blackpython890 / s5 Goto Github PK

This repository is the work link and the continuation of model implementation part of Computer Vision Project. Kindly click on link to know more about the project.

In this module we will train the model with MNIST dataset and to achieve 99.40% accuracy with below mentioned contrained.

• Parameters <=10K
• Epoch <= 15
• Target : 99.40% not once but for consistently 4 epoch
• No Fully connected Layer.

And to achieve this we have a constraint of not using more than 10K parameters.

Our Model is Based on Convolution Neural Network ( CNN ) .

So we had divided the problems into multiple parts and then tried to train the model by correcting features step by step sequentially(mentioned below) till the target accuracy is achieved.

• Building Lighter Model.
• Batch Normalization.
• Use Dropout in proper place.
• Global Average Pooling.
• Image Augmentation techniques.
• Learning Rate(LR) Scheduler.

We had successfully achieve the target accuracy of 99.40% .

1. Parameters : 9736
2. Best Train Accuracy : 98.09%
3. Best Test Accuracy : 99.45%

And we have maintained the model to not "OVERFIT". We have achieved better accuracy by doing slightly image rotation and step size of LR to reduce after 5 step. ( We obtained this by hit and trail and upon rigorous analysing ).

Well "Excess of Anything is not Bad. It is Worst ....!!". Same thing happened,but it is worth for someone, who is in a learning phase.

But tried multiple ways to achieve the target. While in the middle of doing so , we have learnt may things.

• Choose the batch size wisely.
• To achieve high accuracy , first epoch should give good accuracy.
• Never apply Relu , Dropout , Batch Normalization at the last convolution Layer.