Bug: ValueError: numpy.dtype size changed, may indicate binary incompatibility. Expected 96 from C header, got 88 from PyObject

[4Sheetal]

What happened?

I am trying to use yamnet model but library issue is coming

Relevant code

import tensorflow as tf
import numpy as np
import tensorflow_hub as hub
import tensorflow_io as tfio
import random
# For monitoring RAM usage
from sklearn.metrics import classification_report
from collections import Counter
import sys

sys.setrecursionlimit(3000)

# Set random seeds for reproducibility
SEED = 42
tf.random.set_seed(SEED)
np.random.seed(SEED)
random.seed(SEED)

# Paths to the dataset
audio_files = {
    'Rov': "/content/drive/MyDrive/rovdata/guassian16kfolder/rov_16k.wav"
}

# YAMNet model handle
yamnet_model_handle = 'https://tfhub.dev/google/yamnet/1'
yamnet_model = hub.load(yamnet_model_handle)


# Function to load and resample audio to 16 kHz mono (taking only 10% of duration)
def load_wav_16k_mono(filename):
    """Load a WAV file, convert it to 16 kHz mono, and return a tensor."""
    file_contents = tf.io.read_file(filename)
    wav, sample_rate = tf.audio.decode_wav(file_contents, desired_channels=1)
    wav = tf.squeeze(wav, axis=-1)
    sample_rate = tf.cast(sample_rate, dtype=tf.int64)
    wav = tfio.audio.resample(wav, rate_in=sample_rate, rate_out=16000)
    return wav

# Define classes of interest
my_classes = ['Rov', 'ship', 'Mammal', 'Ambient']
class_to_idx = {cls: idx for idx, cls in enumerate(my_classes)}


# Lazy loading with RAM monitoring
data = []
for label, file_path in audio_files.items():


    print(f"🔹 Loading {label} audio...")
    wav_data = load_wav_16k_mono(file_path)

    if wav_data is not None:  # Ensure valid data
        label_idx = class_to_idx[label]
        data.append((wav_data, label_idx))

#print(f"✅ Loaded {len(data)} audio files before hitting RAM limit.")

# Generator function
def generator():
    """Yields (audio waveform, label) pairs."""
    for wav, label in data:
        yield wav, label

# Create dataset using generator
main_ds = tf.data.Dataset.from_generator(
    generator,
    output_signature=(
        tf.TensorSpec(shape=(None,), dtype=tf.float32),  # Audio waveform
        tf.TensorSpec(shape=(), dtype=tf.int32)          # Label
    )
)

def extract_embedding(wav_data, label):
    wav_data = tf.reshape(wav_data, [-1])
    scores, embeddings, spectrogram = yamnet_model(wav_data)
    num_embeddings = tf.shape(embeddings)[0]
    return embeddings, tf.repeat(label, num_embeddings)

# Extract embeddings using YAMNet
#def extract_embedding(wav_data, label):
    """Run YAMNet to extract embeddings from the audio waveform."""

    # Ensure YAMNet gets a 1D tensor (No extra batch dimension)
    #wav_data = tf.reshape(wav_data, [-1])

    #a Run YAMNet model
    #scores, embeddings, spectrogram = yamnet_model(wav_data)

    # Get number of embeddings
    #num_embeddings = tf.shape(embeddings)[0]

    #return embeddings, tf.repeat(label, num_embeddings)

# Apply mapping and unbatch
main_ds = main_ds.map(extract_embedding).unbatch()

print("✅ Dataset prepared successfully.")

Relevant log output

mai_ds_file has found

tensorflow_hub Version

0.13.0.dev (unstable development build)

TensorFlow Version

2.8 (latest stable release)

Other libraries

!pip install tensorflow-io
!pip install scikit-learn
!pip install tensorflow==2.17.1

Python Version

3.x

OS

Linux

[stevemessick]

@4Sheetal According to this stackoverflow thread, you might be able to solve the problem by reinstalling numpy as version 1.26.3 or 1.26.4.

Also, if you do a Google search for the error message you'll get an AI-generated summary of various fixes to try.