Mobile deployment (TFLite, Core ML) in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding TFLite Model Conversion

Which of the following statements about converting a TensorFlow model to TensorFlow Lite (TFLite) format is correct?

A TFLite models can only run on Android devices, not iOS.

B TFLite conversion always requires retraining the model from scratch.

C TFLite conversion can include optimizations like quantization to reduce model size and improve speed.

D TFLite conversion automatically converts all Python code in the model to C++.

💡 Hint

Think about what optimizations help models run faster and smaller on mobile devices.

TFLite conversion supports optimizations such as quantization, which reduces model size and speeds up inference on mobile devices. It does not require retraining, works on multiple platforms, and does not convert Python code.

❓ Predict Output

intermediate

Output Shape After TFLite Model Inference

Given a TFLite model that takes an input image of shape (1, 224, 224, 3) and outputs a tensor of shape (1, 1000), what does the output represent?

A A batch of 1 image with 1000 class probabilities.

B 1000 images each of size 1x224x224x3.

C A single scalar value representing the predicted class index.

D A 4D tensor representing feature maps for 1000 layers.

💡 Hint

Consider typical classification model outputs and their shapes.

The output shape (1, 1000) means one image batch with 1000 class scores, typical for classification tasks like ImageNet.

❓ Model Choice

advanced

Choosing a Model for Core ML Deployment

You want to deploy a computer vision model on iOS using Core ML. Which model architecture is best suited for mobile deployment considering speed and size?

A MobileNetV2 with depthwise separable convolutions.

B VGG-19 with large fully connected layers.

C ResNet-152 with 60 million parameters.

D DenseNet-201 with dense connections.

💡 Hint

Think about models designed specifically for mobile and embedded devices.

MobileNetV2 uses depthwise separable convolutions to reduce size and computation, making it ideal for mobile deployment with Core ML.

❓ Hyperparameter

advanced

Optimizing Quantization Parameters for TFLite

When applying post-training quantization to a TensorFlow model before converting to TFLite, which hyperparameter setting most directly affects the model's size reduction?

A Choosing the number of training epochs.

B Changing the batch size during inference.

C Adjusting the learning rate during training.

D Selecting the bit-width for weights and activations (e.g., 8-bit vs 16-bit).

💡 Hint

Quantization reduces precision to shrink model size.

The bit-width used in quantization directly controls how much the model size shrinks; lower bit-width means smaller size.

🔧 Debug

expert

Debugging Core ML Model Input Shape Mismatch

You converted a TensorFlow model to Core ML but get an error when running inference: "Input shape mismatch: expected (1, 224, 224, 3) but got (3, 224, 224)". What is the most likely cause?

A The Core ML model expects channels last format but input is channels first.

B The batch dimension is missing in the input tensor provided to Core ML.

C The input image size is incorrect; it should be 224x224 pixels.

D The model was converted without specifying the output layer.

💡 Hint

Check if the input tensor includes the batch size dimension.

Core ML models typically expect a batch dimension. The error shows the input lacks the batch size (1), causing the mismatch.