DeepEX

[TOC]

Overview

DeepEX is a universal convenient frame with keras and Tensorflow,

You can get well-known Wide&Deep model such as DeepFM here.

Or, you can define you custom model use this frame.

How to Install

• For Linux:

pip install deepex

• For Windows:

You need active cmd first and,

pip install deepex

Notice: This frame needs keras and tensorflow, maybe it has some problem on windows with python 2.x because of tensorflow.

DeepEX Class API

In the functional API, given some parmeters, you can instantiate a DeepEX object, via:

DeepEX(data = None, feature_dim=None, category_index=None,embedding_dict_size=1000,
      embedding_size=64, depths_size = [1024,256,64],class_num=2,
      aggregate_flag=False, metrics=None, optimizer='Adam', activation='relu',
      embedding_way='dense')

Only data is necessary, other parmeters have default value.

Arguments

• data: np.array

• feature_dim: integer, feature dimension

• category_index:

  1. Can be a 2D list, like [A,B], A and B also a list, if len(A)>1, that means the element of A belong to a field, and will be input embedding layer together.

  2. Can be a integer, it specific feature_dim % integer == 0, that means split feature as equal intervals with category_index

  3. If None, all of feature will be embedding

  4. NEW UPDATE: Can be a mix type, like [A,B,c], A and B is list but c is a integer. It can achive more flexible way of split field, A and B will be split as col.1 and the rest of element will be split as col.2 with parameter c. Notice: c must be the last element, and the count of remaining element can be evenly divisible by c

• embedding_dict_size: embedding dict size of categroy feature

• embedding_size: embedding size, it make output size like (?, len(category_index), embedding_size)

• depths_size: network of deep part parameter, last dimension means fc7 shape

• class_num: multi class or binary class, if class_num < 2, you will get a regression model

• aggregate_flag:

  • if True, first_order and second_order of FM part output as (?,1)
  • if False, output as (?, len(category_index)) and (?, embedding_size)

• metrics: can recive custom metrics, if None, binary class use AUC, multi class use auccary

• optimizer: Network optimizer, default adam, see optimizer.

• activation: Deep part activation, default relu, see activations.

• embedding_way: How network to do embedding, Embedding layer or Dense layer

Methods

get_embedding_layer

embedding input data as class parameter.

return

• inputs: A list, which elements are Input layer, prepare to deep model
• numerics: A list, which elements are numeric feature tensor
• embeddings: A list, which elements are categroy feature embedding tensor
• embedding_layer: A tensor, which is concate numeric feature tensor and categroy feature embedding tensor

---

get_wide_model

get fm part

return

• A tensor, shape depends on class parameter aggregate_flag

---

get_deep_model

get deep part

return

• inputs: A list, which elements are Input layer, prepare to deep model
• model: A tensor, which is also a keras functional layer

---

deepfm

get deepfm model, fc7 which is the last layer before classifier, it will be a class variable

return

• model: A keras functional model

---

auc

auc(y_true, y_pred)

A custom metrics, when class_num=2, use this metrics to evaluate model

---

fit

fit(model, y, save_model_path = None, batch_size=None, epochs=1, verbose=1,
    callbacks=None,validation_split=0.0, validation_data=None,
    shuffle=True, class_weight=None,sample_weight=None, 
    initial_epoch=0, steps_per_epoch=None, validation_steps=None)

fit data to train model

Arguments

• model: a DeepEX model
• save_model_path: A string, where model to save, if None, model will not be saved
• others: see document keras fit

get_fc7_output

get_fc7_output(self, model_path = None, layer_name = 'fc7', data = None)

Get model's fc7 layer output, it can use for other operation, such as model ensemble

Arguments

• model_path: A string, only model_path is necessary, tell function where model is, the model file should be saved use keras.models.Model.save() function.
• layer_name: A string, default 'fc7'
• data: numpy array, NOTICE: if you declare a DeepEX object use same parameter just like model will be load (actually just need [data] and [category_index] are same), this parametre can be None, data will read from self.data. HOWEVER, if you declare other way, you need split data format as model input

return

• intermediate_output: Numpy array(s) of intermediate outputs.

How to Use

This class is very easy to use, three steps to go：

# step 1, declare DeepEX object
deepEX = DeepEX(...)

# step 2, get model you want
model = deepEX.deepfm()

# step 3, train and save
deepEX.fit(...)

example

from deepex import *
import numpy as np

samples = 100000    # set samples num
feat_dim = 10   # set feat_dim 
cate = np.random.randint(1,6,samples)   # set a categroy feat randomly
x = np.random.random((samples,feat_dim))    # generate feat randomly
x[:,3] = cate   # chose a column to be categroy feat
y = np.random.randint(0,2,samples)  # generate label

# declare DeepEX objects
deepEX = DeepEX(data = x, feature_dim=feat_dim, category_index=[[0,1],4], embedding_dict_size=1000, 
embedding_size=64, depths_size = [1024,256,64], class_num=2, 
aggregate_flag=False, metrics=None, optimizer='Adam', activation='relu', embedding_way='emb')

model = deepEX.deepfm()  # get DeepFM
plot_model(model,'deepFM.png',show_shapes=True) # show model graph

# train deepfm
path = None
deepEX.fit(model, y, save_model_path=path, batch_size=None, epochs=1, verbose=1, callbacks=None,validation_split=0.0, validation_data=None, shuffle=True, class_weight=None,sample_weight=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None)

# get fc7 output tensor
fc7 = deepEX.get_fc7_output(model_path=path, layer_name='fc7', data=deepEX.data_split)