eyyub / numpy-convnet

A small and pure Numpy Convolutional Neural Network library.

numpy-convnet

A small and pure Numpy Convolutional Neural Network library I wrote in order to understand backprop through conv layers.

@eyyub_s

Example

Here is an example on how to build a not-so-deep convnet that uses strided convolution:

  layers = [
      Conv((4, 4, 1, 20), strides=2, activation=lkrelu, filter_init=lambda shp: np.random.normal(size=shp)),
      Conv((5, 5, 20, 40), strides=2, activation=lkrelu, filter_init=lambda shp:  np.random.normal(size=shp)),
      Flatten((5, 5, 40)),
      FullyConnected((5*5*40, 100), activation=sigmoid, weight_init=lambda shp: np.random.normal(size=shp)),
      FullyConnected((100, 10), activation=linear, weight_init=lambda shp: np.random.normal(size=shp))
  ]
  net = Network(layers, lr=0.001, loss=cross_entropy)

You can see that it somehow takes inspiration from Keras.

Requirements

• Python 2.7
• Numpy

numpy-convnet's Issues

No module named 'layer'

Add open source license?

Hello there. You didn't include an open source license in your repository, which makes it legally problematic to reuse this code. Just wondering if you'd be willing to add one. You can do this in less than five minutes.

If this code is for tutorial purposes, I would suggest the Unlicense, otherwise maybe BSD or MIT.
https://unlicense.org/

Kind regards,

• Maxime

why this is not abs() ?

https://github.com/Eyyub/numpy-convnet/blob/3db2621f7937862ad2333c16b2b245e0e35e7412/nn/layer/conv.py#L12

Query regarding training

Thank you very much for providing the code. I have a small query with regard to training.
In the network.py file... How ca we loop over the class instance of train_step() continuously to train our input and how are the weights updated and how to store the final weights.
Could you kindly guide me on these queries.
Thank you very much

lz or la?