[WEEK 6 — Predicting the pieces of furniture in living rooms]
Group Members: Mohammed ALI, Aybüke Yalçıner, HaticeAcar.
Last week, we made some research to get more accuracy. And we found some steps like below:
- Adding more layers
- Some image processing skills
- Changing convolution neural network parameters
- Adding more hidden layers to the multilayer feed-forward neural network that is at the end of the neural network
- Collecting more training data
These are some strategies that we found.
This week, we started to apply the strategies, that we found, one by one. To start with, we use the power of image processing and change the structure of the Convolutional Neural Network. Let us remember the structure of the base convolutional neural network.
Figure 1 is our base convolutional neural network’s structure that we got 56.8 accuracies with it. We use 50 as batch size and 5 epochs. We use ReLU and Softmax functions as activation function.
Figure 2 is our new Convolutional Neural Network’s structure. In this structure, we use some image processing skills and we canceled some layers from the first structure in Figure 1. We use 4 as set per epoch and we have 100 epochs. And we use ReLU and Sigmoid functions as activation function.
The accuracies of models are shown in Figure 3. Although Figure 2 has fewer layers than Figure 1, it gives more accuracy and it takes less time. So from now, we will continue with Figure 2 instead of Figure 1.
Our second trial is adding dropout layers. As we found, adding dropout layers increase the accuracy but accuracy goes down instead of up. So we see that there are no positive effects of adding dropout layers. The accuracy with dropout layers comes 55.19 so we ignore this situation.
And as another option, we collect some more data then apply it with the same structure. Because of more image time is longer than others but accuracy is better so it’s not so important for us.
Our fourth trial is editing the train — validation — test datasets, adding additional layers and playing with parameters.
Figure 5 is the structure of our newest Convolutional Neural Network. We add 1 more convolutional layer, 1 more max pooling layer and 1 more dense layer to the structure. In addition to these, we also add 3 batch normalization layers and 3 dropout layers to the structure. In addition, we use 128 as batch size, mean squared error as loss function. And we use “Adam” as an optimizer.
At the end of this trial, the accuracy comes 69.15 and the loss comes 4.58 for our test dataset.
