TF Model Save/Load — Avoid Wrong Version in Serving

import tensorflow as tf
from tensorflow.keras import models, layers

# io.thecodeforge: Standard Model Serialization
# Create a simple model
model = models.Sequential([layers.Dense(10, input_shape=(5,))])
model.compile(optimizer='adam', loss='mse')

# 1. Save as a directory (SavedModel format - Recommended for Production)
# Version directory is mandatory for TF Serving compatibility
model.save('forge_production_v1/2')

# 2. Save as a single file (H5 format - For simple sharing only)
model.save('forge_legacy_model.h5')

# Loading back
new_model = models.load_model('forge_production_v1/2')
print("Model loaded successfully!")

# Inspect serving signature
import subprocess
result = subprocess.run(['saved_model_cli', 'show', '--dir', 'forge_production_v1/2', '--all'], capture_output=True, text=True)
print(result.stdout)

# io.thecodeforge: Automated Checkpoint Strategy
checkpoint_path = "training_checkpoints/forge_model_{epoch:02d}"

cp_callback = tf.keras.callbacks.ModelCheckpoint(
    filepath=checkpoint_path,
    save_weights_only=False, # Save full model for easy resume
    save_best_only=True,     # Keeps only the version with the lowest validation loss
    monitor='val_loss',
    verbose=1
)

early_stop = tf.keras.callbacks.EarlyStopping(
    monitor='val_loss',
    patience=5,
    restore_best_weights=True
)

# The model will now save its 'progress' after every epoch
# model.fit(train_data, train_labels, epochs=50,
#           validation_data=(val_data, val_labels),
#           callbacks=[cp_callback, early_stop])

package io.thecodeforge.ml;

import org.tensorflow.SavedModelBundle;
import org.tensorflow.Session;
import org.tensorflow.Tensor;

public class ModelLoader {
    /**
     * io.thecodeforge: Loading a Python-trained SavedModel in Java
     */
    public static void loadAndPredict(String modelPath) {
        try (SavedModelBundle model = SavedModelBundle.load(modelPath, "serve")) {
            Session session = model.session();
            // Logic for wrapping inputs into Tensors and running session.runner()
            System.out.println("Forge Model successfully loaded in Java runtime.");
        }
    }
}

-- io.thecodeforge: Registering Model Artifacts
INSERT INTO io.thecodeforge.model_artifacts (
    version_tag,
    format_type,
    storage_path,
    final_accuracy,
    created_at
) VALUES (
    'v1.2.0-prod',
    'SavedModel',
    '/mnt/storage/models/forge_production_v1/2/',
    0.9421,
    CURRENT_TIMESTAMP
);

# io.thecodeforge: Production Model Serving
FROM tensorflow/serving:latest

# Set the model name
ENV MODEL_NAME=forge_model

# Copy the SavedModel directory into the container
# The path /models/model_name/version_number/ is mandatory for TF Serving
COPY /forge_production_v1/2 /models/forge_model/2

# Expose the gRPC and REST ports
EXPOSE 8500
EXPOSE 8501