Why randomize weights of the last layer in Yolo Step-by-Step.ipynb?

[jorgeMariano]

Hi, First of all, thank you for sharing this amazing work. Could you explain, why you do this operation?

layer   = model.layers[-4] # the last convolutional layer
weights = layer.get_weights()

new_kernel = np.random.normal(size=weights[0].shape)/(GRID_H*GRID_W)
new_bias   = np.random.normal(size=weights[1].shape)/(GRID_H*GRID_W)

layer.set_weights([new_kernel, new_bias])

Thanks in advance.

[ngrayluna]

Recall that we are using a pretrained model which was trained on a large data set. We are using this model via transfer learning.  To do this, we need to customize the pretrained model. Since we have a small data set (e.g. the raccoon data set), we are going with feature-extraction approach.

The fine-extraction approach involves adding a new classifier on top of the pretrained model. And recall that when we train a model from scratch we initialize our weights to random values.

You can learn more about both types of transfer learning on the TensorFlow Tutorial's page: https://www.tensorflow.org/tutorials/images/transfer_learning