-
Notifications
You must be signed in to change notification settings - Fork 96
/
seq2seq.py
190 lines (157 loc) · 8.72 KB
/
seq2seq.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
import os
import numpy as np
import pandas as pd
from datetime import date, timedelta
from sklearn import metrics
# from sklearn.preprocessing import MinMaxScaler, LabelEncoder
import tensorflow as tf
import keras
from keras.models import Sequential, Model
from keras.layers import *
from keras import optimizers
import gc
from Utils import *
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' # suppress tf warnings
timesteps = 365
df, promo_df, items, stores = load_unstack('all')
# data after 2015
df = df[pd.date_range(date(2014,6,1), date(2017,8,15))]
promo_df = promo_df[pd.date_range(date(2014,6,1), date(2017,8,31))]
promo_df = promo_df[df[pd.date_range(date(2017,1,1), date(2017,8,15))].max(axis=1)>0]
df = df[df[pd.date_range(date(2017,1,1), date(2017,8,15))].max(axis=1)>0]
promo_df = promo_df.astype('int')
df_test = pd.read_csv("test.csv", usecols=[0, 1, 2, 3, 4], dtype={'onpromotion': bool},
parse_dates=["date"]).set_index(['store_nbr', 'item_nbr', 'date'])
item_nbr_test = df_test.index.get_level_values(1)
item_nbr_train = df.index.get_level_values(1)
item_inter = list(set(item_nbr_train).intersection(set(item_nbr_test)))
df = df.loc[df.index.get_level_values(1).isin(item_inter)]
promo_df = promo_df.loc[promo_df.index.get_level_values(1).isin(item_inter)]
df_index = df.index
del item_nbr_test, item_nbr_train, item_inter, df_test; gc.collect()
train_data = train_generator(df, promo_df, items, stores, timesteps, date(2017, 7, 9),
n_range=380, day_skip=1, batch_size=1000, aux_as_tensor=True, reshape_output=2)
Xval, Yval = create_dataset(df, promo_df, items, stores, timesteps, date(2017, 7, 26),
aux_as_tensor=True, reshape_output=2)
Xtest, _ = create_dataset(df, promo_df, items, stores, timesteps, date(2017, 8, 16),
aux_as_tensor=True, is_train=False, reshape_output=2)
w = (Xval[7][:, 0, 2] * 0.25 + 1) / (Xval[7][:, 0, 2] * 0.25 + 1).mean()
del df, promo_df; gc.collect()
# Note
# current best: add item_mean, dim: 50, all as tensor ~ 3500 (~3630 in new cv)
print('1*100, train on private 7, nrange 380 timestep 200, data 1000*1500 \n')
latent_dim = 100
# seq input
seq_in = Input(shape=(timesteps, 1))
is0_in = Input(shape=(timesteps, 1))
promo_in = Input(shape=(timesteps+16, 1))
yearAgo_in = Input(shape=(timesteps+16, 1))
quarterAgo_in = Input(shape=(timesteps+16, 1))
item_mean_in = Input(shape=(timesteps, 1))
store_mean_in = Input(shape=(timesteps, 1))
# store_family_mean_in = Input(shape=(timesteps, 1))
weekday_in = Input(shape=(timesteps+16,), dtype='uint8')
weekday_embed_encode = Embedding(7, 4, input_length=timesteps+16)(weekday_in)
# weekday_embed_decode = Embedding(7, 4, input_length=timesteps+16)(weekday_in)
dom_in = Input(shape=(timesteps+16,), dtype='uint8')
dom_embed_encode = Embedding(31, 4, input_length=timesteps+16)(dom_in)
# dom_embed_decode = Embedding(31, 4, input_length=timesteps+16)(dom_in)
# weekday_onehot = Lambda(K.one_hot, arguments={'num_classes': 7}, output_shape=(timesteps+16, 7))(weekday_in)
# aux input
cat_features = Input(shape=(timesteps+16, 6))
item_family = Lambda(lambda x: x[:, :, 0])(cat_features)
item_class = Lambda(lambda x: x[:, :, 1])(cat_features)
item_perish = Lambda(lambda x: x[:, :, 2])(cat_features)
store_nbr = Lambda(lambda x: x[:, :, 3])(cat_features)
store_cluster = Lambda(lambda x: x[:, :, 4])(cat_features)
store_type = Lambda(lambda x: x[:, :, 5])(cat_features)
# store_in = Input(shape=(timesteps+16,), dtype='uint8')
family_embed = Embedding(33, 8, input_length=timesteps+16)(item_family)
class_embed = Embedding(337, 8, input_length=timesteps+16)(item_class)
store_embed = Embedding(54, 8, input_length=timesteps+16)(store_nbr)
cluster_embed = Embedding(17, 3, input_length=timesteps+16)(store_cluster)
type_embed = Embedding(5, 2, input_length=timesteps+16)(store_type)
# Encoder
encode_slice = Lambda(lambda x: x[:, :timesteps, :])
encode_features = concatenate([promo_in, yearAgo_in, quarterAgo_in, weekday_embed_encode,
family_embed, Reshape((timesteps+16,1))(item_perish), store_embed, cluster_embed, type_embed], axis=2)
encode_features = encode_slice(encode_features)
# conv_in = Conv1D(8, 5, padding='same')(concatenate([seq_in, encode_features], axis=2))
# conv_raw = concatenate([seq_in, encode_slice(quarterAgo_in), encode_slice(yearAgo_in), item_mean_in], axis=2)
# conv_in = Conv1D(8, 5, padding='same')(conv_raw)
conv_in = Conv1D(4, 5, padding='same')(seq_in)
# conv_in_deep = Conv1D(2, 2, padding='causal', dilation_rate=1)(seq_in)
# conv_in_deep = Conv1D(2, 2, padding='causal', dilation_rate=2)(conv_in_deep)
# conv_in_deep = Conv1D(2, 2, padding='causal', dilation_rate=4)(conv_in_deep)
# conv_in_deep = Conv1D(2, 2, padding='causal', dilation_rate=8)(conv_in_deep)
# conv_in_quarter = Conv1D(4, 5, padding='same')(encode_slice(quarterAgo_in))
# conv_in_year = Conv1D(4, 5, padding='same')(encode_slice(yearAgo_in))
# conv_in = concatenate([conv_in_seq, conv_in_deep, conv_in_quarter, conv_in_year])
x_encode = concatenate([seq_in, encode_features, conv_in, item_mean_in], axis=2)
# store_mean_in, is0_in, store_family_mean_in], axis=2)
# encoder1 = CuDNNGRU(latent_dim, return_state=True, return_sequences=True)
# encoder2 = CuDNNGRU(latent_dim, return_state=True, return_sequences=False)
# encoder3 = CuDNNGRU(latent_dim, return_state=True, return_sequences=False)
encoder = CuDNNGRU(latent_dim, return_state=True)
print('Input dimension:', x_encode.shape)
_, h= encoder(x_encode)
# s1, h1 = encoder1(x_encode)
# s1 = Dropout(0.25)(s1)
# s2, h2 = encoder2(s1)
# _, h3 = encoder3(s2)
# Connector
h = Dense(latent_dim, activation='tanh')(h)
# h1 = Dense(latent_dim, activation='tanh')(h1)
# h2 = Dense(latent_dim, activation='tanh')(h2)
# Decoder
previous_x = Lambda(lambda x: x[:, -1, :])(seq_in)
decode_slice = Lambda(lambda x: x[:, timesteps:, :])
decode_features = concatenate([promo_in, yearAgo_in, quarterAgo_in, weekday_embed_encode,
family_embed, Reshape((timesteps+16,1))(item_perish), store_embed, cluster_embed, type_embed], axis=2)
decode_features = decode_slice(decode_features)
# decode_idx_train = np.tile(np.arange(16), (Xtrain.shape[0], 1))
# decode_idx_val = np.tile(np.arange(16), (Xval.shape[0], 1))
# decode_idx = Input(shape=(16,))
# decode_id_embed = Embedding(16, 4, input_length=16)(decode_idx)
# decode_features = concatenate([decode_features, decode_id_embed])
# aux_features = concatenate([dom_embed_decode, store_embed_decode, family_embed_decode], axis=2)
# aux_features = decode_slice(aux_features)
# decoder1 = CuDNNGRU(latent_dim, return_state=True, return_sequences=True)
# decoder2 = CuDNNGRU(latent_dim, return_state=True, return_sequences=False)
# decoder3 = CuDNNGRU(latent_dim, return_state=True, return_sequences=False)
decoder = CuDNNGRU(latent_dim, return_state=True, return_sequences=False)
# decoder_dense1 = Dense(128, activation='relu')
decoder_dense2 = Dense(1, activation='relu')
# dp = Dropout(0.25)
slice_at_t = Lambda(lambda x: tf.slice(x, [0,i,0], [-1,1,-1]))
for i in range(16):
previous_x = Reshape((1,1))(previous_x)
features_t = slice_at_t(decode_features)
# aux_t = slice_at_t(aux_features)
decode_input = concatenate([previous_x, features_t], axis=2)
# output_x, h1 = decoder1(decode_input, initial_state=h1)
# output_x = dp(output_x)
# output_x, h2 = decoder2(output_x, initial_state=h2)
# output_x, h3 = decoder3(output_x, initial_state=h3)
output_x, h = decoder(decode_input, initial_state=h)
# aux input
# output_x = concatenate([output_x, aux_t], axis=2)
# output_x = Flatten()(output_x)
# decoder_dense1 = Dense(64, activation='relu')
# output_x = decoder_dense1(output_x)
# output_x = dp(output_x)
output_x = decoder_dense2(output_x)
# gather outputs
if i == 0: decoder_outputs = output_x
elif i > 0: decoder_outputs = concatenate([decoder_outputs, output_x])
previous_x = output_x
model = Model([seq_in, is0_in, promo_in, yearAgo_in, quarterAgo_in, weekday_in, dom_in, cat_features, item_mean_in, store_mean_in], decoder_outputs)
# rms = optimizers.RMSprop(lr=0.002)
model.compile(optimizer='adam', loss='mean_squared_error')
history = model.fit_generator(train_data, steps_per_epoch=1500, workers=5, use_multiprocessing=True, epochs=18, verbose=2,
validation_data=(Xval, Yval, w))
# val_pred = model.predict(Xval)
# cal_score(Yval, val_pred)
test_pred = model.predict(Xtest)
make_submission(df_index, test_pred, 'seq-private_only-7.csv')
# model.save('save_models/seq2seq_model-withput-promo-2')