Validating Model

[jonathanloganmoran]

First of all, thank you for providing this model! I am working on a term project for a computational neuroscience course and wanted to study biases in emotion detector nets-- and your model seems like a good starting point for my research. Now onto my issue..

I want to test the model's accuracy against the JEFFE and FER2013 validation image sets. Can I pass in images (.png format) to be manually classified? I was able to successfully train the model using the FER2013 dataset, but I don't understand how to validate the model's accuracy.

I am fairly new at machine learning and just starting out with TF/TFLearn/OpenCV-- so thank you for the help!

[isseu]

You would need to modify poc.py file to be able to input an image. I don't include the trained model, only the architecture so is not going to work without training. This is a old project and I loose the trained model.

[jonathanloganmoran]

Can you guide me in doing so? I was able to successfully train the model, which resulted in the output Gudi_model_100_epochs_20000_faces.meta file. I am unsure of how to pass in the test_set_fer2013.py data to validate the model (produce an accurate confusion matrix using the plot_emotion_matrix.py code).

Again I am using the FER2013 dataset and want to produce a confusion matrix based on the FER2013 test dataset. I am guessing that the manual_poc.py file contains the necessary code, but I am unsure where to start. Thank you!

[jonathanloganmoran]

Hey @isseu,

Moved this into my forked repo to avoid cluttering + stay organized with my commits. Would really appreciate the guidance.. Here's the link: Issue #1: Unable to Load Trained Weights

Thanks!