A neural network designed from scratch, using only numpy and matplotlib library.
The code is availble at GitHub
Our code demos are Dense_sample.py and Conv_sample.py.
In Dense_sample.py, you can choose different optimizers by one line among the code as below:
model.config(optimizer=optimizers.SGD, loss=losses.Crossentropy, lr=0.01, metric=metrics.categorical_accuracy)
model.config(optimizer=optimizers.BCD, loss=losses.Crossentropy, lr=0.1, metric=metrics.categorical_accuracy)
model.config(optimizer=optimizers.BCD_V2, loss=losses.Crossentropy, lr=0.1, metric=metrics.categorical_accuracy)In Conv_sample.py, only optimizer.SGD is supported, because the other BCD optimizers don't support layers.maxpool2d.
Our coding structure design is like this:
+------------------------------------------------------------------+
| models |
+----------------+-----------------------+---------+---------+-+---+
^ ^ ^ ^ |
| | | | v
+----------------+-----------------+ +---+---+ +---+--+ +----+-+---+
| layers | |metrics| |losses| |optimizers|
+--------------+--------+----------+ +-------+ +------+ +----------+
|linear(Dense) | Conv2d | MaxPool2D|
+--------+-----+--------++---------+
^ ^
| |
+-----+-----+ +------+-----+
|activations| |initializers|
+-----------+ +------------+