Gem Classifier Project by Sapir Dahan

Overview

This project aims to classify 75 different types of gem images using a neural network model. Given the limited dataset, transfer learning is employed using a pre-trained ResNet50 model. The project is implemented using TensorFlow.

Project Structure

The project consists of two main parts:

1. Training: Training the gem classifier model using transfer learning.
2. Inference: Using the trained model to classify new gem images.

Training

The training process involves the following steps:

1. Examine and Understand the Data: Analyze the dataset to understand its structure and contents.
2. Build an Input Pipeline: Prepare the data for training, including data augmentation and batching.
3. Compose the Model:
   • Load the pre-trained ResNet50 base model.
   • Stack the classification layers on top of the base model.
4. Train the Model:
   • Feature Extraction: Freeze the pre-trained layers and train the new classification layers.
   • Fine-Tuning: Unfreeze some of the pre-trained layers and jointly train them with the new layers.
5. Evaluate the Model: Assess the performance of the model on the test set.

Hyperparameters

• BATCH_SIZE: 32
• IMG_SIZE: (224, 224)
• test_split: 0.1
• validation_split: 0.1
• initial_epochs: 300
• fine_tune_epochs: 2000
• base_learning_rate: 0.0001

Inference

The inference process involves the following steps:

1. Load the Model and Class Names:
   • Load the pre-trained model using tf.keras.models.load_model.
   • Load the class names from a pickle file.
2. Process the Image:
   • Load and resize the input image to (224, 224).
   • Convert the image to a NumPy array and reshape it for model input.
3. Make Predictions:
   • Use the model to predict the class of the input image.
   • Extract the top 5 predictions and their confidence levels.
4. Visualize the Results:
   • Create a bar plot showing the top 5 predicted classes and their confidence levels.
   • Display the input image.

How to Use

Training

1. Ensure you have TensorFlow installed.
2. Prepare your dataset and adjust the hyperparameters if necessary.
3. Run the Gem_Classifier_Train.ipynb notebook to train the model.

Inference

1. Ensure you have TensorFlow and the necessary dependencies installed.
2. Place your input images in the specified directory.
3. Run the Gem_Classifier_Inference.ipynb notebook to classify new gem images.

Dependencies

• TensorFlow
• NumPy
• Matplotlib
• Pillow
• Pickle