im2col | Python implemention to illustrated im2col

My intention is to do preparation of sound wave file, arrange train process, and test process via sound.c. ran into error during compiling darknet. need your help!

make: gcc: command not found Makefile: 175: recipe for target 'obj/sound.o' failed make: *** [obj/sound.o] Error 127 UBUNTU LTS 18.04 CUDA 11.1

@wbcalex-desktop:~$ sudo apt install gcc [sudo] password for wbcalex: Reading package lists... Done Building dependency tree
Reading state information... Done gcc is already the newest version (4:7.4.0-1ubuntu2.3). The following package was automatically installed and is no longer required: linux-hwe-5.4-headers-5.4.0-47 Use 'sudo apt autoremove' to remove it. 0 upgraded, 0 newly installed, 0 to remove and 0 not upgraded. desktop:~$ cd darknet_tmp desktop:~/darknet_tmp$ make gcc -Iinclude/ -I3rdparty/stb/include -DGPU -I/usr/local/cuda/include/ -DCUDNN -Wall -Wfatal-errors -Wno-unused-result -Wno-unknown-pragmas -fPIC -Ofast -DGPU -DCUDNN -I/usr/local/cudnn/include -fPIC -c ./src/sound.c -o obj/sound.o make: gcc: Command not found Makefile:175: recipe for target 'obj/sound.o' failed make: *** [obj/sound.o] Error 127

I'm making 'conditional mini pooling'.

However, applying this to Keras' CNN results in the following error:

TypeError: The added layer must be an instance of class Layer. Found: <main.MiniPooling2D object at 0x0000020E0BDD8B00>

it is my condition minipooling

MiniPooling

I'm trying to train YOLOv4 with darknet on a computing cluster. But when I make the darknet, it occured that:

I am writing to request guidance in optimizing my solution / method "CalculateConvolutionOutputTensor__im2col". I would like help determining the best strategy for moving beyond my naive approach; offerings of intuition about any relevant GPU processes and how they apply (e.g., bank conflicts); and help interpreting the above profile in terms of what I can tweak.

A first run of the method takes 0.774 seconds using a GeForce 2080 Ti. I have included a screenshot of the Nsight Compute profile of the only CUDA C++ kernel I have written: im2col.

I could have each GPU thread access shared memory instead of global memory. I could transfer GPU "heap" variables to kernel "stack" instead of dereferencing for every thread and in-kernel for-loop iteration. I could put small parameters into arrays in GPU memory and pass single pointers to those arrays. I could use a more sophisticated version of im2col.

I would prefer not to use cuDNN 7.6.5; when I use cuDNN 7.6.5 and write the statement "cudnnCreate(&cudnnHandle);", Nsight Compute suggests that method cuModuleGetFunction returns CUDA_ERROR_NOT_FOUND.

The procedure I used to create this project was to create a new CUDA 10.2 Runtime project using Visual Studio Community 2019, rename the default source file to "main.cu", replace all contents with the first code block below, add "CalculateConvolutionOutputTensor__im2col.h" to my project, add the second code block below, add "CalculateConvolutionOutputTensor__im2col.cu" to my project, add the third code block below, and add "cublas.lib;" to Project Properties -> Linker -> Input -> Additional Dependencies.