vanessaguarino/mnist-confidence-thresholded-evaluation

MNIST Diffusion Assessment Results

This dataset contains a FiftyOne export of an evaluation performed on MNIST-style images generated by a diffusion model.

• Content: Generated MNIST digit images
• Predictions: Includes IDK (uncertain) predictions, which are used to help filter out low-quality or ambiguous generated images
• Purpose: Quality assessment and filtering of generated samples
• Split: Test / Evaluation
• Format: FiftyOneDataset

This dataset contains 500 MNIST-style images generated unconditionally using a UNet/DDPM diffusion model. Each sample is enriched with embeddings, pseudo-labels, and quality metrics to facilitate evaluation and downstream analysis.

Below is a screenshot of the FiftyOne App used for interactive inspection of generated images, embeddings, and brain metrics:

Dataset Contents

• Images: Generated MNIST digits
• UNet Embeddings: Extracted from the bottleneck during generation
• Pseudo-labels: Predictions from a pretrained 11-class classifier
• Thresholded Predictions: Applying the optimized IDK cascade for uncertainty-aware labeling
• FiftyOne Metrics:
  • Uniqueness: Measures distinctiveness of each generated sample
  • Representativeness: Assesses coverage in the UNet embedding feature space
• UMAP Embeddings: Dimensionality-reduced representation for visual inspection

Purpose

• Evaluate the quality and diversity of generated MNIST digits
• Identify low-confidence or ambiguous samples using the IDK cascade
• Explore clustering patterns in embedding space

Key Observations

• Unconditionally generated images are visually consistent but can exhibit high uncertainty (e.g., ambiguous or exaggerated features)
• The IDK class helps filter out low-quality or ambiguous digits
• Embedding-based metrics (uniqueness and representativeness) complement but do not fully capture image quality
• UMAP visualization highlights clusters of digits and identifies outliers

Split

• Test / Evaluation (exported from FiftyOne; nominally labeled as train in the dataset export due to API constraints)

---

We can download the dataset snapshot using snapshot_download, then create a new FiftyOne dataset from the local files. Finally, we launch the FiftyOne app to visually confirm that all samples, metadata, and computed metrics have been preserved correctly.

from huggingface_hub import snapshot_download
HF_REPO_ID = "vanessaguarino/mnist-confidence-thresholded-evaluation"

# Download the snapshot from Hugging Face
local_dir = snapshot_download(repo_id=HF_REPO_ID, repo_type="dataset")
fiftyone_dataset_name = "restored_flower_assessment"

if fiftyone_dataset_name in fo.list_datasets():
    fo.delete_dataset(fiftyone_dataset_name)

# Load it back into FiftyOne
restored_dataset = fo.Dataset.from_dir(
    dataset_dir=local_dir, 
    dataset_type=fo.types.FiftyOneDataset,
    name=fiftyone_dataset_name
)

# Verify
print(restored_dataset)
fo.launch_app(restored_dataset)