When using generator for representative dataset in quantization it "Failed to convert value into readable tensor"

Question

I am quantizing a model. The model takes 224x224 input.

I preprocess the data with a build-in function preprocess_input() which subtracts some center pixels.

Now when using a simple image with this preprocessing function in the representative_dataset_gen() everything works fine

def representative_dataset_gen():
        pfad='./000001.jpg'
        img = cv2.imread(pfad)
        img = np.expand_dims(img,0).astype(np.float32) 
        img = preprocess_input(img) 
        yield [img]

But when I use a generator function to use more than one image

def prepare(img):
    img = np.expand_dims(img,0).astype(np.float32)
    img = preprocess_input(img)
    return arg
      
    
repDatagen=tf.keras.preprocessing.image.ImageDataGenerator(preprocessing_function=prepare)
datagen=repDatagen.flow_from_directory(folderpath,target_size=size,batch_size=1)


def representative_dataset_gen():
  for _ in range(10):
    img = datagen.next()
    yield [img]

I get following error:

ValueError: Failed to convert value into readable tensor.

My guess: This is due to ImageDataGenerator(preprocessing_function=prepare) . In the tensorflow description it says:

function that will be applied on each input. The function will run after the image is resized and augmented. The function should take one argument: one image (Numpy tensor with rank 3), and should output a Numpy tensor with the same shape.

I tried to adjust the shape of the img output of the "prepare" function without and with np.squeez(). This results in either (1,244,244,3) or (224,224,3). But I still get the error. I also tried tf.convert_to_tensor() with the same error.

def prepare(img):
    img = np.expand_dims(img,0).astype(np.float32)
    img = preprocess_input(img, version=2)
    img = np.squeeze(img)
    arg = tf.convert_to_tensor(img, dtype=tf.float32)
    return arg

Does anyone know how I have to prepare the output to get the correct tensor?

Thanks

Answer 1

Thank you, you are amazing and your post save my day, I apply quantization on YOLO model, the code:

#https://github.com/ultralytics/ultralytics/issues/6676
#https://github.com/onnx/onnx-tensorflow
#https://github.com/onnx/onnx-tensorflow/blob/main/example/onnx_to_tf.py
#https://docs.ultralytics.com/modes/export/#usage-examples
#https://coral.ai/docs/edgetpu/models-intro/#compatibility-overview

import tensorflow as tf
import numpy as np
import os 
import cv2

def representative_data_gen():
  image_paths = "/..../images/"
  op_ = 0

  for image_path in os.listdir(image_paths):
    op_ = op_ + 1 
    image = cv2.imread("/.../images/"+image_path)    
    print("------111-------", image.shape)  
    image = cv2.resize(image, (640, 640))
    print("------222-------", image.shape)  
    image = np.array(image).astype(np.float32) / 255.0 
    print("------333-------", image.shape)  
    image = np.transpose(image, (2, 0, 1))  # (H, W, C) -> (C, H, W)
    print("------444-------", image.shape)  
    image = np.expand_dims(image, axis=0)  # Add batch dimension

    print("------555-------", image.shape)  
    if op_ == 100:
      break
  
    print(len(image), len(image[0]), len(image[0][0]), len(image[0][0][0]))
    yield [image] #[np.array(image[0][0], dtype=np.float32)]# [image]
    print(" what next ")
   


# ... (Load your TensorFlow model)
converter = tf.lite.TFLiteConverter.from_saved_model('/..../teston')

converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
converter.inference_input_type = tf.int8  
converter.inference_output_type = tf.int8  
converter.representative_dataset = representative_data_gen  # Provide a data generator function

tflite_quant_model = converter.convert()

# Save the quantized model
with open('/..../teston/quantized_model2.tflite', 'wb') as f:
    f.write(tflite_quant_model)

Answer 2

The problem was actually the generator output. The output is not a np.array() . Converting with np.asarray() was not possible as it throws an error.

So the answer is quite simple:

the data_generator output is a tuple

so simply yield only the first element img[0].

---

The code in total:

size=(224,224)

def prepare(img):
    img = np.expand_dims(img,0).astype(np.float32)
    img = preprocess_input(img, version=2)
    return img
      
repDatagen=tf.keras.preprocessing.image.ImageDataGenerator(preprocessing_function=prepare)
datagen=repDatagen.flow_from_directory(folderpath,target_size=size,batch_size=1)

def representative_dataset_gen():
  for _ in range(10):
    img = datagen.next()
    yield [img[0]]
    #yield [np.array(img[0], dtype=np.float32)] # also possible