AI Practices with Tensorflow & Blender

Create training data for the model by scripting automatic coordinates for the cube's vertex positions, using random values to reflect equal distances, and save the results in a json file.

Create a python script to generate 3d cube vertices from random samples inside a for loop using enumerate, scale to 1–10, and save to cube_data.json using json and numpy.

Create and inspect cube vertex data by generating cube_data.json, then sort the coordinates of the vertices to visualize each cube's coordinate arrays more clearly.

Learn how a generative adversarial network trains a generator and a discriminator to reproduce cube coordinates, with the model unaware of what it creates yet learning to build a cube.

Import json and TensorFlow, load cube_data.json into data, convert to tensor, and build train_dataset using from_tensor_slices and batch with 32 for efficient model training.

Build a generator model to produce cube coordinates from a 100-dimensional latent space. Use Sequential network with dense layers and leaky relu to reshape into 8x3 coordinates for 3d figures.

Create a discriminator that distinguishes real from generated data using Conv1D layers with 32 and 64 filters, LeakyReLu activations, Flatten, Dense with 1 neuron, and a sigmoid output.

Compile the generator and discriminator with Adam optimizers, set learning_rate 0.0002 and beta_1 0.5, and define generator_loss and discriminator_loss using binary_crossentropy, ones_like, and zeros_like.

Construct a training loop for a gan with for loops over num_epochs and train_dataset, using tf.GradientTape to update the discriminator and generator with real_data_batch, generated_data_batch, and noise.

Train models to generate eight coordinates representing a cube. Print the generated tensor to the console and convert to a numpy array with tolist for Blender visualization.

Generate a cube in blender from model-produced coordinates by running a GAN script via a subprocess, and use the bpy library to access blender elements to create the figure.

Learn to write a blender script in vscode using bpy, subprocess, and json to create a mesh, link it to a collection, and build a 3D figure from vertex data.

Obtain cube generator data, set script_path, and run the generator via subprocess.run using python3.9. Parse stdout with json.loads to fill my_verts and my_faces, then run in Blender to generate figures.

Run a Blender script to generate a figure and observe how training data and epochs shape vertex coordinates, biasing toward cubes with more data and epochs.

Execute a Blender script in the scripting view to generate 3d coordinates from 100 cubes using 100 epochs, then compare multiple figures in layout to refine the cube form.

Increase num_epochs to 300 to let the model learn from 100 cubes, run the Blender script, and compare results as figures become more cube-like.

We modify the script to generate 500 data points, set num_epochs to 100, run the Blender script, observe distorted cube figures, and plan to increase epochs in the next video.

Increase num_epochs to 300 and run the Blender script on 500 elements, then switch to layout to compare results. Notice cube-like figures with a generally stable structure and fewer distortions.

Modify the script to export coordinates without training, run in Blender, and observe that the generated figures remain random and far from cube-shaped, highlighting quick testing without training.

Apply what you learned to build models that learn to generate the information needed to create 3d figures.