isoyangyang / interview Goto Github PK

CS50 AI Project 5 traffic

Author: Pekka Iso-Markku

\

Project involves the creation of a sequential DL model to classify images of road signs.

\

Part 1: load_data() function

\

Part 2: get_model() function

\

A problem I encountered was the number of samples used in the model.fit stage. For example:

\

Epoch 10/10

583/583 [==============================] - 5s 9ms/step - loss: 0.0165

• accuracy: 0.9950

\

Above is terminal output where one can see that the number of samples in the training data is 583 (0.7/0.3 train test split) even though the number of samples in the whole data set is >25k.

\

I've checked both functions whether there is an issue somewhere, for instance using len(labels), but I always see that the amounts are correct. Do not know what is the issue here.

\

I started off using one Conv-Pool layer with number of channels in the conv-layer set to 32. Also, inspired by the lecture notes I right away included a hidden layer and Dropout.

\

After tinkering around with the number of channels I got the best results with Conv-Conv-Pool Conv-Pool with a drop-out of 0.5 to prevent overfitting.[ ]{.Apple-converted-space}

\

Model performed with .9939 accuracy on the test set.