NNoM has provide a few evaluation interfaces. Thye can either do runtime statistic or model evaluations.
These API are print though the standard printf(), thus a terminal/console is needed.
All these API must not be called before the model has been compiled.
void model_stat(nnom_model_t *m);To print the runtime statistic of the last run. Check the below example for the availble statistics.
Arguments
- ** m:** the model to print.
Notes
It is recommended to run the mode once after compiling to gain these runtime statistic.
Example
Print running stat..
Layer(#) - Time(us) ops(MACs) ops/us
--------------------------------------------------------
#1 Input - 9 0
#2 Conv2D - 8292 172800 20.83
#3 MaxPool - 5753 0
#4 Conv2D - 50095 3612672 72.11
#5 MaxPool - 3617 0
#6 Conv2D - 35893 2654208 73.94
#7 MaxPool - 1206 0
#8 Conv2D - 4412 171072 38.77
#9 GL_SumPool - 30 0
#10 Softmax - 2 0
#11 Output - 0 0
Summary.
Total ops (MAC): 6610752
Prediction time :109309us
Efficiency 60.47 ops/us
Total Memory cost (Network and NNoM): 32876
void model_io_format(nnom_model_t *m);To print the Qformat of input and output tensor, aswell as the output shape.
Arguments
- ** m:** the model to print.
Notes
Qformat: integer bit.decimal bit
Example
Print layer input/output..
Layer(#) - Input(Qnm) Output(Qnm) Oshape
----------------------------------------------------------
#1 Input - 0. 7 0. 7 ( 28, 28, 1,)
#2 Conv2D - 0. 7 3. 4 ( 26, 26, 16,)
#3 Conv2D - 3. 4 3. 4 ( 24, 24, 32,)
#4 MaxPool - 3. 4 3. 4 ( 12, 12, 32,)
#5 Conv2D - 3. 4 3. 4 ( 10, 10, 64,)
#6 Conv2D - 3. 4 3. 4 ( 8, 8, 64,)
#7 MaxPool - 3. 4 3. 4 ( 4, 4, 64,)
#8 Flatten - 3. 4 3. 4 (1024, )
#9 Dense - 3. 4 4. 3 ( 64, )
#10 Dense - 4. 3 5. 2 ( 10, )
#11 Softmax - 5. 2 0. 7 ( 10, )
#12 Output - 0. 7 0. 7 ( 10, )
int32_t nnom_predict(nnom_model_t *m, uint32_t *label, float *prob);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
- The predicted label in digit.
- Error codes if model is failed to run.
Note
The input buffer of the model must be feeded before calling this method.
nnom_predict_t *prediction_create(nnom_model_t *m, int8_t *buf_prediction,
size_t label_num, size_t top_k_size);This method create a prediction instance, which record mutiple parameters in the evaluation process.
Arguments
- m: the model to run prediction (evaluation).
- buf_prediction: the output buffer of the model, which should be the output of Softmax. Size equal to the size of class.
- label_num: the number of labels (the number of classifications).
- top_k_size: the Top-k that wants to evaluate.
Return
- The prediction instance.
Note
Check later examples.
nnom_status_t prediction_run(nnom_predict_t *pre,
uint32_t true_label, uint32_t* predict_label, float* prob)To run a prediction with the new data (feeded by user to the input_buffer which passed to Input layer).
Arguments
- pre: the prediction instance created by
prediction_create(). - true_label: true label of this data.
- predict_label: the predicted label of this data (top-1 results).
- prob: the probability of this label.
Return
- nnom_status_t
void prediction_end(nnom_predict_t *pre);To mark the prediction has done.
Arguments
- pre: the prediction instance created by
prediction_create().
void predicetion_delete(nnom_predict_t *pre);To free all resources.
Arguments
- pre: the prediction instance created by
prediction_create().
void prediction_matrix(nnom_predict_t *pre);To print a confusion matrix when the prediction is done.
Arguments
- pre: the prediction instance created by
prediction_create().
Example
Confusion matrix:
predic 0 1 2 3 4 5 6 7 8 9 10
actual
0 | 395 1 0 0 2 0 0 0 0 0 21 | 94%
1 | 0 355 4 7 1 0 0 0 0 3 35 | 87%
2 | 0 3 325 2 1 0 7 29 3 2 53 | 76%
3 | 0 33 1 335 1 0 0 0 0 2 34 | 82%
4 | 6 0 1 0 371 3 0 0 0 0 31 | 90%
5 | 0 0 2 0 6 347 0 0 0 0 41 | 87%
6 | 0 1 5 8 0 0 322 4 0 0 56 | 81%
7 | 0 3 23 0 3 0 9 330 1 1 32 | 82%
8 | 0 0 5 2 0 0 0 0 343 4 57 | 83%
9 | 0 40 4 10 1 0 0 0 0 304 43 | 75%
10 | 4 61 16 34 28 17 14 6 12 37 6702 | 96%
void prediction_top_k(nnom_predict_t *pre);To print a Top-k when the prediction is done.
Arguments
- pre: the prediction instance created by
prediction_create().
Example
Top 1 Accuracy: 92.03%
Top 2 Accuracy: 96.39%
Top 3 Accuracy: 97.38%
Top 4 Accuracy: 97.85%
Top 5 Accuracy: 98.13%
Top 6 Accuracy: 98.40%
Top 7 Accuracy: 98.59%
Top 8 Accuracy: 98.88%
Top 9 Accuracy: 99.14%
Top 10 Accuracy: 99.60%
void prediction_summary(nnom_predict_t *pre);To print a summary when the prediction is done.
Arguments
- pre: the prediction instance created by
prediction_create().
Example
Prediction summary:
Test frames: 11005
Test running time: 1598 sec
Model running time: 1364908 ms
Average prediction time: 124026 us
Average effeciency: 53.30 ops/us
Average frame rate: 8.0 Hz
Evaluate a model using prediction_* APIs
The model needs to be compiled before it is being evaluated.
The evaluation gose through a few steps
- Create a instance using
prediction_create() - Feed the data one by one to the input buffer, then call
prediction_run()with the true label. - When all data has predicted, call
prediction_end()
Then you can use prediction_matrix(), prediction_top_k(), and prediction_summary() to see the results.
In addition, you can call model_stat() to see the performance of the last prediction.
After all, call prediction_delete() to release all memory.
Log from Key-word Spotting Example.
msh >
\ | /
- RT - Thread Operating System
/ | \ 4.0.0 build Mar 28 2019
2006 - 2018 Copyright by rt-thread team
RTT Control Block Detection Address is 0x20000b3c
msh >
INFO: Start compile...
Layer Activation output shape ops memory mem life-time
----------------------------------------------------------------------------------------------
Input - - ( 62, 12, 1) 0 ( 744, 744, 0) 1 - - - - - - -
Conv2D - ReLU - ( 60, 10, 32) 172800 ( 744,19200, 1152) 1 1 - - - - - -
MaxPool - - ( 30, 9, 32) 0 (19200, 8640, 0) 1 - 1 - - - - -
Conv2D - ReLU - ( 28, 7, 64) 3612672 ( 8640,12544, 2304) 1 1 - - - - - -
MaxPool - - ( 14, 6, 64) 0 (12544, 5376, 0) 1 - 1 - - - - -
Conv2D - ReLU - ( 12, 4, 96) 2654208 ( 5376, 4608, 3456) 1 1 - - - - - -
MaxPool - - ( 6, 3, 96) 0 ( 4608, 1728, 0) 1 - 1 - - - - -
Conv2D - - ( 6, 3, 11) 171072 ( 1728, 198, 396) 1 1 - - - - - -
GL_SumPool - - ( 1, 1, 11) 0 ( 198, 11, 44) 1 - 1 - - - - -
Softmax - - ( 1, 1, 11) 0 ( 11, 11, 0) - 1 - - - - - -
Output - - ( 11, 1, 1) 0 ( 11, 11, 0) 1 - - - - - - -
----------------------------------------------------------------------------------------------
INFO: memory analysis result
Block0: 3456 Block1: 8640 Block2: 19200 Block3: 0 Block4: 0 Block5: 0 Block6: 0 Block7: 0
Total memory cost by network buffers: 31296 bytes
msh >nn
nn_stat
msh >nn_stat
Print running stat..
Layer(#) - Time(us) ops(MACs) ops/us
--------------------------------------------------------
#1 Input - 9 0
#2 Conv2D - 8294 172800 20.83
#3 MaxPool - 5750 0
#4 Conv2D - 50089 3612672 72.12
#5 MaxPool - 3619 0
#6 Conv2D - 35890 2654208 73.95
#7 MaxPool - 1204 0
#8 Conv2D - 4411 171072 38.78
#9 GL_SumPool - 30 0
#10 Softmax - 3 0
#11 Output - 0 0
Summary.
Total ops (MAC): 6610752
Prediction time :109299us
Efficiency 60.48 ops/us
Total Memory cost (Network and NNoM): 32876
msh >pre
predict
msh >predict
Please select the NNoM binary test file and use Ymodem-128/1024 to send.
CCCC
Prediction done.
Prediction summary:
Test frames: 11005
Test running time: 1598 sec
Model running time: 1364908 ms
Average prediction time: 124026 us
Average effeciency: 53.30 ops/us
Average frame rate: 8.0 Hz
Top 1 Accuracy: 92.03%
Top 2 Accuracy: 96.39%
Top 3 Accuracy: 97.38%
Top 4 Accuracy: 97.85%
Top 5 Accuracy: 98.13%
Top 6 Accuracy: 98.40%
Top 7 Accuracy: 98.59%
Top 8 Accuracy: 98.88%
Top 9 Accuracy: 99.14%
Top 10 Accuracy: 99.60%
Confusion matrix:
predict 0 1 2 3 4 5 6 7 8 9 10
actual
0 | 395 1 0 0 2 0 0 0 0 0 21 | 94%
1 | 0 355 4 7 1 0 0 0 0 3 35 | 87%
2 | 0 3 325 2 1 0 7 29 3 2 53 | 76%
3 | 0 33 1 335 1 0 0 0 0 2 34 | 82%
4 | 6 0 1 0 371 3 0 0 0 0 31 | 90%
5 | 0 0 2 0 6 347 0 0 0 0 41 | 87%
6 | 0 1 5 8 0 0 322 4 0 0 56 | 81%
7 | 0 3 23 0 3 0 9 330 1 1 32 | 82%
8 | 0 0 5 2 0 0 0 0 343 4 57 | 83%
9 | 0 40 4 10 1 0 0 0 0 304 43 | 75%
10 | 4 61 16 34 28 17 14 6 12 37 6702 | 96%
msh >
OO: command not found.
msh >
msh >nn
nn_stat
msh >nn_stat
Print running stat..
Layer(#) - Time(us) ops(MACs) ops/us
--------------------------------------------------------
#1 Input - 9 0
#2 Conv2D - 8292 172800 20.83
#3 MaxPool - 5753 0
#4 Conv2D - 50095 3612672 72.11
#5 MaxPool - 3617 0
#6 Conv2D - 35893 2654208 73.94
#7 MaxPool - 1206 0
#8 Conv2D - 4412 171072 38.77
#9 GL_SumPool - 30 0
#10 Softmax - 2 0
#11 Output - 0 0
Summary.
Total ops (MAC): 6610752
Prediction time :109309us
Efficiency 60.47 ops/us
Total Memory cost (Network and NNoM): 32876
msh >