Haiping
|
090dc1eee3
|
Merge pull request #1190 from dogvane/master
解决keras模式下,使用GPU训练时会爆显存的bug。
|
2 years ago |
dogvane
|
baf620a3e8
|
解决keras模式下,使用GPU训练时会爆显存的bug。
观察到的现象是,一些模型增大batchsize后,会在首个epoch的中途爆显存不足,只要过了一个epoch后,就能完整训练。同样的batchsize在python下能设置大得多的值。
最后使用最小训练代码分析出,是每个step之后,图片加载到显存里的数据没有释放导致的。
在寻找释放显存接口没有结果的时候,直接使用了GC.Collect();可以让显存主动回收。
因此当前的修复方案是在每个step里,都执行一次 GC.Collect(); 用来释放显存资源。
|
2 years ago |
dogvane
|
5e4f53077f
|
修正图片左右和上下翻转的问题,并增加对应测试用例。
|
2 years ago |
dogvane
|
ba8f0b084f
|
add DepthwiseConv2D (深度可分离卷积)
|
2 years ago |
Haiping
|
43c3705183
|
Merge pull request #1189 from Wanglongzhi2001/master
feat: add the implementation of class_weight in model.fit
|
2 years ago |
Wanglongzhi2001
|
a1c64effcf
|
feat: add the implementation of class_weight in model.fit
|
2 years ago |
Haiping
|
f16902dd0b
|
Merge pull request #1188 from hchen2020/master
Allow Model to cache weights.
|
2 years ago |
hchen
|
0f02885dfb
|
Allow Model to cache weights.
|
2 years ago |
Haiping
|
0ee9d424e5
|
Merge pull request #1187 from Wanglongzhi2001/master
feat: add the implementation of sample_weight in model.fit
|
2 years ago |
Wanglongzhi2001
|
f5af07ce5e
|
feat: add the implementation of sample_weight in model.fit
|
2 years ago |
Haiping
|
15763dfbbe
|
Merge pull request #1186 from Beacontownfc/ragged
Improve RaggedTensor
|
2 years ago |
Beacontownfc
|
02bfb9af17
|
improve raggedtensor
|
2 years ago |
Haiping Chen
|
eb4c1f4fb0
|
Release v0.110.4.
|
2 years ago |
Haiping
|
8e02682637
|
Merge pull request #1182 from lingbai-kong/imdbfix
fix: adjust imdb dataset loader for faster loading speed
|
2 years ago |
lingbai-kong
|
9fb847991a
|
fix: adjust imdb dataset loader for faster loading speed
|
2 years ago |
Haiping
|
3811e4e140
|
Merge pull request #1180 from Wanglongzhi2001/master
fix: fix EarlyStopping
|
2 years ago |
Wanglongzhi2001
|
f809f6eace
|
fix: fix EarlyStopping
|
2 years ago |
Wanglongzhi2001
|
9552d4cb7a
|
feat: add np.less and np.greater binding
|
2 years ago |
Haiping Chen
|
725ec1e55f
|
Optimize imdb.load_data
|
2 years ago |
Haiping
|
74afa14333
|
Merge pull request #1176 from lingbai-kong/ndarrayload
fix: improve unpickler speed with BufferedStream
|
2 years ago |
lingbai-kong
|
56e389154c
|
improve unpickler speed with BufferedStream
|
2 years ago |
Haiping
|
c814fe121a
|
Merge pull request #1175 from lingbai-kong/ndarrayload
optimize: temporal complexity of Imdb dataset loader
|
2 years ago |
Haiping
|
eb49be09ff
|
Merge pull request #1173 from novikov-alexander/alnovi/gradient_tests
Gradient tests which were dependant only on cached_session
|
2 years ago |
Alexander Novikov
|
57feb65dbc
|
comment IndexedSlices test
|
2 years ago |
lingbai-kong
|
628b2ce736
|
optimize temporal complexity of Imdb dataset loader
|
2 years ago |
Alexander Novikov
|
a9dad3ce11
|
fixme labels
|
2 years ago |
Alexander Novikov
|
adef5bcdc5
|
gradient tests
|
2 years ago |
Haiping
|
9ddff69963
|
Merge pull request #1172 from novikov-alexander/alnovi/cached_session
cached_session for graph tests
|
2 years ago |
Alexander Novikov
|
9d71cad96e
|
using and no IEnumerable
|
2 years ago |
Alexander Novikov
|
ae50fa93ba
|
fix fleaky test boundary continue
|
2 years ago |
Alexander Novikov
|
c33a3a7444
|
cached_session for graph tests
|
2 years ago |
Haiping
|
a91e35884e
|
Merge pull request #1171 from asaf92/patch-1
fix: model does not stop on stop_training == true #1170
|
2 years ago |
Asaf Agami
|
1142828855
|
fix: model does not stop on stop_training == true
|
2 years ago |
Haiping
|
179c3f0216
|
Merge pull request #1169 from lingbai-kong/ndarrayload
add: loading pickled npy file for imdb dataset loader
|
2 years ago |
lingbai-kong
|
f57a6fe6ed
|
optimize the time complexity of Imdb dataset loader
|
2 years ago |
lingbai-kong
|
28c77f53d6
|
implement Imdb dataset loader
|
2 years ago |
lingbai-kong
|
10f6819f08
|
Merge branch 'master' into ndarrayload
# Conflicts:
# src/TensorFlowNET.Core/Tensorflow.Binding.csproj
# src/TensorFlowNET.Keras/Datasets/Imdb.cs
|
2 years ago |
lingbai-kong
|
ea978bbf21
|
optimize code structure of reconstruction ndarray from pickled npy file
|
2 years ago |
lingbai-kong
|
9d10daf30f
|
add reconstruction and setstate of NDArray for loading pickled npy file.
|
2 years ago |
Haiping
|
70d681c020
|
Merge pull request #1168 from Wanglongzhi2001/master
feat: implement GRU layer
|
2 years ago |
Wanglongzhi2001
|
7b077eac7e
|
feat: implement GRU layer
|
2 years ago |
Haiping Chen
|
ba1ddb4448
|
Set SGD default value.
|
2 years ago |
Haiping
|
b18343ca26
|
Merge pull request #1165 from lingbai-kong/imdbfix
fix: validate dataset of `Imdb` do not load bug & add: custom `Imdb` path
|
2 years ago |
Haiping
|
fed53f1f76
|
Merge pull request #1164 from Wanglongzhi2001/master
fix: partially fix the bug of load_model
|
2 years ago |
lingbai-kong
|
8e3ba22c83
|
fix: validate dataset of `Imdb` do not load bug & add: custom `Imdb` path
|
2 years ago |
Wanglongzhi2001
|
f679af67e6
|
fix: partially fix the bug of load_model
|
2 years ago |
Wanglongzhi2001
|
f5eb4ff0a0
|
fix: partially fix the bug of load_model
|
2 years ago |
Haiping
|
adeed05f64
|
Merge pull request #1162 from Wanglongzhi2001/master
fix: remove the reflection in the implemention of Bidirectional
|
2 years ago |
“Wanglongzhi2001”
|
6d3f134637
|
fix: remove the reflection in the implemention of Bidirectional
|
2 years ago |
Haiping
|
48403a55e2
|
Merge pull request #1161 from Wanglongzhi2001/master
fix: add the momentum parameter's implemention of SGD
|
2 years ago |