Skip to content

Latest commit

 

History

History
135 lines (121 loc) · 6.82 KB

README.md

File metadata and controls

135 lines (121 loc) · 6.82 KB
Raw

Traffic : Observations

I built the model by adding a convolutional 2D layer. For which I pass these arguments :

 tf.keras.layers.Conv2D(64, kernel_size=(3, 3), activation='relu', input_shape=(IMG_WIDTH, IMG_HEIGHT, 3))

I will begin with 64 nodes and a kernel size of 3. In addition to the convolution, I will Max pool the results and then flatten. I will finally add a hidden layer with a 128 units and a dropout of 0.5. Following the results, I will do the necessary adjustments. My model was able to compile with these layers but the accuracy was incredibly low :

Epoch 1/10
500/500 [==============================] - 4s 7ms/step - loss: 5.1708 - accuracy: 0.0524   
Epoch 2/10
500/500 [==============================] - 4s 7ms/step - loss: 3.5904 - accuracy: 0.0561
Epoch 3/10
500/500 [==============================] - 4s 8ms/step - loss: 3.5433 - accuracy: 0.0546
Epoch 4/10
500/500 [==============================] - 4s 8ms/step - loss: 3.5201 - accuracy: 0.0551
Epoch 5/10
500/500 [==============================] - 4s 8ms/step - loss: 3.5098 - accuracy: 0.0552
Epoch 6/10
500/500 [==============================] - 4s 8ms/step - loss: 3.5050 - accuracy: 0.0571
Epoch 7/10
500/500 [==============================] - 4s 8ms/step - loss: 3.5026 - accuracy: 0.0571
Epoch 8/10
500/500 [==============================] - 4s 8ms/step - loss: 3.5015 - accuracy: 0.0561
Epoch 9/10
500/500 [==============================] - 4s 8ms/step - loss: 3.5009 - accuracy: 0.0571
Epoch 10/10
500/500 [==============================] - 4s 8ms/step - loss: 3.5005 - accuracy: 0.0571
333/333 - 1s - loss: 3.4946 - accuracy: 0.0553 - 1s/epoch - 3ms/step

For my second take, I increased the Max pooling pool size to 3 x 3. And decreased the hidden layer's nodes to 64. However the results has not improved the slightest, in the contrary, the accuracy decreased by 0.0051.

For the third take, I decided to add another convolution step. The parameters for the initial one were :

  • 64 filters
  • 3 x 3 kernel size
  • 3 x 3 MaxPooling pool size

The second convolution had :

  • 32 filters
  • 3 x 3 kernel size
  • 3 x 3 Pool size

In addition, I added two extra hidden layers one of them with a dropout of 0.2. And changed the previously added hidden layer's dropout to 0.2 as well.

With these layers this is how our Neural Network performed :

Epoch 1/10
500/500 [==============================] - 4s 8ms/step - loss: 3.4260 - accuracy: 0.1296
Epoch 2/10
500/500 [==============================] - 4s 8ms/step - loss: 2.2432 - accuracy: 0.3452
Epoch 3/10
500/500 [==============================] - 4s 8ms/step - loss: 1.8181 - accuracy: 0.4228
Epoch 4/10
500/500 [==============================] - 4s 8ms/step - loss: 1.5802 - accuracy: 0.4966
Epoch 5/10
500/500 [==============================] - 4s 8ms/step - loss: 1.3426 - accuracy: 0.5674
Epoch 6/10
500/500 [==============================] - 4s 8ms/step - loss: 1.1356 - accuracy: 0.6321
Epoch 7/10
500/500 [==============================] - 4s 8ms/step - loss: 0.9701 - accuracy: 0.6775
Epoch 8/10
500/500 [==============================] - 4s 8ms/step - loss: 0.8603 - accuracy: 0.7050
Epoch 9/10
500/500 [==============================] - 4s 8ms/step - loss: 0.8046 - accuracy: 0.7329
Epoch 10/10
500/500 [==============================] - 4s 8ms/step - loss: 0.7112 - accuracy: 0.7640
333/333 - 1s - loss: 0.5386 - accuracy: 0.8224 - 1s/epoch - 3ms/step

A great improvement from my initial take. However, more is needed to achive the 95% accuracy and higher.

Reading the following article on the importance of the keras.layers.Conv2D parameters. The author recommended to start with a lower number of layers for the first Convolution and then going up if needed. So instead of starting with 64 layers and then 34 I started with 32 and then went for 64. And also I rectified my second max pooling to a (2 x 2) pool size. These were the results after these changes :

Epoch 1/10
500/500 [==============================] - 4s 7ms/step - loss: 3.5033 - accuracy: 0.1351   
Epoch 2/10
500/500 [==============================] - 4s 7ms/step - loss: 2.0187 - accuracy: 0.3967
Epoch 3/10
500/500 [==============================] - 4s 7ms/step - loss: 1.4337 - accuracy: 0.5243
Epoch 4/10
500/500 [==============================] - 4s 7ms/step - loss: 1.1122 - accuracy: 0.6249
Epoch 5/10
500/500 [==============================] - 4s 7ms/step - loss: 0.8552 - accuracy: 0.7082
Epoch 6/10
500/500 [==============================] - 4s 7ms/step - loss: 0.6860 - accuracy: 0.7650
Epoch 7/10
500/500 [==============================] - 4s 7ms/step - loss: 0.5665 - accuracy: 0.8081
Epoch 8/10
500/500 [==============================] - 4s 7ms/step - loss: 0.5192 - accuracy: 0.8323
Epoch 9/10
500/500 [==============================] - 4s 7ms/step - loss: 0.4531 - accuracy: 0.8503
Epoch 10/10
500/500 [==============================] - 4s 7ms/step - loss: 0.3931 - accuracy: 0.8719
333/333 - 1s - loss: 0.2605 - accuracy: 0.9289 - 916ms/epoch - 3ms/step

Closer than ever !

Since the results were promising, I thought the only thing left was to fiddle around with the number of neurons in the Dense layers. Knowing that the number of neurons in the first Dense layer should be the biggest and the closest to the size of the image after the Max Pooling. I went with a size of 256 (128 x 2) and I for the rest I kept dividing them by 2. Stopping at a dense layer with 32 neurons and no Dropout because the number was already low and close to the number of Categories (42). So adding a dropout would increase our losses for nothing.

And here were my final results :

Epoch 1/10
500/500 [==============================] - 4s 7ms/step - loss: 2.7262 - accuracy: 0.3046
Epoch 2/10
500/500 [==============================] - 4s 7ms/step - loss: 1.0958 - accuracy: 0.6609
Epoch 3/10
500/500 [==============================] - 4s 7ms/step - loss: 0.5993 - accuracy: 0.8130
Epoch 4/10
500/500 [==============================] - 4s 7ms/step - loss: 0.3854 - accuracy: 0.8859
Epoch 5/10
500/500 [==============================] - 4s 7ms/step - loss: 0.2905 - accuracy: 0.9174
Epoch 6/10
500/500 [==============================] - 4s 7ms/step - loss: 0.2213 - accuracy: 0.9382
Epoch 7/10
500/500 [==============================] - 4s 7ms/step - loss: 0.2144 - accuracy: 0.9459
Epoch 8/10
500/500 [==============================] - 4s 7ms/step - loss: 0.1783 - accuracy: 0.9545
Epoch 9/10
500/500 [==============================] - 4s 7ms/step - loss: 0.1649 - accuracy: 0.9581
Epoch 10/10
500/500 [==============================] - 4s 7ms/step - loss: 0.1415 - accuracy: 0.9619
333/333 - 1s - loss: 0.1156 - accuracy: 0.9724 - 927ms/epoch - 3ms/step

🎉 The perfomance was even better than the one demonstrated by the course admins ! 🎉

🎥 Demo Recording