inference_hf.py

from transformers import  Wav2Vec2FeatureExtractor, Wav2Vec2Model
import sys
import torch
import torchinfo
hf_path = sys.argv[1]
audio_file = sys.argv[2]
extract = Wav2Vec2FeatureExtractor.from_pretrained(hf_path)
hf_model = Wav2Vec2Model.from_pretrained(hf_path)

torchinfo.summary(hf_model)

hf_model.eval()

import torchaudio

waveform, sample_rate = torchaudio.load(audio_file)
if waveform.shape[0] > 1:
    waveform = torch.mean(waveform, dim=0, keepdim=True)

with torch.no_grad():
    feat = extract(waveform.squeeze().numpy(), sampling_rate=sample_rate, return_tensors="pt", padding=True)
    out = hf_model(feat.input_values,output_hidden_states=True)
    last_hidden_states = out.last_hidden_state
    cnn_out = out.extract_features
    hidden_states = out.hidden_states
    print("CNN features shape:", cnn_out.shape)
    print("Hidden states length:", hidden_states.__len__())
    print("Last hidden states shape:", last_hidden_states.shape)


#normalize features for visualization
import numpy as np
last_hidden_states = (last_hidden_states - last_hidden_states.min()) / (last_hidden_states.max() - last_hidden_states.min())
cnn_out = [(feat - feat.min()) / (feat.max() - feat.min()) for feat in cnn_out]


#apply log scaling
last_hidden_states = torch.log1p(last_hidden_states * 100)
cnn_out = [torch.log1p(feat * 100) for feat in cnn_out]

#visualize output feature maps
import matplotlib.pyplot as plt
plt.figure(figsize=(12, 6))
plt.subplot(2, 1, 1)
plt.title("Last Hidden States")
plt.imshow(last_hidden_states[0].cpu().numpy().T, aspect='auto', origin='lower')
plt.colorbar()
plt.subplot(2, 1, 2)
plt.title("CNN Features")
plt.imshow(cnn_out[-1].cpu().numpy().T, aspect='auto', origin='lower')
plt.colorbar()
plt.tight_layout()
plt.savefig("hf_wav2vec2_features.png")
plt.close()