>GNNv3.ONLINEBOOTING

[ai/ml online workbench]

Configure.Train.Predict.

A Matrix-native interface for FNN regression workflows, parameter search, data validation, and ONNX prediction.

Open ConsoleRun Prediction
01

[ABOUT THE PROJECT]

What this workbench does

This project turns the Python regression workflow into an online AI/ML tool for configuring experiments, running predictions, and explaining the model pipeline.

Project explanation

[gnnv3 online]

GNNv3 Online connects a Next.js interface to a FastAPI backend that serves the saved FNN model pipeline. Prediction always uses the deployed model shown here; changing the training command fields does not change live prediction.

Training is used when you need a new or better model. The website prepares the FNN command and explains the settings, but it does not launch heavy runs from the browser. Long Ax or grid experiments should be run manually on a machine sized for training.

Deployed prediction model

[best so far]

This is the best model currently deployed for prediction. To use a different model, replace the server artifacts and redeploy.

Statusbest reached so far
Run folderTrial_3
ModelFNN
TargetddGexp
Seed15
Input size10 numeric + 1 categorical
Hidden layers5
Nodes32
ActivationLeakyReLU
Dropout0.0
Batch size256
LossSmoothL1Loss
Learning rate0.001
OptimizerAdam
Epochs363
Folds5
Shaperectangle
ScalerPowerTransformer
EncoderOneHotEncoder

Controlled execution

The browser prepares FNN commands and runs prediction. Heavy training is kept outside the website because the VPS is expected to return to 2 vCPUs.

Reusable artifacts

A successful training run produces files that can be reused later: model weights or ONNX export, scaler, encoder, final parameters, and losses.

Search strategy

Grid search tests a fixed list of combinations. Ax search proposes new combinations based on earlier results to find lower validation loss faster.

Next.js UIFastAPI backendONNX RuntimeAx SearchGrid Search
02

[PARAMETER INDEX]

Every control and why it exists

Use this index when preparing a run command. These controls do not alter the deployed prediction model unless a new trained model is exported and deployed.

FNN Run Setup

[surface]
Model
Fixed to FNN. The website is focused on the neural network regression path only.
Mode
Controls the operation: train, test, or predict.
Ax Search
Switches FNN training from manual grid combinations to Ax-guided hyperparameter search. Large Ax runs should be executed manually, not launched from the website.
Feature file
Path or uploaded file containing the input table.
Output folder
Folder where training artifacts, logs, parameters, and losses are stored.
Target
The supervised label column used for training and evaluation, usually ddGexp.

Data Columns

[schema]
Numeric features
Comma-separated numeric columns used directly after scaling.
Categorical features
Comma-separated categorical columns converted with the selected encoder.
Row names
Optional identifier columns kept for tracking rows through prediction output.
Scaler
Normalizes numeric feature ranges so training is more stable.
Encoding
Converts categorical values into model-readable numeric inputs.

Training Control

[fit]
Ax trials
Maximum number of Ax-suggested training attempts. Keep this small for command preparation and run large experiments manually.
Seed
Random seed for reproducible splits and model initialization.
Folds
Number of k-fold validation splits used to estimate model performance.
Test size
Holdout fraction used when a train/test split is required.
Batch sizes
Candidate mini-batch sizes used during neural network training.
Epochs
Maximum training passes through the dataset for each candidate model.

FNN Hyperparameters

[network]
Learning rates
Candidate optimizer step sizes. Smaller values are steadier, larger values move faster.
Hidden layers
Candidate number of internal network layers.
Nodes
Candidate number of neurons per hidden layer.
Dropout
Regularization rate that randomly disables units during training.
Activations
Nonlinear functions tested inside the network, such as Tanh or LeakyReLU.
Loss
Training objective used to score model error.
Optimizer
Algorithm that updates model weights, such as Adam.
Shape
Controls whether layer widths stay rectangular, grow, or shrink.
03

[CONFIGURE TRAINING]

Manual training command builder

These values only build a terminal command for manual train or test runs. They do not change the deployed prediction model shown in section 01.

FNN run surface

[AX]
command.preview[live]
python gml.py \
  --method FNN \
  --mode train \
  --use_Ax \
  --ax_trials 300 \
  --feature_file data/features.tsv \
  --scaler PowerTransformer \
  --encoding onehot \
  --reproduce 5 \
  --foldername runs/Trial_1_AX_300 \
  --test_size 0.2 \
  --n_folds 5 \
  --batch_sizes 16,64,256 \
  --learning_rates 0.0001,0.001,0.01 \
  --n_epochs 500 \
  --hidden_layers 1,2,3,5 \
  --num_nodes 16,24,32 \
  --dropout_rates 0.0,0.1 \
  --activation_functions Sigmoid,Tanh,LeakyReLU \
  --loss_function SmoothL1Loss \
  --optimizers Adam \
  --net_shape rectangle

FNN hyperparameters

[matrix]
04

[PROCESS FLOW]

How the model workflow is used

Use the online workbench in two paths: prepare train or test commands for manual execution, and run prediction online when a saved model is ready.

01.input

Prepare the table

Start with a TSV or CSV file that contains the selected numeric and categorical feature columns. Training also needs the target column.

02.configure

Configure the run

Choose train, test, or predict, then set the FNN feature names, scaler, encoding, seed, and output folder.

03.train

Train when you need artifacts

Training creates artifacts, but the website only prepares the command. Large runs should be launched manually from a terminal on suitable hardware.

04.predict

Predict with saved artifacts

Prediction loads the saved model pipeline, validates the uploaded file, transforms the data, and appends the PredY output column.

Mode meaning

[operator guide]
train

Prepares the command for fitting a new FNN model from labeled data and writing model artifacts.

test

Prepares the command for evaluating a saved run folder against known target values.

predict

Uses an existing saved model to generate predictions for new feature rows.

05

[DATA SCHEMA]

Prepare compatible input

Prediction and training depend on exact column names. Keep the file tab-separated, keep headers stable, and include the selected feature columns.

file.format

Tab-separated table

Use one header row. CSV is accepted for prediction, but TSV remains the workflow default.

train.required

Training columns

Include numeric features, categorical features, and the target column selected in the console.

predict.required

Prediction columns

Include the feature columns only. Target values are optional. The API appends PredY.

names.strict

Exact names

Comma-separated parameter lists cannot contain spaces. Column names must match the file header.

06

[PREDICTION PIPELINE]

Upload data and predict

Send a feature file to the FastAPI backend. Prediction uses the deployed Trial_3 FNN model, not the manual training command values.

Input file

[empty]
Template
runtime.log[live]

> api.loading

> deployed.model=Trial_3

> model.none

> features.pending

> search=ax trials=300

> seed=5 folds=5

> training.manual=copy-command

> prediction.empty

07

[RUN STATUS]

Results and status

Prediction results render as a compact table. Training status reflects command preparation only, not a hosted training job.

System state

[attention]
Backendloading
Trainingmanual command only
Predictionempty
ComparisonAx/Grid comparison needs manually loaded run metrics

Prediction preview

[empty]
[empty]

No prediction output yet. Upload a compatible feature file in step 06.