Views
No views yet
1from datasets import load_dataset
2from transformers import AutoProcessor
3
4from optimum.neuron import NeuronModelForCTC, pipeline
5
6
7dataset = load_dataset("hf-internal-testing/librispeech_asr_demo", "clean", split="validation")
8dataset = dataset.sort("id")
9sampling_rate = dataset.features["audio"].sampling_rate
10
11# model_id = "hf-internal-testing/tiny-random-Wav2Vec2Model"
12model_id = "facebook/wav2vec2-large-960h-lv60-self"
13processor = AutoProcessor.from_pretrained(model_id)
14input_shapes = {"batch_size": 1, "audio_sequence_length": 100000}
15compiler_args = {"auto_cast": "matmul", "auto_cast_type": "bf16"}
16model = NeuronModelForCTC.from_pretrained(
17 model_id,
18 export=True,
19 disable_neuron_cache=True,
20 **input_shapes,
21 **compiler_args,
22)
23model.save_pretrained("wav2vec2_neuron")