soundbyte 0.1.3

Learn, Train, Deploy: An Academic-friendly DL Toolkit for Accelerated Learning and Prototyping

SoundByte

An Academic-friendly DL Toolkit for Accelerated Learning and Prototyping

Key Features

Feature	Description
Configuration-Driven Design	Define entire experiments via simple JSON config files—no hardcoding needed.
Maximum Modularity	Swap any model, optimizer, or dataloader without touching core source code.
Plugin Architecture	Register and auto-discover custom components with ease.
Custom Minibatch Logic	Implement tailored forward/backward passes for specialized training needs.
Type Safety	Built-in Pydantic validation ensures safe and correct configurations.
Multiple Paradigms	Supports both supervised classification and knowledge distillation out-of-the-box.
Production Ready	Includes logging, error handling, and output serialization for robust use.
Extensible	Clean, documented interfaces make it easy to plug in new ideas.
Console Command	Run full experiments using the simple soundbyte command-line interface.

---

Installation

• Python 3.11 or higher
• conda (for environment management)

   pip install soundbyte

---

📝 Changelog — Version v0.1.3

---

Improved Experiment Execution

Before (v0.1.2)
To run an experiment, you had to specify the paradigm explicitly:

soundbyte supervised_classification --json_config /path/to/json

Now (v0.1.3)
You can run any experiment with a single streamlined command:

🔹 From the Command Line

soundbyte run examples/configs/classification_example.json
soundbyte run examples/configs/distillation_example.json
soundbyte run examples/configs/custom_logic_example.json

🔹 From a Python Script

python examples/run_classification.py
python examples/run_distillation.py
python examples/run_toy_experiment.py

(+) MTSQ Quantization Algorithm Support 🔗Paper

---

🧩 Custom Module Integration Made Easy

Before (v0.1.2)
You had to reference custom modules in your JSON config:

{
  ...
  "model": "path/to/custom_model.py"
  ...
}

Now (v0.1.3)
You can plug in your modules using decorators and registration logic.

✅ Example: Registering a Custom Model

from soundbyte.plugins.registry import register
from soundbyte.core.interfaces import ModelOps
import torch.nn as nn

@register('model_ops', 'my_custom_model')
class MyCustomModel(nn.Module, ModelOps):
    def __init__(self, num_classes=10, hidden_size=128):
        super().__init__()
        self.fc = nn.Linear(784, hidden_size)
        self.classifier = nn.Linear(hidden_size, num_classes)

    def forward(self, x):
        x = x.view(x.size(0), -1)
        x = torch.relu(self.fc(x))
        return self.classifier(x)

    def get_model(self):
        return self

    def get_config(self):
        return {
            "num_classes": self.num_classes,
            "hidden_size": self.hidden_size
        }

---

You can now register other components similarly:

Component Type	Registry Key	Example Use Case
Loss Functions	penalty_ops	Custom loss
Optimizers	control_ops	Custom optimizer
Schedulers	schedule_ops	Custom learning rate scheduler
Training Pipeline	train_ops	Full custom training logic
Evaluation Logic	audit_ops	Custom evaluation/reporting

---

Quick Start

1. Run a Classification Example

# Using the command-line interface
soundbyte run examples/configs/classification_example.json

# Using Python script
python examples/run_classification.py

2. Run a Knowledge Distillation Example

# Using CLI
soundbyte run examples/configs/distillation_example.json

# Using Python script
python examples/run_distillation.py

3. Run Custom Logic Example

# Example with custom minibatch logic
soundbyte run examples/configs/custom_logic_example.json

# Comprehensive toy experiment
python examples/run_toy_experiment.py

Command Line Interface

The toolkit provides a comprehensive CLI through the soundbyte command:

# Run experiments
soundbyte run <config.json>

# Validate configurations
soundbyte validate <config.json>

# List available components
soundbyte list-components
soundbyte list-components --type model_ops

# Show experiment information
soundbyte info <config.json>

# Initialize example configurations
soundbyte init --template classification
soundbyte init --template distillation

# Override parameters
soundbyte run config.json --override "train_ops.params.max_epochs=20"

---

Modular Component System

Plugin Registry

Components are automatically discovered and registered using decorators:

from soundbyte.plugins.registry import register

@register('model_ops')
class MyCustomModel(nn.Module, ModelOps):
    # Implementation

Configuration Management

Experiments are defined through hierarchical JSON configurations with Pydantic validation:

{
  "name": "my_experiment",
  "data_ops": {"name": "cifar10", "params": {"batch_size": 128}},
  "model_ops": {"name": "resnet18", "params": {"num_classes": 10}},
  "control_ops": {"name": "adam", "params": {"lr": 0.001}}
}

Custom Minibatch Logic

SoundByte supports custom forward and backward pass logic through Python files:

Configuration

{
  "data_ops": {
    "name": "cifar10",
    "params": {"batch_size": 64},
    "train_minibatch_logic": "path/to/custom_logic.py",
    "val_minibatch_logic": "path/to/custom_logic.py"
  }
}

Implementation

def custom_minibatch_logic(idx, minibatch, model, loss_fn, optimizer, scheduler, device):
    """
    Custom training logic for specialized requirements.

    Args:
        idx: Batch index
        minibatch: Data batch (data, targets)
        model: Neural network model
        loss_fn: Loss function
        optimizer: Optimizer
        scheduler: Learning rate scheduler
        device: Device for computation

    Returns:
        tuple: (outputs, loss)
    """
    data, targets = minibatch
    data, targets = data.to(device), targets.to(device)

    optimizer.zero_grad()
    outputs = model(data)
    loss = loss_fn.compute_loss(outputs, targets)

    # Custom logic - skip gradients for every second batch
    if idx % 2 == 0:
        loss.backward()
        optimizer.step()

    return outputs, loss

---

Available Components

SoundByte follows a plugin-based architecture. Each operation type (data, model, loss, optimizer, etc.) supports modular, swappable components.

Datasets

Component	Description
cifar10	CIFAR-10 dataset with preprocessing
mnist	MNIST dataset with normalization

Models

Component	Description
simple_convnet	Lightweight CNN for CIFAR-10
mnist_net	Specialized network for MNIST
resnet18	ResNet-18 architecture
vgg16	VGG-16 architecture
densenet121	DenseNet-121 architecture
mobilenet_v2	MobileNet V2 architecture

Loss Functions

Component	Description
cross_entropy	Standard cross-entropy loss
focal_loss	Focal loss for imbalanced datasets
label_smoothing	Cross-entropy with label smoothing
knowledge_distillation	KL divergence-based distillation loss
mse_loss	Mean squared error for regression
huber_loss	Robust Huber loss
bce_loss	Binary cross-entropy
bce_with_logits	BCE with logits

Optimizers

Component	Description
adam	Adam optimizer
adamw	AdamW with weight decay
sgd	Stochastic gradient descent
rmsprop	RMSprop optimizer
adagrad	Adagrad optimizer
adadelta	Adadelta optimizer
nadam	NAdam optimizer
radam	RAdam optimizer

Schedulers

Component	Description
step_lr	Step learning rate decay
multi_step_lr	Multi-step decay
exponential_lr	Exponential decay
cosine_annealing_lr	Cosine annealing
cosine_annealing_warm_restarts	Cosine annealing with warm restarts
reduce_lr_on_plateau	Adaptive reduction on performance plateau
cyclic_lr	Cyclic learning rate
linear_lr	Linear learning rate schedule
polynomial_lr	Polynomial decay

Traininig Paradigms

Component	Description
classification	Standard supervised training
distillation	Knowledge distillation training

Evaluation Paradigms

Component	Description
classification	Accuracy, precision, recall, F1
distillation	Evaluation with agreement and KL-divergence metrics
regression	MSE, MAE, R² for regression tasks

Additional components are currently under development and will be added soon.

---

Supported Training Paradigms

1. Supervised Classification

Standard supervised learning with comprehensive evaluation metrics.

Example Configuration:

{
  "name": "cifar10_classification",
  "data_ops": {"name": "cifar10", "params": {"batch_size": 128}},
  "model_ops": {"name": "simple_convnet", "params": {"num_classes": 10}},
  "penalty_ops": {"name": "cross_entropy", "params": {}},
  "control_ops": {"name": "adam", "params": {"lr": 0.001}},
  "train_ops": {"name": "classification", "params": {"max_epochs": 50}},
  "audit_ops": {"name": "classification", "params": {"metrics": ["accuracy", "f1"]}}
}

2. Knowledge Distillation

Teacher-student training with configurable temperature and loss combination.

Example Configuration:

{
  "name": "knowledge_distillation",
  "data_ops": {"name": "cifar10", "params": {"batch_size": 128}},
  "teacher_model_ops": {"name": "resnet18", "params": {"num_classes": 10}},
  "student_model_ops": {"name": "simple_convnet", "params": {"num_classes": 10}},
  "distillation_penalty_ops": {
    "name": "knowledge_distillation", 
    "params": {"temperature": 4.0, "alpha": 0.5}
  },
  "train_ops": {"name": "distillation", "params": {"max_epochs": 50}},
  "audit_ops": {"name": "distillation", "params": {"compute_agreement": true}}
}

Programmatic Usage

Basic Experiment Execution

from soundbyte import run_experiment, load_config

# Load and run experiment
config = load_config("my_experiment.json")
results = run_experiment("my_experiment.json")

print(f"Test Accuracy: {results['test_metrics']['accuracy']:.2f}%")

Advanced Usage with ExperimentRunner

from soundbyte import ExperimentRunner
from soundbyte.config.experiment import ExperimentConfig

# Create configuration programmatically
config_dict = {
    "name": "programmatic_experiment",
    "data_ops": {"name": "mnist", "params": {"batch_size": 64}},
    "model_ops": {"name": "mnist_net", "params": {"num_classes": 10}},
    # ... other configuration
}

config = ExperimentConfig(**config_dict)
runner = ExperimentRunner(config)
results = runner.run()

Custom Component Registration

from soundbyte.plugins.registry import register
from soundbyte.core.interfaces import ModelOps
import torch.nn as nn

@register('model_ops', 'my_custom_model')
class MyCustomModel(nn.Module, ModelOps):
    def __init__(self, num_classes=10, hidden_size=128):
        super().__init__()
        self.fc = nn.Linear(784, hidden_size)
        self.classifier = nn.Linear(hidden_size, num_classes)

    def forward(self, x):
        x = x.view(x.size(0), -1)
        x = torch.relu(self.fc(x))
        return self.classifier(x)

    def get_model(self):
        return self

    def get_config(self):
        return {"num_classes": self.num_classes, "hidden_size": self.hidden_size}

---

Extending the Toolkit

Adding New Components

1. Create Component Class

   from soundbyte.core.interfaces import ModelOps
   from soundbyte.plugins.registry import register
   
   @register('model_ops', 'my_new_model')
   class MyNewModel(ModelOps):
       # Implement required methods
   

2. Update Configuration

   {
     "model_ops": {"name": "my_new_model", "params": {"custom_param": 42}}
   }
   

Adding New Training Paradigms

1. Implement Trainer Interface

   from soundbyte.core.interfaces import TrainOps
   from soundbyte.plugins.registry import register
   
   @register('train_ops', 'my_new_paradigm')
   class MyNewParadigmTrainer(TrainOps):
       def train(self, model, data_ops, control_ops, penalty_ops, audit_ops, device, **kwargs):
           # Implement training logic
   

2. Create Specialized Components

3. Update Configuration Schema if Needed

Custom Logic Examples

The toolkit includes several custom minibatch logic examples:

• Gradient Skipping: Skip gradients for specific batches
• Gradient Accumulation: Accumulate gradients over multiple batches
• Mixup: Data augmentation with sample mixing
• Cutout: Random masking augmentation
• Label Smoothing: Custom label smoothing implementation
• Focal Loss: Custom focal loss for class imbalance
• Adversarial Training: FGSM-based adversarial examples

---

Experiment Tracking

Update in progress