Merge pull request #23 from majianjia/dev

Change nnom_predic() interfaces

docs/api_evaluation.md

@@ -52,19 +52,21 @@ Total Memory cost (Network and NNoM): 32876
 
 ---
 
-## nnom_predic_one()
+## nnom_predic()
 
 ~~~C
-int32_t nnom_predic_one(nnom_model_t *m);
+int32_t nnom_predic(nnom_model_t *m, uint32_t *label, float *prob);
 ~~~
 
-To predict one set of input data. This is the standalone API which does not require `printf()` to print results but only return the predicted label. 
+A standalone evaluation method, run single prodiction, return probability and top-1 label. 
 
 This method is basicly `model_run()` + `index(top-1)`
 
 **Arguments**
 
 - **m:** the model to run prediction (evaluation).
+- **label:** the variable to store top-1 label.
+- **prob:** the variable to store probability. Range from 0~1.
 
 **Return**
 

docs/example_mnist_simple_cn.md

@@ -123,6 +123,8 @@ prediction start..
 Time: 62 tick
 Truth label: 8
 Predicted label: 8
+Probability: 100%
+
 ~~~
 
 额，如果恶心到你了，那我道歉...
@@ -134,6 +136,7 @@ Predicted label: 8
 - 此次预测的时间，这里用了 `62 tick`，我这是相当于 62ms
 - 这张图片的真实数字是 `8`
 - 网络计算的这张照片的数字 `8`
+- 可能性是100%
 
 赶快去试试，其他的 9 张图片吧。
 

docs/index.md

@@ -38,9 +38,15 @@ Nowadays, neural networks are **wider**, **deeper**, and **denser**.
 >[3] Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4700-4708).
 
 
-**If you would like to try those more up-to-date, decent and complex structures on MCU** 
+After 2014, the development of Neural Networks are more focus on structure optimising to improve efficiency and performance, which is more important to the small footprint platforms such as MCUs. 
+However, the available NN libs for MCU are too low-level which make it sooooo difficult to use with these complex strucures. 
+Therefore, we build NNoM to help developers to manage the structures, memory and parameters, even with the automatic tools for faster and simpler deploying.
 
-**NNoM can help you to build them with only a few lines of C codes**, same as you did with Python in [**Keras**](https://keras.io/)
+Now with NNoM, you are free to play with these more up-to-date, decent and complex structures on MCU.
+
+With [**Keras**](https://keras.io/) and our tools, deploying a model only takes a few line of codes.
+
+Please do check the [examples](https://github.com/majianjia/nnom/tree/master/examples).
 
 ---
 

examples/mnist-simple/mcu/main.c

@@ -53,7 +53,8 @@ void print_img(int8_t * buf)
 void mnist(int argc, char** argv)
 {
 	uint32_t tick, time;
-	int32_t result;
+	uint32_t predic_label;
+	float prob;
 	int32_t index = atoi(argv[1]);
 
 	if(index >= TOTAL_IMAGE || argc != 2)
@@ -67,15 +68,16 @@ void mnist(int argc, char** argv)
 
 	// copy data and do prediction
 	memcpy(nnom_input_data, (int8_t*)&img[index][0], 784);
-	result = nnom_predic_one(model);
+	nnom_predic(model, &predic_label, &prob);
 	time = rt_tick_get() - tick;
 
 	//print original image to console
 	print_img((int8_t*)&img[index][0]);
 
 	printf("Time: %d tick\n", time);
 	printf("Truth label: %d\n", label[index]);
-	printf("Predicted label: %d\n", result);
+	printf("Predicted label: %d\n", predic_label);
+	printf("Probability: %d%%\n", (int)(prob*100));
 }
 
 FINSH_FUNCTION_EXPORT(mnist, mnist(4) );

inc/nnom_utils.h

@@ -73,7 +73,7 @@ void prediction_summary(nnom_predic_t *pre);
 // this api test one set of data, return the prediction
 // return the predicted label
 // return NN_ARGUMENT_ERROR if parameter error
-int32_t nnom_predic_one(nnom_model_t *m); 
+nnom_status_t nnom_predic(nnom_model_t *m, uint32_t *label, float *prob);
 
 void model_stat(nnom_model_t *m);
 

mkdocs.yml

@@ -13,8 +13,8 @@ nav:
   - Merges: 'api_merge.md'
   - Evaluation: 'api_evaluation.md'
 - Legacy Guide:
-  - A Temporary Guide: 'A Temporary Guide to NNoM.md'
-  - Porting and Optimisation Guide: 'Porting and Optimisation Guide.md'
+  - A Temporary Guide: 'A_Temporary_Guide_to_NNoM.md'
+  - Porting and Optimisation Guide: 'Porting_and_Optimisation_Guide.md'
   - Old Readme: 'legacy_README.md'
 - 中文: 
   - RT-Thread: 'rt-thread_guide.md'

src/nnom_utils.c

@@ -223,9 +223,9 @@ void prediction_summary(nnom_predic_t *pre)
 
 // stand alone prediction API
 // this api test one set of data, return the prediction
-int32_t nnom_predic_one(nnom_model_t *m)
+nnom_status_t nnom_predic(nnom_model_t *m, uint32_t *label, float *prob)
 {
-	int32_t max_val, max_index;
+	int32_t max_val, max_index, sum;
 	int8_t *output;
 
 	if (!m)
@@ -239,15 +239,21 @@ int32_t nnom_predic_one(nnom_model_t *m)
 	// Top 1
 	max_val = output[0];
 	max_index = 0;
+	sum = max_val;
 	for (uint32_t i = 1; i < shape_size(&m->tail->out->shape); i++)
 	{
 		if (output[i] > max_val)
 		{
 			max_val = output[i];
 			max_index = i;
 		}
+		sum += output[i];
 	}
-	return max_index;
+	// send results
+	*label = max_index;
+	*prob  = (float)max_val/(float)sum; 
+
+	return NN_SUCCESS;
 }
 
 static void layer_stat(nnom_layer_t *layer)