manual mods post export_weights to accomodate different input_dim for…

… model05 & model19_check3, both ctests work, updated opus nnet.h patch

CMakeLists.txt

@@ -511,11 +511,12 @@ add_test(NAME bbfm_sc_bpf_loss
 
 # C Port of Core Encoder/decoder
 
-# we test by comparing loss of features_in/features_out, can all happen at rate Rs
+# we test by comparing loss of features_in/features_out, can all happen at rate Rs.  We load model05.bin weights as 
+# compiled-in weights are for model19_check3
 add_test(NAME c_encoder_model5
          COMMAND sh -c "cd ${CMAKE_SOURCE_DIR}; 
          ./inference.sh model05/checkpoints/checkpoint_epoch_100.pth wav/brian_g8sez.wav /dev/null; \
-         cat features_in.f32 | ${CMAKE_CURRENT_BINARY_DIR}/src/test_rade_enc 1 0 > z_c.f32; \
+         cat features_in.f32 | ${CMAKE_CURRENT_BINARY_DIR}/src/test_rade_enc 1 0 ${CMAKE_SOURCE_DIR}/bin/model05.bin > z_c.f32; \
          python3 stateful_encoder.py model05/checkpoints/checkpoint_epoch_100.pth features_in.f32 /dev/null --read_latent z_c.f32 --loss_test 0.2")
          set_tests_properties(c_encoder_model5 PROPERTIES PASS_REGULAR_EXPRESSION "PASS")
 
@@ -524,7 +525,7 @@ add_test(NAME c_encoder_model19_check3
          COMMAND sh -c "cd ${CMAKE_SOURCE_DIR}; 
          ./inference.sh model05/checkpoints/checkpoint_epoch_100.pth wav/brian_g8sez.wav /dev/null --write_latent z.f32; \
          cat features_in.f32 | python3 radae_txe.py | python3 radae_rxe.py > features_out.f32; \
-         cat features_in.f32 | ${CMAKE_CURRENT_BINARY_DIR}/src/test_rade_enc 3 1 ${CMAKE_CURRENT_BINARY_DIR}/model19_check3.bin > z_c.f32; \
+         cat features_in.f32 | ${CMAKE_CURRENT_BINARY_DIR}/src/test_rade_enc 3 1 > z_c.f32; \
          cat z_c.f32 | python3 radae_txe.py --bypass_enc | python3 radae_rxe.py > features_rx_out.f32; \
          python3 loss.py features_in.f32 features_out.f32 --features_hat2 features_rx_out.f32 --compare")
          set_tests_properties(c_encoder_model19_check3 PROPERTIES PASS_REGULAR_EXPRESSION "PASS")

cmake/BuildOpus.cmake

@@ -15,7 +15,7 @@ if(APPLE AND BUILD_OSX_UNIVERSAL)
 ExternalProject_Add(build_opus_x86
     DOWNLOAD_EXTRACT_TIMESTAMP NO
     BUILD_IN_SOURCE 1
-    PATCH_COMMAND sh -c "patch dnn/nnet.h < ${CMAKE_SOURCE_DIR}/src/opus-export.diff"
+    PATCH_COMMAND sh -c "patch dnn/nnet.h < ${CMAKE_SOURCE_DIR}/src/opus-nnet.h.diff"
     CONFIGURE_COMMAND ${CONFIGURE_COMMAND} --host=x86_64-apple-darwin --target=x86_64-apple-darwin CFLAGS=-arch\ x86_64\ -O2\ -mmacosx-version-min=10.11
     BUILD_COMMAND $(MAKE)
     INSTALL_COMMAND ""
@@ -24,7 +24,7 @@ ExternalProject_Add(build_opus_x86
 ExternalProject_Add(build_opus_arm
     DOWNLOAD_EXTRACT_TIMESTAMP NO
     BUILD_IN_SOURCE 1
-    PATCH_COMMAND sh -c "patch dnn/nnet.h < ${CMAKE_SOURCE_DIR}/src/opus-export.diff"
+    PATCH_COMMAND sh -c "patch dnn/nnet.h < ${CMAKE_SOURCE_DIR}/src/opus-nnet.h.diff"
     CONFIGURE_COMMAND ${CONFIGURE_COMMAND} --host=aarch64-apple-darwin --target=aarch64-apple-darwin CFLAGS=-arch\ arm64\ -O2\ -mmacosx-version-min=10.11
     BUILD_COMMAND $(MAKE)
     INSTALL_COMMAND ""
@@ -57,7 +57,7 @@ set_target_properties(opus PROPERTIES
 else(APPLE AND BUILD_OSX_UNIVERSAL)
 ExternalProject_Add(build_opus
     BUILD_IN_SOURCE 1
-    PATCH_COMMAND sh -c "patch dnn/nnet.h < ${CMAKE_SOURCE_DIR}/src/opus-export.diff"
+    PATCH_COMMAND sh -c "patch dnn/nnet.h < ${CMAKE_SOURCE_DIR}/src/opus-nnet.h.diff"
     CONFIGURE_COMMAND ${CONFIGURE_COMMAND}
     BUILD_COMMAND $(MAKE)
     INSTALL_COMMAND ""

export_rade_weights.py

@@ -154,7 +154,7 @@ def c_export(args, model):
     # constants
     constants_writer.header.write(
 f"""
-#define RADE_NUM_FEATURES {model.feature_dim}
+#define RADE_FRAMES_PER_STEP {model.enc_stride}
 
 #define RADE_LATENT_DIM {args.latent_dim}
 

src/opus-export.diff → src/opus-nnet.h.diff

@@ -15,6 +15,10 @@
 > void RADE_EXPORT compute_generic_conv1d_dilation(const LinearLayer *layer, float *output, float *mem, const float *input, int input_size, int dilation, int activation, int arch);
 > void RADE_EXPORT compute_glu(const LinearLayer *layer, float *output, const float *input, int arch);
 > void RADE_EXPORT compute_gated_activation(const LinearLayer *layer, float *output, const float *input, int activation, int arch);
+96c98
+< int parse_weights(WeightArray **list, const void *data, int len);
+---
+> int RADE_EXPORT parse_weights(WeightArray **list, const void *data, int len);
 108c110
 < int linear_init(LinearLayer *layer, const WeightArray *arrays,
 ---

src/rade_constants.h

@@ -7,7 +7,7 @@
 #include "nnet.h"
 
 
-#define RADE_NUM_FEATURES 21
+#define RADE_FRAMES_PER_STEP 4
 
 #define RADE_LATENT_DIM 80
 

src/rade_enc_data.c

@@ -227860,10 +227860,10 @@ const WeightArray radeenc_arrays[] = {
     {NULL, 0, 0, NULL}
 };
 #endif /* USE_WEIGHTS_FILE */
-#include <stdio.h>
+
 #ifndef DUMP_BINARY_WEIGHTS
-int init_radeenc(RADEEnc *model, const WeightArray *arrays) {
-    if (linear_init(&model->enc_dense1, arrays, "enc_dense1_bias", NULL, NULL,"enc_dense1_weights_float", NULL, NULL, NULL, 84, 64)) return 1;
+int init_radeenc(RADEEnc *model, const WeightArray *arrays, int input_dim) {
+    if (linear_init(&model->enc_dense1, arrays, "enc_dense1_bias", NULL, NULL,"enc_dense1_weights_float", NULL, NULL, NULL, input_dim, 64)) return 1;
     if (linear_init(&model->enc_zdense, arrays, "enc_zdense_bias", NULL, NULL,"enc_zdense_weights_float", NULL, NULL, NULL, 864, 80)) return 1;
     if (linear_init(&model->enc_gru1_input, arrays, "enc_gru1_input_bias", "enc_gru1_input_subias", "enc_gru1_input_weights_int8","enc_gru1_input_weights_float", "enc_gru1_input_weights_idx", NULL, "enc_gru1_input_scale", 64, 192)) return 1;
     if (linear_init(&model->enc_gru1_recurrent, arrays, "enc_gru1_recurrent_bias", "enc_gru1_recurrent_subias", "enc_gru1_recurrent_weights_int8","enc_gru1_recurrent_weights_float", NULL, NULL, "enc_gru1_recurrent_scale", 64, 192)) return 1;

src/rade_enc_data.h

@@ -98,6 +98,6 @@ struct RADEEnc {
     LinearLayer enc_conv5;
 };
 
-int init_radeenc(RADEEnc *model, const WeightArray *arrays);
+int init_radeenc(RADEEnc *model, const WeightArray *arrays, int input_dim);
 
 #endif /* RADE_ENC_DATA_H */

src/test_rade_enc.c

@@ -32,11 +32,12 @@ int main(int argc, char *argv[])
     int nb_total_features = 36;
     int num_features = 20;
     int num_used_features = 20;
-    int frames_per_step = 4; // TODO dig this out of network somehow
-
+    int frames_per_step = RADE_FRAMES_PER_STEP;
+    
     if (auxdata) {
         num_features += 1;
     }
+    int input_dim = num_features*frames_per_step;
 
     int fd;
     void *data;
@@ -60,21 +61,20 @@ int main(int argc, char *argv[])
         for (int i=0;i<nb_arrays;i++) {
             fprintf(stderr, "found %s: size %d\n", list[i].name, list[i].size);
         }
-        if (init_radeenc(&enc_model, list) != 0) {
+        if (init_radeenc(&enc_model, list, input_dim) != 0) {
             fprintf(stderr, "Error initialising encoder model from %s\n", argv[3]);
             exit(1);       
         }
-    } else if (init_radeenc(&enc_model, radeenc_arrays) != 0) {
+    } else if (init_radeenc(&enc_model, radeenc_arrays, input_dim) != 0) {
         fprintf(stderr, "Error initialising built-in encoder model\n");
         exit(1);        
     }
     rade_init_encoder(&enc_state);
 
-    int n_features_in = enc_model.enc_dense1.nb_inputs;
     assert(enc_model.enc_zdense.nb_outputs == RADE_LATENT_DIM);
 
     float features_read[frames_per_step*nb_total_features];
-    float features[n_features_in];
+    float features[input_dim];
     float z[RADE_LATENT_DIM];
 
     // From celt/cpu_support.h:
@@ -90,8 +90,8 @@ int main(int argc, char *argv[])
     // This auto-magically selects best arch
     // arch = opus_select_arch();
 
-    fprintf(stderr, "arch: %d bottleneck: %d auxdata: %d n_features_in: %d n_z_out: %d\n", 
-            arch, bottleneck, auxdata, n_features_in, enc_model.enc_zdense.nb_outputs);
+    fprintf(stderr, "arch: %d bottleneck: %d auxdata: %d input_dim: %d n_z_out: %d\n", 
+            arch, bottleneck, auxdata, input_dim, enc_model.enc_zdense.nb_outputs);
     int nb_feature_vecs = 0;
     size_t to_read, nb_read;
     to_read = frames_per_step*nb_total_features;

weight-exchange/wexchange/c_export/c_writer.py

@@ -40,7 +40,7 @@ def __init__(self,
                  model_struct_name="Model",
                  nnet_header="nnet.h"):
         """
-        Writer class for creating souce and header files for weight exports to C
+        Writer class for creating source and header files for weight exports to C
 
         Parameters:
         -----------