forked from chelw97/MPD_challenge
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
299 lines (263 loc) · 15.2 KB
/
model.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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
from lightgbm import LGBMRanker
from sklearn.pipeline import Pipeline,make_pipeline
from sklearn.preprocessing import FunctionTransformer,StandardScaler
from sklearn.model_selection import train_test_split,cross_validate,cross_val_score,GridSearchCV,cross_val_predict
from sklearn.multiclass import OneVsRestClassifier
from sklearn.decomposition import NMF
from sklearn.feature_extraction.text import CountVectorizer
from lightfm import LightFM
from lightfm.evaluation import recall_at_k
import os
import numpy as np
import pickle
import pandas as pd
import scipy.sparse as sp
import glob
from utils.__init__ import get_config
config=get_config()
def define_Hyper_pare():
cv_num=3
num_class=8
num_repeat=1
lgbm_para=None
lgbm_para ={
'boosting_type': 'gbdt',
'num_leaves': 128,
'max_depth': 8,
# 'min_child_weight':1,
'min_child_samples':110,
'feature_fraction': 1.,
'subsample': 1.,
'subsample_freq':0,
'reg_alpha': 0.,
'reg_lambda': 0.,
'learning_rate': 0.05,
'objective': 'multiclass',
'n_estimators': 200,
'subsample_for_bin':20000
}
fm_para={
# 'no_components':10,
'no_components':200, 'loss':'warp', 'learning_rate':0.02, 'max_sampled':400, 'random_state':1, 'user_alpha':1e-05
}
fm_para_text={
# 'no_components':10,
'no_components':200,
'loss':'warp',
'learning_rate':0.03,
'max_sampled':400,
'random_state':1,
'user_alpha':1e-05,
}
return cv_num,num_class,lgbm_para,fm_para,fm_para_text
cv_num,num_class,lgbm_para,fm_para,fm_para_text=define_Hyper_pare()
class TwoStageModel:
def __init__(self):
self.first_stage_models=self.define_first_stage_models(fm_para,fm_para_text)
self.second_stage_model=LGBMRanker()
def define_first_stage_models(self,fm_para=None,fm_para_text=None):
models={}
if fm_para:
models['lightfm']=LightFM(**fm_para)
if fm_para_text:
models['lightfm_text']=LightFM(**fm_para)
return models
def predict(self,X):
prob_train_list=[]
for model_name in self.first_stage_models.keys():
prob_for_second_stage_model=self.first_stage_models_predict(X,model_name)
prob_train_list.append(prob_for_second_stage_model)
prob_train_mean=np.mean(prob_train_list)
second_stage_model_name='./checkpoints/lgbm.sav'
self.second_stage_model=pickle.load(open(second_stage_model_name, 'rb'))
prob=pd.DataFrame(self.second_stage_model.predict(prob_train_mean),index=X.index)
return prob
def first_stage_models_predict(self,X,model_name):
saved_model_name='./checkpoints/%s_%s.sav'%(model_name)
model = pickle.load(open(saved_model_name, 'rb'))
prob= pd.DataFrame(model.transform(X.values),index=X.index)
return prob
def train(self,playlist,tracks,map_train,map_val1,map_val2,val1_pids,val2_pids,play_list_test):
# self.train_first_stage_fm_models('lightfm',self.first_stage_models['lightfm'],map_train,map_val1,map_val2,val1_pids,val2_pids)
# self.train_first_stage_fm_text_models('lightfm_text',self.first_stage_models['lightfm_text'],playlist,tracks,map_train,map_val1,map_val2,val1_pids,val2_pids)
user_seen = set(zip(map_train.pid, map_train.tid))
print('saving candidates')
# self.save_candidates(
# map_train,
# val1_pids.values,
# map_val1.pid.value_counts(),
# './res/ii_candidate.csv',
# map_val1,
# user_seen
# )
# self.save_candidates(
# map_train,
# val2_pids.values,
# map_val2.pid.value_counts(),
# './res/iii_candidate.csv',
# map_val2,
# user_seen
# )
# self.save_candidates(
# map_train,
# play_list_test.pid.values.reshape([-1,1]),
# play_list_test.set_index('pid').num_holdouts,
# 'res/test_candidate.csv',
# None,
# user_seen
# )
# print('creating lightfm features')
self.create_lightfm_features(pd.read_csv('res/ii_candidate.csv'),'new_data/ii_lightfm_features.csv')#modefied!!
self.create_lightfm_features(pd.read_csv('res/iii_candidate.csv'),'new_data/iii_lightfm_features.csv')
self.create_lightfm_features(pd.read_csv('res/test_candidate.csv'),'new_data/test_lightfm_features.csv')
def train_first_stage_fm_models(self,model_name,model,map_train,map_val1,map_val2,val1_pids,val2_pids):
print('\n----------->>fitting fm model: ',model_name)
saved_model_name='./checkpoints/%s.sav'%(model_name)
train_X_sparse = sp.coo_matrix(
(np.ones(map_train.shape[0]), (map_train.pid, map_train.tid)),
shape=(config['num_playlists'], config['num_tracks'])
)
val_sparse = sp.coo_matrix(
(np.ones(map_val1.shape[0]), (map_val1.pid, map_val1.tid)),
shape=(config['num_playlists'], config['num_tracks'])
)
#-------------------------------------------train the fm model-------------------------------------------
best_recall = 0
for i in range(config['epochs_stage1']):
model.fit_partial(train_X_sparse, epochs=config['steps_per_epoch_epoch_stage1'])
recall=recall_at_k(model, val_sparse, k=config['top_k_stage1']).mean()
print('best_recall:',best_recall,'current_recal:',recall)
if recall > best_recall:
pickle.dump(model, open(saved_model_name, 'wb'))
best_recall = recall
def train_first_stage_fm_text_models(self,model_name,model,playlist,tracks,map_train,map_val1,map_val2,val1_pids,val2_pids):
print('\n----------->>fitting fm_text model: ',model_name)
playlist_name = playlist.set_index('pid').name.sort_index()
playlist_name = playlist_name.reindex(np.arange(config['num_playlists'])).fillna('')#expand from train size to train+test size
vectorizer = CountVectorizer(max_features=20000)
user_features = vectorizer.fit_transform(playlist_name)
user_features = sp.hstack([sp.eye(config['num_playlists']), user_features])#??
pickle.dump(user_features, open('./new_data/user_features.pkl', 'wb'))
print('saved user features')
saved_model_name='./checkpoints/%s.sav'%(model_name)
train_X_sparse = sp.coo_matrix(
(np.ones(map_train.shape[0]), (map_train.pid, map_train.tid)),
shape=(config['num_playlists'], config['num_tracks'])
)
zeros_pids = np.array(list(set(val1_pids.values.T.tolist()[0]).difference(map_train.pid.unique())))#pid of val1-val1^train
val_sparse = sp.coo_matrix(
(np.ones(map_val1.shape[0]), (map_val1.pid, map_val1.tid)),
shape=(config['num_playlists'], config['num_tracks'])
)
# no_zeros_pids = np.array(list(set(val1_pids.values).difference(zeros_pids))[:1000])#pid of val1^train
#-------------------------------------------train the fm model-------------------------------------------
best_recall = 0
for i in range(config['epochs_stage1']):
model.fit_partial(train_X_sparse, epochs=config['steps_per_epoch_epoch_stage1'], user_features=user_features)
recall=recall_at_k(model, val_sparse,user_features=user_features, k=config['top_k_stage1']).mean()
# recall_no_zeros=recall_at_k(model, no_zeros_pids[['pid']], k=config['top_k_stage1']).mean()
print('best_recall:',best_recall,'current_recal:',recall)
if recall > best_recall:
pickle.dump(model, open(saved_model_name, 'wb'))
best_recall = recall
def train_second_stage_model(self,train_X,train_Y):
print('\n----------->>fitting 2nd model: ')
#1.-------------------------------------------train the model-------------------------------------------
saved_model_name='./checkpoints/lgbm.sav'
model.fit(train_X.values,train_Y.values.ravel())
pickle.dump(model, open(saved_model_name, 'wb'))
prob_train=model.predict(train_X)
return prob_train
def save_candidates(self,map_train,target_pids, df_size, file_name, df=None,user_seen=None):
'''
df_size:{pid:counts(pid)}
df:map
'''
print('0')
model = pickle.load(open('./checkpoints/lightfm.sav', 'rb'))
model_text = pickle.load(open('./checkpoints/lightfm_text.sav', 'rb'))
user_features = pickle.load(open('new_data/user_features.pkl', 'rb'))
target_pids=target_pids.T.tolist()[0]
print('1')
target_pids_text = list(set(target_pids).difference(map_train.pid))
target_pids_no_text = list(set(target_pids).difference(target_pids_text))
res_notext={}
res_text={}
res={}
print('2')
print(len(target_pids_text))
print(len(target_pids_no_text))
c=1
num_candi_limits=1000
for i in target_pids_text:
scores=model_text.predict(i, np.arange(config['num_tracks']),user_features=user_features)
res[i]=np.argsort(-scores)[:num_candi_limits]
if c%1000==0:
pickle.dump(res, open('./res/candidate%s.pkl'%c, 'wb'))
res={}
c+=1
print(c)
del model_text,user_features
print('2.5')
for i in target_pids_no_text:
scores=model.predict(i, np.arange(config['num_tracks']))
res[i]=np.argsort(-scores)[:num_candi_limits]
if c%1000==0:
pickle.dump(res, open('./res/candidate%s.pkl'%c, 'wb'))
res={}
c+=1
print(c)
pickle.dump(res, open('./res/candidate%s.pkl'%c, 'wb'))
del model
res={}
file_names=glob.glob('./res/candidate*.pkl')
for file_name_tmp in file_names:
res_tmp=pickle.load(open(file_name_tmp, 'rb'))
res.update(res_tmp)
print('3')
if df is not None:
val_tracks = df.groupby('pid').tid.apply(set).to_dict() #{pid:[tracks]}
pids = []
tids = []
targets = []
for pid in target_pids:
l = min(max(df_size[pid] * 15, 700 + df_size[pid]),num_candi_limits)
#l = 2000
pids += [pid] * l
tids += list(res[pid][:l])
if df is not None:
tracks_t = val_tracks[pid]
targets += [i in tracks_t for i in res[pid][:l]]#tracs_true & track_predict_topl
print('4')
candidates = pd.DataFrame()
candidates['pid'] = np.array(pids)
candidates['tid'] = np.array(tids)
if df is not None:
candidates['target'] = np.array(targets).astype(int)
index = []
for pid, tid in candidates[['pid', 'tid']].values:
index.append((pid, tid) not in user_seen)
print('5')
candidates = candidates[index]
candidates.to_csv(file_name, index = None)
print('saved:',file_name)
def create_lightfm_features(self,df,path):
user_features = pickle.load(open('new_data/user_features.pkl', 'rb'))
model = pickle.load(open('./checkpoints/lightfm.sav', 'rb'))
model_text = pickle.load(open('./checkpoints/lightfm_text.sav', 'rb'))
_user_repr_biases,_user_repr=model_text.get_user_representations(user_features)
df['pid_bias'] = model.user_biases[df.pid]
df['tid_bias'] = model.item_biases[df.tid]
pid_embeddings = model.user_embeddings[df.pid]
tid_embeddings = model.item_embeddings[df.tid]
df['lightfm_dot_product'] = (pid_embeddings * tid_embeddings).sum(axis=1)
df['lightfm_prediction'] = df['lightfm_dot_product'] + df['pid_bias'] + df['tid_bias']
df['lightfm_rank'] = df.groupby('pid').lightfm_prediction.rank(ascending=False)
df['pid_bias_text'] = _user_repr_biases[df.pid]
df['tid_bias_text'] = model_text.item_biases[df.tid]
pid_embeddings = _user_repr[df.pid]
tid_embeddings = model_text.item_embeddings[df.tid]
df['lightfm_dot_product_text'] = (pid_embeddings * tid_embeddings).sum(axis=1)
df['lightfm_prediction_text'] = df['lightfm_dot_product_text'] + df['pid_bias_text'] + df['tid_bias_text']
df['lightfm_rank_text'] = df.groupby('pid').lightfm_prediction_text.rank(ascending=False)
df.to_csv(path, index = None)