Temperature Scaling Optimizer

Temperature Scaling Optimizer is to calibrate your neural network and visualize how well-calibrated it in Pytorch.
This library is based on the below papers.

• On Calibration of Modern Neural Networks.
• temperature_scaling

Table of Contents

1. Features
2. Installation
3. Examples

Features

• Optimize for calibrating your neural network compatible with Cross Validation
• Visualize how well-calibrated your neural network is

Installation

pip install -r requirements.txt

And you also install pytorch from here

Examples

To Optimize

import temp_opt as topt

model_dict = {
        model_1: DataLoader_1,
        model_2: DataLoader_2,
        model_3: DataLoader_3
        }
label_store = topt.label_stores.LogitsAndLabelsStore(topt.label_stores.PredictingTable(model_dict))
lbfgs_opt = topt.optimizers.LBFGSOptimizer(label_store, topt.trainers.TemperatureScaleTrainer())
lbfgs_opt.run()

To Predict with Temperature Scaling

import torch
import torchvision.models as models
import temp_opt as topt

model = models.resnet18(pretrained=True)
temperature = 5.32  # set an optimized temperature value 
predictor = topt.predictors.TemperatureScalePredictor(model, temperature)
inputs = torch.Tensor(34, 3, 32, 32)
print(predictor(inputs))

To Visualize

import matplotlib.pyplot as plt
import temp_opt as topt

model_dict = {
        model_1: DataLoader_1,
        model_2: DataLoader_2,
        model_3: DataLoader_3
        }
label_store = topt.label_stores.LogitsAndLabelsStore(topt.label_stores.PredictingTable(model_dict))
plotter = topt.visualizers.CalibationPlotter()
plotter.plot(label_store)
plt.show()

You can visualize your neural network as in the diagram below