Clean the decode loop of the whisper example.

candle-examples/examples/whisper/main.rs

@@ -2,7 +2,8 @@
 // https://github.com/openai/whisper/blob/main/whisper/model.py
 // TODO:
 // - kv-cache support?
-// - language detection?
+// - Language detection?
+// - Batch size greater than 1.
 
 use anyhow::{Error as E, Result};
 use candle::{DType, Device, Tensor};
@@ -31,6 +32,10 @@ const LOGPROB_THRESHOLD: f64 = -1.0;
 const TEMPERATURES: [f64; 6] = [0.0, 0.2, 0.4, 0.6, 0.8, 1.0];
 const COMPRESSION_RATIO_THRESHOLD: f64 = 2.4;
 
+// Tokenizer dependent bits.
+const SOT_TOKEN: u32 = 50257;
+const EOT_TOKEN: u32 = 50256;
+
 #[derive(Parser, Debug)]
 #[command(author, version, about, long_about = None)]
 struct Args {
@@ -83,11 +88,12 @@ impl Decode {
         let sample_len = model.config.n_text_ctx / 2;
         let mut sum_logprob = 0f64;
         let no_speech_prob = f64::NAN;
-        // TODO: 50257 is the start of transcipt token, be more principled about get initial tokens
+        let mut tokens = vec![SOT_TOKEN];
-        let mut tokens: Vec<u32> = vec![50257];
         for _i in 0..sample_len {
             let tokens_t = Tensor::new(tokens.as_slice(), &mel.device())?;
-            // Insert a batch dim.
+
+            // The model expects a batch dim but this inference loop does not handle
+            // it so we add it at this point.
             let tokens_t = tokens_t.unsqueeze(0)?;
             let logits = model.decoder.forward(&tokens_t, &audio_features)?;
             let logits = logits.squeeze(0)?;
@@ -112,11 +118,10 @@ impl Decode {
                 .softmax(logits.rank() - 1)?
                 .get(next_token as usize)?
                 .to_scalar::<f32>()? as f64;
-            sum_logprob += prob.ln();
+            if next_token == EOT_TOKEN || tokens.len() > model.config.n_text_ctx {
-            // 50256 is the eot token, TODO: parameterize this.
-            if next_token == 50256 || tokens.len() > model.config.n_text_ctx {
                 break;
             }
+            sum_logprob += prob.ln();
         }
         let text = self
             .tokenizer
@@ -136,14 +141,17 @@ impl Decode {
 
     fn decode_with_fallback(&mut self, segment: &Tensor) -> Result<DecodingResult> {
         for (i, &t) in TEMPERATURES.iter().enumerate() {
-            let dr: DecodingResult = self.decode(segment, t)?;
+            let dr: Result<DecodingResult> = self.decode(segment, t);
             if i == TEMPERATURES.len() - 1 {
-                return Ok(dr);
+                return dr;
             }
-            let needs_fallback = dr.compression_ratio > COMPRESSION_RATIO_THRESHOLD
+            // On errors, we try again with a different temperature.
-                || dr.avg_logprob < LOGPROB_THRESHOLD;
+            if let Ok(dr) = dr {
-            if !needs_fallback || dr.no_speech_prob > NO_SPEECH_THRESHOLD {
+                let needs_fallback = dr.compression_ratio > COMPRESSION_RATIO_THRESHOLD
-                return Ok(dr);
+                    || dr.avg_logprob < LOGPROB_THRESHOLD;
+                if !needs_fallback || dr.no_speech_prob > NO_SPEECH_THRESHOLD {
+                    return Ok(dr);
+                }
             }
         }
         unreachable!()
@@ -195,7 +203,7 @@ fn main() -> Result<()> {
             duration: segment_duration,
             dr,
         };
-        println!("{seek} {segment:?}");
+        println!("{seek}: {segment:?}");
         segments.push(segment)
     }
     Ok(())

candle-examples/examples/whisper/model.rs

@@ -1,3 +1,5 @@
+// We use anyhow rather than candle errors as it provides better support for getting the backtrace
+// back when using RUST_LIB_BACKTRACE=1.
 use anyhow::Result;
 use candle::{safetensors::SafeTensors, DType, Device, Shape, Tensor};
 use std::collections::HashMap;