vad : rename lstm weights for consistency

Author: danbev

src/whisper.cpp

@@ -4388,10 +4388,10 @@ struct whisper_vad_model {
     struct ggml_tensor * encoder_3_bias;    // [128]
 
     // LSTM decoder tensors
-    struct ggml_tensor * lstm_weight_ih;    // [512, 128] input-to-hidden
-    struct ggml_tensor * lstm_bias_ih;      // [512]
-    struct ggml_tensor * lstm_weight_hh;    // [512, 128] hidden-to-hidden
-    struct ggml_tensor * lstm_bias_hh;      // [512]
+    struct ggml_tensor * lstm_ih_weight;    // [512, 128] input-to-hidden
+    struct ggml_tensor * lstm_ih_bias;      // [512]
+    struct ggml_tensor * lstm_hh_weight;    // [512, 128] hidden-to-hidden
+    struct ggml_tensor * lstm_hh_bias;      // [512]
 
     // Final conv layer
     struct ggml_tensor * final_conv_weight; // [1, 128, 1]
@@ -4536,12 +4536,12 @@ static ggml_tensor * whisper_vad_build_lstm_layer(ggml_context* ctx0,
     ggml_set_input(c_in);
 
     // Create operations using the input-to-hidden weights.
-    struct ggml_tensor * inp_gate = ggml_mul_mat(ctx0, model.lstm_weight_ih, x_t);
-    inp_gate = ggml_add(ctx0, inp_gate, model.lstm_bias_ih);
+    struct ggml_tensor * inp_gate = ggml_mul_mat(ctx0, model.lstm_ih_weight, x_t);
+    inp_gate = ggml_add(ctx0, inp_gate, model.lstm_ih_bias);
 
     // Create operations using the hidden-to-hidden weights.
-    struct ggml_tensor * hid_gate = ggml_mul_mat(ctx0, model.lstm_weight_hh, h_in);
-    hid_gate = ggml_add(ctx0, hid_gate, model.lstm_bias_hh);
+    struct ggml_tensor * hid_gate = ggml_mul_mat(ctx0, model.lstm_hh_weight, h_in);
+    hid_gate = ggml_add(ctx0, hid_gate, model.lstm_hh_bias);
 
     // Create add operation to get preactivations for all gates.
     struct ggml_tensor * out_gate = ggml_add(ctx0, inp_gate, hid_gate);
@@ -4883,25 +4883,25 @@ whisper_vad_context * whisper_vad_init_from_file_with_params_no_state(
         const int hstate_dim = hparams.lstm_hidden_size * 4;
 
         // LSTM weights - input to hidden
-        model.lstm_weight_ih = create_tensor(
+        model.lstm_ih_weight = create_tensor(
             VAD_TENSOR_LSTM_WEIGHT_IH,
             ggml_new_tensor_2d(ctx, GGML_TYPE_F32, hparams.lstm_hidden_size, hstate_dim)
         );
 
         // LSTM weights - hidden to hidden
-        model.lstm_weight_hh = create_tensor(
+        model.lstm_hh_weight = create_tensor(
             VAD_TENSOR_LSTM_WEIGHT_HH,
             ggml_new_tensor_2d(ctx, GGML_TYPE_F32, hparams.lstm_hidden_size, hstate_dim)
         );
 
         // LSTM bias - input to hidden
-        model.lstm_bias_ih = create_tensor(
+        model.lstm_ih_bias = create_tensor(
             VAD_TENSOR_LSTM_BIAS_IH,
             ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hstate_dim)
         );
 
         // LSTM bias - hidden to hidden
-        model.lstm_bias_hh = create_tensor(
+        model.lstm_hh_bias = create_tensor(
             VAD_TENSOR_LSTM_BIAS_HH,
             ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hstate_dim)
         );
@@ -5130,39 +5130,39 @@ whisper_vad_context * whisper_vad_init_from_file_with_params_no_state(
 
         {
             // Print as F32
-            struct ggml_tensor * tensor = model.lstm_weight_ih;
+            struct ggml_tensor * tensor = model.lstm_ih_weight;
             std::vector<float> read_b(ggml_nbytes(tensor));
             ggml_backend_tensor_get(tensor, read_b.data(), 0, ggml_nbytes(tensor));
             for (int i = 0; i < 10; i++) {
-                WHISPER_LOG_INFO("%s: lstm_weight_ih: [%d]: %f\n", __func__, i, read_b[i]);
+                WHISPER_LOG_INFO("%s: lstm_ih_weight: [%d]: %f\n", __func__, i, read_b[i]);
             }
         }
         {
             // Print as F32
-            struct ggml_tensor * tensor = model.lstm_bias_ih;
+            struct ggml_tensor * tensor = model.lstm_ih_bias;
             std::vector<float> read_b(ggml_nbytes(tensor));
             ggml_backend_tensor_get(tensor, read_b.data(), 0, ggml_nbytes(tensor));
             for (int i = 0; i < 10; i++) {
-                WHISPER_LOG_INFO("%s: lstm_bias_ih: [%d]: %f\n", __func__, i, read_b[i]);
+                WHISPER_LOG_INFO("%s: lstm_ih_bias: [%d]: %f\n", __func__, i, read_b[i]);
             }
         }
 
         {
             // Print as F32
-            struct ggml_tensor * tensor = model.lstm_weight_hh;
+            struct ggml_tensor * tensor = model.lstm_hh_weight;
             std::vector<float> read_b(ggml_nbytes(tensor));
             ggml_backend_tensor_get(tensor, read_b.data(), 0, ggml_nbytes(tensor));
             for (int i = 0; i < 10; i++) {
-                WHISPER_LOG_INFO("%s: lstm_weight_hh: [%d]: %f\n", __func__, i, read_b[i]);
+                WHISPER_LOG_INFO("%s: lstm_hh_weight: [%d]: %f\n", __func__, i, read_b[i]);
             }
         }
         {
             // Print as F32
-            struct ggml_tensor * tensor = model.lstm_bias_hh;
+            struct ggml_tensor * tensor = model.lstm_hh_bias;
             std::vector<float> read_b(ggml_nbytes(tensor));
             ggml_backend_tensor_get(tensor, read_b.data(), 0, ggml_nbytes(tensor));
             for (int i = 0; i < 10; i++) {
-                WHISPER_LOG_INFO("%s: lstm_bias_hh: [%d]: %f\n", __func__, i, read_b[i]);
+                WHISPER_LOG_INFO("%s: lstm_hh_bias: [%d]: %f\n", __func__, i, read_b[i]);
             }
         }