tensorflow2.0

import pandas as pd
import numpy as np
import tensorflow as tf
from sklearn.preprocessing import LabelEncoder
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score
from tensorflow import keras
from tqdm import tqdm

tf.get_logger().setLevel('ERROR')

Loading Data

data = pd.read_csv('../data/data.csv')

data_X = data.iloc[:,2:]
data_y = data.click.values

data_X

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

</style>

	hour	C1	banner_pos	site_id	site_domain	site_category	app_id	app_domain	app_category	device_id	...	device_type	device_conn_type	C14	C15	C16	C17	C18	C19	C20	C21
0	14102100	1005	0	1fbe01fe	f3845767	28905ebd	ecad2386	7801e8d9	07d7df22	a99f214a	...	1	2	15705	320	50	1722	0	35	-1	79
1	14102100	1005	0	4dd0a958	79cf0c8d	f028772b	ecad2386	7801e8d9	07d7df22	a99f214a	...	1	0	20352	320	50	2333	0	39	-1	157
2	14102100	1005	0	543a539e	c7ca3108	3e814130	ecad2386	7801e8d9	07d7df22	a99f214a	...	1	0	20352	320	50	2333	0	39	-1	157
3	14102100	1005	0	8cbacf0b	a434fa42	f028772b	ecad2386	7801e8d9	07d7df22	a99f214a	...	1	0	19772	320	50	2227	0	687	100075	48
4	14102100	1005	0	f282ab5a	61eb5bc4	f028772b	ecad2386	7801e8d9	07d7df22	a99f214a	...	1	0	18993	320	50	2161	0	35	-1	157
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
99995	14102101	1005	0	1fbe01fe	f3845767	28905ebd	ecad2386	7801e8d9	07d7df22	a99f214a	...	1	0	15705	320	50	1722	0	35	100084	79
99996	14102101	1005	1	d9750ee7	98572c79	f028772b	ecad2386	7801e8d9	07d7df22	a99f214a	...	1	0	17614	320	50	1993	2	1063	-1	33
99997	14102101	1005	0	85f751fd	c4e18dd6	50e219e0	febd1138	82e27996	0f2161f8	a99f214a	...	1	0	21611	320	50	2480	3	297	100111	61
99998	14102101	1005	0	1fbe01fe	f3845767	28905ebd	ecad2386	7801e8d9	07d7df22	a99f214a	...	1	0	15699	320	50	1722	0	35	100084	79
99999	14102101	1005	0	1fbe01fe	f3845767	28905ebd	ecad2386	7801e8d9	07d7df22	a99f214a	...	1	0	15706	320	50	1722	0	35	-1	79

100000 rows × 22 columns

可以看到测试的数据全都是类别特征, 其实实际的业务场景中几乎也都是类别型的特征
这里我们给特征进行Label Encode

data_X = data_X.apply(LabelEncoder().fit_transform)

data_X

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

</style>

	hour	C1	banner_pos	site_id	site_domain	site_category	app_id	app_domain	app_category	device_id	...	device_type	device_conn_type	C14	C15	C16	C17	C18	C19	C20	C21
0	0	2	0	110	823	1	712	28	0	5703	...	1	1	128	3	2	36	0	1	0	18
1	0	2	0	303	403	16	712	28	0	5703	...	1	0	303	3	2	103	0	2	0	31
2	0	2	0	334	668	3	712	28	0	5703	...	1	0	303	3	2	103	0	2	0	31
3	0	2	0	543	563	16	712	28	0	5703	...	1	0	234	3	2	76	0	26	53	10
4	0	2	0	924	316	16	712	28	0	5703	...	1	0	210	3	2	71	0	1	0	31
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
99995	1	2	0	110	823	1	712	28	0	5703	...	1	0	128	3	2	36	0	1	59	18
99996	1	2	1	825	510	16	712	28	0	5703	...	1	0	173	3	2	60	2	30	0	6
99997	1	2	0	519	658	5	767	31	2	5703	...	1	0	407	3	2	130	3	13	77	13
99998	1	2	0	110	823	1	712	28	0	5703	...	1	0	122	3	2	36	0	1	59	18
99999	1	2	0	110	823	1	712	28	0	5703	...	1	0	129	3	2	36	0	1	0	18

100000 rows × 22 columns

每一个特征都独立进行了label 编码， 这种好处是可以直接进行embedding
当我们embedding共享权值的时候， 可以给每列特征的label加入之前特征的类别总和，来达到所有特征的label
这也是所有模型代码中 offset 的作用

e.g. field_dims = [2, 4, 2], offsets = [0, 2, 6]

所以，实际look up table中
0 - 1行 对应 特征 X0, 即 field_dims[0]
2 - 5行 对应 特征 X1, 即 field_dims[1]
6 - 7行 对应 特征 X2, 即 field_dims[2]
但实际特征取值 forward(self, x) 的 x大小 只在自身词表内取值
比如：X1取值1，对应Embedding内行数就是 offsets[X1] + X1 = 2 + 1 = 3

fields = data_X.max().values + 1 # 模型输入的feature_fields

fields

array([    2,     6,     6,   987,   872,    18,   769,    62,    19,
        8544, 47309,  2606,     4,     4,   448,     5,     6,   141,
           4,    38,   144,    33], dtype=int64)

tmp_X, test_X, tmp_y, test_y = train_test_split(data_X, data_y, test_size = 0.2, random_state=42, stratify=data_y)
train_X, val_X, train_y, val_y = train_test_split(tmp_X, tmp_y, test_size = 0.25, random_state=42, stratify=tmp_y)

训练过程

数据是avazu数据的随机10万条
优化器统一Adam， lr = 0.001
epoch 为 1, batch_size = 32
主要的目的是跑通所有的模型
epoch多几次, 调调参数对稍微复杂的网络有好处


tips : 类别特征embedding等价于一层没有bias项的全连接，所以模型中几乎都用embedding来模拟LR线性过程

LR

from model import LR

model = LR.LogisticRegression(feature_fields = fields)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 5s 86us/sample - loss: 0.4232 - auc: 0.7029 - val_loss: 0.4135 - val_auc: 0.7317





<tensorflow.python.keras.callbacks.History at 0x243c491bdd8>

FM

from model import FM

model = FM.FactorizationMachine(feature_fields = fields, embed_dim = 8)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 14s 229us/sample - loss: 0.4195 - auc_1: 0.7145 - val_loss: 0.4058 - val_auc_1: 0.7435





<tensorflow.python.keras.callbacks.History at 0x243cb3ecc18>

FFM

from model import FFM

model = FFM.FieldFactorizationMachine(feature_fields = fields, embed_dim = 8)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 216s 4ms/sample - loss: 0.4079 - auc_2: 0.7364 - val_loss: 0.4018 - val_auc_2: 0.7529





<tensorflow.python.keras.callbacks.History at 0x243d1f00908>

AFM

from model import AFM

model = AFM.AttentionalFactorizationMachine(feature_fields = fields, embed_dim = 8, attn_size = 8, dropout = 0.2)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 20s 325us/sample - loss: 0.4267 - auc_3: 0.6965 - val_loss: 0.4119 - val_auc_3: 0.7317





<tensorflow.python.keras.callbacks.History at 0x243ec9d36d8>

DeepFM

from model import DeepFM

model = DeepFM.DeepFM(feature_fields = fields, embed_dim = 8, mlp_dims = [32,16], dropout=0.2)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 17s 288us/sample - loss: 0.4250 - auc_4: 0.7027 - val_loss: 0.4068 - val_auc_4: 0.7410





<tensorflow.python.keras.callbacks.History at 0x243cc90a438>

xDeepFM

from model import xDeepFM

model = xDeepFM.xDeepFM(feature_fields = fields, embed_dim = 8, mlp_dims = (32, 16), 
                        dropout = 0.3, cross_layer_sizes = (16, 16))

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 20s 340us/sample - loss: 0.4291 - auc_5: 0.6972 - val_loss: 0.4099 - val_auc_5: 0.7386





<tensorflow.python.keras.callbacks.History at 0x243f5982128>

PNN

from model import PNN

model = PNN.PNN(feature_fields=fields, embed_dim=8, mlp_dims=[32,16], dropout=0.2, method='inner')

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 15s 246us/sample - loss: 0.4352 - auc_6: 0.6846 - val_loss: 0.4120 - val_auc_6: 0.7339





<tensorflow.python.keras.callbacks.History at 0x243fb894080>

DCN

from model import DCN

model = DCN.DeepCrossNet(feature_fields=fields, embed_dim=8, num_layers=3, mlp_dims=[32, 16], dropout=0.2)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 14s 241us/sample - loss: 0.4152 - auc_7: 0.7203 - val_loss: 0.4049 - val_auc_7: 0.7458





<tensorflow.python.keras.callbacks.History at 0x243fa8b5b00>

AutoInt

from model import AutoInt

model = AutoInt.AutoInt(feature_fields=fields, embed_dim=16, head_num=4, attn_layers=3, mlp_dims=(32,16), dropout=0.2)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 29s 477us/sample - loss: 0.4283 - auc_8: 0.7040 - val_loss: 0.4064 - val_auc_8: 0.7437





<tensorflow.python.keras.callbacks.History at 0x243858b2be0>

FiBiNet

from model import FiBiNET

model = FiBiNET.FiBiNET(feature_fields=fields, embed_dim=8, reduction_ratio=2, pooling='mean')

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 75s 1ms/sample - loss: 0.4149 - auc_9: 0.7225 - val_loss: 0.4094 - val_auc_9: 0.7445





<tensorflow.python.keras.callbacks.History at 0x2439817def0>

DCNv2

from model import DCNv2

model = DCNv2.DeepCrossNetv2(feature_fields = fields, embed_dim = 16, layer_num = 2,
                             mlp_dims = (32, 16), dropout = 0.1, cross_method = 'Matrix')

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 39s 645us/sample - loss: 0.4122 - auc_10: 0.7278 - val_loss: 0.4039 - val_auc_10: 0.7517





<tensorflow.python.keras.callbacks.History at 0x24421bdd198>

model = DCNv2.DeepCrossNetv2(feature_fields = fields, embed_dim = 16, layer_num = 2,
                             mlp_dims = (32, 16), dropout = 0.1, cross_method = 'Mix')

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

Train on 60000 samples, validate on 20000 samples
60000/60000 [==============================] - 30s 507us/sample - loss: 0.4142 - auc_11: 0.7233 - val_loss: 0.4066 - val_auc_11: 0.7464





<tensorflow.python.keras.callbacks.History at 0x24424dac080>

DIFM

from model import DIFM

model = DIFM.DIFM(feature_fields=fields, embed_dim=8, head_num=2, dropout=0.1)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

1875/1875 [==============================] - 11s 6ms/step - loss: 0.4154 - auc: 0.7222 - val_loss: 0.4100 - val_auc: 0.7392





<tensorflow.python.keras.callbacks.History at 0x1d2a7f34e20>

AFN

from model import AFN

model = AFN.AFN(feature_fields=fields, embed_size=8, hidden_size=256, dropout=0.1)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

1875/1875 [==============================] - 14s 8ms/step - loss: 0.4203 - auc: 0.7150 - val_loss: 0.4073 - val_auc: 0.7446





<tensorflow.python.keras.callbacks.History at 0x206acfd5940>

ONN

from model import ONN

model = ONN.ONN(feature_fields=fields, embed_dim=8, mlp_dims=[64, 32], dropout=0.1)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=1)

1875/1875 [==============================] - 137s 73ms/step - loss: 0.4377 - auc: 0.7004 - val_loss: 0.4053 - val_auc: 0.7478





<tensorflow.python.keras.callbacks.History at 0x24f6a8b6a90>

序列模型

DIN

Deep Interest Net在预测的时候，对用户不同的行为的注意力是不一样的

在生成User embedding的时候，加入了Activation Unit Layer.这一层产生了每个用户行为的权重乘上相应的物品embedding，从而生产了user interest embedding的表示

实际例子： Amazon Book数据 10K

每条数据记录会有用户的行为数据

只保留了商品特征，以及历史上的商品hist的特征.

# 预处理好的数据
# 处理的函数在AmazonDataPreprocress.py中
# 原始数据为.txt文件

data = pd.read_csv('../data/amazon-books-100k-preprocessed.csv', index_col = 0)

data

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

</style>

	hist_cate_0	hist_cate_1	hist_cate_2	hist_cate_3	hist_cate_4	hist_cate_5	hist_cate_6	hist_cate_7	hist_cate_8	hist_cate_9	...	hist_cate_32	hist_cate_33	hist_cate_34	hist_cate_35	hist_cate_36	hist_cate_37	hist_cate_38	hist_cate_39	cateID	label
0	142	142	142	142	142	0	0	0	0	0	...	0	0	0	0	0	0	0	0	751	0
1	142	142	142	142	142	0	0	0	0	0	...	0	0	0	0	0	0	0	0	142	1
2	142	142	142	142	97	142	142	0	0	0	...	0	0	0	0	0	0	0	0	1094	0
3	142	142	142	142	97	142	142	0	0	0	...	0	0	0	0	0	0	0	0	142	1
4	142	142	142	142	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	142	0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
99995	142	142	142	142	142	751	0	0	0	0	...	0	0	0	0	0	0	0	0	142	1
99996	142	142	142	142	142	142	142	142	142	142	...	0	0	0	0	0	0	0	0	142	0
99997	142	142	142	142	142	142	142	142	142	142	...	0	0	0	0	0	0	0	0	607	1
99998	142	142	142	142	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	142	0
99999	142	142	142	142	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	142	1

100000 rows × 42 columns

data_X = data.iloc[:,:-1]
data_y = data.label.values

fields = data_X.max().max()

fields

tmp_X, test_X, tmp_y, test_y = train_test_split(data_X, data_y, test_size = 0.2, random_state=42, stratify=data_y)
train_X, val_X, train_y, val_y = train_test_split(tmp_X, tmp_y, test_size = 0.25, random_state=42, stratify=tmp_y)

from model import DIN

model = DIN.DeepInterestNet(feature_dim=fields, embed_dim=8, mlp_dims=[64,32], dropout=0.2)

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), 
              loss = 'binary_crossentropy', metrics=[keras.metrics.AUC()])

model.fit(train_X.values, train_y, batch_size=32, validation_data=(val_X.values, val_y), epochs=2)

Train on 60000 samples, validate on 20000 samples
Epoch 1/2
60000/60000 [==============================] - 14s 235us/sample - loss: 0.6788 - auc: 0.5817 - val_loss: 0.6751 - val_auc: 0.5981
Epoch 2/2
60000/60000 [==============================] - 11s 188us/sample - loss: 0.6687 - auc: 0.6080 - val_loss: 0.6744 - val_auc: 0.5921





<tensorflow.python.keras.callbacks.History at 0x1abfc62fb00>

DIEN

相比于DIN， DIEN的改动：

1）关注兴趣的演化过程，提出了兴趣进化网络，用序列模型做的， DIN中用户兴趣之间是相互独立的，但实际上的兴趣是不断进化的

2）设计了一个兴趣抽取层，加入了一个二分类模型来辅助计算兴趣抽取的准确性

3）用序列模型表达用户的兴趣动态变化性

实际的数据用例和DIN一样

data = pd.read_csv('../data/amazon-books-100k-preprocessed.csv', index_col=0)

data

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

</style>

	hist_cate_0	hist_cate_1	hist_cate_2	hist_cate_3	hist_cate_4	hist_cate_5	hist_cate_6	hist_cate_7	hist_cate_8	hist_cate_9	...	hist_cate_32	hist_cate_33	hist_cate_34	hist_cate_35	hist_cate_36	hist_cate_37	hist_cate_38	hist_cate_39	cateID	label
0	142	142	142	142	142	0	0	0	0	0	...	0	0	0	0	0	0	0	0	751	0
1	142	142	142	142	142	0	0	0	0	0	...	0	0	0	0	0	0	0	0	142	1
2	142	142	142	142	97	142	142	0	0	0	...	0	0	0	0	0	0	0	0	1094	0
3	142	142	142	142	97	142	142	0	0	0	...	0	0	0	0	0	0	0	0	142	1
4	142	142	142	142	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	142	0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
99995	142	142	142	142	142	751	0	0	0	0	...	0	0	0	0	0	0	0	0	142	1
99996	142	142	142	142	142	142	142	142	142	142	...	0	0	0	0	0	0	0	0	142	0
99997	142	142	142	142	142	142	142	142	142	142	...	0	0	0	0	0	0	0	0	607	1
99998	142	142	142	142	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	142	0
99999	142	142	142	142	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	142	1

100000 rows × 42 columns

fields = data.max().max()
data_X = data.iloc[:,:-1]
data_y = data.label.values

from model.DIEN import DeepInterestEvolutionNet, auxiliary_sample

tmp_X, test_X, tmp_y, test_y = train_test_split(data_X, data_y, test_size = 0.2, random_state=42, stratify=data_y)
train_X, val_X, train_y, val_y = train_test_split(tmp_X, tmp_y, test_size = 0.25, random_state=42, stratify=tmp_y)

train_X_neg = auxiliary_sample(train_X)

train_X = train_X.values
val_X = val_X.values
test_X = test_X.values

train_loader = tf.data.Dataset.from_tensor_slices((train_X, train_X_neg, train_y)).shuffle(len(train_X)).batch(128)

val_loader  =tf.data.Dataset.from_tensor_slices((val_X, val_y)).batch(128)

model = DeepInterestEvolutionNet(feature_dim=fields, embed_dim=4, mlp_dims=[32,32], dropout=0.2, gru_type = 'GRU')
optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)

epoches = 3
for epoch in range(epoches):
    epoch_train_loss = tf.keras.metrics.Mean()
    for batch, (x, neg_x, y) in tqdm(enumerate(train_loader)):
        with tf.GradientTape() as tape:
            out, aux_loss = model(x, neg_x)
            loss = tf.keras.losses.binary_crossentropy(y, out)
            loss = tf.reduce_mean(loss) + tf.cast(aux_loss, tf.float32)
            loss = tf.reduce_mean(loss)
        grads = tape.gradient(loss, model.trainable_variables)
        optimizer.apply_gradients(grads_and_vars = zip(grads, model.trainable_variables))
        epoch_train_loss(loss)
    epoch_val_loss = tf.keras.metrics.Mean()
    for batch, (x, y) in tqdm(enumerate(val_loader)):
        out,_ = model(x)
        loss = tf.keras.losses.binary_crossentropy(y, out)
        loss = tf.reduce_mean(loss)
        epoch_val_loss(loss)
    print('EPOCH : %s, train loss : %s, val loss: %s' % (epoch,
                                                         epoch_train_loss.result().numpy(),
                                                         epoch_val_loss.result().numpy()))

469it [01:42,  4.58it/s]
157it [00:11, 14.24it/s]
0it [00:00, ?it/s]

EPOCH : 0, train loss : 1.9061264, val loss: 0.69325197


469it [01:43,  4.55it/s]
157it [00:11, 14.19it/s]
0it [00:00, ?it/s]

EPOCH : 1, train loss : 0.80915856, val loss: 0.6931492


469it [01:42,  4.57it/s]
157it [00:11, 14.26it/s]

EPOCH : 2, train loss : 0.7702951, val loss: 0.693148

Name		Name	Last commit message	Last commit date
parent directory ..
model		model
README.md		README.md
example.ipynb		example.ipynb

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

Loading Data

训练过程

LR

FM

FFM

AFM

DeepFM

xDeepFM

PNN

DCN

AutoInt

FiBiNet

DCNv2

DIFM

AFN

ONN

序列模型

DIN

DIEN

FilesExpand file tree

tensorflow2.0

Directory actions

More options

Directory actions

More options

Latest commit

History

tensorflow2.0

Folders and files

parent directory

README.md

Loading Data

训练过程

LR

FM

FFM

AFM

DeepFM

xDeepFM

PNN

DCN

AutoInt

FiBiNet

DCNv2

DIFM

AFN

ONN

序列模型

DIN

DIEN