-- coding: utf-8 -- """Untitled11.ipynb Automatically generated by Colab. Original file…
- User
-- coding: utf-8 --
"""Untitled11.ipynb
Automatically generated by Colab.
Original file is located at https://colab.research.google.com/drive/1_QKSuEIOHM1Ih7B797i6XC-CIrx7N6JZ """
-- coding: utf-8 --
"""96%cnnECHONLPipynb"""
#dataset import os os.environ['KAGGLE_CONFIG_DIR'] = "/content"
!kaggle datasets download -d ejlok1/toronto-emotional-speech-set-tess !unzip *.zip && rm *.zip
Import necessary modules
import pandas as pd import numpy as np import os import seaborn as sns import matplotlib.pyplot as plt import librosa import librosa.display from IPython.display import Audio import warnings warnings.filterwarnings('ignore')
Load the dataset
paths = [] labels = [] for dirname, , filenames in os.walk('/content/TESS Toronto emotional speech set data'): for filename in filenames: paths.append(os.path.join(dirname, filename)) label = filename.split('')[-1] label = label.split('.')[0] labels.append(label.lower()) if len(paths) == 2800: break print('Dataset is Loaded')
Check dataset length and sample paths
print(f'Number of samples: {len(paths)}') print('Sample paths:', paths[:5]) print('Sample labels:', labels[:5])
Create a dataframe
df = pd.DataFrame() df['audio_paths'] = paths df['labels'] = labels print(df.head()) print(df['labels'].value_counts())
Exploratory Data Analysis
sns.countplot(data=df, x='labels') plt.show()
Function to plot the waveform with time and amplitude labels
def waveplot(data, sr, emotion): plt.figure(figsize=(10, 4)) plt.title(f'Waveform of {emotion.capitalize()}', size=16, color='navy') librosa.display.waveshow(data, sr=sr, alpha=0.8, color='mediumseagreen')
# Adding time axis plt.xlabel("Time (seconds)", fontsize=12) # Adding amplitude axis plt.ylabel("Amplitude", fontsize=12) plt.grid(True, linestyle='--', alpha=0.7) plt.show()Function to plot a unique spectrogram with both time and frequency axis and detailed color bar
def spectogram(data, sr, emotion): stft = librosa.stft(data, n_fft=2048, hop_length=512) stft_db = librosa.amplitude_to_db(abs(stft))
plt.figure(figsize=(10, 4)) plt.title(f'Spectrogram of {emotion.capitalize()}', size=16, color='darkred') # Display the spectrogram with a unique color map librosa.display.specshow(stft_db, sr=sr, hop_length=512, x_axis='time', y_axis='log', cmap='plasma') # Add color bar with labels in decibels (dB) cbar = plt.colorbar(format="%+2.0f dB") cbar.set_label('Amplitude (dB)', fontsize=12) # Add time and frequency axis labels plt.xlabel("Time (seconds)", fontsize=12) plt.ylabel("Frequency (Hz)", fontsize=12) plt.grid(True, linestyle='--', alpha=0.7) plt.tight_layout() plt.show()Display sample audio waves and spectrograms with unique styling
emotions = ['fear', 'angry', 'disgust', 'neutral', 'sad', 'ps', 'happy'] for emotion in emotions: path = np.array(df['audio_paths'][df['labels'] == emotion])[0] data, sampling_rate = librosa.load(path, sr=16000)
# Plot waveform waveplot(data, sampling_rate, emotion) # Plot spectrogram spectogram(data, sampling_rate, emotion) # Play the audio sample display(Audio(path))Create label to integer mapping
label_map = {label: idx for idx, label in enumerate(df['labels'].unique())} inverse_label_map = {idx: label for label, idx in label_map.items()}
print('Label to Integer Mapping:', label_map) print('Integer to Label Mapping:', inverse_label_map)
Convert labels in the dataframe to integers
df['labels'] = df['labels'].map(label_map) print(df.head(2))
import torch import torchaudio from torch.utils.data import Dataset
class SpeechEmotionDataset(Dataset): def init(self, df, processor, max_length=32000): self.df = df self.processor = processor self.max_length = max_length
def __len__(self): return len(self.df) def __getitem__(self, idx): audio_path = self.df.iloc[idx]['audio_paths'] label = self.df.iloc[idx]['labels'] # This is already an integer # Use librosa for loading and resampling speech, sr = librosa.load(audio_path, sr=16000) # Pad or truncate the speech to the required length if len(speech) > self.max_length: speech = speech[:self.max_length] else: speech = np.pad(speech, (0, self.max_length - len(speech)), 'constant') # Preprocess the audio file inputs = self.processor(speech, sampling_rate=16000, return_tensors="pt", padding=True, truncation=True, max_length=self.max_length) input_values = inputs.input_values.squeeze() return {'input_values': input_values, 'labels': torch.tensor(label, dtype=torch.long)} # Label is already integerSplit the DataFrame into training and testing sets
from sklearn.model_selection import train_test_split train_df, test_df = train_test_split(df, test_size=0.2, random_state=42)
from transformers import Wav2Vec2Processor, Wav2Vec2ForSequenceClassification
Load the Wav2Vec2 processor and model for sequence classification
processor = Wav2Vec2Processor.from_pretrained('facebook/wav2vec2-base') model = Wav2Vec2ForSequenceClassification.from_pretrained('facebook/wav2vec2-base', num_labels=len(label_map)) train_dataset = SpeechEmotionDataset(train_df, processor, max_length=32000) # Pass max_length explicitly test_dataset = SpeechEmotionDataset(test_df, processor, max_length=32000) # Pass max_length explicitly
Create DataLoader for training and testing datasets
from torch.utils.data import DataLoader train_dataloader = DataLoader(train_dataset, batch_size=8, shuffle=True) test_dataloader = DataLoader(test_dataset, batch_size=8, shuffle=False)
Training pipeline setup
from transformers import Trainer, TrainingArguments from sklearn.metrics import accuracy_score, precision_recall_fscore_support
Define compute metrics function for accuracy and F1 score
def compute_metrics(pred): labels = pred.label_ids preds = np.argmax(pred.predictions, axis=1) accuracy = accuracy_score(labels, preds) precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='weighted')
return { 'accuracy': accuracy, 'precision': precision, 'recall': recall, 'f1': f1 }training_args = TrainingArguments( output_dir='./results', evaluation_strategy='epoch', save_strategy='epoch', learning_rate=2e-5, per_device_train_batch_size=16, per_device_eval_batch_size=16, num_train_epochs=3, weight_decay=0.01, report_to=[], )
Initialize the Trainer
trainer = Trainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=test_dataset, compute_metrics=compute_metrics )
Start training
trainer.train()
Evaluate the model and print results
results = trainer.evaluate() print(results)
Save the model's state_dict instead of the entire model
torch.save(model.state_dict(), 'cnn_model_full_final.pth')
Test Prediction
import torchaudio import numpy as np import torch
Preprocess audio
def preprocess_audio(file_path, processor, max_length=32000): audio, sr = torchaudio.load(file_path) audio = torchaudio.transforms.Resample(orig_freq=sr, new_freq=16000)(audio).squeeze().numpy() audio = audio[:max_length] if len(audio) > max_length else np.pad(audio, (0, max_length - len(audio)), 'constant') return processor(audio, sampling_rate=16000, return_tensors="pt", padding=True, truncation=True, max_length=max_length).input_values.squeeze()
Predict emotion
def predict_emotion(file_path, model, processor, label_map): inputs = preprocess_audio(file_path, processor) logits = model(inputs.unsqueeze(0)).logits predicted_class_id = torch.argmax(logits, dim=-1).item() return {v: k for k, v in label_map.items()}[predicted_class_id]
from google.colab import files
Upload file
uploaded = files.upload() file_path = list(uploaded.keys())[0]
Prediction
predicted_emotion = predict_emotion(file_path, model, processor, label_map) print(f'The predicted emotion is: {predicted_emotion}')in this code i want graph of model of the accuracy precision and etc of the model don't print the table of these instead i want 2 dimension of each
- Assistant
To visualize the accuracy, precision, recall, and F1 score of your model across training epochs, you can extract these metrics from the
Trainer'slog_historyand plot them usingmatplotlib. Below, I'll guide you through the necessary steps to achieve this:-
Extract Metrics from
log_history: TheTrainerobject maintains alog_historyattribute that contains logs of metrics at various stages during training and evaluation. We'll extract the relevant metrics from this history. -
Plot the Metrics: Using
matplotlib, we'll plot these metrics to visualize the model's performance over each epoch.
Step-by-Step Implementation
1. Modify Training Arguments
Ensure that your
TrainingArgumentsare set up to log evaluation metrics at the end of each epoch:training_args = TrainingArguments( output_dir='./results', evaluation_strategy='epoch', # Evaluate at the end of each epoch save_strategy='epoch', learning_rate=2e-5, per_device_train_batch_size=16, per_device_eval_batch_size=16, num_train_epochs=3, weight_decay=0.01, report_to=['none'], # Disable reporting to external systems )Note: Setting
report_to=['none']ensures that logs are only stored locally and not sent to external systems like TensorBoard or WandB. If you wish to use these platforms, adjust accordingly.2. Train the Model
Proceed with training as you have:
# Initialize the Trainer trainer = Trainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=test_dataset, compute_metrics=compute_metrics ) # Start training trainer.train() # Evaluate the model and print results results = trainer.evaluate() print(results)3. Extract and Plot Metrics
After training, use the following code to extract metrics from
log_historyand plot them:import matplotlib.pyplot as plt # Initialize lists to store metrics epochs = [] accuracy = [] precision = [] recall = [] f1 = [] # Iterate through the log history for log in trainer.state.log_history: # Check if the log contains evaluation metrics if 'epoch' in log and 'eval_accuracy' in log: epochs.append(log['epoch']) accuracy.append(log['eval_accuracy']) precision.append(log['eval_precision']) recall.append(log['eval_recall']) f1.append(log['eval_f1']) # Plotting the metrics plt.figure(figsize=(12, 8)) # Plot Accuracy plt.plot(epochs, accuracy, marker='o', label='Accuracy') # Plot Precision plt.plot(epochs, precision, marker='s', label='Precision') # Plot Recall plt.plot(epochs, recall, marker='^', label='Recall') # Plot F1 Score plt.plot(epochs, f1, marker='d', label='F1 Score') # Adding titles and labels plt.title('Model Performance Metrics per Epoch', fontsize=16) plt.xlabel('Epoch', fontsize=14) plt.ylabel('Score', fontsize=14) plt.xticks(epochs) # Ensure x-axis ticks are at each epoch plt.ylim(0, 1) # Assuming metrics are between 0 and 1 plt.grid(True, linestyle='--', alpha=0.5) plt.legend(fontsize=12) plt.tight_layout() # Display the plot plt.show()4. Full Integration
Here’s how you can integrate the plotting code into your existing script:
# ... [Your existing code above] ... # Initialize the Trainer trainer = Trainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=test_dataset, compute_metrics=compute_metrics ) # Start training trainer.train() # Evaluate the model and print results results = trainer.evaluate() print(results) # Plotting the metrics after training import matplotlib.pyplot as plt # Initialize lists to store metrics epochs = [] accuracy = [] precision = [] recall = [] f1 = [] # Iterate through the log history for log in trainer.state.log_history: # Check if the log contains evaluation metrics if 'epoch' in log and 'eval_accuracy' in log: epochs.append(log['epoch']) accuracy.append(log['eval_accuracy']) precision.append(log['eval_precision']) recall.append(log['eval_recall']) f1.append(log['eval_f1']) # Plotting the metrics plt.figure(figsize=(12, 8)) # Plot Accuracy plt.plot(epochs, accuracy, marker='o', label='Accuracy') # Plot Precision plt.plot(epochs, precision, marker='s', label='Precision') # Plot Recall plt.plot(epochs, recall, marker='^', label='Recall') # Plot F1 Score plt.plot(epochs, f1, marker='d', label='F1 Score') # Adding titles and labels plt.title('Model Performance Metrics per Epoch', fontsize=16) plt.xlabel('Epoch', fontsize=14) plt.ylabel('Score', fontsize=14) plt.xticks(epochs) # Ensure x-axis ticks are at each epoch plt.ylim(0, 1) # Assuming metrics are between 0 and 1 plt.grid(True, linestyle='--', alpha=0.5) plt.legend(fontsize=12) plt.tight_layout() # Display the plot plt.show() # Save the model's state_dict instead of the entire model torch.save(model.state_dict(), 'cnn_model_full_final.pth') # ... [Rest of your existing code] ...Example Plot
After running the above code, you should see a plot similar to the one below (actual values will depend on your training results):
[image: Model Performance Metrics per Epoch] open
Note: Replace the image link with an actual example if available.
Additional Recommendations
-
Increase Number of Epochs: Training for only 3 epochs might not provide a comprehensive view of the model's learning curve. Consider increasing the number of epochs to 10 or more to better observe trends.
-
Use Callbacks for Advanced Logging: If you require more advanced logging (e.g., step-wise metrics), consider implementing custom callbacks by subclassing
TrainerCallback. -
TensorBoard Integration: For more sophisticated and interactive visualizations, integrate TensorBoard by setting
report_to=['tensorboard']inTrainingArgumentsand using the%tensorboardmagic in Colab. -
Save Metrics for Future Reference: Consider saving the extracted metrics to a file (e.g., CSV) for future analysis or comparison with other models.
Conclusion
By following the above steps, you can effectively visualize your model's performance metrics across training epochs, allowing for better insights into the training process and model behavior. This visualization aids in diagnosing issues like overfitting, underfitting, or learning plateaus.
If you encounter any issues or have further questions, feel free to ask!
-