Implemention of a full backprop algorithm using only numpy. I assume sigmoid activation across all layers and a single value in the output layer.
I trained it on the Bike Sharing dataset.
Use the following features from the data:
temp atemp hum windspeed weekday
The response variable is, raw["success"] = raw["cnt"] > (raw["cnt"].describe()["mean"]).
The architecture of the network : [5, 40, 30, 10, 7, 5, 3, 1].
Use batch_size=8, and train it for 100 epochs on the train set.
Then, ploted loss per epoch.