Add Readme file for metric results

benckmarks.md

@@ -0,0 +1,8 @@
+**Metrics on Mozilla-Foundation/Common-Voice-17-0 dataset**
+
+| Model                                | Metric Value |
+| ------------------------------------ | ------------ |
+| distil-whisper/distil-large-v3       | 120.48       |
+| distil-whisper/distil-large-v3 + Hqq | 120.88       |
+
+Hqq: https://github.com/mobiusml/hqq/

whisperplus/audio_eval.py

@@ -1,72 +1,83 @@
 import torch
 from datasets import load_dataset
 from evaluate import load
-from hqq.core.quantize import HQQBackend, HQQLinear
-from hqq.utils.patching import prepare_for_inference
+from pipelines.whisper import SpeechToTextPipeline
 from tqdm import tqdm
 from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, BitsAndBytesConfig, HqqConfig, pipeline
 from transformers.pipelines.pt_utils import KeyDataset
 
-from whisperplus.pipelines.whisper import SpeechToTextPipeline
+model_id = "distil-whisper/distil-large-v3"
+processor = AutoProcessor.from_pretrained(model_id)
 
-HQQLinear.set_backend(HQQBackend.PYTORCH)  # Pytorch backend
-HQQLinear.set_backend(HQQBackend.PYTORCH_COMPILE)  # Compiled Pytorch via dynamo
-HQQLinear.set_backend(HQQBackend.ATEN)  # C++ Aten/CUDA backend (set automatically by default if available)
 
-model_id = "distil-whisper/distil-large-v3"
+def hqq_load_model():
+    from hqq.core.quantize import HQQBackend, HQQLinear
+    from hqq.utils.patching import prepare_for_inference
 
-hqq_config = HqqConfig(
-    nbits=4,
-    group_size=64,
-    quant_zero=False,
-    quant_scale=False,
-    axis=0,
-    offload_meta=False,
-)  # axis=0 is used by default
+    HQQLinear.set_backend(HQQBackend.PYTORCH)  # Pytorch backend
+    HQQLinear.set_backend(HQQBackend.PYTORCH_COMPILE)  # Compiled Pytorch via dynamo
+    HQQLinear.set_backend(
+        HQQBackend.ATEN)  # C++ Aten/CUDA backend (set automatically by default if available)
 
-bnb_config = BitsAndBytesConfig(
-    load_in_4bit=True,
-    bnb_4bit_quant_type="nf4",
-    bnb_4bit_compute_dtype=torch.bfloat16,
-    bnb_4bit_use_double_quant=True,
-)
+    hqq_config = HqqConfig(
+        nbits=4,
+        group_size=64,
+        quant_zero=False,
+        quant_scale=False,
+        axis=0,
+        offload_meta=False,
+    )  # axis=0 is used by default
 
-model = SpeechToTextPipeline(model_id="distil-whisper/distil-large-v3", quant_config=hqq_config)
+    model = SpeechToTextPipeline(model_id="distil-whisper/distil-large-v3", quant_config=hqq_config)
+    model = model.model
+    return model
 
-model = model.model
 
-processor = AutoProcessor.from_pretrained(model_id)
+def base_load_model():
+    model = AutoModelForSpeechSeq2Seq.from_pretrained(
+        model_id,
+        torch_dtype=torch.bfloat16,
+        low_cpu_mem_usage=True,
+        use_safetensors=True,
+        device="cuda:0",
+    )
+
+    return model
+
+
+def calculate_metrics(model, processor, model_id):
+    pipe = pipeline(
+        "automatic-speech-recognition",
+        model=model,
+        torch_dtype=torch.bfloat16,
+        tokenizer=processor.tokenizer,
+        feature_extractor=processor.feature_extractor,
+        model_kwargs={"use_flash_attention_2": True},
+    )
+
+    wer_metric = load("wer")
+
+    common_voice_test = load_dataset(
+        "mozilla-foundation/common_voice_17_0",  # mozilla-foundation/common_voice_17_0
+        "dv",
+        split="test")
+
+    all_predictions = []
+
+    # run streamed inference
+    for prediction in tqdm(
+            pipe(
+                KeyDataset(common_voice_test, "audio"),
+                max_new_tokens=128,
+                generate_kwargs={"task": "transcribe"},
+                batch_size=32,
+            ),
+            total=len(common_voice_test),
+    ):
+        all_predictions.append(prediction["text"])
+
+    wer_ortho = 100 * wer_metric.compute(
+        references=common_voice_test["sentence"], predictions=all_predictions)
 
-pipe = pipeline(
-    "automatic-speech-recognition",
-    model=model,
-    torch_dtype=torch.bfloat16,
-    tokenizer=processor.tokenizer,
-    feature_extractor=processor.feature_extractor,
-    model_kwargs={"use_flash_attention_2": True},
-)
-
-wer_metric = load("wer")
-
-common_voice_test = load_dataset(
-    "mozilla-foundation/common_voice_17_0",  # mozilla-foundation/common_voice_17_0
-    "dv",
-    split="test")
-
-all_predictions = []
-
-# run streamed inference
-for prediction in tqdm(
-        pipe(
-            KeyDataset(common_voice_test, "audio"),
-            max_new_tokens=128,
-            generate_kwargs={"task": "transcribe"},
-            batch_size=32,
-        ),
-        total=len(common_voice_test),
-):
-    all_predictions.append(prediction["text"])
-
-wer_ortho = 100 * wer_metric.compute(references=common_voice_test["sentence"], predictions=all_predictions)
-
-print(f"WER: {wer_ortho:.2f}%")
+    print(f"WER: {wer_ortho:.2f}%")
+    return wer_ortho