search

dzyswy / sesr

sesr super-efficient super resolution
MIT License
1 stars 0 forks source link

Onnx conversion code #1

Open xushuanglong opened 2 weeks ago

xushuanglong commented 2 weeks ago

大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk244’,'unk245'等 not support 代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx

xushuanglong commented 2 weeks ago

Hello boss, it seems that your model only supports images of 64X64 size. I think my model can support images of other sizes, such as 400X400. How should I write it, and it seems that some operators need to be removed. There is a problem with directly generating onnx to rknn code, such as' unk244 ',' unk245 ', etc., which are not supported

dzyswy commented 2 weeks ago

我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。

---Original--- From: @.> Date: Sat, Sep 14, 2024 14:55 PM To: @.>; Cc: @.***>; Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)

大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk244’,'unk245'等 not support 代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>

xushuanglong commented 2 weeks ago

那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛

---Original--- From: @.> Date: Sat, Sep 14, 2024 15:04 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。

---Original--- From: @.> Date: Sat, Sep 14, 2024 14:55 PM To: @.>; Cc: @.***>; Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)

大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk244’,'unk245'等 not support 代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.>

dzyswy commented 2 weeks ago

是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。

---Original--- From: @.> Date: Sat, Sep 14, 2024 15:11 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛

---Original--- From: @.> Date: Sat, Sep 14, 2024 15:04 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。

---Original---
From: @.>
Date: Sat, Sep 14, 2024 14:55 PM
To: @.>;
Cc: @.***>;
Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)

大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk244’,'unk245'等 not support
代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

xushuanglong commented 2 weeks ago

是的，就是因为导出的onnx模型设置输入是6464，因此rknn就支持6464,我自己转换有点问题😂

---Original--- From: @.> Date: Sat, Sep 14, 2024 15:16 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。

---Original--- From: @.> Date: Sat, Sep 14, 2024 15:11 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛

---Original---
From: @.>
Date: Sat, Sep 14, 2024 15:04 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。

---Original---
From: @.>
Date: Sat, Sep 14, 2024 14:55 PM
To: @.>;
Cc: @.***>;
Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)

大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk244’,'unk245'等 not support
代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.>

xushuanglong commented 2 weeks ago

您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64

---Original--- From: @.> Date: Sat, Sep 14, 2024 15:16 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。

---Original--- From: @.> Date: Sat, Sep 14, 2024 15:11 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛

---Original---
From: @.>
Date: Sat, Sep 14, 2024 15:04 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。

---Original---
From: @.>
Date: Sat, Sep 14, 2024 14:55 PM
To: @.>;
Cc: @.***>;
Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)

大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk244’,'unk245'等 not support
代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.>

dzyswy commented 2 weeks ago

不好意思，我目前还不知道。

---Original--- From: @.> Date: Sat, Sep 14, 2024 16:11 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64

---Original--- From: @.> Date: Sat, Sep 14, 2024 15:16 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。

---Original---
From: @.>
Date: Sat, Sep 14, 2024 15:11 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛

---Original---
From: @.>
Date: Sat, Sep 14, 2024 15:04 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。    

---Original---    
From: ***@***.***>    
Date: Sat, Sep 14, 2024 14:55 PM    
To: ***@***.***>;    
Cc: ***@***.***>;    
Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)    

大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support    
 代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx    

—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.>

xushuanglong commented 2 weeks ago

好的，谢谢您的帮助😁

---Original--- From: @.> Date: Sat, Sep 14, 2024 16:25 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

不好意思，我目前还不知道。

---Original--- From: @.> Date: Sat, Sep 14, 2024 16:11 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64

---Original---
From: @.>
Date: Sat, Sep 14, 2024 15:16 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。

---Original---
From: @.>
Date: Sat, Sep 14, 2024 15:11 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛    

---Original---    
From: ***@***.***>    
Date: Sat, Sep 14, 2024 15:04 PM    
To: ***@***.***>;    
Cc: ***@***.******@***.***>;    
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)    

 我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。     

 ---Original---     
 From: ***@***.***>     
 Date: Sat, Sep 14, 2024 14:55 PM     
 To: ***@***.***>;     
 Cc: ***@***.***>;     
 Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)     

 大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support     
  代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx     

 —     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
 You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>     
—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you authored the thread.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.>

xushuanglong commented 1 week ago

嗯嗯，我下载github上sesr训练完生成的onnx和贵公司算法工程师生成的onnx不一样，应该是贵公司算法工程师修改并自定义onnx了，我只能利用您代码上的onnx进行修改，修改了onnx输入和输出尺寸和部分节点，这样控制了图片输入大小。

---Original--- From: @.> Date: Sat, Sep 14, 2024 16:25 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

不好意思，我目前还不知道。

---Original--- From: @.> Date: Sat, Sep 14, 2024 16:11 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64

---Original---
From: @.>
Date: Sat, Sep 14, 2024 15:16 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。

---Original---
From: @.>
Date: Sat, Sep 14, 2024 15:11 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛    

---Original---    
From: ***@***.***>    
Date: Sat, Sep 14, 2024 15:04 PM    
To: ***@***.***>;    
Cc: ***@***.******@***.***>;    
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)    

 我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。     

 ---Original---     
 From: ***@***.***>     
 Date: Sat, Sep 14, 2024 14:55 PM     
 To: ***@***.***>;     
 Cc: ***@***.***>;     
 Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)     

 大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support     
  代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx     

 —     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
 You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>     
—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you authored the thread.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.>

dzyswy commented 1 week ago

好的，谢谢提醒。

---Original--- From: @.> Date: Wed, Sep 18, 2024 17:15 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

嗯嗯，我下载github上sesr训练完生成的onnx和贵公司算法工程师生成的onnx不一样，应该是贵公司算法工程师修改并自定义onnx了，我只能利用您代码上的onnx进行修改，修改了onnx输入和输出尺寸和部分节点，这样控制了图片输入大小。

---Original--- From: @.> Date: Sat, Sep 14, 2024 16:25 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

不好意思，我目前还不知道。

---Original---
From: @.>
Date: Sat, Sep 14, 2024 16:11 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64

---Original---
From: @.>
Date: Sat, Sep 14, 2024 15:16 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。    

---Original---    
From: ***@***.***>    
Date: Sat, Sep 14, 2024 15:11 PM    
To: ***@***.***>;    
Cc: ***@***.******@***.***>;    
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)    

 那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛     

 ---Original---     
 From: ***@***.***>     
 Date: Sat, Sep 14, 2024 15:04 PM     
 To: ***@***.***>;     
 Cc: ***@***.******@***.***>;     
 Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)     

  我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。      

  ---Original---      
  From: ***@***.***>      
  Date: Sat, Sep 14, 2024 14:55 PM      
  To: ***@***.***>;      
  Cc: ***@***.***>;      
  Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)      

  大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support      
   代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx      

  —      
  Reply to this email directly, view it on GitHub, or unsubscribe.      
  You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>      
 —     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
 You are receiving this because you authored the thread.Message ID: ***@***.***>     
—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you commented.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.>

xushuanglong commented 4 days ago

大佬您好，我这边64x64超分到256x256,在rk3588上有150ms左右,6，7帧的样子，请问你那边部署速度怎么样，我看到文章说在arm ethos-n78 npu上，这个模型应该能干到大几百帧，跟我情况完全不一样哇。

---Original--- From: @.> Date: Thu, Sep 19, 2024 10:14 AM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

好的，谢谢提醒。

---Original--- From: @.> Date: Wed, Sep 18, 2024 17:15 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

嗯嗯，我下载github上sesr训练完生成的onnx和贵公司算法工程师生成的onnx不一样，应该是贵公司算法工程师修改并自定义onnx了，我只能利用您代码上的onnx进行修改，修改了onnx输入和输出尺寸和部分节点，这样控制了图片输入大小。

---Original---
From: @.>
Date: Sat, Sep 14, 2024 16:25 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

不好意思，我目前还不知道。

---Original---
From: @.>
Date: Sat, Sep 14, 2024 16:11 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64    

---Original---    
From: ***@***.***>    
Date: Sat, Sep 14, 2024 15:16 PM    
To: ***@***.***>;    
Cc: ***@***.******@***.***>;    
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)    

 是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。     

 ---Original---     
 From: ***@***.***>     
 Date: Sat, Sep 14, 2024 15:11 PM     
 To: ***@***.***>;     
 Cc: ***@***.******@***.***>;     
 Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)     

  那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛      

  ---Original---      
  From: ***@***.***>      
  Date: Sat, Sep 14, 2024 15:04 PM      
  To: ***@***.***>;      
  Cc: ***@***.******@***.***>;      
  Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)      

   我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。       

   ---Original---       
   From: ***@***.***>       
   Date: Sat, Sep 14, 2024 14:55 PM       
   To: ***@***.***>;       
   Cc: ***@***.***>;       
   Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)       

   大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support       
    代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx       

   —       
   Reply to this email directly, view it on GitHub, or unsubscribe.       
   You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>       
  —      
  Reply to this email directly, view it on GitHub, or unsubscribe.      
  You are receiving this because you authored the thread.Message ID: ***@***.***>      
 —     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
 You are receiving this because you commented.Message ID: ***@***.***>     
—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you authored the thread.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.>

dzyswy commented 4 days ago

我测试是至少可以跑到60hz的，但用的时间不固定。有时是8ms，有时是16ms。我用了rga做预处理，cpu做的后处理。内存分配是用的是dma内存。

---Original--- From: @.> Date: Tue, Sep 24, 2024 16:14 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

大佬您好，我这边64x64超分到256x256,在rk3588上有150ms左右,6，7帧的样子，请问你那边部署速度怎么样，我看到文章说在arm ethos-n78 npu上，这个模型应该能干到大几百帧，跟我情况完全不一样哇。

---Original--- From: @.> Date: Thu, Sep 19, 2024 10:14 AM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

好的，谢谢提醒。

---Original---
From: @.>
Date: Wed, Sep 18, 2024 17:15 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

嗯嗯，我下载github上sesr训练完生成的onnx和贵公司算法工程师生成的onnx不一样，应该是贵公司算法工程师修改并自定义onnx了，我只能利用您代码上的onnx进行修改，修改了onnx输入和输出尺寸和部分节点，这样控制了图片输入大小。

---Original---
From: @.>
Date: Sat, Sep 14, 2024 16:25 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

不好意思，我目前还不知道。    

---Original---    
From: ***@***.***>    
Date: Sat, Sep 14, 2024 16:11 PM    
To: ***@***.***>;    
Cc: ***@***.******@***.***>;    
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)    

 您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64     

 ---Original---     
 From: ***@***.***>     
 Date: Sat, Sep 14, 2024 15:16 PM     
 To: ***@***.***>;     
 Cc: ***@***.******@***.***>;     
 Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)     

  是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。      

  ---Original---      
  From: ***@***.***>      
  Date: Sat, Sep 14, 2024 15:11 PM      
  To: ***@***.***>;      
  Cc: ***@***.******@***.***>;      
  Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)      

   那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛       

   ---Original---       
   From: ***@***.***>       
   Date: Sat, Sep 14, 2024 15:04 PM       
   To: ***@***.***>;       
   Cc: ***@***.******@***.***>;       
   Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)       

    我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。        

    ---Original---        
    From: ***@***.***>        
    Date: Sat, Sep 14, 2024 14:55 PM        
    To: ***@***.***>;        
    Cc: ***@***.***>;        
    Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)        

    大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support        
     代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx        

    —        
    Reply to this email directly, view it on GitHub, or unsubscribe.        
    You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>        
   —       
   Reply to this email directly, view it on GitHub, or unsubscribe.       
   You are receiving this because you authored the thread.Message ID: ***@***.***>       
  —      
  Reply to this email directly, view it on GitHub, or unsubscribe.      
  You are receiving this because you commented.Message ID: ***@***.***>      
 —     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
 You are receiving this because you authored the thread.Message ID: ***@***.***>     
—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you commented.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.>

xushuanglong commented 3 days ago

好的大佬，请问就是64x64到256x256的rknn哈，我用rk3588测试是114ms左右,您说的rga和dma内存方法用的是其他c++代码吗，您的sesr_rknn_demo.cpp用的是opencv

---Original--- From: @.> Date: Tue, Sep 24, 2024 16:32 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

我测试是至少可以跑到60hz的，但用的时间不固定。有时是8ms，有时是16ms。我用了rga做预处理，cpu做的后处理。内存分配是用的是dma内存。

---Original--- From: @.> Date: Tue, Sep 24, 2024 16:14 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

大佬您好，我这边64x64超分到256x256,在rk3588上有150ms左右,6，7帧的样子，请问你那边部署速度怎么样，我看到文章说在arm ethos-n78 npu上，这个模型应该能干到大几百帧，跟我情况完全不一样哇。

---Original---
From: @.>
Date: Thu, Sep 19, 2024 10:14 AM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

好的，谢谢提醒。

---Original---
From: @.>
Date: Wed, Sep 18, 2024 17:15 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

嗯嗯，我下载github上sesr训练完生成的onnx和贵公司算法工程师生成的onnx不一样，应该是贵公司算法工程师修改并自定义onnx了，我只能利用您代码上的onnx进行修改，修改了onnx输入和输出尺寸和部分节点，这样控制了图片输入大小。    

---Original---    
From: ***@***.***>    
Date: Sat, Sep 14, 2024 16:25 PM    
To: ***@***.***>;    
Cc: ***@***.******@***.***>;    
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)    

 不好意思，我目前还不知道。     

 ---Original---     
 From: ***@***.***>     
 Date: Sat, Sep 14, 2024 16:11 PM     
 To: ***@***.***>;     
 Cc: ***@***.******@***.***>;     
 Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)     

  您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64      

  ---Original---      
  From: ***@***.***>      
  Date: Sat, Sep 14, 2024 15:16 PM      
  To: ***@***.***>;      
  Cc: ***@***.******@***.***>;      
  Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)      

   是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。       

   ---Original---       
   From: ***@***.***>       
   Date: Sat, Sep 14, 2024 15:11 PM       
   To: ***@***.***>;       
   Cc: ***@***.******@***.***>;       
   Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)       

    那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛        

    ---Original---        
    From: ***@***.***>        
    Date: Sat, Sep 14, 2024 15:04 PM        
    To: ***@***.***>;        
    Cc: ***@***.******@***.***>;        
    Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)        

     我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。         

     ---Original---         
     From: ***@***.***>         
     Date: Sat, Sep 14, 2024 14:55 PM         
     To: ***@***.***>;         
     Cc: ***@***.***>;         
     Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)         

     大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support         
      代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx         

     —         
     Reply to this email directly, view it on GitHub, or unsubscribe.         
     You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>         
    —        
    Reply to this email directly, view it on GitHub, or unsubscribe.        
    You are receiving this because you authored the thread.Message ID: ***@***.***>        
   —       
   Reply to this email directly, view it on GitHub, or unsubscribe.       
   You are receiving this because you commented.Message ID: ***@***.***>       
  —      
  Reply to this email directly, view it on GitHub, or unsubscribe.      
  You are receiving this because you authored the thread.Message ID: ***@***.***>      
 —     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
 You are receiving this because you commented.Message ID: ***@***.***>     
—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you authored the thread.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.>

dzyswy commented 3 days ago

是的，我们的工程代码。不能上传的。😅

---Original--- From: @.> Date: Wed, Sep 25, 2024 10:24 AM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

好的大佬，请问就是64x64到256x256的rknn哈，我用rk3588测试是114ms左右,您说的rga和dma内存方法用的是其他c++代码吗，您的sesr_rknn_demo.cpp用的是opencv

---Original--- From: @.> Date: Tue, Sep 24, 2024 16:32 PM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

我测试是至少可以跑到60hz的，但用的时间不固定。有时是8ms，有时是16ms。我用了rga做预处理，cpu做的后处理。内存分配是用的是dma内存。

---Original---
From: @.>
Date: Tue, Sep 24, 2024 16:14 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

大佬您好，我这边64x64超分到256x256,在rk3588上有150ms左右,6，7帧的样子，请问你那边部署速度怎么样，我看到文章说在arm ethos-n78 npu上，这个模型应该能干到大几百帧，跟我情况完全不一样哇。

---Original---
From: @.>
Date: Thu, Sep 19, 2024 10:14 AM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

好的，谢谢提醒。    

---Original---    
From: ***@***.***>    
Date: Wed, Sep 18, 2024 17:15 PM    
To: ***@***.***>;    
Cc: ***@***.******@***.***>;    
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)    

 嗯嗯，我下载github上sesr训练完生成的onnx和贵公司算法工程师生成的onnx不一样，应该是贵公司算法工程师修改并自定义onnx了，我只能利用您代码上的onnx进行修改，修改了onnx输入和输出尺寸和部分节点，这样控制了图片输入大小。     

 ---Original---     
 From: ***@***.***>     
 Date: Sat, Sep 14, 2024 16:25 PM     
 To: ***@***.***>;     
 Cc: ***@***.******@***.***>;     
 Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)     

  不好意思，我目前还不知道。      

  ---Original---      
  From: ***@***.***>      
  Date: Sat, Sep 14, 2024 16:11 PM      
  To: ***@***.***>;      
  Cc: ***@***.******@***.***>;      
  Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)      

   您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64       

   ---Original---       
   From: ***@***.***>       
   Date: Sat, Sep 14, 2024 15:16 PM       
   To: ***@***.***>;       
   Cc: ***@***.******@***.***>;       
   Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)       

    是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。        

    ---Original---        
    From: ***@***.***>        
    Date: Sat, Sep 14, 2024 15:11 PM        
    To: ***@***.***>;        
    Cc: ***@***.******@***.***>;        
    Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)        

     那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛         

     ---Original---         
     From: ***@***.***>         
     Date: Sat, Sep 14, 2024 15:04 PM         
     To: ***@***.***>;         
     Cc: ***@***.******@***.***>;         
     Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)         

      我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。          

      ---Original---          
      From: ***@***.***>          
      Date: Sat, Sep 14, 2024 14:55 PM          
      To: ***@***.***>;          
      Cc: ***@***.***>;          
      Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)          

      大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support          
       代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx          

      —          
      Reply to this email directly, view it on GitHub, or unsubscribe.          
      You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>          
     —         
     Reply to this email directly, view it on GitHub, or unsubscribe.         
     You are receiving this because you authored the thread.Message ID: ***@***.***>         
    —        
    Reply to this email directly, view it on GitHub, or unsubscribe.        
    You are receiving this because you commented.Message ID: ***@***.***>        
   —       
   Reply to this email directly, view it on GitHub, or unsubscribe.       
   You are receiving this because you authored the thread.Message ID: ***@***.***>       
  —      
  Reply to this email directly, view it on GitHub, or unsubscribe.      
  You are receiving this because you commented.Message ID: ***@***.***>      
 —     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
 You are receiving this because you authored the thread.Message ID: ***@***.***>     
—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you commented.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.>

xushuanglong commented 3 days ago

好的，谢谢您的帮助，请问你们还有x2的onnx嘛，只实现x4嘛

---Original--- From: @.> Date: Wed, Sep 25, 2024 11:13 AM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

是的，我们的工程代码。不能上传的。😅

---Original--- From: @.> Date: Wed, Sep 25, 2024 10:24 AM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

好的大佬，请问就是64x64到256x256的rknn哈，我用rk3588测试是114ms左右,您说的rga和dma内存方法用的是其他c++代码吗，您的sesr_rknn_demo.cpp用的是opencv

---Original---
From: @.>
Date: Tue, Sep 24, 2024 16:32 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

我测试是至少可以跑到60hz的，但用的时间不固定。有时是8ms，有时是16ms。我用了rga做预处理，cpu做的后处理。内存分配是用的是dma内存。

---Original---
From: @.>
Date: Tue, Sep 24, 2024 16:14 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

大佬您好，我这边64x64超分到256x256,在rk3588上有150ms左右,6，7帧的样子，请问你那边部署速度怎么样，我看到文章说在arm ethos-n78 npu上，这个模型应该能干到大几百帧，跟我情况完全不一样哇。    

---Original---    
From: ***@***.***>    
Date: Thu, Sep 19, 2024 10:14 AM    
To: ***@***.***>;    
Cc: ***@***.******@***.***>;    
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)    

 好的，谢谢提醒。     

 ---Original---     
 From: ***@***.***>     
 Date: Wed, Sep 18, 2024 17:15 PM     
 To: ***@***.***>;     
 Cc: ***@***.******@***.***>;     
 Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)     

  嗯嗯，我下载github上sesr训练完生成的onnx和贵公司算法工程师生成的onnx不一样，应该是贵公司算法工程师修改并自定义onnx了，我只能利用您代码上的onnx进行修改，修改了onnx输入和输出尺寸和部分节点，这样控制了图片输入大小。      

  ---Original---      
  From: ***@***.***>      
  Date: Sat, Sep 14, 2024 16:25 PM      
  To: ***@***.***>;      
  Cc: ***@***.******@***.***>;      
  Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)      

   不好意思，我目前还不知道。       

   ---Original---       
   From: ***@***.***>       
   Date: Sat, Sep 14, 2024 16:11 PM       
   To: ***@***.***>;       
   Cc: ***@***.******@***.***>;       
   Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)       

    您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64        

    ---Original---        
    From: ***@***.***>        
    Date: Sat, Sep 14, 2024 15:16 PM        
    To: ***@***.***>;        
    Cc: ***@***.******@***.***>;        
    Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)        

     是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。         

     ---Original---         
     From: ***@***.***>         
     Date: Sat, Sep 14, 2024 15:11 PM         
     To: ***@***.***>;         
     Cc: ***@***.******@***.***>;         
     Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)         

      那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛          

      ---Original---          
      From: ***@***.***>          
      Date: Sat, Sep 14, 2024 15:04 PM          
      To: ***@***.***>;          
      Cc: ***@***.******@***.***>;          
      Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)          

       我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。           

       ---Original---           
       From: ***@***.***>           
       Date: Sat, Sep 14, 2024 14:55 PM           
       To: ***@***.***>;           
       Cc: ***@***.***>;           
       Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)           

       大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support           
        代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx           

       —           
       Reply to this email directly, view it on GitHub, or unsubscribe.           
       You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>           
      —          
      Reply to this email directly, view it on GitHub, or unsubscribe.          
      You are receiving this because you authored the thread.Message ID: ***@***.***>          
     —         
     Reply to this email directly, view it on GitHub, or unsubscribe.         
     You are receiving this because you commented.Message ID: ***@***.***>         
    —        
    Reply to this email directly, view it on GitHub, or unsubscribe.        
    You are receiving this because you authored the thread.Message ID: ***@***.***>        
   —       
   Reply to this email directly, view it on GitHub, or unsubscribe.       
   You are receiving this because you commented.Message ID: ***@***.***>       
  —      
  Reply to this email directly, view it on GitHub, or unsubscribe.      
  You are receiving this because you authored the thread.Message ID: ***@***.***>      
 —     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
 You are receiving this because you commented.Message ID: ***@***.***>     
—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you authored the thread.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.>

dzyswy commented 3 days ago

2倍放大的效果和传统算法比差别不大。没有进行移植。

---Original--- From: @.> Date: Wed, Sep 25, 2024 11:18 AM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

好的，谢谢您的帮助，请问你们还有x2的onnx嘛，只实现x4嘛

---Original--- From: @.> Date: Wed, Sep 25, 2024 11:13 AM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

是的，我们的工程代码。不能上传的。😅

---Original---
From: @.>
Date: Wed, Sep 25, 2024 10:24 AM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

好的大佬，请问就是64x64到256x256的rknn哈，我用rk3588测试是114ms左右,您说的rga和dma内存方法用的是其他c++代码吗，您的sesr_rknn_demo.cpp用的是opencv

---Original---
From: @.>
Date: Tue, Sep 24, 2024 16:32 PM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

我测试是至少可以跑到60hz的，但用的时间不固定。有时是8ms，有时是16ms。我用了rga做预处理，cpu做的后处理。内存分配是用的是dma内存。    

---Original---    
From: ***@***.***>    
Date: Tue, Sep 24, 2024 16:14 PM    
To: ***@***.***>;    
Cc: ***@***.******@***.***>;    
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)    

 大佬您好，我这边64x64超分到256x256,在rk3588上有150ms左右,6，7帧的样子，请问你那边部署速度怎么样，我看到文章说在arm ethos-n78 npu上，这个模型应该能干到大几百帧，跟我情况完全不一样哇。     

 ---Original---     
 From: ***@***.***>     
 Date: Thu, Sep 19, 2024 10:14 AM     
 To: ***@***.***>;     
 Cc: ***@***.******@***.***>;     
 Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)     

  好的，谢谢提醒。      

  ---Original---      
  From: ***@***.***>      
  Date: Wed, Sep 18, 2024 17:15 PM      
  To: ***@***.***>;      
  Cc: ***@***.******@***.***>;      
  Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)      

   嗯嗯，我下载github上sesr训练完生成的onnx和贵公司算法工程师生成的onnx不一样，应该是贵公司算法工程师修改并自定义onnx了，我只能利用您代码上的onnx进行修改，修改了onnx输入和输出尺寸和部分节点，这样控制了图片输入大小。       

   ---Original---       
   From: ***@***.***>       
   Date: Sat, Sep 14, 2024 16:25 PM       
   To: ***@***.***>;       
   Cc: ***@***.******@***.***>;       
   Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)       

    不好意思，我目前还不知道。        

    ---Original---        
    From: ***@***.***>        
    Date: Sat, Sep 14, 2024 16:11 PM        
    To: ***@***.***>;        
    Cc: ***@***.******@***.***>;        
    Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)        

     您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64         

     ---Original---         
     From: ***@***.***>         
     Date: Sat, Sep 14, 2024 15:16 PM         
     To: ***@***.***>;         
     Cc: ***@***.******@***.***>;         
     Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)         

      是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。          

      ---Original---          
      From: ***@***.***>          
      Date: Sat, Sep 14, 2024 15:11 PM          
      To: ***@***.***>;          
      Cc: ***@***.******@***.***>;          
      Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)          

       那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛           

       ---Original---           
       From: ***@***.***>           
       Date: Sat, Sep 14, 2024 15:04 PM           
       To: ***@***.***>;           
       Cc: ***@***.******@***.***>;           
       Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)           

        我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。            

        ---Original---            
        From: ***@***.***>            
        Date: Sat, Sep 14, 2024 14:55 PM            
        To: ***@***.***>;            
        Cc: ***@***.***>;            
        Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)            

        大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support            
         代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx            

        —            
        Reply to this email directly, view it on GitHub, or unsubscribe.            
        You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>            
       —           
       Reply to this email directly, view it on GitHub, or unsubscribe.           
       You are receiving this because you authored the thread.Message ID: ***@***.***>           
      —          
      Reply to this email directly, view it on GitHub, or unsubscribe.          
      You are receiving this because you commented.Message ID: ***@***.***>          
     —         
     Reply to this email directly, view it on GitHub, or unsubscribe.         
     You are receiving this because you authored the thread.Message ID: ***@***.***>         
    —        
    Reply to this email directly, view it on GitHub, or unsubscribe.        
    You are receiving this because you commented.Message ID: ***@***.***>        
   —       
   Reply to this email directly, view it on GitHub, or unsubscribe.       
   You are receiving this because you authored the thread.Message ID: ***@***.***>       
  —      
  Reply to this email directly, view it on GitHub, or unsubscribe.      
  You are receiving this because you commented.Message ID: ***@***.***>      
 —     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
 You are receiving this because you authored the thread.Message ID: ***@***.***>     
—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you commented.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.>

xushuanglong commented 3 days ago

好的，再次感谢您的回复

---Original--- From: @.> Date: Wed, Sep 25, 2024 11:21 AM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

2倍放大的效果和传统算法比差别不大。没有进行移植。

---Original--- From: @.> Date: Wed, Sep 25, 2024 11:18 AM To: @.>; Cc: @.**@.>; Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

好的，谢谢您的帮助，请问你们还有x2的onnx嘛，只实现x4嘛

---Original---
From: @.>
Date: Wed, Sep 25, 2024 11:13 AM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)

是的，我们的工程代码。不能上传的。😅

---Original---
From: @.>
Date: Wed, Sep 25, 2024 10:24 AM
To: @.>;
Cc: @.**@.>;
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1) 

好的大佬，请问就是64x64到256x256的rknn哈，我用rk3588测试是114ms左右,您说的rga和dma内存方法用的是其他c++代码吗，您的sesr_rknn_demo.cpp用的是opencv    

---Original---    
From: ***@***.***>    
Date: Tue, Sep 24, 2024 16:32 PM    
To: ***@***.***>;    
Cc: ***@***.******@***.***>;    
Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)    

 我测试是至少可以跑到60hz的，但用的时间不固定。有时是8ms，有时是16ms。我用了rga做预处理，cpu做的后处理。内存分配是用的是dma内存。     

 ---Original---     
 From: ***@***.***>     
 Date: Tue, Sep 24, 2024 16:14 PM     
 To: ***@***.***>;     
 Cc: ***@***.******@***.***>;     
 Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)     

  大佬您好，我这边64x64超分到256x256,在rk3588上有150ms左右,6，7帧的样子，请问你那边部署速度怎么样，我看到文章说在arm ethos-n78 npu上，这个模型应该能干到大几百帧，跟我情况完全不一样哇。      

  ---Original---      
  From: ***@***.***>      
  Date: Thu, Sep 19, 2024 10:14 AM      
  To: ***@***.***>;      
  Cc: ***@***.******@***.***>;      
  Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)      

   好的，谢谢提醒。       

   ---Original---       
   From: ***@***.***>       
   Date: Wed, Sep 18, 2024 17:15 PM       
   To: ***@***.***>;       
   Cc: ***@***.******@***.***>;       
   Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)       

    嗯嗯，我下载github上sesr训练完生成的onnx和贵公司算法工程师生成的onnx不一样，应该是贵公司算法工程师修改并自定义onnx了，我只能利用您代码上的onnx进行修改，修改了onnx输入和输出尺寸和部分节点，这样控制了图片输入大小。        

    ---Original---        
    From: ***@***.***>        
    Date: Sat, Sep 14, 2024 16:25 PM        
    To: ***@***.***>;        
    Cc: ***@***.******@***.***>;        
    Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)        

     不好意思，我目前还不知道。         

     ---Original---         
     From: ***@***.***>         
     Date: Sat, Sep 14, 2024 16:11 PM         
     To: ***@***.***>;         
     Cc: ***@***.******@***.***>;         
     Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)         

      您也不清楚onnx是怎么转换的嘛，是的，onnx的输入被设置了64x64,因此rknn就支持64x64          

      ---Original---          
      From: ***@***.***>          
      Date: Sat, Sep 14, 2024 15:16 PM          
      To: ***@***.***>;          
      Cc: ***@***.******@***.***>;          
      Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)          

       是我们公司的算法工程师帮我转换的。模型的原始输入确实是64x64的。但是，用ncnn进行量化运行推理，可以支持任意分辨率。用rknn就只支持固定64x64。           

       ---Original---           
       From: ***@***.***>           
       Date: Sat, Sep 14, 2024 15:11 PM           
       To: ***@***.***>;           
       Cc: ***@***.******@***.***>;           
       Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)           

        那可以请问您的onnx是如何生成的嘛，我生成的onnx不能转成rknn模型，我用的sesr算法是tensorflow框架，您也用的是这个框架的嘛            

        ---Original---            
        From: ***@***.***>            
        Date: Sat, Sep 14, 2024 15:04 PM            
        To: ***@***.***>;            
        Cc: ***@***.******@***.***>;            
        Subject: Re: [dzyswy/sesr] Onnx conversion code (Issue #1)            

         我也注意到这个问题了，但我也正在入门😅。ncnn的量化不存在这个问题，可以自适应分辨率。目前，我还不知道怎么解决，可能需要研究rknn的框架的资料。             

         ---Original---             
         From: ***@***.***>             
         Date: Sat, Sep 14, 2024 14:55 PM             
         To: ***@***.***>;             
         Cc: ***@***.***>;             
         Subject: [dzyswy/sesr] Onnx conversion code (Issue #1)             

         大佬您好，您的模型貌似只支持64X64大小的图片，我想我的模型能够支持其他尺寸大小的图片，比如400X400,我应该怎么写呢，而且好像要删除部分算子。我直接代码生成onnx转rknn有问题，比如‘unk__244’,'unk__245'等 not support             
          代码： python -m tf2onnx.convert --saved-model logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32 --opset 11 --output logs/x2_models/SESR_m5_f16_x2_fs400_collapsedTraining_FP32/model.onnx             

         —             
         Reply to this email directly, view it on GitHub, or unsubscribe.             
         You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>             
        —            
        Reply to this email directly, view it on GitHub, or unsubscribe.            
        You are receiving this because you authored the thread.Message ID: ***@***.***>            
       —           
       Reply to this email directly, view it on GitHub, or unsubscribe.           
       You are receiving this because you commented.Message ID: ***@***.***>           
      —          
      Reply to this email directly, view it on GitHub, or unsubscribe.          
      You are receiving this because you authored the thread.Message ID: ***@***.***>          
     —         
     Reply to this email directly, view it on GitHub, or unsubscribe.         
     You are receiving this because you commented.Message ID: ***@***.***>         
    —        
    Reply to this email directly, view it on GitHub, or unsubscribe.        
    You are receiving this because you authored the thread.Message ID: ***@***.***>        
   —       
   Reply to this email directly, view it on GitHub, or unsubscribe.       
   You are receiving this because you commented.Message ID: ***@***.***>       
  —      
  Reply to this email directly, view it on GitHub, or unsubscribe.      
  You are receiving this because you authored the thread.Message ID: ***@***.***>      
 —     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
 You are receiving this because you commented.Message ID: ***@***.***>     
—    
Reply to this email directly, view it on GitHub, or unsubscribe.    
You are receiving this because you authored the thread.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: @.>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.>