Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

AI训练复现不成功的原因?Valid Loss超大 #13

Open
linchengsi opened this issue Jul 4, 2019 · 3 comments
Open

AI训练复现不成功的原因?Valid Loss超大 #13

linchengsi opened this issue Jul 4, 2019 · 3 comments

Comments

@linchengsi
Copy link

linchengsi commented Jul 4, 2019

您好!我下载了您的代码然后直接运行了HEVC-Complexity-Reduction-master\ETH-CNN_Training_AI里面的train_CNN_CTU64.py。因为readme文件里面说是可以直接运行的?所以我也没有进行修改。
然而我得到的结果图跟原本就在models文件夹里面的很不一样,具体如下
image
您提供的结果(原本就在models文件夹里面的)其中一张如下
image
可以看到我的valid loss等等很不正常,数值超大,但是我刚刚进行代码的学习暂时不是很清楚造成问题的原因是什么。请问您能解释一下吗?感谢!

(我的配置是windows10 GTX1060MQ python3.7 tensorflow1.13.1 使用pycharm运行)
还有我的邮箱是[email protected]

@tianyili2017
Copy link
Owner

您好,
程序确实可以直接运行,但因为全部数据库太大,附带的数据文件只是所有数据的一小部分,用于测试程序是否能跑通。如果迭代100万次训练,很容易出现过拟合。我也准备重新下载一次、运行,看结果是否像您那样?如果希望复现论文结果,推荐您按照训练说明,由原始YUV视频和标签生成完整数据文件,再训练。谢谢您指出本程序的一个问题,关注我们的工作!

@tianyili2017
Copy link
Owner

您好。我重新运行一遍,结果和您的图像很相似,说明在小样本上确实会出现过拟合。之后,准备在训练说明里标注一下:demo sets只用于使程序运行,可以用来验证Tensorflow等环境是否正确安装,但不用于复现论文结果。欲复现与论文接近的结果,请先生成完整训练数据。谢谢您理解!

@zhoushuairan
Copy link

zhoushuairan commented Aug 12, 2019

您好,我下载您的代码之后运行HEVC-Complexity-Reduction-master\HM-16.5_Test_AI\build\HM_vc10.sln 结果出现如下信息 Tensor("Conv2D:0", shape=(?, 1, 1, 1), dtype=float32)
Tensor("ResizeNearestNeighbor:0", shape=(?, 16, 16, 1), dtype=float32)
Tensor("LeakyRelu:0", shape=(?, 4, 4, 16), dtype=float32)
Tensor("LeakyRelu_1:0", shape=(?, 2, 2, 24), dtype=float32)
Tensor("LeakyRelu_2:0", shape=(?, 1, 1, 32), dtype=float32)
Tensor("Conv2D_4:0", shape=(?, 2, 2, 1), dtype=float32)
Tensor("ResizeNearestNeighbor_1:0", shape=(?, 32, 32, 1), dtype=float32)
Tensor("LeakyRelu_3:0", shape=(?, 8, 8, 16), dtype=float32)
Tensor("LeakyRelu_4:0", shape=(?, 4, 4, 24), dtype=float32)
Tensor("LeakyRelu_5:0", shape=(?, 2, 2, 32), dtype=float32)
Tensor("Conv2D_8:0", shape=(?, 4, 4, 1), dtype=float32)
Tensor("ResizeNearestNeighbor_2:0", shape=(?, 64, 64, 1), dtype=float32)
Tensor("LeakyRelu_6:0", shape=(?, 16, 16, 16), dtype=float32)
Tensor("LeakyRelu_7:0", shape=(?, 8, 8, 24), dtype=float32)
Tensor("LeakyRelu_8:0", shape=(?, 4, 4, 32), dtype=float32)
Tensor("concat:0", shape=(?, 2688), dtype=float32)
Tensor("cond/Merge:0", shape=(?, 64), dtype=float32)
Tensor("cond_1/Merge:0", shape=(?, 48), dtype=float32)
Tensor("cond_2/Merge:0", shape=(?, 1), dtype=float32)
Tensor("cond_3/Merge:0", shape=(?, 128), dtype=float32)
Tensor("cond_4/Merge:0", shape=(?, 96), dtype=float32)
Tensor("cond_5/Merge:0", shape=(?, 4), dtype=float32)
Tensor("cond_7/Merge:0", shape=(?, 256), dtype=float32)
Tensor("cond_8/Merge:0", shape=(?, 192), dtype=float32)
Tensor("cond_9/Merge:0", shape=(?, 16), dtype=float32)
D:\HM\HM-16.5_Test_AI\bin\vc10\x64\Release\BasketballPass_416x240_50.yuv frame 501/501 416x240

Predicting Time : 7.986 sec.


HM software: Encoder Version [16.5] (including RExt)[Windows][VS 1900][64 bit]

python video_to_cu_depth.py D:\HM\HM-16.5_Test_AI\bin\vc10\x64\Release\BasketballPass_416x240_50.yuv 416 240 32

Input File : D:\HM\HM-16.5_Test_AI\bin\vc10\x64\Release\BasketballPass_416x240_50.yuv
Bitstream File : str.bin
Reconstruction File : rec.yuv
Real Format : 416x240 50Hz
Internal Format : 416x240 50Hz
Sequence PSNR output : Linear average only
Sequence MSE output : Disabled
Frame MSE output : Disabled
Cabac-zero-word-padding : Enabled
Frame/Field : Frame based coding
Frame index : 0 - 99 (100 frames)
Profile : main
CU size / depth / total-depth : 64 / 4 / 4
RQT trans. size (min / max) : 4 / 32
Max RQT depth inter : 3
Max RQT depth intra : 3
Min PCM size : 8
Motion search range : 64
Intra period : 1
Decoding refresh type : 0
QP : 32.00
Max dQP signaling depth : 0
Cb QP Offset : 0
Cr QP Offset : 0
QP adaptation : 0 (range=0)
GOP size : 1
Input bit depth : (Y:8, C:8)
MSB-extended bit depth : (Y:8, C:8)
Internal bit depth : (Y:8, C:8)
PCM sample bit depth : (Y:8, C:8)
Intra reference smoothing : Enabled
diff_cu_chroma_qp_offset_depth : -1
extended_precision_processing_flag : Disabled
implicit_rdpcm_enabled_flag : Disabled
explicit_rdpcm_enabled_flag : Disabled
transform_skip_rotation_enabled_flag : Disabled
transform_skip_context_enabled_flag : Disabled
cross_component_prediction_enabled_flag: Disabled
high_precision_offsets_enabled_flag : Disabled
persistent_rice_adaptation_enabled_flag: Disabled
cabac_bypass_alignment_enabled_flag : Disabled
log2_sao_offset_scale_luma : 0
log2_sao_offset_scale_chroma : 0
Cost function: : Lossy coding (default)
RateControl : 0
Max Num Merge Candidates : 5

TOOL CFG: IBD:0 HAD:1 RDQ:1 RDQTS:1 RDpenalty:0 SQP:0 ASR:0 FEN:1 ECU:0 FDM:1 CFM:0 ESD:0 RQT:1 TransformSkip:1 TransformSkipFast:1 TransformSkipLog2MaxSize:2 Slice: M=0 SliceSegment: M=0 CIP:0 SAO:1 PCM:0 TransQuantBypassEnabled:0 WPP:0 WPB:0 PME:2 WaveFrontSynchro:0 WaveFrontSubstreams:1 ScalingList:0 TMVPMode:1 AQpS:0 SignBitHidingFlag:1 RecalQP:0

Non-environment-variable-controlled macros set as follows:

                            RExt__DECODER_DEBUG_BIT_STATISTICS =   0
                                  RExt__HIGH_BIT_DEPTH_SUPPORT =   0
                        RExt__HIGH_PRECISION_FORWARD_TRANSFORM =   0
                                    O0043_BEST_EFFORT_DECODING =   0

               Input ChromaFormatIDC =   4:2:0
   Output (internal) ChromaFormatIDC =   4:2:0

``
于是我单独运行D:\HM\HM-16.5_Test_AI\bin下文件video_to_cu_depth.py文件 出现
File"video_to_cu_depth.py",line 120,in
assert len(sys.argv)==5
AssertionError
请问是为什么呢???如果你能回答 很感谢!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants