Presentation is loading. Please wait.

Presentation is loading. Please wait.

Sketch Object Prediction

Similar presentations


Presentation on theme: "Sketch Object Prediction"— Presentation transcript:

1 Sketch Object Prediction
Luciano Ricotta

2 Executive Summary Project Intentions Tasks Performed:
Train a CNN for sketch object identification and optimize its performance Tasks Performed: Data Preparation Generalize Implementation Optimize performance Assessment of Project More computing power is needed to use all 345 categories Successful in generalizing the code and optimizing performance

3 Approaches: Program Keisuke Irie’s implementation of a CNN and ensemble learning had only 4 image classifications Step 1: Debugging the code Step 2: Preparing the rest of the data Step 3: Generalizing the code for N categories Step 4: Optimizing the code by varying: Number of categories Sample size Epochs Convolution layers Batch size Neurons CNN architecture

4 Approaches: Platform Personal computer for debugging the code
Euler for data preparation Used python’s multiprocessing utility to run program in parallel. Allocated 20 CPUs for a significant speed up Euler for training the CNN Data files (not raw) totaled to 382 GB, and each object file was between 0.5 to 1 GB A lot of RAM was required (more than the 12 GB in my laptop)

5 Results: Image Prediction
Number of categories range between 2 and 20 Varying neurons, batch size, and convolution layers does not have much of an effect, although an increase in neurons has a slight accuracy boost. The accuracy generally likes to stay between 80 and 90 for high number of categories and between 90 and 100 for low number of categories. Number of categories 2 4 6 8 12 16 20 Average Accuracy

6 Results: Country Prediction
Number of categories range between 4 and 16 Varying neurons, batch size, and convolution layers does not have much of an effect, although an increase in neurons has a slight performance boost. The accuracy is not significantly affected by any of these factors. Number of categories 4 8 12 16 20 Average Accuracy

7 Discussion: Moving forward
More computing power could let us: Include the other 300+ categories - which could result in poor accuracy, but… This would push us to implement the a common CNN architecture such as the AlexNet architecture to compensate for the increased number of categories.


Download ppt "Sketch Object Prediction"

Similar presentations


Ads by Google