Ported to new MFCC

.gitignore

@@ -1,6 +1,5 @@
 BUILD*
 *.dat
-nntool_script
 *_emul
 samples/speech_dataset
 *__pycache__*

MFCC_params_LARGE.h → MFCC_params.h

@@ -1,18 +1,10 @@
 #define	SAMPLERATE                16000
-#define	PREEMP_FACTOR               0.0
 #define	FRAME_SIZE                  640
 #define	FRAME_STEP                  320
-#define	N_FRAME                      49
 #define	N_FFT                      1024
-#define	USE_RADIX_4                   0
 #define	DATA_TYPE                     0
-#define	USE_POWER                     1
-#define	OUT_FFT                       0
 #define	MFCC_BANK_CNT                40
 #define	FMIN                       20.0
 #define	FMAX                     4000.0
 #define	MFCC_COEFF_CNT              494
 #define	N_DCT                        40
-#define	OUTPUT_MELSPECTROGRAM         0
-#define	LIFTER_COEFF                  0
-#define	USE_DB                        0

MFCCmodel.c

@@ -1,37 +1,27 @@
 #include "AutoTilerLib.h"
 #include "AutoTilerLibTypes.h"
-#include "MfccGenerator.h"
-#ifdef SMALL
-  #include "MFCC_params_SMALL.h"
-#endif
-#ifdef MEDIUM
-  #include "MFCC_params_MEDIUM.h"
-#endif
-#ifdef LARGE
-  #include "MFCC_params_LARGE.h"
-#endif
-
+#include "DSP_Generators.h"
+#include "MFCC_params.h"
 
 void MFCCConfiguration(unsigned int L1Memory)
 {
   SetInlineMode(ALWAYS_INLINE);
-	SetSymbolDynamics();
+  SetSymbolDynamics();
 
-  SetUsedFilesNames(0, 1, "MfccBasicKernels.h");
+  SetUsedFilesNames(0, 3, "MfccBasicKernels.h", "CmplxFunctions.h", "PreProcessing.h");
   SetGeneratedFilesNames("MFCCKernels.c", "MFCCKernels.h");
 
   SetL1MemorySize(L1Memory);
 }
 
 int main(int argc, char **argv)
 {
-  	if (TilerParseOptions(argc, argv)) GenTilingError("Failed to initialize or incorrect output arguments directory.\n");
+    if (TilerParseOptions(argc, argv)) GenTilingError("Failed to initialize or incorrect output arguments directory.\n");
 
     // Set Auto Tiler configuration, given shared L1 memory is 51200
-    MFCCConfiguration(51200);
+    MFCCConfiguration(64*1024);
     // Load FIR basic kernels
     LoadMFCCLibrary();
-    MFCC_Generator("MFCC",        0, N_FRAME, FRAME_SIZE, FRAME_STEP, N_FFT, MFCC_COEFF_CNT, N_DCT, PREEMP_FACTOR, LIFTER_COEFF, USE_RADIX_4, USE_POWER, DATA_TYPE, MFCC_BANK_CNT, 0, 0);
-  //MFCC_Generator("MFCC_single", 0,       1, FRAME_SIZE, FRAME_STEP, N_FFT, MFCC_COEFF_CNT, N_DCT, PREEMP_FACTOR, LIFTER_COEFF, USE_RADIX_4, USE_POWER, DATA_TYPE, MFCC_BANK_CNT, 0, 0);
+    MFCC_Generator("MFCC", 0, 49, FRAME_SIZE, FRAME_STEP, N_FFT, MFCC_BANK_CNT, MFCC_COEFF_CNT, N_DCT, 0, 0, 0, USE_POWER, DATA_TYPE, 1, 0);
     GenerateTilingCode();
 }

Makefile

@@ -56,17 +56,14 @@ QUANT_BITS=8
 BUILD_DIR=BUILD
 
 #MODEL_NE16 ?= 0
-NNTOOL_SCRIPT_PARAMETRIC=model/nntool_script_params
-MODEL_SUFFIX = _$(QUANT_BITS)BIT
+NNTOOL_SCRIPT=model/nntool_script
+MODEL_SUFFIX = _SQ$(QUANT_BITS)
 ifeq ($(MODEL_NE16), 1)
-	NNTOOL_SCRIPT_PARAMETRIC=model/nntool_script_ne16_params
+	NNTOOL_SCRIPT=model/nntool_script_ne16
 	MODEL_SUFFIX = _NE16
 endif
-ifeq ($(DUMP_TENSORS), 1)
-	NNTOOL_SET_GRAPH_DUMP = set graph_dump_tensor 7
-endif
 NNTOOL_EXTRA_FLAGS = -q
-MODEL_BUILD = BUILD_MODEL_$(QUANT_BITS)BIT
+MODEL_BUILD = BUILD_MODEL$(MODEL_SUFFIX)
 
 CLUSTER_STACK_SIZE=4096
 CLUSTER_SLAVE_STACK_SIZE=1024
@@ -83,26 +80,24 @@ endif
 FREQ_FC?=250
 VOLTAGE?=1.2
 
-include common/model_decl.mk
 ifeq ($(USE_POWER), 1)
 	# override the tflite model name to the one which expects power MFCC -> more efficient
-	TRAINED_TFLITE_MODEL=model/$(MODEL_PREFIX)_power.tflite
+	TRAINED_MODEL=model/$(MODEL_PREFIX)_power.tflite
+else
+	TRAINED_MODEL=model/$(MODEL_PREFIX).tflite
 endif
+include common/model_decl.mk
+include mfcc_model.mk
 
 
 ifeq ($(WITH_MFCC), 1)
 	APP_SRCS    += main_with_mfcc.c $(MODEL_GEN_C) $(MODEL_COMMON_SRCS) $(CNN_LIB)
-	APP_SRCS    += $(GAP_LIB_PATH)/wav_io/wavIO.c $(MFCC_KER_SRCS) $(MFCCBUILD_DIR)/MFCCKernels.c
+	APP_SRCS    += $(GAP_LIB_PATH)/wav_io/wavIO.c $(MFCCBUILD_DIR)/MFCCKernels.c $(TILER_DSP_KERNEL_PATH)/LUT_Tables/TwiddlesDef.c $(TILER_DSP_KERNEL_PATH)/LUT_Tables/RFFTTwiddlesDef.c $(TILER_DSP_KERNEL_PATH)/LUT_Tables/SwapTablesDef.c
+	APP_SRCS  	+= $(TILER_DSP_KERNEL_PATH)/MfccBasicKernels.c $(TILER_DSP_KERNEL_PATH)/FFT_Library.c $(TILER_DSP_KERNEL_PATH)/CmplxFunctions.c $(TILER_DSP_KERNEL_PATH)/PreProcessing.c 
 	APP_CFLAGS  += -I$(MFCC_GENERATOR) -I$(MFCCBUILD_DIR)
 else
 	APP_SRCS    += main.c $(MODEL_GEN_C) $(MODEL_COMMON_SRCS) $(CNN_LIB)
 endif
-ifeq ($(USE_POWER), 0)
-	APP_CFLAGS += -DUSE_ABS
-endif
-ifeq ($(USE_HIGH_PREC), 1)
-	APP_CFLAGS += -DHIGH_PREC_FFT
-endif
 ifeq ($(PMSIS_OS), freertos)
 	APP_CFLAGS += -DFREERTOS
 endif
@@ -111,7 +106,7 @@ APP_CFLAGS += -O3 -s -mno-memcpy -fno-tree-loop-distribute-patterns
 APP_CFLAGS += -I. -I$(MODEL_COMMON_INC) -I$(TILER_EMU_INC) -I$(TILER_INC) -I$(MODEL_BUILD) $(CNN_LIB_INCLUDE)
 APP_CFLAGS += -DAT_MODEL_PREFIX=$(MODEL_PREFIX) $(MODEL_SIZE_CFLAGS)
 APP_CFLAGS += -DSTACK_SIZE=$(CLUSTER_STACK_SIZE) -DSLAVE_STACK_SIZE=$(CLUSTER_SLAVE_STACK_SIZE) -DFREQ_FC=$(FREQ_FC) -DFREQ_CL=$(FREQ_CL) -DVOLTAGE=$(VOLTAGE)
-APP_CFLAGS += -DAT_IMAGE=$(IMAGE) -DAT_WAV=$(WAV_PATH) #-DWRITE_WAV #-DPRINT_AT_INPUT #-DPRINT_WAV 
+APP_CFLAGS += -DAT_IMAGE=$(IMAGE) -DAT_WAV=$(WAV_PATH)
 
 ifeq ($(USE_MIC_SENSORBOARD), 1)
 	APP_CFLAGS += -DFROM_SENSOR -DSILENT
@@ -127,26 +122,18 @@ generate_samples:
 	python3 utils/generate_samples_images.py --dct_coefficient_count $(DCT_COUNT) --window_size_ms $(FRAME_SIZE_ms) --window_stride_ms $(FRAME_STEP_ms)
 
 test_accuracy:
-	python3 utils/test_accuracy_emul.py --tflite_model $(TRAINED_TFLITE_MODEL) --dct_coefficient_count $(DCT_COUNT) --window_size_ms $(FRAME_SIZE_ms) --window_stride_ms $(FRAME_STEP_ms) --test_with_wav $(WITH_MFCC) --use_power_spectrogram $(USE_POWER) --use_high_prec $(USE_HIGH_PREC)
+	python3 utils/test_accuracy_emul.py --tflite_model $(TRAINED_MODEL) --dct_coefficient_count $(DCT_COUNT) --window_size_ms $(FRAME_SIZE_ms) --window_stride_ms $(FRAME_STEP_ms) --test_with_wav $(WITH_MFCC) --use_power_spectrogram $(USE_POWER)
 
 test_accuracy_tflite:
-	python3 utils/test_accuracy_tflite.py --tflite_model $(TRAINED_TFLITE_MODEL) --dct_coefficient_count $(DCT_COUNT) --window_size_ms $(FRAME_SIZE_ms) --window_stride_ms $(FRAME_STEP_ms) --use_power_spectrogram $(USE_POWER)
+	python3 utils/test_accuracy_tflite.py --tflite_model $(TRAINED_MODEL) --dct_coefficient_count $(DCT_COUNT) --window_size_ms $(FRAME_SIZE_ms) --window_stride_ms $(FRAME_STEP_ms) --use_power_spectrogram $(USE_POWER)
 
 # all depends on the model
-all:: model mfcc_model
-
-clean:: clean_model clean_mfcc_model
+all::  $(MFCCBUILD_DIR)/MFCCKernels.c model
 
-clean_at_model:
-	$(RM) $(MODEL_GEN_EXE)
+clean:: clean_model clean_mfcc_code
 
 at_model_disp:: $(MODEL_BUILD) $(MODEL_GEN_EXE)
 	$(MODEL_GEN_EXE) -o $(MODEL_BUILD) -c $(MODEL_BUILD) $(MODEL_GEN_EXTRA_FLAGS) --debug=Disp
 
-at_model:: $(MODEL_BUILD) $(MODEL_GEN_EXE)
-	$(MODEL_GEN_EXE) -o $(MODEL_BUILD) -c $(MODEL_BUILD) $(MODEL_GEN_EXTRA_FLAGS)
-
 include common/model_rules.mk
-include mfcc_model.mk
-
 include $(GAP_SDK_HOME)/tools/rules/pmsis_rules.mk

MfccConfig.json

@@ -0,0 +1,12 @@
+{
+	"n_fft": 1024,
+	"frame_size": 640,
+	"frame_step": 320,
+	"sample_rate": 16000,
+	"mfcc_bank_cnt": 40,
+	"n_dct": 40,
+	"fmin": 20.0,
+	"fmax": 4000.0,
+	"dtype": "fix16",
+	"use_tf_mfcc": true
+}

MfccConfig_HighPrec.json

@@ -0,0 +1,12 @@
+{
+	"n_fft": 1024,
+	"frame_size": 640,
+	"frame_step": 320,
+	"sample_rate": 16000,
+	"mfcc_bank_cnt": 40,
+	"n_dct": 40,
+	"fmin": 20.0,
+	"fmax": 4000.0,
+	"dtype": "fix32_scal",
+	"use_tf_mfcc": true
+}