NeMo SALM scripts

pzelasko · pzelasko · commit aacbb0ca5f46 · 2025-07-07T07:07:47.000-07:00
Signed-off-by: Piotr Żelasko &lt;pzelasko@nvidia.com&gt;
diff --git a/nemo_asr/run_eval_salm.py b/nemo_asr/run_eval_salm.py
@@ -0,0 +1,258 @@
+import argparse
+
+import io
+import os
+import torch
+import evaluate
+import soundfile
+import lhotse
+
+from tqdm import tqdm
+from normalizer import data_utils
+import numpy as np
+
+from nemo.collections.asr.models import ASRModel
+import time
+
+
+from nemo.collections.speechlm2.models.salm import SALM
+from omegaconf import OmegaConf
+from pathlib import Path
+from transformers import GenerationConfig
+
+
+
+wer_metric = evaluate.load("wer")
+
+
+class ToAudio(torch.utils.data.Dataset):
+    def __getitem__(self, cuts):
+        audios, audio_lens = cuts.load_audio(collate=True)
+        return {"cuts": cuts, "audios": audios, "audio_lens": audio_lens}
+
+
+def setup_dloader(audio_files, batch_size, num_workers):
+    cuts = lhotse.CutSet([lhotse.Recording.from_file(p).to_cut() for p in audio_files])
+    return torch.utils.data.DataLoader(
+        dataset=ToAudio(),
+        sampler=lhotse.dataset.DynamicCutSampler(cuts, max_cuts=batch_size),
+        num_workers=num_workers,
+        batch_size=None,
+    )
+
+
+def transcribe(model, dloader) -> list[str]:
+    hyps = []
+    eos_tokens = torch.tensor([model.text_eos_id])
+    for batch_idx, batch in enumerate(dloader):
+        answer_ids = model.generate(
+            prompts=[
+                [
+                    {"role": "user", "slots": {"message": f"Transcribe the following: {model.audio_locator_tag}"}}
+                ]
+            ] * len(batch["cuts"]),
+            audios=batch["audios"].to(model.device, non_blocking=True),
+            audio_lens=batch["audio_lens"].to(model.device, non_blocking=True),
+            generation_config=GenerationConfig(
+                max_new_tokens=128,
+                bos_token_id=model.text_bos_id,
+                eos_token_id=eos_tokens,
+                pad_token_id=model.text_pad_id,
+            ),
+        )
+        answer_ids = [parse_hyp(ans, eos_tokens) for ans in answer_ids.cpu()]
+        hyps.extend(model.tokenizer.ids_to_text(ans).strip() for ans in answer_ids)
+    return hyps
+
+
+def parse_hyp(answer: torch.Tensor, eos_tokens):
+    end = (answer == torch.isin(answer, eos_tokens)).nonzero(as_tuple=True)[0]
+    if end.numel() == 0:
+        return answer
+    end = end[0]
+    return answer[:end]
+
+
+def main(args):
+
+    DATA_CACHE_DIR = os.path.join(os.getcwd(), "audio_cache")
+    DATASET_NAME = args.dataset
+    SPLIT_NAME = args.split
+
+    CACHE_DIR = os.path.join(DATA_CACHE_DIR, DATASET_NAME, SPLIT_NAME)
+    if not os.path.exists(CACHE_DIR):
+        os.makedirs(CACHE_DIR)
+
+    torch.set_float32_matmul_precision("medium")
+
+    device = torch.device(f"cuda:{args.device}")
+    model = SALM.from_pretrained(args.model_id).eval().to(torch.bfloat16).to(device)
+
+    dataset = data_utils.load_data(args)
+
+    def download_audio_files(batch):
+
+        # download audio files and write the paths, transcriptions and durations to a manifest file
+        audio_paths = []
+        durations = []
+
+        for id, sample in zip(batch["id"], batch["audio"]):
+
+            # first step added here to make ID and wav filenames unique
+            # several datasets like earnings22 have a hierarchical structure
+            # for eg. earnings22/test/4432298/281.wav, earnings22/test/4450488/281.wav
+            # lhotse uses the filename (281.wav) here as unique ID to create and name cuts
+            # ref: https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/collation.py#L186
+            id = id.replace('/', '_').removesuffix('.wav')
+
+            audio_path = os.path.join(CACHE_DIR, f"{id}.wav")
+
+            if "array" in sample:
+                audio_array = np.float32(sample["array"])
+                sample_rate = 16000
+
+            elif "bytes" in sample: # added to be compatible with latest datasets library (3.x.x) that produces byte stream
+                with io.BytesIO(sample["bytes"]) as audio_file:
+                    audio_array, sample_rate = soundfile.read(audio_file, dtype="float32")
+
+            else:
+                raise ValueError("Sample must have either 'array' or 'bytes' key")
+
+            if not os.path.exists(audio_path):
+                os.makedirs(os.path.dirname(audio_path), exist_ok=True)
+                soundfile.write(audio_path, audio_array, sample_rate)
+
+            audio_paths.append(audio_path)
+            durations.append(len(audio_array) / sample_rate)
+
+
+        batch["references"] = batch["norm_text"]
+        batch["audio_filepaths"] = audio_paths
+        batch["durations"] = durations
+
+        return batch
+
+
+    if args.max_eval_samples is not None and args.max_eval_samples > 0:
+        print(f"Subsampling dataset to first {args.max_eval_samples} samples !")
+        dataset = dataset.take(args.max_eval_samples)
+
+    dataset = data_utils.prepare_data(dataset)
+
+    # prepraing the offline dataset
+    dataset = dataset.map(download_audio_files, batch_size=args.batch_size, batched=True, remove_columns=["audio"])
+
+    # Write manifest from daraset batch using json and keys audio_filepath, duration, text
+
+    all_data = {
+        "audio_filepaths": [],
+        "durations": [],
+        "references": [],
+    }
+
+    data_itr = iter(dataset)
+    for data in tqdm(data_itr, desc="Downloading Samples"):
+        for key in all_data:
+            all_data[key].append(data[key])
+
+    # Sort audio_filepaths and references based on durations values
+    sorted_indices = sorted(range(len(all_data["durations"])), key=lambda k: all_data["durations"][k], reverse=True)
+    all_data["audio_filepaths"] = [all_data["audio_filepaths"][i] for i in sorted_indices]
+    all_data["references"] = [all_data["references"][i] for i in sorted_indices]
+    all_data["durations"] = [all_data["durations"][i] for i in sorted_indices]
+
+
+    total_time = 0
+    for _ in range(2): # warmup once and calculate rtf
+        if _ == 0:
+            audio_files = all_data["audio_filepaths"][:args.batch_size * 4] # warmup with 4 batches
+        else:
+            audio_files = all_data["audio_filepaths"]
+        dloader = setup_dloader(audio_files=audio_files, batch_size=args.batch_size, num_workers=1)
+        with torch.inference_mode():
+            start_time = time.time()
+            transcriptions = transcribe(model, dloader)
+            end_time = time.time()
+        if _ == 1:
+            total_time += end_time - start_time
+    total_time = total_time
+
+    # normalize transcriptions with English normalizer
+    if isinstance(transcriptions, tuple) and len(transcriptions) == 2:
+        transcriptions = transcriptions[0]
+    predictions = [data_utils.normalizer(pred) for pred in transcriptions]
+
+    avg_time = total_time / len(all_data["audio_filepaths"])
+
+    # Write manifest results (WER and RTFX)
+    manifest_path = data_utils.write_manifest(
+        all_data["references"],
+        predictions,
+        args.model_id,
+        args.dataset_path,
+        args.dataset,
+        args.split,
+        audio_length=all_data["durations"],
+        transcription_time=[avg_time] * len(all_data["audio_filepaths"]),
+    )
+
+    print("Results saved at path:", os.path.abspath(manifest_path))
+
+    wer = wer_metric.compute(references=all_data['references'], predictions=predictions)
+    wer = round(100 * wer, 2)
+
+    audio_length = sum(all_data["durations"])
+    rtfx = audio_length / total_time
+    rtfx = round(rtfx, 2)
+
+    print("RTFX:", rtfx)
+    print("WER:", wer, "%")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with NVIDIA NeMo.",
+    )
+    parser.add_argument(
+        '--dataset_path', type=str, default='esb/datasets', help='Dataset path. By default, it is `esb/datasets`'
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name. *E.g.* `'librispeech_asr` for the LibriSpeech ASR dataset, or `'common_voice'` for Common Voice. The full list of dataset names "
+        "can be found at `https://huggingface.co/datasets/esb/datasets`",
+    )
+    parser.add_argument(
+        "--split",
+        type=str,
+        default="test",
+        help="Split of the dataset. *E.g.* `'validation`' for the dev split, or `'test'` for the test split.",
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=-1,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    parser.add_argument(
+        "--batch_size", type=int, default=32, help="Number of samples to go through each streamed batch.",
+    )
+    parser.add_argument(
+        "--max_eval_samples",
+        type=int,
+        default=None,
+        help="Number of samples to be evaluated. Put a lower number e.g. 64 for testing this script.",
+    )
+    parser.add_argument(
+        "--no-streaming",
+        dest='streaming',
+        action="store_false",
+        help="Choose whether you'd like to download the entire dataset or stream it during the evaluation.",
+    )
+    args = parser.parse_args()
+    parser.set_defaults(streaming=True)
+
+    main(args)
diff --git a/nemo_asr/run_salm.sh b/nemo_asr/run_salm.sh
@@ -0,0 +1,95 @@
+#!/bin/bash
+
+export PYTHONPATH="..":$PYTHONPATH
+
+MODEL_IDs=(
+    nvidia/canary-qwen-2.5b
+)
+BATCH_SIZE=192
+DEVICE_ID=0
+
+num_models=${#MODEL_IDs[@]}
+
+for (( i=0; i<${num_models}; i++ ));
+do
+    MODEL_ID=${MODEL_IDs[$i]}
+
+    python run_eval_salm.py \
+        --model_id=${MODEL_ID} \
+        --dataset_path="hf-audio/esb-datasets-test-only-sorted" \
+        --dataset="ami" \
+        --split="test" \
+        --device=${DEVICE_ID} \
+        --batch_size=${BATCH_SIZE} \
+        --max_eval_samples=-1
+
+    python run_eval_salm.py \
+        --model_id=${MODEL_ID} \
+        --dataset_path="hf-audio/esb-datasets-test-only-sorted" \
+        --dataset="earnings22" \
+        --split="test" \
+        --device=${DEVICE_ID} \
+        --batch_size=${BATCH_SIZE} \
+        --max_eval_samples=-1
+
+    python run_eval_salm.py \
+        --model_id=${MODEL_ID} \
+        --dataset_path="hf-audio/esb-datasets-test-only-sorted" \
+        --dataset="gigaspeech" \
+        --split="test" \
+        --device=${DEVICE_ID} \
+        --batch_size=${BATCH_SIZE} \
+        --max_eval_samples=-1
+
+    python run_eval_salm.py \
+        --model_id=${MODEL_ID} \
+        --dataset_path="hf-audio/esb-datasets-test-only-sorted" \
+        --dataset="librispeech" \
+        --split="test.clean" \
+        --device=${DEVICE_ID} \
+        --batch_size=${BATCH_SIZE} \
+        --max_eval_samples=-1
+
+    python run_eval_salm.py \
+        --model_id=${MODEL_ID} \
+        --dataset_path="hf-audio/esb-datasets-test-only-sorted" \
+        --dataset="librispeech" \
+        --split="test.other" \
+        --device=${DEVICE_ID} \
+        --batch_size=${BATCH_SIZE} \
+        --max_eval_samples=-1
+
+    python run_eval_salm.py \
+        --model_id=${MODEL_ID} \
+        --dataset_path="hf-audio/esb-datasets-test-only-sorted" \
+        --dataset="spgispeech" \
+        --split="test" \
+        --device=${DEVICE_ID} \
+        --batch_size=${BATCH_SIZE} \
+        --max_eval_samples=-1
+
+    python run_eval_salm.py \
+        --model_id=${MODEL_ID} \
+        --dataset_path="hf-audio/esb-datasets-test-only-sorted" \
+        --dataset="tedlium" \
+        --split="test" \
+        --device=${DEVICE_ID} \
+        --batch_size=${BATCH_SIZE} \
+        --max_eval_samples=-1
+
+    python run_eval_salm.py \
+        --model_id=${MODEL_ID} \
+        --dataset_path="hf-audio/esb-datasets-test-only-sorted" \
+        --dataset="voxpopuli" \
+        --split="test" \
+        --device=${DEVICE_ID} \
+        --batch_size=${BATCH_SIZE} \
+        --max_eval_samples=-1
+
+    # Evaluate results
+    RUNDIR=`pwd` && \
+    cd ../normalizer && \
+    python -c "import eval_utils; eval_utils.score_results('${RUNDIR}/results', '${MODEL_ID}')" && \
+    cd $RUNDIR
+
+done
