nnl_automl_sketch_extrudability_cnn

Sketch Extrudability Classifier

Task

This model performs binary image classification on line sketches. The goal is to predict whether a sketch represents a shape that can be directly extruded into a 3D form.

Dataset:

`sunkaiwen/sketch2stl-sketches-image'

Target variable: extrudable

• 0 = not extrudable
• 1 = extrudable

The model uses only the sketch image as input.

Input and Preprocessing

Images are line sketches from the Sketch2STL dataset.

• Input size: 128 x 128
• Images are converted to RGB
• Images are resized to 128 x 128
• Pixel values are converted to PyTorch tensors

Model Architecture

A compact convolutional neural network (CNN) was implemented in PyTorch. The architecture consists of convolutional blocks with batch normalization, ReLU activation, and max pooling, followed by adaptive average pooling and fully connected classification layers.

A 4 x 4 adaptive pooling output was used to retain more spatial information from the sketches before classification.

AutoML Search

Optuna was used to search over both neural-network architecture choices and training hyperparameters.

The search used a fixed computational budget of:

• 10 Optuna trials
• maximum 20 epochs per trial
• early stopping with patience of 5 epochs

Search Space

Architecture parameters:

• convolutional blocks: 2–3
• base filters: 16, 32, or 64

Training hyperparameters:

• dropout: 0.2–0.5
• learning rate: 1e-4, 3e-4, 1e-3, or 3e-3
• batch size: 16 or 32

Compute Budget

The AutoML search used 10 Optuna trials with a maximum of 20 epochs per trial and early stopping with patience 5. Training was performed using a CPU.

Known Failure Modes

The model may have difficulty distinguishing sketches with similar geometric structure. The validation and test sets are small, with only six images in the final test set, so the reported metrics are sensitive to individual predictions.

Best Configuration

• convolutional blocks: 3
• base filters: 32
• dropout: 0.384
• learning rate: 0.003
• batch size: 16

Performance

• Best validation score: 0.8000
• Test accuracy: 0.5000
• Weighted F1: 0.4857

Limitations

The validation and test sets are small, particularly the test set with only six images. As a result, the reported metrics are sensitive to individual predictions and should be interpreted cautiously.

This model was developed as part of an AutoML experiment comparing compact neural-network architectures and hyperparameter configurations.