Zero2Neuro Examples
The following examples explore the essential features of Zero2Neuro, inluding different data formats and network architectures.
- Exclusive-Or: simple 2-input, 1-output logic problem.
- Problem type: logistic regression
- Input data: binary feature vector
- Model type: fully-connected
- Data file type: csv
- Loss function: mean squared error
- Iris: UCI iris flower classification problem
- Problem type: classification
- Input data: feature vector
- Model type: fully-connected
- Data file type: csv
- 4 input features
- 3 classes specified using strings for categorical variable values
- Loss function: sparse categorical cross-entropy
- Metric: sparse categorical accuracy
- Cross-Validation for generating training/validation/testing data sets
- Wisconsin Diagnostic Breast Cancer
- Problem type: classification
- Input data: feature vector
- Model type: fully-connected
- Data file type: csv
- 30 input features
- Estimate the probability of being in class M
- Loss function: binary cross-entropy
- Metric: binary accuracy
- Cross-Validation for generating training/validation/testing data sets
- Daily Rainfall Retrieval: Oklahoma Mesonet rainfall estimation from other weather sensors
- Problem type: both regression and classification
- Input data: feature vector
- Model type: fully-connected
- Data file type: csv
- ?? input features
- Can be configured as either a continuous estimation problem or a rain/no rain classification problem
- Loss functions: mean absolute error (mae) / Sparse categorical cross-entropy
- Metrics: n/a / Sparse categorical accuracy
- Cross-Validation for generating training/validation/testing data sets
- Predicting Peptide Binding from Amino Acid Sequences
- Problem type: both regression and classification
- Input data: strings of amino acid sequences
- Model type: examples with fully-connected, convolutional neural network (1D), and recurrent neural network
- Tokenization by character (one character per amino acid)
- Embedding: translate integer token ID to a vector
- Data file type: csv
- One input feature: variable length string
- One output: either scaled binding affinity (0 … 1) or binary class
- Loss functions: mean absolute error or binary cross-entropy
- Metrics: mse/mae or binary accuracy
- N-fold cross-validation
- TF Datasets
- Example of translating CSV file to TF Dataset folds
- Use of TF Dataset folds instead of CSV file
- Sentiment140 Natural Language Processing
- Problem type: classification
- Input data: strings of natural language
- Model type: recurrent neural network
- Tokenization by whitespace (each token is a unique word)
- Embedding: translate integer token ID to a vector
- Data file type: csv
- One input feature: string containing social media post
- One output: whether the sentiment was positive (1) or negative (0)
- Loss functions: binary cross-entropy
- Metrics: binary accuracy
- N-fold cross validation
- Core 50 Image Recognition: Object classification problem from images
- Problem type: classification
- Input data: images
- Model type: convolutional neural network (2D)
- Data file type: csv that refers to a set of images (tabular-indirect)
- Note: the image data are not included in this repository
- Images are 128x128 in size
- Output: 2 classes (though there are 10 available in the full data set)
- Loss function: sparse categorical cross-entropy
- Metrics: sparse categorical accuracy
- Cross-Validation
- Chesapeake Bay Pixel Classification: Pixel classification from satellite images
- Problem type: classification
- Input data: images
- Model type: U-Net (2D)
- Data file type: TF-Datasets that contain the images and labels for each pixel directly
- Note: The datasets are not included in this repository
- Images are 256x256 with 26 channels
- Output: 7 classes
- Loss function: Sparse Categorical Cross-entropy
- Metrics: Sparse Categorical Accuracy
Test Examples
In addition to the standard examples, we also have a set of tests that exercise different Zero2Neuro features. See Test Example Documentation