|
| 1 | +import argparse |
| 2 | +import os |
| 3 | +import torch |
| 4 | +from torch.nn.attention import sdpa_kernel, SDPBackend |
| 5 | +from transformers import AutoConfig, AutoModel, AutoModelForCTC, AutoProcessor, MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING |
| 6 | +import evaluate |
| 7 | +from normalizer import data_utils |
| 8 | +import time |
| 9 | +from tqdm import tqdm |
| 10 | + |
| 11 | +wer_metric = evaluate.load("wer") |
| 12 | +torch.set_float32_matmul_precision('high') |
| 13 | + |
| 14 | +def main(args): |
| 15 | + model = AutoModel.from_pretrained(args.model_id, torch_dtype=torch.bfloat16, trust_remote_code=True).to(args.device) |
| 16 | + processor = AutoProcessor.from_pretrained("openai/whisper-large-v3") |
| 17 | + model_input_name = processor.model_input_names[0] |
| 18 | + |
| 19 | + if model.can_generate(): |
| 20 | + gen_kwargs = {"max_new_tokens": args.max_new_tokens} |
| 21 | + # for multilingual Whisper-checkpoints we see a definitive WER boost by setting the language and task args |
| 22 | + # print(model.generation_config) |
| 23 | + # if getattr(model.generation_config, "is_multilingual"): |
| 24 | + # gen_kwargs["language"] = "en" |
| 25 | + # gen_kwargs["task"] = "transcribe" |
| 26 | + elif args.max_new_tokens: |
| 27 | + raise ValueError("`max_new_tokens` should only be set for auto-regressive models, but got a CTC model.") |
| 28 | + |
| 29 | + if args.torch_compile: |
| 30 | + model.forward = torch.compile(model.forward, mode=args.compile_mode, fullgraph=True) |
| 31 | + if model.can_generate(): |
| 32 | + # enable static k/v cache for autoregressive models |
| 33 | + model.generation_config.cache_implementation = "static" |
| 34 | + |
| 35 | + def benchmark(batch, min_new_tokens=None): |
| 36 | + # Load audio inputs |
| 37 | + audios = [audio["array"] for audio in batch["audio"]] |
| 38 | + minibatch_size = len(audios) |
| 39 | + |
| 40 | + # START TIMING |
| 41 | + start_time = time.time() |
| 42 | + |
| 43 | + # 1. Pre-Processing |
| 44 | + # 1.1 Pad audios to max batch size if using torch compile to prevent re-compilations |
| 45 | + padding_size = None |
| 46 | + if minibatch_size != args.batch_size and args.torch_compile: |
| 47 | + padding_size = args.batch_size - minibatch_size |
| 48 | + padding_audios = [audios[-1] for _ in range(padding_size)] |
| 49 | + audios.extend(padding_audios) |
| 50 | + |
| 51 | + if not model.can_generate(): #or len(audios[0]) > processor.feature_extractor.n_samples: |
| 52 | + # 1.2 Either CTC pre-processing (normalize to mean 0, std 1), or long-form Whisper processing |
| 53 | + inputs = processor( |
| 54 | + audios, |
| 55 | + sampling_rate=16_000, |
| 56 | + truncation=False, |
| 57 | + padding="longest", |
| 58 | + return_tensors="pt", |
| 59 | + return_attention_mask=True, |
| 60 | + ) |
| 61 | + else: |
| 62 | + # 1.3 Standard Whisper processing: pad audios to 30-seconds and converted to log-mel |
| 63 | + inputs = processor(audios, sampling_rate=16_000, return_tensors="pt", device=args.device) |
| 64 | + |
| 65 | + inputs = inputs.to(args.device) |
| 66 | + inputs[model_input_name] = inputs[model_input_name].to(torch.bfloat16) |
| 67 | + |
| 68 | + # 2. Model Inference |
| 69 | + with sdpa_kernel(SDPBackend.MATH if args.torch_compile else SDPBackend.FLASH_ATTENTION): |
| 70 | + if model.can_generate(): |
| 71 | + # 2.1 Auto-regressive generation for encoder-decoder models |
| 72 | + pred_ids = model.generate(**inputs, **gen_kwargs, min_new_tokens=min_new_tokens) |
| 73 | + else: |
| 74 | + # 2.2. Single forward pass for CTC |
| 75 | + with torch.no_grad(): |
| 76 | + logits = model(**inputs).logits |
| 77 | + pred_ids = logits.argmax(-1) |
| 78 | + |
| 79 | + # 3. Post-processing |
| 80 | + # 3.1 Strip padded ids from predictions |
| 81 | + if padding_size is not None: |
| 82 | + pred_ids = pred_ids[:-padding_size, ...] |
| 83 | + |
| 84 | + # 3.2 Convert token ids to text transcription |
| 85 | + pred_text = processor.batch_decode(pred_ids, skip_special_tokens=True) |
| 86 | + |
| 87 | + # END TIMING |
| 88 | + runtime = time.time() - start_time |
| 89 | + |
| 90 | + # normalize by minibatch size since we want the per-sample time |
| 91 | + batch["transcription_time_s"] = minibatch_size * [runtime / minibatch_size] |
| 92 | + |
| 93 | + # normalize transcriptions with English normalizer |
| 94 | + batch["predictions"] = [data_utils.normalizer(pred) for pred in pred_text] |
| 95 | + batch["references"] = batch["norm_text"] |
| 96 | + return batch |
| 97 | + |
| 98 | + if args.warmup_steps is not None: |
| 99 | + dataset = data_utils.load_data(args) |
| 100 | + dataset = data_utils.prepare_data(dataset) |
| 101 | + |
| 102 | + num_warmup_samples = args.warmup_steps * args.batch_size |
| 103 | + if args.streaming: |
| 104 | + warmup_dataset = dataset.take(num_warmup_samples) |
| 105 | + else: |
| 106 | + warmup_dataset = dataset.select(range(min(num_warmup_samples, len(dataset)))) |
| 107 | + warmup_dataset = iter(warmup_dataset.map(benchmark, batch_size=args.batch_size, batched=True, fn_kwargs={"min_new_tokens": args.max_new_tokens})) |
| 108 | + |
| 109 | + for _ in tqdm(warmup_dataset, desc="Warming up..."): |
| 110 | + continue |
| 111 | + |
| 112 | + dataset = data_utils.load_data(args) |
| 113 | + if args.max_eval_samples is not None and args.max_eval_samples > 0: |
| 114 | + print(f"Subsampling dataset to first {args.max_eval_samples} samples!") |
| 115 | + if args.streaming: |
| 116 | + dataset = dataset.take(args.max_eval_samples) |
| 117 | + else: |
| 118 | + dataset = dataset.select(range(min(args.max_eval_samples, len(dataset)))) |
| 119 | + dataset = data_utils.prepare_data(dataset) |
| 120 | + |
| 121 | + dataset = dataset.map( |
| 122 | + benchmark, batch_size=args.batch_size, batched=True, remove_columns=["audio"], |
| 123 | + ) |
| 124 | + |
| 125 | + all_results = { |
| 126 | + "audio_length_s": [], |
| 127 | + "transcription_time_s": [], |
| 128 | + "predictions": [], |
| 129 | + "references": [], |
| 130 | + } |
| 131 | + result_iter = iter(dataset) |
| 132 | + for result in tqdm(result_iter, desc="Samples..."): |
| 133 | + for key in all_results: |
| 134 | + all_results[key].append(result[key]) |
| 135 | + |
| 136 | + # Write manifest results (WER and RTFX) |
| 137 | + manifest_path = data_utils.write_manifest( |
| 138 | + all_results["references"], |
| 139 | + all_results["predictions"], |
| 140 | + args.model_id, |
| 141 | + args.dataset_path, |
| 142 | + args.dataset, |
| 143 | + args.split, |
| 144 | + audio_length=all_results["audio_length_s"], |
| 145 | + transcription_time=all_results["transcription_time_s"], |
| 146 | + ) |
| 147 | + print("Results saved at path:", os.path.abspath(manifest_path)) |
| 148 | + |
| 149 | + wer = wer_metric.compute( |
| 150 | + references=all_results["references"], predictions=all_results["predictions"] |
| 151 | + ) |
| 152 | + wer = round(100 * wer, 2) |
| 153 | + rtfx = round(sum(all_results["audio_length_s"]) / sum(all_results["transcription_time_s"]), 2) |
| 154 | + print("WER:", wer, "%", "RTFx:", rtfx) |
| 155 | + |
| 156 | + |
| 157 | +if __name__ == "__main__": |
| 158 | + parser = argparse.ArgumentParser() |
| 159 | + |
| 160 | + parser.add_argument( |
| 161 | + "--model_id", |
| 162 | + type=str, |
| 163 | + required=True, |
| 164 | + help="Model identifier. Should be loadable with 🤗 Transformers", |
| 165 | + ) |
| 166 | + parser.add_argument( |
| 167 | + "--dataset_path", |
| 168 | + type=str, |
| 169 | + default="esb/datasets", |
| 170 | + help="Dataset path. By default, it is `esb/datasets`", |
| 171 | + ) |
| 172 | + parser.add_argument( |
| 173 | + "--dataset", |
| 174 | + type=str, |
| 175 | + required=True, |
| 176 | + help="Dataset name. *E.g.* `'librispeech_asr` for the LibriSpeech ASR dataset, or `'common_voice'` for Common Voice. The full list of dataset names " |
| 177 | + "can be found at `https://huggingface.co/datasets/esb/datasets`", |
| 178 | + ) |
| 179 | + parser.add_argument( |
| 180 | + "--split", |
| 181 | + type=str, |
| 182 | + default="test", |
| 183 | + help="Split of the dataset. *E.g.* `'validation'` for the dev split, or `'test'` for the test split.", |
| 184 | + ) |
| 185 | + parser.add_argument( |
| 186 | + "--device", |
| 187 | + type=int, |
| 188 | + default=-1, |
| 189 | + help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.", |
| 190 | + ) |
| 191 | + parser.add_argument( |
| 192 | + "--batch_size", |
| 193 | + type=int, |
| 194 | + default=16, |
| 195 | + help="Number of samples to go through each streamed batch.", |
| 196 | + ) |
| 197 | + parser.add_argument( |
| 198 | + "--max_eval_samples", |
| 199 | + type=int, |
| 200 | + default=None, |
| 201 | + help="Number of samples to be evaluated. Put a lower number e.g. 64 for testing this script.", |
| 202 | + ) |
| 203 | + parser.add_argument( |
| 204 | + "--no-streaming", |
| 205 | + dest="streaming", |
| 206 | + action="store_false", |
| 207 | + help="Choose whether you'd like to download the entire dataset or stream it during the evaluation.", |
| 208 | + ) |
| 209 | + parser.add_argument( |
| 210 | + "--max_new_tokens", |
| 211 | + type=int, |
| 212 | + default=None, |
| 213 | + help="Maximum number of tokens to generate (for auto-regressive models).", |
| 214 | + ) |
| 215 | + parser.add_argument( |
| 216 | + "--torch_compile", |
| 217 | + action="store_true", |
| 218 | + help="Whether to JIT compile the forward pass of the model.", |
| 219 | + ) |
| 220 | + parser.add_argument( |
| 221 | + "--compile_mode", |
| 222 | + type=str, |
| 223 | + default="max-autotune", |
| 224 | + help="Mode for torch compiling model forward pass. Can be either 'default', 'reduce-overhead', 'max-autotune' or 'max-autotune-no-cudagraphs'.", |
| 225 | + ) |
| 226 | + parser.add_argument( |
| 227 | + "--warmup_steps", |
| 228 | + type=int, |
| 229 | + default=10, |
| 230 | + help="Number of warm-up steps to run before launching the timed runs.", |
| 231 | + ) |
| 232 | + args = parser.parse_args() |
| 233 | + parser.set_defaults(streaming=False) |
| 234 | + |
| 235 | + main(args) |
0 commit comments