VAE Trainer
This page allows you to train a VAE model using HT-SELEX data.
Accessing the VAE Trainer Page
Navigate to the VAE Trainer by clicking the VAE Trainer
link in the top menu or the navigation bar.
Training Process
Initiating a New Training Job
Click the + Add a New Training Job
button located at the top left of the page.
Setting the parameters
Set the preprocessing parameters for your training job, then click the Next
button.
Preprocessing Parameters
- Model Type: The type of the model. Choose
RaptGen
. - Experiment Name: Assign a name to your experiment.
- Target Length: Specify the sequence length, including adapter length. The
Estimate
button activates after loading SELEX sequences (⑦). - Adapters: Set sequence adapters (constant region at 3' or 5' end). The
Estimate
button becomes available after setting theTarget Length
parameter (③). - Filtering Tolernace: Set the tolerance for sequence filtering, which is the allowable difference between the target length and the length of the sequence.
- Minimum Count: Define the minimum sequence count. Sequences below this threshold will be filtered out.
- SELEX sequences: Upload your training sequences in
.fasta
or.fastq
format.
Next, set the training parameters and click the Train
button.
Training Parameters
- Reiteration of Training: Specify how many times to repeat training with the same dataset. (Not epoch, but the number of training)
- Device: Select
CPU
orcuda:X
for GPU (if available). - Seed Value: Set a seed for reproducibility.
Generate Random Seed
button in the right will generate a random seed. - Maximum Number of Epochs: Set the maximum training epochs.
- Early Stopping Patience: Define how many epochs without improvement of the validation loss before stopping the training.
- Beta Weighting Epochs: Specify epochs for increasing beta value from 0 to 1.
- Force Matching Epochs: Specify epochs for force-matching. During the force-matching phase, the profile HMM model will be forced to have less penalty on match-to-match state transition score.
- Match Cost: Define the intensity of the force-matching. The larger value will force the profile HMM model to have less penalty match-to-match state transition score.
- pHMM model length: Set the profile HMM model length. Default is random region length.
Training the model
After submission, the job appears in the Running Jobs list. You can monitor progress, stop, delete, or rename jobs from this list.
- To stop, delete or rename the job, you can click the
Stop
,Delete
, orRename
button, respectively. - The current status of the experiment section shows the job's progress and total number of jobs.
- The Training Parameters section displays the job's training parameters.
- The Job information section provides details about the training job.
Adding the Trained Model to the Model List
Once training completes, add the model to the model list by clicking Add to Viewer Dataset
button.
Name your model in the pop-up modal and click the Add to Viewer Dataset
button.
The trained model will then be available in the Viewer, GMM Trainer and Bayes Optimization pages.
Next Step
Proceed to the GMM Trainer page to train a GMM model.
If you have already trained a GMM model or have a initializing dataset in BO module, you can proceed to the Bayesian Optimization page.