pyTorchClassification/furniture训练记录 at master · chaffeechenyefei/pyTorchClassification · GitHub

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
##20190813##
#Config
ResNet50 + Avg + FC1024 + Relu + DictLayer 10*128(alpha=0.05) + Fc128 + Softmax 128
#Loss
Fc128 -> softmax + ce
DictLayer 1280 -> triplet loss 0.1
DictLayer -> dictloss 0.05

--> valid_top5 97.725 | valid_top1 83.608 | valid_f1 0.833 | valid_loss 0.587

code source
/home/ubuntu/pytorch/2class_clf_pytorch/


##20190814##
#Config
ResNet50 + Avg + FC1024 + ReLU + FC1024 + ReLU + FC128 + softmax 128
#Loss
FC128 -> softmax + ce

--> valid_top5 97.170 | valid_top1 82.428 | valid_f1 0.821 | valid_loss 0.632
:: We saw a increasement by using DictLayer


##20190815##
#Config
ResNet50 + Avg + FC1024 + Leaky + FC1024 + Leaky + FC128 + softmax 128
#Loss

--> valid_top5 97.465 | valid_top1 82.740 | valid_f1 0.824 | valid_loss 0.607

##20190815##
#Config
ResNet50 + Avg + FC1024 + Leaky + FC1024 + Leaky + FC128 + softmax 128
#Loss
-> model.initial | valid_f1 0.84


##20190823##
#Config
ResNetBBoxMask::
ResNet50 -> 4 coords -> MaskLayer -> newInput -> ResNet50 + Avg + FC1024 + Leaky + FC1024 + Leaky + FC128 + softmax 128
#Loss
Fc128 -> softmax + ce 0.5
DictLayer 1024 -> triplet loss
"{'valid_f1': 0.7470779236985641, 'valid_loss': 1.3797708, 'valid_top1': 74.73519897460938, 'valid_top5': 93.8704605102539, 'step': 25000, 'dt': '2019-08-23T02:27:08.308639'}"

##20190823## -> furniture_resnet50
#Config
ResNet50 + Avg + FC1024 + Leaky + FC1024 + Leaky + FC128 + softmax 128
#Loss
valid_top5 97.204 | valid_top1 83.521 | valid_f1 0.834 | valid_loss 0.632 (max 0.84) -> initial


##20190910
inceptionv4_netvlad: Raw Inceptionv4 + conv(1536,32) + netvlad(32,2048)
valid_top5 82.759 | valid_top1 52.107 | valid_loss 1.858 | valid_f1 0.413
valid_top5 96.043 | valid_top1 95.324 | valid_f1 0.931 | valid_loss 0.345

##20190910
inceptionv4_netvlad: yiheng best_trained[freezed] + conv(1535,256) + netvlad(256,64)
valid_top5 92.912 | valid_top1 68.582 | valid_loss 1.138 | valid_f1 0.561 ==> model_freeze.pth

inceptionv4_netvlad: yiheng best_trained + conv(1535,256) + netvlad(256,64)  [best model from freezed one]
valid_top5 86.973 | valid_top1 56.130 | valid_loss 1.599 | valid_f1 0.475 [Epoch 1]

inceptionv4_netvlad: yiheng best_trained + conv(1535,256) + netvlad(256,128)
valid_top5 94.828 | valid_top1 69.540 | valid_loss 1.114 | valid_f1 0.601 ==> model_freeze_conv256_k128.pth

inceptionv4_netvlad: yiheng best_trained2
valid_top5 93.870 | valid_top1 69.732 | valid_loss 1.088 | valid_f1 0.574 ==> max_valid_model_7146.pth

inceptionv4_netvlad: yiheng best_trained2 + conv(1535,256) + netvlad(256,128)
valid_top5 95.594 | valid_top1 71.456 | valid_loss 1.068 | valid_f1 0.620 ==> model_freeze_conv256_k128_v2.pth
best: 72.22% not stored!

Conclusion: netvlad can be used as traditional toolbox where features are fixed and learned from cnn.
直接attention模型，结果不收敛。freeze cls_net, freeze att_net.radius=0.5，依然无效。
必须先进行attention模型的预训练。


##Notes:
location_recommend_model_v6_5city：基于location_scorecard_190912的结果，一个company唯一确定一个location
location_recommend_model_v6_5city_191113：基于location_scorecard_1901113的结果，一个company唯一确定一个location。新增了ww的location。
上述虽保证了一个company一个location，但存在location没有匹配上任何company的现象。
下一步主要解决这个问题，放开一个company一个location的限制。

location_recommend_model_v6_191114: 基于location_scorecard_1901113的结果，一个company可以映射多个location。


nohup python3 -u main_location_company.py --model location_recommend_model_v5 --run_root result/location_recommend_model_v5 --lr 0.01 --cos_sim_loss --testStep 20000 > nohup.out 2> nohup.err &

python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6 --lr 0.01

python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6 --lr 0.01 --mode validate --ckpt model_deep_wide.final --query_location

python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_5city --lr 0.01
python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_5city --lr 0.01 --mode validate

nohup python3 -u main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_5city --lr 0.01 --mode train >lrm_5c.out 2>lrm_5c.err &
python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_5city --lr 0.01 --mode predict_test --ckpt model_loss_best.pt

python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_5city --lr 0.01 --mode predict_sub --ckpt model_loss_best.pt

nohup python3 -u main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_5city_191113 --lr 0.01 --mode train >lrm_5c.out 2>lrm_5c.err &
python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_5city_191113 --lr 0.01 --mode predict_test --ckpt model_loss_best.pt
python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_5city_191113 --lr 0.01 --mode predict_test --ckpt model_loss_best.pt --apps _191113.csv
python3 main_location_company_model_based_reason.py --apps _191113.csv --pre_name sampled_ww_


nohup python3 -u main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_191114 --lr 0.01 --mode train --apps _191114.csv >lrm_5c.out 2>lrm_5c.err &
python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_191114 --lr 0.01 --mode predict_test --ckpt model_loss_best.pt


python3 linkCompanyAndLocation_v2_focus_on_location.py --run_root ~/location_recommender_system/ --geo_bit 7 --dist_thresh 500 --apps _191114.csv
python3 get_csv_for_training_and_testing.py --run_root ~/location_recommender_system/ --app_date _191114 --ratio 0.8
python3 get_sub_recommend_reason_after_similarity.py --run_root ~/location_recommender_system/ --ls_card location_scorecard_191113.csv --apps _191113.csv --sampled --ww
python3 get_embedding_feature_region.py --path /home/ubuntu/location_recommender_system/ --maxK 50 --model location_recommend_region_model_v4 --run_root result/location_RSRBv4_191114/


python3 get_csv_for_training_and_testing_region_model.py --run_root /home/ubuntu/location_recommender_system --app_date _191114
python3 main_location_intelligence_region.py --run_root result/location_RSRBv1_191114 --model location_recommend_region_model_v1 --lr 0.01 --mode train --trainStep 1000 --batch-size 2 --workers 4
nohup python3 -u main_location_intelligence_region.py --run_root result/location_RSRBv1_191114 --model location_recommend_region_model_v1 --lr 0.01 --mode train --trainStep 1000 --batch-size 4 >mlir_1.out 2>mlir_1.err &
python3 main_location_intelligence_region.py --run_root result/location_RSRBv1_191114 --model location_recommend_region_model_v1 --lr 0.01 --mode train --trainStep 1000 --batch-size 1

python3 main_location_intelligence_region.py --run_root result/location_RSRBv2_191114 --model location_recommend_region_model_v2 --lr 0.01 --mode train --trainStep 1000 --batch-size 1


nohup python3 -u main_location_intelligence_region.py --run_root result/location_RSRBv0_191114 --model location_recommend_region_model_v0 --lr 0.01 --mode train --trainStep 1000 --batch-size 4 >mlir_0.out 2>mlir_0.err &
nohup python3 -u main_location_intelligence_region.py --run_root result/location_RSRBv1_191114 --model location_recommend_region_model_v1 --lr 0.01 --mode train --trainStep 1000 --batch-size 4 >mlir_1.out 2>mlir_1.err &
nohup python3 -u main_location_intelligence_region.py --run_root result/location_RSRBv2_191114 --model location_recommend_region_model_v2 --lr 0.01 --mode train --trainStep 1000 --batch-size 4 >mlir_2.out 2>mlir_2.err &
nohup python3 -u main_location_intelligence_region.py --run_root result/location_RSRBv3_191114 --model location_recommend_region_model_v3 --lr 0.01 --mode train --trainStep 1000 --batch-size 4 --n-epochs 160 >mlir_3.out 2>mlir_3.err &
nohup python3 -u main_location_intelligence_region.py --run_root result/location_RSRBv4_191114 --model location_recommend_region_model_v4 --lr 0.01 --mode train --trainStep 1000 --batch-size 4 --n-epochs 80 >mlir_4.out 2>mlir_4.err &
python3 -u main_location_intelligence_region.py --run_root result/location_RSRBv5_191114 --model location_recommend_region_model_v5 --lr 0.01 --mode train --trainStep 1000 --batch-size 4 --n-epochs 80
nohup python3 -u main_location_intelligence_region.py --run_root result/location_RSRBv5_191114 --model location_recommend_region_model_v5 --lr 0.01 --mode train --trainStep 1000 --batch-size 4 --n-epochs 160 >mlir_5.out 2>mlir_5.err &

python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_5city_191113 --lr 0.01 --mode predict_salesforce --ckpt model_loss_best.pt --apps _191113.csv --query_location
==OR==
python3 main_location_company.py --model location_recommend_model_v6 --run_root result/location_recommend_model_v6_5city_191113 --lr 0.01 --mode predict_salesforce --ckpt model_loss_best.pt --apps _191113.csv
cp *.csv ~/location_recommender_system/
python3 main_location_company_model_based_reason.py --apps _191113.csv --pre_name sampled_ww_
python3 get_sub_recommend_reason_after_similarity.py --run_root ~/location_recommender_system/ --ls_card location_scorecard_191113.csv --apps _191113.csv --sampled --ww