Prepare Data

Prepare your training dataset

Begin by capturing images with your camera and performing the necessary preprocessing steps. Then, annotate the captured images to identify the materials present in each one. These annotations allow the model to learn which materials it must recognize during training. You must save annotations in COCO JSON format.

The GitHub repository provides the required preprocessing scripts to prepare the model for retraining. These scripts convert your annotated images into TFRecords, the required input format for TensorFlow models.

The preprocessing scripts of the repository let you perform the following actions:

• Convert image annotations to the COCO-annotated JSON file format, which is necessary for labels and metadata on a dataset.
• Clean and prepare a COCO-annotated JSON file for training an ML model, ensuring that your dataset is clean, consistent, and ready for effective model training.
• Filter out irrelevant or noisy annotations that could negatively impact the training process.
• Adjust annotations to ensure they are in the optimal format for model training.
• Verify that all annotated images exist and are not corrupted.
• Merge COCO-annotated JSON files into a single file and convert it into TFRecord files, the required format for training with the Mask R-CNN model.

Before proceeding with the retraining pipeline, run the preprocessing scripts with your data on your local workstation, remote server, or database. Once your TFRecords are ready, upload them to a Cloud Storage bucket. The bucket can have two locations, one for the training dataset and another for the validation dataset.

---

Configure the training job

Configure your model training by customizing values in the CircularNET_Vertex_AI_ReTraining_v1.ipynb script, including your project ID, bucket URI, and region. Provide the following information in the corresponding script variables:

• input_train_data_path: path to the TFRecords of your training dataset in the Cloud Storage bucket.
• input_validation_data_path: path to the TFRecords of your validation dataset in the Cloud Storage bucket.
• init_checkpoint_path: path to the initial checkpoints with the model's weights. You can use the open-source initial checkpoints in this variable from the configuration file.
• config_file_path: path to the configuration file containing parameters for fine-tuning the training. You can use the open-source configuration file in this variable.
• service_account: name of the Vertex AI service account with permissions to perform training jobs.

The script's placeholders, such as PROJECT_ID, REGION, and STAGING_BUCKET, must be replaced with your Google Cloud project, region, and Cloud Storage bucket, respectively.

Note: You can also customize other script variables, such as num_classes, which refers to the number of categories for materials or other classes in your annotated images.