updates

Author: matrn

README.md

@@ -36,6 +36,7 @@ Library supports many operations but more development is needed because currentl
 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
+![GUI](./data/assets/C-digit-recognition-example-of-usage.jpg)
 
 ## Control
 - `left mouse button & drag` to draw
@@ -67,16 +68,11 @@ Even though Python is much slower than C, Python-digit-recognition is faster. Th
 # ToDo
 - [ ] 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
-- [x] Write everything into README
 - [ ] Choose license
-- [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
@@ -87,4 +83,5 @@ Even though Python is much slower than C, Python-digit-recognition is faster. Th
 - [MNIST database](http://yann.lecun.com/exdb/mnist/)
 - [George Hotz's AI notebook](https://github.com/geohot/ai-notebooks/blob/master/mnist_from_scratch.ipynb)
 - [Michael Nielsen's book about neural networks and deep learning](http://neuralnetworksanddeeplearning.com/chap2.html)
-- [Artcile about backpropagation from Brilliant.org](https://brilliant.org/wiki/backpropagation/)
+- [Artcile about backpropagation from Brilliant.org](https://brilliant.org/wiki/backpropagation/)
+- [Epoch vs Batch Size vs Iterations](https://towardsdatascience.com/epoch-vs-iterations-vs-batch-size-4dfb9c7ce9c9)

build/usr/bin/crecog

@@ -527,9 +527,9 @@ 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;
-	dbgexec(double test_accuracy = MSB_2bytes_to_int16(buf)/100.0);
+	double test_accuracy = MSB_2bytes_to_int16(buf)/100.0;
 
-	dbg("Network was tested with accuracy: %.2f %%\n", test_accuracy);
+	printf("Network was tested with accuracy: %.2f %%\n", test_accuracy);
 
 	// weights and biases
 	for(int i = 0; i < nn_size; i ++){

data/assets/C-digit-recognition-example-of-usage.jpg

@@ -1001,8 +1001,8 @@ matrix_size_t matrix_1ubyteMat_mean(const uint8_t* mat, const matrix_size_t rows
 			break;
 	}
 
-	printf("m_sum: %d\n", m_sum);
-	printf("mr_sum: %d\n", mr_sum);
+	//printf("m_sum: %d\n", m_sum);
+	//printf("mr_sum: %d\n", mr_sum);
 	return round((double)mr_sum / m_sum);
 }
 

lib/ceural/ceural.c

@@ -32,15 +32,15 @@ static void gui_display_results(const uint8_t results[], const double accuracies
 	sprintf(text, "%d", results[0]);
 	gtk_text_buffer_insert_with_tags_by_name(result_buffer, &iter_start, text, -1, "result_format", "top_result", "green", NULL);
 
-	sprintf(text, ", accuracy: %.2f %%\n", accuracies[0]);
+	sprintf(text, ", confidence: %.1f %%\n", accuracies[0]);
 	gtk_text_buffer_insert_with_tags_by_name(result_buffer, &iter_start, text, -1, "result_format", "top_result", NULL);
 
 	// 2. result
-	sprintf(text, "2.Result: %d, accuracy: %.2f %%\n", results[1], accuracies[1]);
+	sprintf(text, "2.Result: %d, confidence: %.1f %%\n", results[1], accuracies[1]);
 	gtk_text_buffer_insert_with_tags_by_name(result_buffer, &iter_start, text, -1, "result_format", "second_result", NULL);
 
 	// 3. result
-	sprintf(text, "3.Result: %d, accuracy: %.2f %%\n", results[2], accuracies[2]);
+	sprintf(text, "3.Result: %d, confidence: %.1f %%\n", results[2], accuracies[2]);
 	gtk_text_buffer_insert_with_tags_by_name(result_buffer, &iter_start, text, -1, "result_format", "third_result", NULL);
 }
 
@@ -105,7 +105,7 @@ static void recognise_input(GtkWidget *widget, gpointer data) {
 
 	img = gdk_pixbuf_to_uint8_grayscale(pixbuf);  // convert image to the grayscale img
 
-	matrix_1ubyteMat_display(img, 20, 20);
+	//matrix_1ubyteMat_display(img, 20, 20);
 	int row_center = matrix_1ubyteMat_mean(img, 20, 20, ROW_AXIS);
 	int col_center = matrix_1ubyteMat_mean(img, 20, 20, COL_AXIS);
 	printf("Center: %dx%d\n", row_center, col_center);
@@ -121,7 +121,7 @@ static void recognise_input(GtkWidget *widget, gpointer data) {
 
 	puts("Move done");
 	dbg("new_size: %dx%d\n", img_final_rows, img_final_cols);
-	dbgexec(matrix_1ubyteMat_display(img_final, img_final_rows, img_final_cols));
+	matrix_1ubyteMat_display(img_final, img_final_rows, img_final_cols);
 
 	pixbuf = uint8_grayscale_to_gdk_pixbuf(img_final, img_final_rows, img_final_cols);
 	gui_display_image(pixbuf);	// show preprocessed image in result window
@@ -130,15 +130,15 @@ static void recognise_input(GtkWidget *widget, gpointer data) {
 	nn_recognise(img_final, out_mat);
 
 	results[0] = matrix_argmax(out_mat);
-	accuracies[0] = matrix_max(out_mat);
+	accuracies[0] = matrix_max(out_mat)*100;
 	*matrix_at_index(out_mat, results[0]) = -1;
 
 	results[1] = matrix_argmax(out_mat);
-	accuracies[1] = matrix_max(out_mat);
+	accuracies[1] = matrix_max(out_mat)*100;
 	*matrix_at_index(out_mat, results[1]) = -1;
 
 	results[2] = matrix_argmax(out_mat);
-	accuracies[2] = matrix_max(out_mat);
+	accuracies[2] = matrix_max(out_mat)*100;
 	*matrix_at_index(out_mat, results[2]) = -1;
 
 	dbg("RESULT: %d\n", results[0]);