ads-percent 0.7 \
--acquisition-type "psp" \
--top-k 50 \
--ssample-input-size 0.7 \
--ssample-label-size 2000 \
--calc-subsample-size 1000 \
--lambdas 0.01 0.01 0.01 0.01 0.01 10 \
--penalty "l21" \
--X-name "[DATANAME]_X*.pkl" \
--y-name "[DATANAME]_y.pkl" \
--model-name "[MODELNAME] (without extension)" \
--ospath "[absolute path to the object files directory (e.g. objectset)]" \
--dspath "[absolute path to the dataset directory (e.g. dataset)]" \
--mdpath "[absolute path to the model directory (e.g. model)]" \
--rspath "[absolute path to the result directory (e.g. result)]" \
--logpath "[absolute path to the log directory (e.g. log)]" \
--batch 50 \
--max-inner-iter 100 \
--num-epochs 10 \
--num-models 10 \
--num-jobs 2 \
The table below summarizes all the command-line arguments that are specific to this framework:
| Argument name | Description | Default Value |
|---|---|---|
| --train | Training the leADS model | True |
| --train-labels | A boolean variable to suggest training leADS using only class-labels data (e.g. "biocyc21_Xe.pkl" and "biocyc21_y.pkl") | False |
| --calc-ads | A boolean variable indicating whether to subsample dataset using active dataset subsampling (ADS) | False |
| --ads-percent | Corresponds the dataset subsampling size (within [0, 1]) | 0.7 |
| --acquisition-type | The acquisition function for estimating the predictive uncertainty (["entropy", "mutual", "variation", "psp"]) | "psp" |
| --top-k | The labels to be considered for variation ratio or psp acquisition functions | 10 |
| --ssample-input-size | Corresponds to the size of random subsampled inputs | 0.7 |
| --ssample-label-size | Corresponds to the size of random subsampled pathway labels | 2000 |
| --calc-subsample-size | The number of samples on which the cost function is computed | 1000 |
| --lambdas | Corresponds to the six hyper-parameters for constraints | 0.01, 0.01, 0.01, 0.01, 0.01, 10 |
| --penalty | The type of regularization term to be applied | l21 |
| --X-name | Input space of multi-label data | biocyc_Xe.pkl |
| --y-name | Pathway space of multi-label data | biocyc_y.pkl |
| --model-name | Corresponds to the name of the model excluding any EXTENSION. The model name will have .pkl extension | leADS |
| --mdpath | Path to store model | Outside source code |
| --rspath | Path to store costs and resulting samples indices | Outside source code |
| --logpath | path to the log directory | Outside source code |
| --ospath | The path to the data object that contains extracted information from the MetaCyc database (biocyc.pkl) | Outside source code |
| --dspath | The path to the datasets | Outside source code |
| --batch | Batch size | 50 |
| --max-inner-iter | Corresponds to the number of inner iteration for logistic regression | 100 |
| --num-epochs | Corresponds to the number of iterations over the training set | 10 |
| --num-models | Corresponds to the number of base learners in an ensemble | 10 |
| --num-jobs | The number of parallel workers | 2 |
| File | Description |
|---|---|
| [MODELNAME].pkl | The trained model |
| [MODELNAME]_cost.txt | This file contains error values between predicted values and expected values |
| [MODELNAME]_samples.pkl | This file contains the sample indices that were produced during training of the model. It is only created if the subsampling flag (--calc-ads) is applied. See Example 1 and 2
|
| log file | This file contains information regarding the run such as time taken to train the model, arguments applied and the files to which the results were stored |
If you wish to train a multi-label dataset (e.g. "biocyc21_Xe.pkl" and "biocyc21_y.pkl") using the subsampling step with variation or psp as an acquisition function, you will need to provide an additional argument --top-k. Run the following command:
python main.py --train --train-labels --calc-ads --ads-percent 0.7 --acquisition-type "psp" --top-k 50 --ssample-input-size 0.7 --ssample-label-size 2000 --calc-subsample-size 1000 --lambdas 0.01 0.01 0.01 0.01 0.01 10 --penalty "l21" --X-name "biocyc21_Xe.pkl" --y-name "biocyc21_y.pkl" --model-name "leADS_retrained_1" --batch 50 --max-inner-iter 5 --num-epochs 10 --num-models 3 --num-jobs 2
After running the command, the output will be saved to the model/, result/ and log/ folders. A short description of the output is given in the table above. The tree structure for the folder with the outputs will look like this:
leADS_materials/
├── objectset/
│ └── ...
├── model/
│ ├── leADS_retrained_1.pkl
│ └── ...
├── dataset/
│ └── ...
├── result/
| ├── leADS_retrained_1_cost.txt
| ├── leADS_retrained_1_samples.pkl
│ └── ...
├── log/
| ├── leADS_events
│ └── ...
└── leADS/
└── ...
To train a multi-label dataset using subsampling and a different acquisition function, execute the following command:
python main.py --train --train-labels --calc-ads --ads-percent 0.7 --acquisition-type "entropy" --ssample-input-size 0.7 --ssample-label-size 2000 --calc-subsample-size 1000 --lambdas 0.01 0.01 0.01 0.01 0.01 10 --penalty "l21" --X-name "biocyc21_Xe.pkl" --y-name "biocyc21_y.pkl" --model-name "leADS_retrained_2" --batch 50 --max-inner-iter 5 --num-epochs 10 --num-models 3 --num-jobs 2
After running the command, the output will be saved to the model/, result/, and log/ folders. A short description of the output is given in the table above. The tree structure for the folder with the outputs will look like this:
leADS_materials/
├── objectset/
│ └── ...
├── model/
│ ├── leADS_retrained_2.pkl
│ └── ...
├── dataset/
│ └── ...
├── result/
| ├── leADS_retrained_2_cost.txt
| ├── leADS_retrained_2_samples.pkl
│ └── ...
├── log/
| ├── leADS_events
│ └── ...
└── leADS/
└── ...
To train a multi-label dataset without the subsampling step, execute the following command:
python main.py --train --train-labels --ssample-input-size 0.7 --ssample-label-size 2000 --calc-subsample-size 1000 --lambdas 0.01 0.01 0.01 0.01 0.01 10 --penalty "l21" --X-name "biocyc21_Xe.pkl" --y-name "biocyc21_y.pkl" --model-name "leADS_retrained_3" --batch 50 --max-inner-iter 5 --num-epochs 10 --num-models 3 --num-jobs 2
After running the command, the output will be saved to the model/, result/, and log/ folders. A short description of the output is given in the table above. The tree structure for the folder with the outputs will look like this:
leADS_materials/
├── objectset/
│ └── ...
├── model/
│ ├── leADS_retrained_3.pkl
│ └── ...
├── dataset/
│ └── ...
├── result/
| ├── leADS_retrained_3_cost.txt
│ └── ...
├── log/
| ├── leADS_events
│ └── ...
└── leADS/
└── ...
To train a multi-label dataset using predefined samples, you need to provide an additional arguments --train-selected-sample and the name of the file in --samples-ids (e.g."leADS_samples.pkl") that is stored in rspath:
python main.py --train --train-labels --train-selected-sample --ssample-input-size 0.7 --ssample-label-size 2000 --calc-subsample-size 1000 --lambdas 0.01 0.01 0.01 0.01 0.01 10 --penalty "l21" --X-name "biocyc21_Xe.pkl" --y-name "biocyc21_y.pkl" --samples-ids "leADS_samples.pkl" --model-name "leADS_retrained_4" --batch 50 --max-inner-iter 5 --num-epochs 10 --num-models 3 --num-jobs 2
After running the command, the output will be saved to the model/, result/, and log/ folders. A short description of the output is given in the table above. The tree structure for the folder with the outputs will look like this:
leADS_materials/
├── objectset/
│ └── ...
├── model/
│ ├── leADS_retrained_4.pkl
│ └── ...
├── dataset/
│ └── ...
├── result/
| ├── leADS_retrained_4_cost.txt
│ └── ...
├── log/
| ├── leADS_events
│ └── ...
└── leADS/
└── ...