DevHeads IoT Integration Server•5d ago

Fixing TensorFlow Lite Dynamic Tensor Error and Optimizing Mask R-CNN on Raspberry Pi

I'm deploying a fine-tuned Mask R-CNN model (ResNet-101) on a Raspberry Pi 4 (4GB RAM) running Raspbian OS, using TensorFlow Lite v2.6 for aerial object detection. During inference, I get the following error:

ValueError: TensorFlow Lite currently supports models with fixed-size input tensors, but the model has dynamic-sized input tensors. Expected input shape: [1, 1024, 1024, 3], but received input shape: [1, 512, 512, 3].

ValueError: TensorFlow Lite currently supports models with fixed-size input tensors, but the model has dynamic-sized input tensors. Expected input shape: [1, 1024, 1024, 3], but received input shape: [1, 512, 512, 3].

Despite resizing all input images to 1024x1024, the model still expects dynamic input shapes. Here's the code for loading and running inference:

import tensorflow as tf
import numpy as np
from PIL import Image

interpreter = tf.lite.Interpreter(model_path="mask_rcnn_model.tflite")
interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

image = Image.open('aerial_image.jpg').resize((1024, 1024))
input_data = np.expand_dims(np.array(image), axis=0).astype(np.float32)

interpreter.set_tensor(input_details[0]['index'], input_data)
interpreter.invoke()
output_data = interpreter.get_tensor(output_details[0]['index'])
print(output_data)

import tensorflow as tf
import numpy as np
from PIL import Image

interpreter = tf.lite.Interpreter(model_path="mask_rcnn_model.tflite")
interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

image = Image.open('aerial_image.jpg').resize((1024, 1024))
input_data = np.expand_dims(np.array(image), axis=0).astype(np.float32)

interpreter.set_tensor(input_details[0]['index'], input_data)
interpreter.invoke()
output_data = interpreter.get_tensor(output_details[0]['index'])
print(output_data)

I'm also facing significant inference delays (~10 seconds per image). I attempted post-training quantization using float16 and int8, but the performance remains suboptimal. What strategies can I use to fix the dynamic tensor error, optimize inference speed on Raspberry Pi, and improve detection accuracy for small objects in aerial imagery?

1 Reply

wafa_ath•5d ago

Hi, the essuie was fixed after i switched to YOLOv4 , maybe because it's lighter model compared to R-CNN and more suitable for raspberry pi4

Gaming

Programming

Fixing TensorFlow Lite Dynamic Tensor Error and Optimizing Mask R-CNN on Raspberry Pi