search

jakc4103 / DFQ

PyTorch implementation of Data Free Quantization Through Weight Equalization and Bias Correction.
MIT License
258 stars 45 forks source link

Could you explain the variable 'data' in main_cls.py line 77. Thanks. #6

Closed EdwardLiuYu closed 4 years ago

jakc4103 commented 4 years ago

It's a dummy input for PyTransformer to build model graph. You can change it to whatever you want with dimensions 4 (batch, channel, height width) .

EdwardLiuYu commented 4 years ago

Ok, I got it. Thanks a lot.

刘子力

nmliuzili@163.com | 签名由网易邮箱大师定制

On 12/16/2019 11:35，jakc4103notifications@github.com wrote：

It's a dummy input for PyTransformer to build model graph. You can change it to whatever you want with dimensions 4 (batch, channel, height width) .

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

EdwardLiuYu commented 4 years ago

Sorry for bother you again. I don’t understand the operation in relation.py line61-78. And why some relations are deleted from ’tmp’?

刘子力

nmliuzili@163.com | 签名由网易邮箱大师定制

On 12/16/2019 11:36，刘子力nmliuzili@163.com wrote： Ok, I got it. Thanks a lot.

刘子力

nmliuzili@163.com | 签名由网易邮箱大师定制

On 12/16/2019 11:35，jakc4103notifications@github.com wrote：

It's a dummy input for PyTransformer to build model graph. You can change it to whatever you want with dimensions 4 (batch, channel, height width) .

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

jakc4103 commented 4 years ago

Sorry for bother you again. I don’t understand the operation in relation.py line61-78. And why some relations are deleted from ’tmp’? | | 刘子力 | | nmliuzili@163.com | 签名由网易邮箱大师定制 On 12/16/2019 11:36，刘子力nmliuzili@163.com wrote： Ok, I got it. Thanks a lot. | | 刘子力 | | nmliuzili@163.com | 签名由网易邮箱大师定制 On 12/16/2019 11:35，jakc4103notifications@github.com wrote： It's a dummy input for PyTransformer to build model graph. You can change it to whatever you want with dimensions 4 (batch, channel, height width) . — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

I apply layer equalization to those at least 3 consecutive conv_layers are connected without branching. Since there is no such structure in segmentation and classification model, It only affect ssd model (which i am currently testing).

EdwardLiuYu commented 4 years ago

Thanks for your answer. I have learned.

刘子力

邮箱：nmliuzili@163.com |

签名由 网易邮箱大师 定制

On 12/16/2019 18:24, jakc4103 wrote:

Sorry for bother you again. I don’t understand the operation in relation.py line61-78. And why some relations are deleted from ’tmp’? | | 刘子力 | | nmliuzili@163.com | 签名由网易邮箱大师定制 On 12/16/2019 11:36，刘子力nmliuzili@163.com wrote： Ok, I got it. Thanks a lot. | | 刘子力 | | nmliuzili@163.com | 签名由网易邮箱大师定制 On 12/16/2019 11:35，jakc4103notifications@github.com wrote： It's a dummy input for PyTransformer to build model graph. You can change it to whatever you want with dimensions 4 (batch, channel, height width) . — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

I apply layer equalization to those at least 3 consecutive conv_layers are connected without branching. Since there is no such structure in segmentation and classification model, It only affect ssd model (which i am currently testing).

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.