avoid using torch_tensor twice

DESCRIPTION

@@ -14,7 +14,7 @@ URL: https://github.com/bnosac/audio.vadsilero
 Encoding: UTF-8
 Imports:
     torch,
-    wav
+    audio
 Suggests:
     av
 RoxygenNote: 7.1.2

NAMESPACE

@@ -2,7 +2,8 @@
 
 S3method(print,VAD)
 export(silero)
+importFrom(audio,load.wave)
+importFrom(torch,autograd_set_grad_mode)
 importFrom(torch,jit_load)
 importFrom(torch,jit_scalar)
 importFrom(torch,torch_tensor)
-importFrom(wav,read_wav)

NEWS.md

@@ -1,6 +1,7 @@
 ## CHANGES IN audio.vadsilero VERSION 0.2
 
 - disable the gradient history recording & set model in evaluation (inference) mode
+- replace package wav with package audio for reading in the wav file
 
 ## CHANGES IN audio.vadsilero VERSION 0.1
 

R/pkg.R

@@ -1,3 +1,3 @@
-#' @importFrom wav read_wav
+#' @importFrom audio load.wave
 #' @importFrom torch torch_tensor jit_load jit_scalar autograd_set_grad_mode
 NULL

R/silero.R

@@ -102,17 +102,19 @@ SILERO <- function(){
 }
 
 predict.SILERO <- function(object, newdata, sample_rate, milliseconds, window = milliseconds * (sample_rate / 1000), threshold = 0.5){
+    sound       <- audio::load.wave(newdata)
+    sample_rate <- attr(sound, which = "rate")
+    n_samples   <- length(sound)
+    sample_rate <- torch::jit_scalar(as.integer(sample_rate))
+
     if(!sample_rate %in% c(8000, 16000)){
         stop("sample_rate should be 8000 or 16000")
     }
     if(!window %in% c(256, 512, 768, 1024, 1536)){
         stop("Unknown combination of milliseconds and sample_rate")
     }
-    sound       <- wav::read_wav(newdata)
-    sound       <- sound[1, ]
+
     test        <- torch::torch_tensor(sound)
-    n_samples   <- length(sound)
-    sample_rate <- torch::jit_scalar(sample_rate)
 
     elements    <- seq.int(from = 1, to = n_samples, by = window)
     out         <- numeric(length = length(elements))
@@ -121,10 +123,12 @@ predict.SILERO <- function(object, newdata, sample_rate, milliseconds, window =
         if((elements[i]+window-1) > n_samples){
             samples <- sound[elements[i]:length(sound)]
             samples <- c(samples, rep(as.numeric(0), times = window - length(samples)))
-            out[i]  <- as.numeric(object$model$forward(torch::torch_tensor(samples), sr = sample_rate))    
+            samples <- torch::torch_tensor(samples)
+            out[i]  <- as.numeric(object$model$forward(samples, sr = sample_rate))    
         }else{
             samples <- test[elements[i]:(elements[i]+window-1)]
-            samples <- torch::torch_tensor(samples)
+            #samples <- torch::torch_tensor(samples)
+            #print(str(samples))
             out[i]  <- as.numeric(object$model$forward(samples, sr = sample_rate))    
         }
     }