vexNet

Training the VexIR2Vec Model

To train VexIR2Vec, two files in .h5 format are needed:

1. Keys File
2. Training Data File

These can be generated using the following scripts:

Generating Keys File

gen_keys_h5.py generates the Keys .h5 file.

  python gen_keys_h5.py <space-separated list of project names> ...

Generate Training Data File

gen_training_data_from_csvs.py generates the training data .h5 file used for VexIR2Vec training.

python gen_training_data_from_csvs.py <space-separated list of project names>

Training the VexNet Model

python vexir2vec_training.py -opt <optimizer> --beta <value> --lr <learning-rate> --gamma <value> --batch_size <batch size> -e <no-of-epochs> -temp <value> --inp_dim <input_dim> --out_dim <output-dim> --best_model_path models/vexir2vec_model/

Key Parameters for vexir2vec_training.py

Parameter	Description
-bs	Batch size (e.g., 1024)
-e	Number of epochs (e.g., 150)
-temp	temperature for NTXent loss (e.g., 0.8)
-inpd	Input dimension (e.g., 144)
-outd	Output embedding dimension (e.g., 100)
-opt	Optimizer to use (e.g., adam)
-lr	Learning rate (e.g., 0.005)
-bmp	Base model path (where to save the trained model)
-tune	Usage of Ray (true/false)
-beta	Beta value

The following config defines Additional training configuration parameters used in vexir2vec_training.py:

config = {
            "activation": "<str: activation function>",  # e.g., 'relu', 'leaky_relu', 'tanh', 'gelu'
            "batch_size": args.batch_size,  # batch size, typically power of 2
            "beta": args.beta,  # e.g., beta1/beta2 for Adam optimizer or loss scaling
            "concat_layer": "<int: 0-num_layers-1>",  # index at which to apply layer concatenation
            "drop_units": "<list[float]: each in 0.0-0.5>",  # dropout rate per layer
            "gamma":  "<float: 0.0-1.0>",  # scheduler decay factor or discount rate
            "hidden": "<list[int]: e.g., [64-1024, 64-1024]>",  # hidden units per layer
            "lr": args.lr,  # learning rate
            "temperature": args.temperature,
            "num_O_layers":"<int: 1-3>",  # number of output layers
            "num_layers": "<int: 1-5>",  # number of hidden layers
            "opt": args.optimizer,  # e.g., 'adam', 'sgd', 'adamw'
            "sched": "Linear_lr",  # e.g., 'Linear_lr', 'StepLR', 'CosineAnnealing'
            "thresh_max": "<int: > thresh_min>",  # upper threshold value
            "thresh_min": "<int: < thresh_max>",  # lower threshold value
        }

• Parameters such as batch_size, lr, beta, temperature, and opt—which have a significant impact on training performance—are passed via argparse. The remaining parameters should be manually configured within the vexir2vec_training.py script.