Upload chula_gino_parkinson.py
Browse files- chula_gino_parkinson.py +881 -0
chula_gino_parkinson.py
ADDED
|
@@ -0,0 +1,881 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# -*- coding: utf-8 -*-
|
| 2 |
+
"""CHULA Gino_Parkinson.ipynb
|
| 3 |
+
|
| 4 |
+
Automatically generated by Colaboratory.
|
| 5 |
+
|
| 6 |
+
Original file is located at
|
| 7 |
+
https://colab.research.google.com/drive/1XPgGZILiBbDji5G0dHoFV7OQaUwGM3HJ
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
!pip install SoundFile transformers scikit-learn
|
| 11 |
+
|
| 12 |
+
from google.colab import drive
|
| 13 |
+
drive.mount('/content/drive')
|
| 14 |
+
|
| 15 |
+
import matplotlib.pyplot as plt
|
| 16 |
+
import numpy as np
|
| 17 |
+
|
| 18 |
+
import os
|
| 19 |
+
import soundfile as sf
|
| 20 |
+
import torch
|
| 21 |
+
import torch.nn as nn
|
| 22 |
+
import torch.nn.functional as F
|
| 23 |
+
from torch.utils.data import Dataset, DataLoader
|
| 24 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor, Wav2Vec2ForSequenceClassification
|
| 25 |
+
from sklearn.model_selection import train_test_split
|
| 26 |
+
import re
|
| 27 |
+
from collections import Counter
|
| 28 |
+
from sklearn.metrics import classification_report
|
| 29 |
+
|
| 30 |
+
# Custom Dataset class
|
| 31 |
+
class DysarthriaDataset(Dataset):
|
| 32 |
+
def __init__(self, data, labels, max_length=100000):
|
| 33 |
+
self.data = data
|
| 34 |
+
self.labels = labels
|
| 35 |
+
self.max_length = max_length
|
| 36 |
+
self.processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
|
| 37 |
+
|
| 38 |
+
def __len__(self):
|
| 39 |
+
return len(self.data)
|
| 40 |
+
|
| 41 |
+
def __getitem__(self, idx):
|
| 42 |
+
try:
|
| 43 |
+
wav_data, _ = sf.read(self.data[idx])
|
| 44 |
+
except:
|
| 45 |
+
print(f"Error opening file: {self.data[idx]}. Skipping...")
|
| 46 |
+
return self.__getitem__((idx + 1) % len(self.data))
|
| 47 |
+
inputs = self.processor(wav_data, sampling_rate=16000, return_tensors="pt", padding=True)
|
| 48 |
+
input_values = inputs.input_values.squeeze(0) # Squeeze the batch dimension
|
| 49 |
+
if self.max_length - input_values.shape[-1] > 0:
|
| 50 |
+
input_values = torch.cat([input_values, torch.zeros((self.max_length - input_values.shape[-1],))], dim=-1)
|
| 51 |
+
else:
|
| 52 |
+
input_values = input_values[:self.max_length]
|
| 53 |
+
|
| 54 |
+
# Remove unsqueezing the channel dimension
|
| 55 |
+
# input_values = input_values.unsqueeze(0)
|
| 56 |
+
|
| 57 |
+
# label = torch.zeros(32,dtype=torch.long)
|
| 58 |
+
# label[self.labels[idx]] = 1
|
| 59 |
+
|
| 60 |
+
### CHANGES: simply return the label as a single integer
|
| 61 |
+
return {"input_values": input_values}, self.labels[idx]
|
| 62 |
+
# return {"input_values": input_values, "audio_path": self.data[idx]}, self.labels[idx]
|
| 63 |
+
###
|
| 64 |
+
|
| 65 |
+
def train(model, dataloader, criterion, optimizer, device, loss_vals, epochs, current_epoch):
|
| 66 |
+
model.train()
|
| 67 |
+
running_loss = 0
|
| 68 |
+
|
| 69 |
+
for i, (inputs, labels) in enumerate(dataloader):
|
| 70 |
+
inputs = {key: value.squeeze().to(device) for key, value in inputs.items()}
|
| 71 |
+
labels = labels.to(device)
|
| 72 |
+
|
| 73 |
+
optimizer.zero_grad()
|
| 74 |
+
logits = model(**inputs).logits
|
| 75 |
+
loss = criterion(logits, labels)
|
| 76 |
+
loss.backward()
|
| 77 |
+
optimizer.step()
|
| 78 |
+
|
| 79 |
+
# append loss value to list
|
| 80 |
+
loss_vals.append(loss.item())
|
| 81 |
+
running_loss += loss.item()
|
| 82 |
+
|
| 83 |
+
if i % 10 == 0: # Update the plot every 10 iterations
|
| 84 |
+
plt.clf() # Clear the previous plot
|
| 85 |
+
plt.plot(loss_vals)
|
| 86 |
+
plt.xlim([0, len(dataloader)*epochs])
|
| 87 |
+
plt.ylim([0, max(loss_vals) + 2])
|
| 88 |
+
plt.xlabel('Training Iterations')
|
| 89 |
+
plt.ylabel('Loss')
|
| 90 |
+
plt.title(f"Training Loss at Epoch {current_epoch + 1}")
|
| 91 |
+
plt.pause(0.001) # Pause to update the plot
|
| 92 |
+
|
| 93 |
+
avg_loss = running_loss / len(dataloader)
|
| 94 |
+
print(f"Average Loss after Epoch {current_epoch + 1}: {avg_loss}\n")
|
| 95 |
+
return avg_loss
|
| 96 |
+
|
| 97 |
+
def predict(model, file_path, processor, device, max_length=100000): ### CHANGES: added max_length as an argument.
|
| 98 |
+
model.eval()
|
| 99 |
+
with torch.no_grad():
|
| 100 |
+
wav_data, _ = sf.read(file_path)
|
| 101 |
+
inputs = processor(wav_data, sampling_rate=16000, return_tensors="pt", padding=True)
|
| 102 |
+
# inputs = {key: value.squeeze().to(device) for key, value in inputs.items()}
|
| 103 |
+
|
| 104 |
+
### NEW CODES HERE
|
| 105 |
+
input_values = inputs.input_values.squeeze(0) # Squeeze the batch dimension
|
| 106 |
+
if max_length - input_values.shape[-1] > 0:
|
| 107 |
+
input_values = torch.cat([input_values, torch.zeros((max_length - input_values.shape[-1],))], dim=-1)
|
| 108 |
+
else:
|
| 109 |
+
input_values = input_values[:max_length]
|
| 110 |
+
input_values = input_values.unsqueeze(0).to(device)
|
| 111 |
+
inputs = {"input_values": input_values}
|
| 112 |
+
###
|
| 113 |
+
|
| 114 |
+
logits = model(**inputs).logits
|
| 115 |
+
# _, predicted = torch.max(logits, dim=0)
|
| 116 |
+
|
| 117 |
+
### NEW CODES HERE
|
| 118 |
+
# Remove the batch dimension.
|
| 119 |
+
logits = logits.squeeze()
|
| 120 |
+
predicted_class_id = torch.argmax(logits, dim=-1).item()
|
| 121 |
+
###
|
| 122 |
+
|
| 123 |
+
# return predicted.item()
|
| 124 |
+
return predicted_class_id
|
| 125 |
+
|
| 126 |
+
def evaluate(model, dataloader, criterion, device):
|
| 127 |
+
model.eval()
|
| 128 |
+
running_loss = 0
|
| 129 |
+
correct_predictions = 0
|
| 130 |
+
total_predictions = 0
|
| 131 |
+
wrong_files = []
|
| 132 |
+
all_labels = []
|
| 133 |
+
all_predictions = []
|
| 134 |
+
|
| 135 |
+
with torch.no_grad():
|
| 136 |
+
for inputs, labels in dataloader:
|
| 137 |
+
inputs = {key: value.squeeze().to(device) for key, value in inputs.items()}
|
| 138 |
+
labels = labels.to(device)
|
| 139 |
+
|
| 140 |
+
logits = model(**inputs).logits
|
| 141 |
+
loss = criterion(logits, labels)
|
| 142 |
+
running_loss += loss.item()
|
| 143 |
+
|
| 144 |
+
_, predicted = torch.max(logits, 1)
|
| 145 |
+
correct_predictions += (predicted == labels).sum().item()
|
| 146 |
+
total_predictions += labels.size(0)
|
| 147 |
+
|
| 148 |
+
wrong_idx = (predicted != labels).nonzero().squeeze().cpu().numpy()
|
| 149 |
+
if wrong_idx.ndim > 0:
|
| 150 |
+
for idx in wrong_idx:
|
| 151 |
+
wrong_files.append(dataloader.dataset.data[idx])
|
| 152 |
+
elif wrong_idx.size > 0:
|
| 153 |
+
wrong_files.append(dataloader.dataset.data[wrong_idx])
|
| 154 |
+
|
| 155 |
+
all_labels.extend(labels.cpu().numpy())
|
| 156 |
+
all_predictions.extend(predicted.cpu().numpy())
|
| 157 |
+
|
| 158 |
+
avg_loss = running_loss / len(dataloader)
|
| 159 |
+
accuracy = correct_predictions / total_predictions
|
| 160 |
+
|
| 161 |
+
return avg_loss, accuracy, wrong_files, np.array(all_labels), np.array(all_predictions)
|
| 162 |
+
|
| 163 |
+
def get_wav_files(base_path):
|
| 164 |
+
wav_files = []
|
| 165 |
+
for subject_folder in os.listdir(base_path):
|
| 166 |
+
subject_path = os.path.join(base_path, subject_folder)
|
| 167 |
+
if os.path.isdir(subject_path):
|
| 168 |
+
for wav_file in os.listdir(subject_path):
|
| 169 |
+
if wav_file.endswith('.wav'):
|
| 170 |
+
wav_files.append(os.path.join(subject_path, wav_file))
|
| 171 |
+
|
| 172 |
+
return wav_files
|
| 173 |
+
|
| 174 |
+
def get_torgo_data(dysarthria_path, non_dysarthria_path):
|
| 175 |
+
dysarthria_files = [os.path.join(dysarthria_path, f) for f in os.listdir(dysarthria_path) if f.endswith('.wav')]
|
| 176 |
+
non_dysarthria_files = [os.path.join(non_dysarthria_path, f) for f in os.listdir(non_dysarthria_path) if f.endswith('.wav')]
|
| 177 |
+
|
| 178 |
+
data = dysarthria_files + non_dysarthria_files
|
| 179 |
+
labels = [1] * len(dysarthria_files) + [0] * len(non_dysarthria_files)
|
| 180 |
+
|
| 181 |
+
train_data, test_data, train_labels, test_labels = train_test_split(data, labels, test_size=0.2, stratify=labels)
|
| 182 |
+
train_data, val_data, train_labels, val_labels = train_test_split(train_data, train_labels, test_size=0.25, stratify=train_labels) # 0.25 x 0.8 = 0.2
|
| 183 |
+
|
| 184 |
+
return train_data, val_data, test_data, train_labels, val_labels, test_labels
|
| 185 |
+
|
| 186 |
+
dysarthria_path = "/content/drive/MyDrive/RECORDINGS_ANALYSIS/SP_ANALYSIS"
|
| 187 |
+
non_dysarthria_path = "/content/drive/MyDrive/RECORDINGS_ANALYSIS/CT_ANALYSIS"
|
| 188 |
+
|
| 189 |
+
dysarthria_files = get_wav_files(dysarthria_path)
|
| 190 |
+
non_dysarthria_files = get_wav_files(non_dysarthria_path)
|
| 191 |
+
|
| 192 |
+
|
| 193 |
+
|
| 194 |
+
data = dysarthria_files + non_dysarthria_files
|
| 195 |
+
labels = [1] * len(dysarthria_files) + [0] * len(non_dysarthria_files)
|
| 196 |
+
|
| 197 |
+
train_data, test_data, train_labels, test_labels = train_test_split(data, labels, test_size=0.2, stratify=labels)
|
| 198 |
+
train_data, val_data, train_labels, val_labels = train_test_split(train_data, train_labels, test_size=0.25, stratify=train_labels) # 0.25 x 0.8 = 0.2
|
| 199 |
+
train_dataset = DysarthriaDataset(train_data, train_labels)
|
| 200 |
+
test_dataset = DysarthriaDataset(test_data, test_labels)
|
| 201 |
+
val_dataset = DysarthriaDataset(val_data, val_labels) # Create a validation dataset
|
| 202 |
+
|
| 203 |
+
train_loader = DataLoader(train_dataset, batch_size=16, drop_last=False)
|
| 204 |
+
test_loader = DataLoader(test_dataset, batch_size=16, drop_last=False)
|
| 205 |
+
validation_loader = DataLoader(val_dataset, batch_size=16, drop_last=False) # Use the validation dataset for the validation_loader
|
| 206 |
+
|
| 207 |
+
""" dysarthria_path = "/content/drive/MyDrive/torgo_data/dysarthria_male/training"
|
| 208 |
+
non_dysarthria_path = "/content/drive/MyDrive/torgo_data/non_dysarthria_male/training"
|
| 209 |
+
|
| 210 |
+
dysarthria_files = [os.path.join(dysarthria_path, f) for f in os.listdir(dysarthria_path) if f.endswith('.wav')]
|
| 211 |
+
non_dysarthria_files = [os.path.join(non_dysarthria_path, f) for f in os.listdir(non_dysarthria_path) if f.endswith('.wav')]
|
| 212 |
+
|
| 213 |
+
data = dysarthria_files + non_dysarthria_files
|
| 214 |
+
labels = [1] * len(dysarthria_files) + [0] * len(non_dysarthria_files)
|
| 215 |
+
|
| 216 |
+
train_data, test_data, train_labels, test_labels = train_test_split(data, labels, test_size=0.2)
|
| 217 |
+
|
| 218 |
+
train_dataset = DysarthriaDataset(train_data, train_labels)
|
| 219 |
+
test_dataset = DysarthriaDataset(test_data, test_labels)
|
| 220 |
+
|
| 221 |
+
train_loader = DataLoader(train_dataset, batch_size=8, drop_last=True)
|
| 222 |
+
test_loader = DataLoader(test_dataset, batch_size=8, drop_last=True)
|
| 223 |
+
validation_loader = DataLoader(test_dataset, batch_size=8, drop_last=True)
|
| 224 |
+
|
| 225 |
+
dysarthria_validation_path = "/content/drive/MyDrive/torgo_data/dysarthria_male/validation"
|
| 226 |
+
non_dysarthria_validation_path = "/content/drive/MyDrive/torgo_data/non_dysarthria_male/validation"
|
| 227 |
+
|
| 228 |
+
dysarthria_validation_files = [os.path.join(dysarthria_validation_path, f) for f in os.listdir(dysarthria_validation_path) if f.endswith('.wav')]
|
| 229 |
+
non_dysarthria_validation_files = [os.path.join(non_dysarthria_validation_path, f) for f in os.listdir(non_dysarthria_validation_path) if f.endswith('.wav')]
|
| 230 |
+
|
| 231 |
+
validation_data = dysarthria_validation_files + non_dysarthria_validation_files
|
| 232 |
+
validation_labels = [1] * len(dysarthria_validation_files) + [0] * len(non_dysarthria_validation_files)"""
|
| 233 |
+
|
| 234 |
+
|
| 235 |
+
|
| 236 |
+
|
| 237 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 238 |
+
|
| 239 |
+
|
| 240 |
+
|
| 241 |
+
|
| 242 |
+
|
| 243 |
+
|
| 244 |
+
|
| 245 |
+
|
| 246 |
+
|
| 247 |
+
|
| 248 |
+
|
| 249 |
+
|
| 250 |
+
# model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h").to(device)
|
| 251 |
+
# model.classifier = nn.Linear(model.config.hidden_size, 2).to(device)
|
| 252 |
+
|
| 253 |
+
### NEW CODES
|
| 254 |
+
# It seems like the classifier layer is excluded from the model's forward method (i.e., model(**inputs)).
|
| 255 |
+
# That's why the number of labels in the output was 32 instead of 2 even when you had already changed the classifier.
|
| 256 |
+
# Instead, huggingface offers the option for loading the Wav2Vec model with an adjustable classifier head on top (by setting num_labels).
|
| 257 |
+
|
| 258 |
+
model = Wav2Vec2ForSequenceClassification.from_pretrained("facebook/wav2vec2-base-960h", num_labels=2).to(device)
|
| 259 |
+
##
|
| 260 |
+
model_path = "/content/dysarthria_classifier1.pth"
|
| 261 |
+
if os.path.exists(model_path):
|
| 262 |
+
print(f"Loading saved model {model_path}")
|
| 263 |
+
model.load_state_dict(torch.load(model_path))
|
| 264 |
+
|
| 265 |
+
criterion = nn.CrossEntropyLoss()
|
| 266 |
+
optimizer = torch.optim.Adam(model.parameters(), lr=1e-5)
|
| 267 |
+
|
| 268 |
+
from torch.optim.lr_scheduler import StepLR
|
| 269 |
+
|
| 270 |
+
scheduler = StepLR(optimizer, step_size=5, gamma=0.1)
|
| 271 |
+
|
| 272 |
+
# dysarthria_validation_path = "/content/drive/MyDrive/RECORDINGS_ANALYSIS/SP_ANALYSIS/testing"
|
| 273 |
+
# non_dysarthria_validation_path = "/content/drive/MyDrive/RECORDINGS_ANALYSIS/CT_ANALYSIS/testing"
|
| 274 |
+
|
| 275 |
+
#dysarthria_validation_files = get_wav_files(dysarthria_validation_path)
|
| 276 |
+
# non_dysarthria_validation_files = get_wav_files(non_dysarthria_validation_path)
|
| 277 |
+
|
| 278 |
+
#validation_data = dysarthria_validation_files + non_dysarthria_validation_files
|
| 279 |
+
#validation_labels = [1] * len(dysarthria_validation_files) + [0] * len(non_dysarthria_validation_files)
|
| 280 |
+
|
| 281 |
+
epochs = 10
|
| 282 |
+
plt.ion()
|
| 283 |
+
fig, ax = plt.subplots()
|
| 284 |
+
x_vals = np.arange(len(train_loader)*epochs)
|
| 285 |
+
loss_vals = []
|
| 286 |
+
for epoch in range(epochs):
|
| 287 |
+
train_loss = train(model, train_loader, criterion, optimizer, device, loss_vals, epochs, epoch)
|
| 288 |
+
print(f"Epoch {epoch + 1}, Train Loss: {train_loss}")
|
| 289 |
+
|
| 290 |
+
val_loss, val_accuracy, wrong_files, true_labels, pred_labels = evaluate(model, validation_loader, criterion, device)
|
| 291 |
+
print(f"Epoch {epoch + 1}, Validation Loss: {val_loss}, Validation Accuracy: {val_accuracy:.2f}")
|
| 292 |
+
print("Misclassified Files")
|
| 293 |
+
for file_path in wrong_files:
|
| 294 |
+
print(file_path)
|
| 295 |
+
|
| 296 |
+
|
| 297 |
+
sentence_pattern = re.compile(r"_(\d+)\.wav$")
|
| 298 |
+
|
| 299 |
+
sentence_counts = Counter()
|
| 300 |
+
for file_path in wrong_files:
|
| 301 |
+
match = sentence_pattern.search(file_path)
|
| 302 |
+
if match:
|
| 303 |
+
sentence_number = int(match.group(1))
|
| 304 |
+
sentence_counts[sentence_number] += 1
|
| 305 |
+
|
| 306 |
+
total_wrong = len(wrong_files)
|
| 307 |
+
print("Total wrong files:", total_wrong)
|
| 308 |
+
print()
|
| 309 |
+
|
| 310 |
+
for sentence_number, count in sentence_counts.most_common():
|
| 311 |
+
percent = count / total_wrong * 100
|
| 312 |
+
print(f"Sentence {sentence_number}: {count} ({percent:.2f}%)")
|
| 313 |
+
scheduler.step()
|
| 314 |
+
print(classification_report(true_labels, pred_labels, target_names=['non_dysarthria', 'dysarthria']))
|
| 315 |
+
audio_file = "/content/drive/MyDrive/torgo_data/dysarthria_male/validation/M01_Session1_0005.wav"
|
| 316 |
+
predicted_label = predict(model, audio_file, train_dataset.processor, device)
|
| 317 |
+
print(f"Predicted label: {predicted_label}")
|
| 318 |
+
|
| 319 |
+
|
| 320 |
+
|
| 321 |
+
|
| 322 |
+
|
| 323 |
+
# Test on a specific audio file
|
| 324 |
+
##audio_file = "/content/drive/MyDrive/torgo_data/dysarthria_male/validation/M01_Session1_0005.wav"
|
| 325 |
+
##predicted_label = predict(model, audio_file, train_dataset.processor, device)
|
| 326 |
+
##print(f"Predicted label: {predicted_label}")
|
| 327 |
+
|
| 328 |
+
torch.save(model.state_dict(), "dysarthria_classifier1.pth")
|
| 329 |
+
print("Predicting...")
|
| 330 |
+
|
| 331 |
+
"""#audio aug"""
|
| 332 |
+
|
| 333 |
+
!pip install audiomentations
|
| 334 |
+
from audiomentations import Compose, PitchShift, TimeStretch
|
| 335 |
+
|
| 336 |
+
augmenter = Compose([
|
| 337 |
+
PitchShift(min_semitones=-2, max_semitones=2, p=0.1),
|
| 338 |
+
TimeStretch(min_rate=0.9, max_rate=1.1, p=0.1)
|
| 339 |
+
])
|
| 340 |
+
|
| 341 |
+
# from torch.optim.lr_scheduler import StepLR
|
| 342 |
+
|
| 343 |
+
# scheduler = StepLR(optimizer, step_size=2, gamma=0.5)
|
| 344 |
+
|
| 345 |
+
from transformers import get_linear_schedule_with_warmup
|
| 346 |
+
|
| 347 |
+
# Define the total number of training steps
|
| 348 |
+
# It is usually the number of epochs times the number of batches per epoch
|
| 349 |
+
num_training_steps = epochs * len(train_loader)
|
| 350 |
+
|
| 351 |
+
# Define the number of warmup steps
|
| 352 |
+
# Usually set to a fraction of total_training_steps such as 0.1 * num_training_steps
|
| 353 |
+
num_warmup_steps = int(num_training_steps * 0.3)
|
| 354 |
+
|
| 355 |
+
# Create the learning rate scheduler
|
| 356 |
+
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)
|
| 357 |
+
|
| 358 |
+
model = Wav2Vec2ForSequenceClassification.from_pretrained("facebook/wav2vec2-base-960h", num_labels=2).to(device)
|
| 359 |
+
##
|
| 360 |
+
model_path = "/content/models/my_model_06/pytorch_model.bin"
|
| 361 |
+
if os.path.exists(model_path):
|
| 362 |
+
print(f"Loading saved model {model_path}")
|
| 363 |
+
model.load_state_dict(torch.load(model_path))
|
| 364 |
+
|
| 365 |
+
criterion = nn.CrossEntropyLoss()
|
| 366 |
+
optimizer = torch.optim.Adam(model.parameters(), lr=1e-5)
|
| 367 |
+
|
| 368 |
+
import numpy as np
|
| 369 |
+
|
| 370 |
+
def trainaug(model, dataloader, criterion, optimizer, device, loss_vals, epochs, current_epoch):
|
| 371 |
+
model.train()
|
| 372 |
+
running_loss = 0
|
| 373 |
+
|
| 374 |
+
for i, (inputs, labels) in enumerate(dataloader):
|
| 375 |
+
inputs = {key: value.squeeze().to(device) for key, value in inputs.items() if torch.is_tensor(value)}
|
| 376 |
+
labels = labels.to(device)
|
| 377 |
+
|
| 378 |
+
# Apply audio augmentation
|
| 379 |
+
augmented_audio = []
|
| 380 |
+
for audio in inputs['input_values']:
|
| 381 |
+
# The augmenter works with numpy arrays, so we need to convert the tensor to a numpy array
|
| 382 |
+
audio_np = audio.cpu().numpy()
|
| 383 |
+
|
| 384 |
+
# Apply the augmentation
|
| 385 |
+
augmented = augmenter(audio_np, sample_rate=16000) # Assuming a sample rate of 16000Hz
|
| 386 |
+
|
| 387 |
+
augmented_audio.append(augmented)
|
| 388 |
+
|
| 389 |
+
# Convert the list of numpy arrays back to a tensor
|
| 390 |
+
inputs['input_values'] = torch.from_numpy(np.array(augmented_audio)).to(device)
|
| 391 |
+
|
| 392 |
+
optimizer.zero_grad()
|
| 393 |
+
logits = model(**inputs).logits
|
| 394 |
+
loss = criterion(logits, labels)
|
| 395 |
+
loss.backward()
|
| 396 |
+
optimizer.step()
|
| 397 |
+
|
| 398 |
+
# append loss value to list
|
| 399 |
+
loss_vals.append(loss.item())
|
| 400 |
+
running_loss += loss.item()
|
| 401 |
+
|
| 402 |
+
if i % 10 == 0: # Update the plot every 10 iterations
|
| 403 |
+
plt.clf() # Clear the previous plot
|
| 404 |
+
plt.plot(loss_vals)
|
| 405 |
+
plt.xlim([0, len(dataloader)*epochs])
|
| 406 |
+
plt.ylim([0, max(loss_vals) + 2])
|
| 407 |
+
plt.xlabel('Training Iterations')
|
| 408 |
+
plt.ylabel('Loss')
|
| 409 |
+
plt.title(f"Training Loss at Epoch {current_epoch + 1}")
|
| 410 |
+
plt.pause(0.001) # Pause to update the plot
|
| 411 |
+
|
| 412 |
+
avg_loss = running_loss / len(dataloader)
|
| 413 |
+
print(f"Average Loss after Epoch {current_epoch + 1}: {avg_loss}\n")
|
| 414 |
+
return avg_loss
|
| 415 |
+
|
| 416 |
+
epochs = 20
|
| 417 |
+
plt.ion()
|
| 418 |
+
fig, ax = plt.subplots()
|
| 419 |
+
x_vals = np.arange(len(train_loader)*epochs)
|
| 420 |
+
loss_vals = []
|
| 421 |
+
for epoch in range(epochs):
|
| 422 |
+
train_loss = trainaug(model, train_loader, criterion, optimizer, device, loss_vals, epochs, epoch)
|
| 423 |
+
print(f"Epoch {epoch + 1}, Train Loss: {train_loss}")
|
| 424 |
+
|
| 425 |
+
val_loss, val_accuracy, wrong_files, true_labels, pred_labels = evaluate(model, validation_loader, criterion, device)
|
| 426 |
+
print(f"Epoch {epoch + 1}, Validation Loss: {val_loss}, Validation Accuracy: {val_accuracy:.2f}")
|
| 427 |
+
print("Misclassified Files")
|
| 428 |
+
for file_path in wrong_files:
|
| 429 |
+
print(file_path)
|
| 430 |
+
|
| 431 |
+
|
| 432 |
+
sentence_pattern = re.compile(r"_(\d+)\.wav$")
|
| 433 |
+
|
| 434 |
+
sentence_counts = Counter()
|
| 435 |
+
for file_path in wrong_files:
|
| 436 |
+
match = sentence_pattern.search(file_path)
|
| 437 |
+
if match:
|
| 438 |
+
sentence_number = int(match.group(1))
|
| 439 |
+
sentence_counts[sentence_number] += 1
|
| 440 |
+
|
| 441 |
+
total_wrong = len(wrong_files)
|
| 442 |
+
print("Total wrong files:", total_wrong)
|
| 443 |
+
print()
|
| 444 |
+
|
| 445 |
+
for sentence_number, count in sentence_counts.most_common():
|
| 446 |
+
percent = count / total_wrong * 100
|
| 447 |
+
print(f"Sentence {sentence_number}: {count} ({percent:.2f}%)")
|
| 448 |
+
scheduler.step()
|
| 449 |
+
print(classification_report(true_labels, pred_labels, target_names=['non_dysarthria', 'dysarthria']))
|
| 450 |
+
audio_file = "/content/drive/MyDrive/torgo_data/dysarthria_male/validation/M01_Session1_0005.wav"
|
| 451 |
+
# predicted_label = predict(model, audio_file, train_dataset.processor, device)
|
| 452 |
+
# print(f"Predicted label: {predicted_label}")
|
| 453 |
+
|
| 454 |
+
|
| 455 |
+
|
| 456 |
+
|
| 457 |
+
|
| 458 |
+
# Test on a specific audio file
|
| 459 |
+
##audio_file = "/content/drive/MyDrive/torgo_data/dysarthria_male/validation/M01_Session1_0005.wav"
|
| 460 |
+
##predicted_label = predict(model, audio_file, train_dataset.processor, device)
|
| 461 |
+
##print(f"Predicted label: {predicted_label}")
|
| 462 |
+
|
| 463 |
+
import re
|
| 464 |
+
from collections import Counter
|
| 465 |
+
import matplotlib.pyplot as plt
|
| 466 |
+
import numpy as np
|
| 467 |
+
from sklearn.metrics import classification_report
|
| 468 |
+
|
| 469 |
+
# Define the pattern to extract the sentence number from the file path
|
| 470 |
+
sentence_pattern = re.compile(r"_(\d+)\.wav$")
|
| 471 |
+
|
| 472 |
+
# Counter for the total number of each sentence type in the dataset
|
| 473 |
+
total_sentence_counts = Counter()
|
| 474 |
+
|
| 475 |
+
for file_path in train_loader.dataset.data: # Access the file paths directly
|
| 476 |
+
match = sentence_pattern.search(file_path)
|
| 477 |
+
if match:
|
| 478 |
+
sentence_number = int(match.group(1))
|
| 479 |
+
total_sentence_counts[sentence_number] += 1
|
| 480 |
+
|
| 481 |
+
epochs = 1
|
| 482 |
+
plt.ion()
|
| 483 |
+
fig, ax = plt.subplots()
|
| 484 |
+
x_vals = np.arange(len(train_loader)*epochs)
|
| 485 |
+
loss_vals = []
|
| 486 |
+
|
| 487 |
+
for epoch in range(epochs):
|
| 488 |
+
# train_loss = trainaug(model, train_loader, criterion, optimizer, device, loss_vals, epochs, epoch)
|
| 489 |
+
# print(f"Epoch {epoch + 1}, Train Loss: {train_loss}")
|
| 490 |
+
|
| 491 |
+
val_loss, val_accuracy, wrong_files, true_labels, pred_labels = evaluate(model, validation_loader, criterion, device)
|
| 492 |
+
print(f"Epoch {epoch + 1}, Validation Loss: {val_loss}, Validation Accuracy: {val_accuracy:.2f}")
|
| 493 |
+
print("Misclassified Files")
|
| 494 |
+
for file_path in wrong_files:
|
| 495 |
+
print(file_path)
|
| 496 |
+
|
| 497 |
+
# Counter for the misclassified sentences
|
| 498 |
+
sentence_counts = Counter()
|
| 499 |
+
|
| 500 |
+
for file_path in wrong_files:
|
| 501 |
+
match = sentence_pattern.search(file_path)
|
| 502 |
+
if match:
|
| 503 |
+
sentence_number = int(match.group(1))
|
| 504 |
+
sentence_counts[sentence_number] += 1
|
| 505 |
+
|
| 506 |
+
print("Total wrong files:", len(wrong_files))
|
| 507 |
+
print()
|
| 508 |
+
|
| 509 |
+
for sentence_number, count in sentence_counts.most_common():
|
| 510 |
+
percent = count / total_sentence_counts[sentence_number] * 100
|
| 511 |
+
print(f"Sentence {sentence_number}: {count} ({percent:.2f}%)")
|
| 512 |
+
|
| 513 |
+
scheduler.step()
|
| 514 |
+
print(classification_report(true_labels, pred_labels, target_names=['non_dysarthria', 'dysarthria']))
|
| 515 |
+
|
| 516 |
+
torch.save(model.state_dict(), "dysarthria_classifier2.pth")
|
| 517 |
+
|
| 518 |
+
save_dir = "models/my_model_06"
|
| 519 |
+
model.save_pretrained(save_dir)
|
| 520 |
+
|
| 521 |
+
"""## Cross testing
|
| 522 |
+
|
| 523 |
+
"""
|
| 524 |
+
|
| 525 |
+
# dysarthria_validation_path = "/content/drive/MyDrive/RECORDINGS_ANALYSIS/SP_ANALYSIS/testing"
|
| 526 |
+
# non_dysarthria_validation_path = "/content/drive/MyDrive/RECORDINGS_ANALYSIS/CT_ANALYSIS/testing"
|
| 527 |
+
|
| 528 |
+
#dysarthria_validation_files = get_wav_files(dysarthria_validation_path)
|
| 529 |
+
# non_dysarthria_validation_files = get_wav_files(non_dysarthria_validation_path)
|
| 530 |
+
|
| 531 |
+
#validation_data = dysarthria_validation_files + non_dysarthria_validation_files
|
| 532 |
+
#validation_labels = [1] * len(dysarthria_validation_files) + [0] * len(non_dysarthria_validation_files)
|
| 533 |
+
|
| 534 |
+
epochs = 1
|
| 535 |
+
plt.ion()
|
| 536 |
+
fig, ax = plt.subplots()
|
| 537 |
+
x_vals = np.arange(len(train_loader)*epochs)
|
| 538 |
+
loss_vals = []
|
| 539 |
+
for epoch in range(epochs):
|
| 540 |
+
#train_loss = train(model, train_loader, criterion, optimizer, device, loss_vals, epochs, epoch)
|
| 541 |
+
#print(f"Epoch {epoch + 1}, Train Loss: {train_loss}")
|
| 542 |
+
|
| 543 |
+
val_loss, val_accuracy, wrong_files, true_labels, pred_labels = evaluate(model, validation_loader, criterion, device)
|
| 544 |
+
print(f"Epoch {epoch + 1}, Validation Loss: {val_loss}, Validation Accuracy: {val_accuracy:.2f}")
|
| 545 |
+
print("Misclassified Files")
|
| 546 |
+
for file_path in wrong_files:
|
| 547 |
+
print(file_path)
|
| 548 |
+
|
| 549 |
+
|
| 550 |
+
sentence_pattern = re.compile(r"_(\d+)\.wav$")
|
| 551 |
+
|
| 552 |
+
sentence_counts = Counter()
|
| 553 |
+
for file_path in wrong_files:
|
| 554 |
+
match = sentence_pattern.search(file_path)
|
| 555 |
+
if match:
|
| 556 |
+
sentence_number = int(match.group(1))
|
| 557 |
+
sentence_counts[sentence_number] += 1
|
| 558 |
+
|
| 559 |
+
total_wrong = len(wrong_files)
|
| 560 |
+
print("Total wrong files:", total_wrong)
|
| 561 |
+
print()
|
| 562 |
+
|
| 563 |
+
for sentence_number, count in sentence_counts.most_common():
|
| 564 |
+
percent = count / total_wrong * 100
|
| 565 |
+
print(f"Sentence {sentence_number}: {count} ({percent:.2f}%)")
|
| 566 |
+
scheduler.step()
|
| 567 |
+
print(classification_report(true_labels, pred_labels, target_names=['non_dysarthria', 'dysarthria']))
|
| 568 |
+
audio_file = "/content/drive/MyDrive/torgo_data/dysarthria_male/validation/M01_Session1_0005.wav"
|
| 569 |
+
predicted_label = predict(model, audio_file, train_dataset.processor, device)
|
| 570 |
+
print(f"Predicted label: {predicted_label}")
|
| 571 |
+
|
| 572 |
+
|
| 573 |
+
|
| 574 |
+
|
| 575 |
+
|
| 576 |
+
# Test on a specific audio file
|
| 577 |
+
##audio_file = "/content/drive/MyDrive/torgo_data/dysarthria_male/validation/M01_Session1_0005.wav"
|
| 578 |
+
##predicted_label = predict(model, audio_file, train_dataset.processor, device)
|
| 579 |
+
##print(f"Predicted label: {predicted_label}")
|
| 580 |
+
|
| 581 |
+
"""## DEBUGGING"""
|
| 582 |
+
|
| 583 |
+
dysarthria_path = "/content/drive/MyDrive/torgo_data/dysarthria_male/training"
|
| 584 |
+
non_dysarthria_path = "/content/drive/MyDrive/torgo_data/non_dysarthria_male/training"
|
| 585 |
+
|
| 586 |
+
dysarthria_files = [os.path.join(dysarthria_path, f) for f in os.listdir(dysarthria_path) if f.endswith('.wav')]
|
| 587 |
+
non_dysarthria_files = [os.path.join(non_dysarthria_path, f) for f in os.listdir(non_dysarthria_path) if f.endswith('.wav')]
|
| 588 |
+
|
| 589 |
+
data = dysarthria_files + non_dysarthria_files
|
| 590 |
+
labels = [1] * len(dysarthria_files) + [0] * len(non_dysarthria_files)
|
| 591 |
+
|
| 592 |
+
train_data, test_data, train_labels, test_labels = train_test_split(data, labels, test_size=0.2)
|
| 593 |
+
|
| 594 |
+
train_dataset = DysarthriaDataset(train_data, train_labels)
|
| 595 |
+
test_dataset = DysarthriaDataset(test_data, test_labels)
|
| 596 |
+
|
| 597 |
+
train_loader = DataLoader(train_dataset, batch_size=4, drop_last=True)
|
| 598 |
+
test_loader = DataLoader(test_dataset, batch_size=4, drop_last=True)
|
| 599 |
+
|
| 600 |
+
|
| 601 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 602 |
+
# model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h").to(device)
|
| 603 |
+
# model.classifier = nn.Linear(model.config.hidden_size, 2).to(device)
|
| 604 |
+
|
| 605 |
+
model = Wav2Vec2ForSequenceClassification.from_pretrained("facebook/wav2vec2-base-960h", num_labels=2).to(device)
|
| 606 |
+
|
| 607 |
+
max_length = 100_000
|
| 608 |
+
processor = train_dataset.processor
|
| 609 |
+
|
| 610 |
+
model.eval()
|
| 611 |
+
audio_file = "/content/drive/MyDrive/torgo_data/dysarthria_male/validation/M01_Session1_0005.wav"
|
| 612 |
+
# predicted_label = predict(model, audio_file, train_dataset.processor, device)
|
| 613 |
+
# print(f"Predicted label: {predicted_label}")
|
| 614 |
+
|
| 615 |
+
wav_data, _ = sf.read(audio_file)
|
| 616 |
+
inputs = processor(wav_data, sampling_rate=16000, return_tensors="pt", padding=True)
|
| 617 |
+
input_values = inputs.input_values.squeeze(0) # Squeeze the batch dimension
|
| 618 |
+
if max_length - input_values.shape[-1] > 0:
|
| 619 |
+
input_values = torch.cat([input_values, torch.zeros((max_length - input_values.shape[-1],))], dim=-1)
|
| 620 |
+
else:
|
| 621 |
+
input_values = input_values[:max_length]
|
| 622 |
+
|
| 623 |
+
input_values = input_values.unsqueeze(0).to(device)
|
| 624 |
+
input_values.shape
|
| 625 |
+
|
| 626 |
+
with torch.no_grad():
|
| 627 |
+
outputs = model(**{"input_values": input_values})
|
| 628 |
+
logits = outputs.logits
|
| 629 |
+
|
| 630 |
+
input_values.shape, logits.shape
|
| 631 |
+
|
| 632 |
+
import torch.nn.functional as F
|
| 633 |
+
# Remove the batch dimension.
|
| 634 |
+
logits = logits.squeeze()
|
| 635 |
+
predicted_class_id = torch.argmax(logits, dim=-1)
|
| 636 |
+
predicted_class_id
|
| 637 |
+
|
| 638 |
+
"""Cross testing
|
| 639 |
+
|
| 640 |
+
##origial code
|
| 641 |
+
"""
|
| 642 |
+
|
| 643 |
+
import os
|
| 644 |
+
import soundfile as sf
|
| 645 |
+
import torch
|
| 646 |
+
import torch.nn as nn
|
| 647 |
+
import torch.nn.functional as F
|
| 648 |
+
from torch.utils.data import Dataset, DataLoader
|
| 649 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor, Wav2Vec2ForSequenceClassification
|
| 650 |
+
from sklearn.model_selection import train_test_split
|
| 651 |
+
|
| 652 |
+
# Custom Dataset class
|
| 653 |
+
class DysarthriaDataset(Dataset):
|
| 654 |
+
def __init__(self, data, labels, max_length=100000):
|
| 655 |
+
self.data = data
|
| 656 |
+
self.labels = labels
|
| 657 |
+
self.max_length = max_length
|
| 658 |
+
self.processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
|
| 659 |
+
|
| 660 |
+
def __len__(self):
|
| 661 |
+
return len(self.data)
|
| 662 |
+
|
| 663 |
+
def __getitem__(self, idx):
|
| 664 |
+
try:
|
| 665 |
+
wav_data, _ = sf.read(self.data[idx])
|
| 666 |
+
except:
|
| 667 |
+
print(f"Error opening file: {self.data[idx]}. Skipping...")
|
| 668 |
+
return self.__getitem__((idx + 1) % len(self.data))
|
| 669 |
+
inputs = self.processor(wav_data, sampling_rate=16000, return_tensors="pt", padding=True)
|
| 670 |
+
input_values = inputs.input_values.squeeze(0) # Squeeze the batch dimension
|
| 671 |
+
if self.max_length - input_values.shape[-1] > 0:
|
| 672 |
+
input_values = torch.cat([input_values, torch.zeros((self.max_length - input_values.shape[-1],))], dim=-1)
|
| 673 |
+
else:
|
| 674 |
+
input_values = input_values[:self.max_length]
|
| 675 |
+
|
| 676 |
+
# Remove unsqueezing the channel dimension
|
| 677 |
+
# input_values = input_values.unsqueeze(0)
|
| 678 |
+
|
| 679 |
+
# label = torch.zeros(32,dtype=torch.long)
|
| 680 |
+
# label[self.labels[idx]] = 1
|
| 681 |
+
|
| 682 |
+
### CHANGES: simply return the label as a single integer
|
| 683 |
+
return {"input_values": input_values}, self.labels[idx]
|
| 684 |
+
###
|
| 685 |
+
|
| 686 |
+
|
| 687 |
+
def train(model, dataloader, criterion, optimizer, device, ax, loss_vals, x_vals, fig,train_loader,epochs):
|
| 688 |
+
model.train()
|
| 689 |
+
running_loss = 0
|
| 690 |
+
|
| 691 |
+
for i, (inputs, labels) in enumerate(dataloader):
|
| 692 |
+
inputs = {key: value.squeeze().to(device) for key, value in inputs.items()}
|
| 693 |
+
labels = labels.to(device)
|
| 694 |
+
|
| 695 |
+
optimizer.zero_grad()
|
| 696 |
+
logits = model(**inputs).logits
|
| 697 |
+
loss = criterion(logits, labels)
|
| 698 |
+
loss.backward()
|
| 699 |
+
optimizer.step()
|
| 700 |
+
|
| 701 |
+
# append loss value to list
|
| 702 |
+
loss_vals.append(loss.item())
|
| 703 |
+
running_loss += loss.item()
|
| 704 |
+
|
| 705 |
+
if i:
|
| 706 |
+
# update plot
|
| 707 |
+
ax.clear()
|
| 708 |
+
ax.set_xlim([0, len(train_loader)*epochs])
|
| 709 |
+
ax.set_xlabel('Training Iterations')
|
| 710 |
+
ax.set_ylim([0, max(loss_vals) + 2])
|
| 711 |
+
ax.set_ylabel('Loss')
|
| 712 |
+
ax.plot(x_vals[:len(loss_vals)], loss_vals)
|
| 713 |
+
fig.canvas.draw()
|
| 714 |
+
plt.pause(0.001)
|
| 715 |
+
|
| 716 |
+
avg_loss = running_loss / len(dataloader)
|
| 717 |
+
print(avg_loss)
|
| 718 |
+
print("\n")
|
| 719 |
+
return avg_loss
|
| 720 |
+
|
| 721 |
+
|
| 722 |
+
|
| 723 |
+
def main():
|
| 724 |
+
dysarthria_path = "/content/drive/MyDrive/RECORDINGS_ANALYSIS/SP_ANALYSIS/training"
|
| 725 |
+
non_dysarthria_path = "/content/drive/MyDrive/RECORDINGS_ANALYSIS/CT_ANALYSIS/training"
|
| 726 |
+
|
| 727 |
+
dysarthria_files = get_wav_files(dysarthria_path)
|
| 728 |
+
non_dysarthria_files = get_wav_files(non_dysarthria_path)
|
| 729 |
+
|
| 730 |
+
data = dysarthria_files + non_dysarthria_files
|
| 731 |
+
labels = [1] * len(dysarthria_files) + [0] * len(non_dysarthria_files)
|
| 732 |
+
|
| 733 |
+
train_data, test_data, train_labels, test_labels = train_test_split(data, labels, test_size=0.2)
|
| 734 |
+
|
| 735 |
+
train_dataset = DysarthriaDataset(train_data, train_labels)
|
| 736 |
+
test_dataset = DysarthriaDataset(test_data, test_labels)
|
| 737 |
+
|
| 738 |
+
train_loader = DataLoader(train_dataset, batch_size=8, drop_last=True)
|
| 739 |
+
test_loader = DataLoader(test_dataset, batch_size=8, drop_last=True)
|
| 740 |
+
validation_loader = DataLoader(test_dataset, batch_size=8, drop_last=True)
|
| 741 |
+
|
| 742 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 743 |
+
# model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h").to(device)
|
| 744 |
+
# model.classifier = nn.Linear(model.config.hidden_size, 2).to(device)
|
| 745 |
+
|
| 746 |
+
### NEW CODES
|
| 747 |
+
# It seems like the classifier layer is excluded from the model's forward method (i.e., model(**inputs)).
|
| 748 |
+
# That's why the number of labels in the output was 32 instead of 2 even when you had already changed the classifier.
|
| 749 |
+
# Instead, huggingface offers the option for loading the Wav2Vec model with an adjustable classifier head on top (by setting num_labels).
|
| 750 |
+
|
| 751 |
+
model = Wav2Vec2ForSequenceClassification.from_pretrained("facebook/wav2vec2-base-960h", num_labels=2).to(device)
|
| 752 |
+
###
|
| 753 |
+
#model_path = "/content/dysarthria_classifier3.pth"
|
| 754 |
+
#if os.path.exists(model_path):
|
| 755 |
+
#print(f"Loading saved model {model_path}")
|
| 756 |
+
#model.load_state_dict(torch.load(model_path))
|
| 757 |
+
|
| 758 |
+
criterion = nn.CrossEntropyLoss()
|
| 759 |
+
optimizer = torch.optim.Adam(model.parameters(), lr=3e-5)
|
| 760 |
+
dysarthria_validation_path = "/content/drive/MyDrive/RECORDINGS_ANALYSIS/SP_ANALYSIS/testing"
|
| 761 |
+
non_dysarthria_validation_path = "/content/drive/MyDrive/RECORDINGS_ANALYSIS/CT_ANALYSIS/testing"
|
| 762 |
+
|
| 763 |
+
dysarthria_validation_files = get_wav_files(dysarthria_validation_path)
|
| 764 |
+
non_dysarthria_validation_files = get_wav_files(non_dysarthria_validation_path)
|
| 765 |
+
|
| 766 |
+
validation_data = dysarthria_validation_files + non_dysarthria_validation_files
|
| 767 |
+
validation_labels = [1] * len(dysarthria_validation_files) + [0] * len(non_dysarthria_validation_files)
|
| 768 |
+
|
| 769 |
+
epochs = 10
|
| 770 |
+
fig, ax = plt.subplots()
|
| 771 |
+
x_vals = np.arange(len(train_loader)*epochs)
|
| 772 |
+
loss_vals = []
|
| 773 |
+
nume = 1
|
| 774 |
+
for epoch in range(epochs):
|
| 775 |
+
train_loss = train(model, train_loader, criterion, optimizer, device, ax, loss_vals, x_vals, fig, train_loader, epoch+1)
|
| 776 |
+
print(f"Epoch {epoch + 1}, Train Loss: {train_loss}")
|
| 777 |
+
|
| 778 |
+
val_loss, val_accuracy, wrong_files = evaluate(model, validation_loader, criterion, device)
|
| 779 |
+
print(f"Epoch {epoch + 1}, Validation Loss: {val_loss}, Validation Accuracy: {val_accuracy:.2f}")
|
| 780 |
+
print("Misclassified Files")
|
| 781 |
+
for file_path in wrong_files:
|
| 782 |
+
print(file_path)
|
| 783 |
+
|
| 784 |
+
|
| 785 |
+
sentence_pattern = re.compile(r"_(\d+)\.wav$")
|
| 786 |
+
|
| 787 |
+
sentence_counts = Counter()
|
| 788 |
+
for file_path in wrong_files:
|
| 789 |
+
match = sentence_pattern.search(file_path)
|
| 790 |
+
if match:
|
| 791 |
+
sentence_number = int(match.group(1))
|
| 792 |
+
sentence_counts[sentence_number] += 1
|
| 793 |
+
|
| 794 |
+
total_wrong = len(wrong_files)
|
| 795 |
+
print("Total wrong files:", total_wrong)
|
| 796 |
+
print()
|
| 797 |
+
|
| 798 |
+
for sentence_number, count in sentence_counts.most_common():
|
| 799 |
+
percent = count / total_wrong * 100
|
| 800 |
+
print(f"Sentence {sentence_number}: {count} ({percent:.2f}%)")
|
| 801 |
+
|
| 802 |
+
|
| 803 |
+
torch.save(model.state_dict(), "dysarthria_classifier4.pth")
|
| 804 |
+
print("Predicting...")
|
| 805 |
+
# Test on a specific audio file
|
| 806 |
+
##audio_file = "/content/drive/MyDrive/torgo_data/dysarthria_male/validation/M01_Session1_0005.wav"
|
| 807 |
+
##predicted_label = predict(model, audio_file, train_dataset.processor, device)
|
| 808 |
+
##print(f"Predicted label: {predicted_label}")
|
| 809 |
+
|
| 810 |
+
def predict(model, file_path, processor, device, max_length=100000): ### CHANGES: added max_length as an argument.
|
| 811 |
+
model.eval()
|
| 812 |
+
with torch.no_grad():
|
| 813 |
+
wav_data, _ = sf.read(file_path)
|
| 814 |
+
inputs = processor(wav_data, sampling_rate=16000, return_tensors="pt", padding=True)
|
| 815 |
+
# inputs = {key: value.squeeze().to(device) for key, value in inputs.items()}
|
| 816 |
+
|
| 817 |
+
### NEW CODES HERE
|
| 818 |
+
input_values = inputs.input_values.squeeze(0) # Squeeze the batch dimension
|
| 819 |
+
if max_length - input_values.shape[-1] > 0:
|
| 820 |
+
input_values = torch.cat([input_values, torch.zeros((max_length - input_values.shape[-1],))], dim=-1)
|
| 821 |
+
else:
|
| 822 |
+
input_values = input_values[:max_length]
|
| 823 |
+
input_values = input_values.unsqueeze(0).to(device)
|
| 824 |
+
inputs = {"input_values": input_values}
|
| 825 |
+
###
|
| 826 |
+
|
| 827 |
+
logits = model(**inputs).logits
|
| 828 |
+
# _, predicted = torch.max(logits, dim=0)
|
| 829 |
+
|
| 830 |
+
### NEW CODES HERE
|
| 831 |
+
# Remove the batch dimension.
|
| 832 |
+
logits = logits.squeeze()
|
| 833 |
+
predicted_class_id = torch.argmax(logits, dim=-1).item()
|
| 834 |
+
###
|
| 835 |
+
|
| 836 |
+
# return predicted.item()
|
| 837 |
+
return predicted_class_id
|
| 838 |
+
def evaluate(model, dataloader, criterion, device):
|
| 839 |
+
model.eval()
|
| 840 |
+
running_loss = 0
|
| 841 |
+
correct_predictions = 0
|
| 842 |
+
total_predictions = 0
|
| 843 |
+
wrong_files = []
|
| 844 |
+
with torch.no_grad():
|
| 845 |
+
for inputs, labels in dataloader:
|
| 846 |
+
inputs = {key: value.squeeze().to(device) for key, value in inputs.items()}
|
| 847 |
+
labels = labels.to(device)
|
| 848 |
+
|
| 849 |
+
logits = model(**inputs).logits
|
| 850 |
+
loss = criterion(logits, labels)
|
| 851 |
+
running_loss += loss.item()
|
| 852 |
+
|
| 853 |
+
_, predicted = torch.max(logits, 1)
|
| 854 |
+
correct_predictions += (predicted == labels).sum().item()
|
| 855 |
+
total_predictions += labels.size(0)
|
| 856 |
+
|
| 857 |
+
wrong_idx = (predicted != labels).nonzero().squeeze().cpu().numpy()
|
| 858 |
+
if wrong_idx.ndim > 0:
|
| 859 |
+
for idx in wrong_idx:
|
| 860 |
+
wrong_files.append(dataloader.dataset.data[idx])
|
| 861 |
+
elif wrong_idx.size > 0:
|
| 862 |
+
wrong_files.append(dataloader.dataset.data[wrong_idx])
|
| 863 |
+
|
| 864 |
+
|
| 865 |
+
avg_loss = running_loss / len(dataloader)
|
| 866 |
+
accuracy = correct_predictions / total_predictions
|
| 867 |
+
return avg_loss, accuracy, wrong_files
|
| 868 |
+
|
| 869 |
+
|
| 870 |
+
|
| 871 |
+
def get_wav_files(base_path):
|
| 872 |
+
wav_files = []
|
| 873 |
+
for subject_folder in os.listdir(base_path):
|
| 874 |
+
subject_path = os.path.join(base_path, subject_folder)
|
| 875 |
+
if os.path.isdir(subject_path):
|
| 876 |
+
for wav_file in os.listdir(subject_path):
|
| 877 |
+
if wav_file.endswith('.wav'):
|
| 878 |
+
wav_files.append(os.path.join(subject_path, wav_file))
|
| 879 |
+
return wav_files
|
| 880 |
+
if __name__ == "__main__":
|
| 881 |
+
main()
|