How to implement a CNN-LSTM using Keras

Question

I am attempting to implement a CNN-LSTM that classifies mel-spectrogram images representing the speech of people with Parkinson's Disease/Healthy Controls. I am trying to implement a pre-existing model (DenseNet-169) with an LSTM model, however I am running into the following error: ValueError: Input 0 of layer zero_padding2d is incompatible with the layer: expected ndim=4, found ndim=3. Full shape received: [None, 216, 1]. Can anyone advise where I'm going wrong?

import librosa
import os
import glob
import IPython.display as ipd
from pathlib import Path
import timeit
import time, sys

%matplotlib inline
import matplotlib.pyplot as plt
import librosa.display

import pandas as pd
from sklearn import datasets, linear_model
from sklearn.model_selection import train_test_split
from matplotlib import pyplot as plt
import numpy as np
import cv2
import seaborn as sns

%tensorflow_version 1.x #version 1 works without problems
import tensorflow

from tensorflow.keras import models
from sklearn.preprocessing import LabelEncoder
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.layers import TimeDistributed

import keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.callbacks import EarlyStopping
from sklearn.metrics import confusion_matrix, plot_confusion_matrix
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dropout, Dense, BatchNormalization, Activation, GaussianNoise, LSTM
from sklearn.metrics import accuracy_score

DATA_DIR = Path('/content/drive/MyDrive/PhD_Project_Experiments/Spontaneous_Dialogue_PD_Dataset') 
diagnosis = [x.name for x in DATA_DIR.glob('*') if x.is_dir()]
diagnosis

def create_paths_ds(paths: Path, label: str) -> list:
    EXTENSION_TYPE = '.wav'
    return [(x, label) for x in paths.glob('*' + EXTENSION_TYPE)]

from collections import Counter

categories_to_use = [
    'Parkinsons_Disease',
    'Healthy_Control',
]

NUM_CLASSES = len(categories_to_use)

print(f'Number of classes: {NUM_CLASSES}')

paths_all_labels = []
for cat in categories_to_use:
    paths_all_labels += create_paths_ds(DATA_DIR / cat, cat)
 
X_train, X_test = train_test_split(paths_all_labels,test_size=0.1, stratify = [paths_all_labels[y][1] for y in range(len(paths_all_labels))] ) #fix stratified sampling for test data
X_train, X_val = train_test_split(X_train, test_size=0.2, stratify = [X_train[y][1] for y in range(len(X_train))] ) 

for i in categories_to_use:
  print('Number of train samples for '+i+': '+ str([X_train[y][1] for y in range(len(X_train))].count(i))) #checks whether train samples are equally divided
  print('Number of test samples for '+i+': '+ str([X_test[y][1] for y in range(len(X_test))].count(i))) #checks whether test samples are equally divided
  print('Number of validation samples for '+i+': '+ str([X_val[y][1] for y in range(len(X_val))].count(i))) #checks whether val samples are equally divided

print(f'Train length: {len(X_train)}')
print(f'Validation length: {len(X_val)}')
print(f'Test length: {len(X_test)}')

def load_and_preprocess_lstm(dataset, SAMPLE_SIZE = 30):
    IMG_SIZE = (216,128) 
    progress=0

    data = []
    labels = []
    for (path, label) in dataset:
        audio, sr = librosa.load(path)
        dur = librosa.get_duration(audio, sr = sr)
        sampleNum = int(dur / SAMPLE_SIZE)
        offset = (dur % SAMPLE_SIZE) / 2
        for i in range(sampleNum):
            audio, sr = librosa.load(path, offset= offset+i, duration=SAMPLE_SIZE)
            sample = librosa.feature.melspectrogram(audio, sr=sr)
            # print(sample.shape)
            sample = cv2.resize(sample, dsize=IMG_SIZE)
            sample = np.expand_dims(sample,-1)
            print(sample.shape)
            data += [(sample, label)]
            labels += [label]
        progress +=1
        print('\r Progress: '+str(round(100*progress/len(dataset))) + '%', end='')
    return data, labels

def retrieve_samples(sample_size, model_type):

    if model_type == 'cnn':
  
        print("\nLoading train samples")
        X_train_samples, train_labels = load_and_preprocess_cnn(X_train,sample_size)
        print("\nLoading test samples")
        X_test_samples, test_labels = load_and_preprocess_cnn(X_test,sample_size)
        print("\nLoading val samples")
        X_val_samples, val_labels = load_and_preprocess_cnn(X_val,sample_size)
        print('\n')

    elif model_type == 'lstm':

        print("\nLoading train samples")
        X_train_samples, train_labels = load_and_preprocess_lstm(X_train,sample_size)
        print("\nLoading test samples")
        X_test_samples, test_labels = load_and_preprocess_lstm(X_test,sample_size)
        print("\nLoading val samples")
        X_val_samples, val_labels = load_and_preprocess_lstm(X_val,sample_size)      
        print('\n')

    elif model_type == "cnnlstm":

        print("\nLoading train samples")
        X_train_samples, train_labels = load_and_preprocess_lstm(X_train,sample_size)
        print("\nLoading test samples")
        X_test_samples, test_labels = load_and_preprocess_lstm(X_test,sample_size)
        print("\nLoading val samples")
        X_val_samples, val_labels = load_and_preprocess_lstm(X_val,sample_size)      
        print('\n')

    print("shape: " + str(X_train_samples[0][0].shape))
    print("number of training samples: "+ str(len(X_train_samples)))
    print("number of validation samples: "+ str(len(X_val_samples)))
    print("number of test samples: "+ str(len(X_test_samples)))


    return X_train_samples, X_test_samples, X_val_samples

def create_cnn_lstm_model(input_shape):

    model = Sequential()
    cnn = tensorflow.keras.applications.DenseNet169(include_top=True, weights=None, input_tensor=None, input_shape=input_shape, pooling=None, classes=2)
    # define LSTM model
    model.add(tensorflow.keras.layers.TimeDistributed(cnn, input_shape=input_shape))
    model.add(LSTM(units = 512, dropout=0.5, recurrent_dropout=0.3, return_sequences = True, input_shape = input_shape))
    model.add(LSTM(units = 512, dropout=0.5, recurrent_dropout=0.3, return_sequences = False))
    model.add(Dense(units=NUM_CLASSES, activation='sigmoid'))#Compile

    model.compile(loss=tensorflow.keras.losses.binary_crossentropy, optimizer='adam', metrics=['accuracy'])
    print(model.summary())

    return model

def create_model_data_and_labels(X_train_samples, X_val_samples, X_test_samples):
    #Prepare samples to work for training the model
    labelizer = LabelEncoder()

    #prepare training data and labels
    x_train = np.array([x[0] for x in X_train_samples])
    y_train = np.array([x[1] for x in X_train_samples])
    y_train = labelizer.fit_transform(y_train) 
    y_train = to_categorical(y_train)

    #prepare validation data and labels
    x_val = np.array([x[0] for x in X_val_samples])
    y_val = np.array([x[1] for x in X_val_samples])
    y_val = labelizer.transform(y_val)
    y_val = to_categorical(y_val)

    #prepare test data and labels
    x_test = np.array([x[0] for x in X_test_samples])
    y_test = np.array([x[1] for x in X_test_samples])
    y_test = labelizer.transform(y_test)
    y_test = to_categorical(y_test)

    return x_train, y_train, x_val, y_val, x_test, y_test, labelizer


#Main loop for testing multiple sample sizes

#choose model type: 'cnn' or 'lstm'
model_type = 'cnnlstm'

n_epochs = 20
patience= 20
es = EarlyStopping(patience=20)
fragment_sizes = [5,10]
start = timeit.default_timer()

ModelData = pd.DataFrame(columns = ['Model Type','Fragment size (s)', 'Time to Compute (s)',  'Early Stopping epoch', 'Training accuracy', 'Validation accuracy', 'Test Accuracy']) #create a DataFrame for storing the results 

conf_matrix_data = []

for i in fragment_sizes:

    start_per_size = timeit.default_timer()

    print(f'\n---------- Model trained on fragments of size: {i} seconds ----------------')
    X_train_samples, X_test_samples, X_val_samples = retrieve_samples(i,model_type)
    x_train, y_train, x_val, y_val, x_test, y_test, labelizer = create_model_data_and_labels(X_train_samples, X_val_samples, X_test_samples)

    if model_type == 'cnn':
        model = create_cnn_model(X_train_samples[0][0].shape)
    elif model_type == 'lstm':
        model = create_lstm_model(X_train_samples[0][0].shape)
    elif model_type == 'cnnlstm':
        model = create_cnn_lstm_model(X_train_samples[0][0].shape)


    history = model.fit(x_train, y_train, 
              batch_size = 8, 
              epochs=n_epochs,
              verbose=1, 
              callbacks=[es],
              validation_data=(x_val, y_val))
    print('Finished training')


    early_stopping_epoch = len(history.history['accuracy'])
    training_accuracy = history.history['accuracy'][early_stopping_epoch-1-patience]
    validation_accuracy = history.history['val_accuracy'][early_stopping_epoch-1-patience]

    plot_data(history, i)

    predictions = model.predict(x_test)
    score = accuracy_score(labelizer.inverse_transform(y_test.argmax(axis=1)), labelizer.inverse_transform(predictions.argmax(axis=1)))

    print('Fragment size = ' + str(i) + ' seconds')
    print('Accuracy on test samples: ' + str(score))
    
    conf_matrix_data += [(predictions, y_test, i)]

    stop_per_size = timeit.default_timer()
    time_to_compute = round(stop_per_size - start_per_size)

    print ('Time to compute: '+str(time_to_compute))

    ModelData.loc[len(ModelData)] = [model_type, i, time_to_compute, early_stopping_epoch, training_accuracy, validation_accuracy, score] #store particular settings configuration, early stoppping epoch and accuracies in dataframe

stop = timeit.default_timer()
print ('\ntime to compute: '+str(stop-start))

Dettmer

Hi, it's better to just provide the minimal required code to be able to reproduce the problem instead of the complete code dump as it would be extremely difficult for someone to go through all of it. Focusing on the function that creates the computation graph, what is your INPUT_SHAPE? – Kerianne

Sorry, what do you mean by computation graph? – Dettmer

computation graph is the model architecture. in your case its Densenet169 instances timedistrubited. I am referring to the create_cnn_lstm_model() function. What is your input_shape that you use as parameter to call the funciton – Kerianne

also, shapes of X_train, y_train would also help – Kerianne

I believe the input_shape is (128, 216, 1) – Dettmer

could you check X_train_samples[0][0].shape and tell me? – Kerianne

The issue here is that you have to have some time axis to time distribute your cnn layer. In this step - tensorflow.keras.layers.TimeDistributed(cnn, input_shape=(128,216,1))), you are passing the 128 dimension axis as time axis. That means each of the cnn is left with (216,1), which is not an image and therefore throws an error because its expecting a 3D tensors and not 2D. – Kerianne

Your input shape needs to be a 4D tensor something like - (10, 128, 216, 1), so that the 10 becomes time axis (for timedistributing), and (128, 216, 1) becomes an image input for the cnn. – Kerianne

The x_train shape is (318, 128, 216, 1), the y_train shape is (318, 2) and X_train_samples[0][0].shape is (128, 216, 1) – Dettmer

Answer 1

I believe the input_shape is (128, 216, 1)

The issue here is that you don't have a time-axis to time distribute your CNN (DenseNet169) layer over.

In this step -

tensorflow.keras.layers.TimeDistributed(cnn, input_shape=(128,216,1)))

You are passing the 128 dimension axis as a time-axis. That means each of the CNN (DenseNet169) is left with a input shape of (216,1), which is not an image and therefore throws an error because it's expecting 3D tensors (images) and not 2D tensors.

Your input shape needs to be a 4D tensor something like - (10, 128, 216, 1), so that the 10 becomes the time axis (for time distributing), and (128, 216, 1) becomes an image input for the CNN (DenseNet169).

---

A solution with ragged tensors and time-distributed layer

IIUC, your data contains n audio files, each file containing a variable number of mel-spectrogram images.

1. You need to use tf.raggedtensors to be able to work with variable tensor shapes as inputs to the model
2. This requires an explicit definition of an Input layer where you set ragged=True
3. This allows you to pass each audio file as a single sample, with variable images, each of which will be time distributed.
4. You will have to use None as the time distributed axis shape while defining the model

1. Creating a dummy dataset

Let's start with a sample dataset -

import tensorflow as tf
from tensorflow.keras import layers, Model, utils, applications


#Assuming there are 5 audio files
num_audio = 5

data = []

#Create a random number of mel-spectrograms for each audio file
for i in range(num_audio):
    n_images = np.random.randint(4,10)
    data.append(np.random.random((n_images,128,216,1)))
    
print([i.shape for i in data])

[(5, 128, 216, 1), 
 (5, 128, 216, 1), 
 (9, 128, 216, 1), 
 (6, 128, 216, 1), 
 (4, 128, 216, 1)]

So, your data should be looking something like this. Here, I have a dummy dataset with 5 audio files, first one has 5 images of shape (128,216,1), while the last one has 4 images of the same shape.

2. Converting them to ragged-tensors

Next, let's convert and store these are ragged tensors. Ragged tensors allow variable-length objects to be stored, in this case, a variable number of images. Read more about them here.

#Convert each set of images (for each audio) to tensors and then a ragged tensor
tensors = [tensorflow.convert_to_tensor(i) for i in data]
X_train = tensorflow.ragged.stack(tensors).to_tensor()

#Creating dummy y_train, one for each audio files
y_train = tensorflow.convert_to_tensor(np.random.randint(0,2,(5,2)))

3. Create a model

I am using a functional API since I find it more readable and works better with an explicit input layer, but you can use input layers in Sequential API as well. Feel free to convert it to your preference.

Notice that I am using (None,128,216,1) as input shape. This creates 5 channels (first implicit one for batches) as - (Batch, audio_files, h, w, channels)

I have a dummy LSTM layer to showcase how the architecture works, feel free to stack more layers. Also, do note, that your DenseNet169 is only returning 2 features. And therefore your TimeDistributed layers is returning (None, None, 2) shaped tensor, where first None is the number of audio files, and the second None is the number of images (time axis). Therefore, do choose your next layers accordingly as 512 LSTM cells may be too much :)

#Create model
inp = layers.Input((None,128,216,1), ragged=True)

cnn = tensorflow.keras.applications.DenseNet169(include_top=True, 
                                                weights=None, 
                                                input_tensor=None, 
                                                input_shape=(128,216,1), #<----- input shape for cnn is just the image
                                                pooling=None, classes=2)


#Feel free to modify these layers!
x = layers.TimeDistributed(cnn)(inp)
x = layers.LSTM(8)(x)
out = layers.Dense(2)(x)

model = Model(inp, out)
model.compile(loss='binary_crossentropy', 
              optimizer='adam', 
              metrics='accuracy')

utils.plot_model(model, show_shapes=True, show_layer_names=False)

4. Train!

The next step is simply to train. Feel free to add your own parameters.

model.fit(X_train, y_train, epochs=2)

Epoch 1/2
WARNING:tensorflow:5 out of the last 5 calls to <function Model.make_train_function.<locals>.train_function at 0x7f8e55b4fe50> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for  more details.
1/1 [==============================] - 37s 37s/step - loss: 3.4057 - accuracy: 0.4000
Epoch 2/2
1/1 [==============================] - 16s 16s/step - loss: 3.3544 - accuracy: 0.4000

Hope that helps.

Kerianne

Thank you for clarifying! The only issue is, the lengths of the audio files vary therefore is it possible to use a variable for the time-axis? – Dettmer

Glad to help. i have updated my answer to show you how the error is avoided with the right input shapes. – Kerianne

For my clarity, what is the training data? What is the 128 and 216 and 1 in the (128,216,1)? – Kerianne

I've generated a series of mel-spectrogram images for each audio file. 128 and 216 are the x and y dimensions of each mel-spectrogram image. – Dettmer

So, each audio file has n images (spectrograms) and you want to model those over time distribuited network? (so time axis is that of an audio file)? – Kerianne

It it necessary to use a Densenet169? Or just a CNN block can be used? – Kerianne

After experimenting with different CNN architectures, for this dataset DenseNet169 achieved the best accuracy so I was aiming to see whether combining DenseNet169 with an LSTM achieved a higher classification accuracy or not. But for the mean time, I'd be happy just to get a simple CNN-LSTM model working – Dettmer

Hi, I have updated my answer with a full implementation of the architecture. Do mark the answer if it helped you and do upvote it if it’s well written to your liking. This encourages me to help solve your future questions :) – Kerianne

Thank you very much for your help - it's greatly appreciated :) – Dettmer