DeepCnn 2025

C++ Libtorch based Interface for creation of convolutional neural network, train and test image classification, without code
Generation of equivalent python code automatically done using Mistral AI (need codestral API key to be stored in mistral_api_key.txt)

Third party dll are available here

Download compiled version of Opencv version 4.10 (file opencv_world4100.dll)
https://opencv.org/blog/opencv-4-10-0/
run opencv installation, and copy opencv_world4100.dll in the directory of DeepCnn2025.exe

Libtorch 2.6 for windows, with cuda 12.6

https://download.pytorch.org/libtorch/cu126/libtorch-win-shared-with-deps-2.6.0%2Bcu126.zip

Copy all DLL files (from Lib subdirectory of libtorch) in the same directory as DeepCnn2025.exe

#Other dependencies : curl json (nlohmann) : installation is easy using vcpkg

Install visual c++ 2022 redistribuable

https://learn.microsoft.com/en-us/cpp/windows/latest-supported-vc-redist?view=msvc-170

Run DeepCnn2025.exe

Run configuration to create a network structure and save it (.configCNN file)

Train the network selecting configCNN file and data directory structure including train and test subdirectories as follows
(see sampledata.zip provided)

directory\train\classe1\image1.jpg
directory\train\classe1\image2.jpg
...
directory\train\classe2\image1.jpg
directory\train\classe2\image1.jpg directory\test\classe1\image1.jpg
directory\test\classe1\image2.jpg
...
directory\test\classe2\image1.jpg
directory\test\classe2\image1.jpg

Wait for training completion (press F5 to stop the process)

Save training result as .pt file

Test image classification selecting one of the test image.

or multiclass image segmentation

It is possible to classify one single image or to create/label patches from images in order to create a train/test directory structure Choose the blob size
Clic on Select image and record patches - select saving folder.
Open the training image containing objects to be localized.
Left clic on several patches of the background (class 0)

Right clic to increment the class number, then left clic on the patches of the current class

Repeat the process for several classes. Close the training image.
Clic on "Create train/test directories" and select the folder where you stored the patches.
Close the Classification test dialog box and train your CNN.
Open Classification test, choose the blob size and the image to be segmented.
Segmented image (_segmented.png) and list of detected patches (csv file) are automatically saved.