updates & cleanup & refactoring

matrn · matrn · commit ad71c5d07f37 · 2021-08-10T17:36:26.000+02:00
diff --git a/.gitignore b/.gitignore
@@ -50,3 +50,6 @@ modules.order
 Module.symvers
 Mkfile.old
 dkms.conf
+
+
+C_digit_recognition*.AppImage
diff --git a/C_digit_recognition-78b7798-x86_64.AppImage b/C_digit_recognition-78b7798-x86_64.AppImage
diff --git a/Makefile b/Makefile
@@ -5,6 +5,10 @@ install_tools:
 	sudo apt install patchelf
 	sudo apt install libopenblas-dev
 
+	#sudo apt install doxygen
+	#sudo apt install graphviz
+	#sudo npm install moxygen -g
+
 
 release:
 	cd ./src && $(MAKE) main
@@ -26,3 +30,4 @@ clean:
 	rm -f ./build/usr/lib/*
 	rm -f ./build/usr/bin/*
 	rm -f ./build/AppRun
+	rm -f C_digit_recognition-*.AppImage 
diff --git a/README.md b/README.md
@@ -1,16 +1,20 @@
 # C-digit-recognition
-Handwritten digits recognition in C using neural network trained using [MNIST database](http://yann.lecun.com/exdb/mnist/)
+Handwritten digits recognition in C using neural network trained using [MNIST database](http://yann.lecun.com/exdb/mnist/).
 
 
 # Installation
 
 ## Linux
-- `sudo apt install libopenblas-dev`
-- `sudo apt install libgtk-3-dev`
-- `make`
 
-## Windows
-🤷‍♂️
+### Binary package
+`.AppImage` file is available under release section.
+
+### Compilation from source
+- `sudo apt install libopenblas-dev` - installs OpenBLAS library
+- `sudo apt install libgtk-3-dev` - installs
+- compilation:
+	- use `cd src && make clean && make main && ./main` to run normal compilation
+	- use `make clean && make release` to generate AppImage binary (you have to install `linuxdeploy` and other dependecies using `make install_tools` first)
 
 
 # Libraries
@@ -19,12 +23,17 @@ Dependency of the libraries is in this order: `GUI -> ceural -> lag`
  - `ceural` - [C neural network library](#Ceural)
 
 ## Lag
+Library supports many operations but more development is needed because currently uses [OpenBLAS](https://github.com/xianyi/OpenBLAS) only for matrix multiplication and matrix transposition.
+
+### Naming
 - `mat` - stands for matrix
 - `ew` - stands for element wise
+
+### Notes
 - Matrix part of the library automatically checks if destination and source is same where shouldn't be same and warns using `assert()`.
 
 ## Ceural
-Ceural library is created for multi-layer networks for MNIST dataset but with small modifications it can be used for other datasets.
+Ceural library is created for multi-layer networks trained using MNIST dataset but with small modifications it can be used for other datasets too. See [Accuracy](#Accuracy) for more info.
 
 # GUI
 
@@ -46,26 +55,28 @@ Preprocessing used in [MNIST](http://yann.lecun.com/exdb/mnist/) database: *The
 8. **neural network forward propagation** - this preprocessed image is fed to the neural network
 
 # Accuracy
-After training the test set accuracy is `94.49 %`. This accuracy not bad considering the test error rate in [MNIST database website](http://yann.lecun.com/exdb/mnist/) of the 2-layer NN. Sadly accuracy is not as good in practice as it's in the test data set.
+After `10` epochs of training with batch size `32` the test set accuracy is `97.47 %` which is not bad considering the test error rate in [MNIST database website](http://yann.lecun.com/exdb/mnist/) of the 2-layer NN. Sadly accuracy is not as good in practice as it's in the test data set 🥺.
 
 Accuracy is calculated using formula [`accuracy = (TP+TN)/(TP+TN+FP+FN)`](https://en.wikipedia.org/wiki/Accuracy_and_precision) which is `accuracy = correct/total`
 
 
 # Performance
-Even though Python is much slower than C, Python-digit-recognition is much faster. The reason behind it is that Python version uses great library NumPy, which is written in C and uses BLAS implementation of matrix operations.
+Even though Python is much slower than C, Python-digit-recognition is faster. The reason behind it is that Python version uses great library NumPy, which is perfectly optimized.
 
 
 # ToDo
-- [ ] Cleanup & document `lag` library
-- [ ] Cleanup & document `ceural` library
+- [ ] add `lag` tests
+- [ ] add `ceural` tests
+- [x] Cleanup & document `lag` library
+- [x] Cleanup & document `ceural` library
 - [ ] Cleanup & document `gui`
 - [ ] Add icons into `gui`
 - [ ] Add command line options to train & test & save & load NN
 - [x] Check NN definition during the NN weights & biases load process
 - [ ] Create documentation
-- [ ] Write everything into README
+- [x] Write everything into README
 - [ ] Choose license
-- [ ] Finish top-level 'Makefile' to create final binary for linux
+- [x] Finish top-level 'Makefile' to create final binary for linux
 - [ ] Create Windows compilation script & test it on Windows
 - [x] Center digit by center of mass of the pixels before feeding it to the neural network from GUI input
 - [ ] Use BLAS (for example [OpenBLAS](https://github.com/xianyi/OpenBLAS)) library for linear algebra to improve speed
diff --git a/build/.gitignore b/build/.gitignore
@@ -1,3 +1,5 @@
 *.AppImage
 AppRun
-.DirIcon
+.DirIcon
+C-digit-recognition.desktop
+C-digit-recognition_icon.png
diff --git a/build/AppRun b/build/AppRun
diff --git a/build/C-digit-recognition.desktop b/build/C-digit-recognition.desktop
diff --git a/build/C-digit-recognition_icon.png b/build/C-digit-recognition_icon.png
diff --git a/build/usr/bin/crecog b/build/usr/bin/crecog
diff --git a/data/data.ceural b/data/data.ceural
diff --git a/lib/ceural/Makefile b/lib/ceural/Makefile
@@ -1,4 +1,4 @@
-CFLAGS+=-g -pedantic -Wall -Werror -O2 -fomit-frame-pointer -DDEBUG #-DCEURAL_DISABLE_EMBEDDED_TRAINED_MODEL
+CFLAGS+=-g -pedantic -Wall -Werror -O2 -fomit-frame-pointer #-DDEBUG #-DCEURAL_DISABLE_EMBEDDED_TRAINED_MODEL
 RESOURCES_FLAGS+=-Wl,--format=binary -Wl,data.ceural -Wl,--format=default   # used for including trained NN model into binary
 LDFLAGS+=-lm -llag ${RESOURCES_FLAGS} -Wl,-R../../obj
 
diff --git a/lib/ceural/ceural.c b/lib/ceural/ceural.c
@@ -527,7 +527,7 @@ ceural_rtn ceural_net_load_from_file(ceural_net_t * nn, const char * filename){
 
 	// train accuracy
 	if(fread(&buf, 2, 1, f) != 1) return CEURAL_PARSE_ERROR;
-	double test_accuracy = MSB_2bytes_to_int16(buf)/100.0;
+	dbgexec(double test_accuracy = MSB_2bytes_to_int16(buf)/100.0);
 
 	dbg("Network was tested with accuracy: %.2f %%\n", test_accuracy);
 
diff --git a/lib/ceural/include/ceural.h b/lib/ceural/include/ceural.h
@@ -12,6 +12,8 @@
 //#define CEURAL_IGNORE_NN_INCOMPATIBILITY  // used when loading weights & biases from file
 //#define CEURAL_DISABLE_EMBEDDED_TRAINED_MODEL   // uncomment to disable embedded trained model load
 
+
+
 #ifndef CEURAL_DISABLE_EMBEDDED_TRAINED_MODEL
 extern uint8_t data_ceural[] asm("_binary_data_ceural_start");
 extern uint8_t data_ceural_end[] asm("_binary_data_ceural_end");
diff --git a/lib/ceural/main.c b/lib/ceural/main.c
@@ -63,14 +63,14 @@ int main(){
 	cstop(forward_phase);
 
 	puts("Loading NN...");
-	//printf("load rtn: %d\n", ceural_net_load_from_file(&nn, "data.ceural"));  // NULL));
-
+	printf("load rtn: %d\n", ceural_net_load_from_file(&nn, "data.ceural"));  // NULL));
+	
 	puts("Testing...");
 	ceural_net_test(&nn, &test_set);
 	
 
 	//ceural_net_backpropagate(&nn, train_images.data[0], train_labels.data[0]);
-	ceural_net_train(&nn, &train_set, 1, 32);
+	ceural_net_train(&nn, &train_set, 5, 32);
 
 	
 	ceural_net_forward(&nn, out, train_images.data[0]);
diff --git a/lib/lag/Makefile b/lib/lag/Makefile
@@ -1,5 +1,5 @@
-CFLAGS+=-g -pedantic -Wall -Werror -DDEBUG #-I /opt/OpenBLAS/include/ -L /opt/OpenBLAS/lib/
-LDFLAGS+=-lm -lopenblas #-Wl,-R/opt/OpenBLAS/lib/
+CFLAGS+=-g -pedantic -Wall -Werror #-DDEBUG 
+LDFLAGS+=-lm -lopenblas
 
 lag.o: matrix.c
 	echo "lag.o"
diff --git a/lib/lag/include/debug.h b/lib/lag/include/debug.h
@@ -4,6 +4,9 @@
 #include <stdio.h>
 #include <time.h>
 
+//#define DEBUG
+//#define TM_DEBUG
+
 
 #ifdef DEBUG
 	#define dbg(fmt, ...) printf(fmt, __VA_ARGS__)
@@ -24,7 +27,6 @@
 #endif
 
 
-//#define TM_DEBUG
 
 typedef clock_t dbg_tm;
 
diff --git a/lib/lag/include/lag.h b/lib/lag/include/lag.h
@@ -17,6 +17,6 @@
 
 
 void random_seed(unsigned int seed);
-
+void blas_set_num_threads(int num);
 
 #endif
diff --git a/lib/lag/include/matrix.h b/lib/lag/include/matrix.h
@@ -18,13 +18,15 @@
 #define MATRIX_USE_BLAS
 
 
+
 #ifdef MATRIX_USE_DOUBLE
 	#define MATRIX_TYPE double
 #endif
 #ifdef MATRIX_USE_FLOAT
 	#define MATRIX_TYPE float
 	#define cblas_dgemm cblas_sgemm
 	#define cblas_dimatcopy cblas_simatcopy
+	#define cblas_domatcopy cblas_somatcopy
 #endif
 
 
diff --git a/lib/lag/lag.c b/lib/lag/lag.c
@@ -7,4 +7,8 @@
  */
 void random_seed(unsigned int seed){
 	srand(seed);
+}
+
+void blas_set_num_threads(int num){
+	openblas_set_num_threads(num);
 }
diff --git a/lib/lag/matrix.c b/lib/lag/matrix.c
@@ -629,7 +629,7 @@ matrix_rtn matrix_multiply_r1ubyteMat(matrix_t* out, matrix_t* a, uint8_t* b, ma
 		int N = cols;
 		int K = a->c;
 		MATRIX_TYPE *b_cp  = malloc(rows*cols*sizeof(MATRIX_TYPE));
-		for(int i = 0; i < rows*cols; i ++) b_cp[i] = b[i];
+		for(int i = 0; i < rows*cols; i ++) b_cp[i] = b[i];   // it's awful but can it be better?
 		cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, M, N, K, 1.0, a->data, K, b_cp, N, 0.0, out->data, N);
 		free(b_cp);
 	#else	
