cnnframe

A simple CNN framework built based on numpy.

---

How to use

Check Here for example

Easy to build a CNN or BP model the way like Keras

from cnnframe.cnn_frame import Model
from cnnframe.layers.conv2d import Conv2D,PoolingLayer,Flatten
from cnnframe.layers.dense import Dense

model = Model()

model.add_layer(Conv2D(kernel=[4, 3, 3, 1], stride=[2, 2], padding='same',activation='relu'))
model.add_layer(PoolingLayer(kernel_size=(2,2)))
model.add_layer(Flatten())
model.add_layer(Dense(30,activation='relu'))
model.add_layer(Dense(ydata2.shape[1],activation='sigmoid'))

model.build(xdata2.shape[1:])
model.summary()
model.train(xdata2, ydata2, train_round=100,plot_loss=True)

---

Layers

All layers imlement methods:

• forward_propagation
• backward_propagation

The layers:

• Conv2D: layer for convolution (kernel, stride ...)

• PoolingLayer: layer for pooling

• Flatten: layer for flattening, connecting convolution layer and dense layer(common bp layer)

• Dense: layer for common BP network

---

Activation Functions

• sigmoid: sigmoid function
• relu: sure you know what it is
• softmax: usually used for the last layer

Or add your function in /cnnframe/acfuns.py.

oh! The backward process.

It means for now it is not quite extremely easy to expand more functions.

Maybe update this some other days.

---

Other test

Convolution effect

How would a input Imgae change after the convolution layers, example here (dependency: PIL, numpy)

---

Stephen Lee, 245885195@qq.com, 2019,5,21