Recurrent Residual Convolutional Neural Network ¶

Installation¶

You will need Python 3.12 or later, and a working JAX installation. For example, you can install JAX with:

In [1]:

%pip install -U "jax[cuda]"

%pip install -U "jax[cuda]"

/home/faldor_google_com/dev/cax/.venv/bin/python3: No module named pip

Note: you may need to restart the kernel to use updated packages.

Then, install CAX from PyPi:

In [2]:

%pip install -U "cax[examples]"

%pip install -U "cax[examples]"

/home/faldor_google_com/dev/cax/.venv/bin/python3: No module named pip

Note: you may need to restart the kernel to use updated packages.

Import¶

In [3]:

import time

import jax
import jax.numpy as jnp
import mediapy
import optax
import torchvision
from flax import nnx
from jax import Array

from cax.core import ComplexSystem
from cax.core.perceive import ConvPerceive
from cax.core.update import ResidualUpdate
from cax.utils.render import clip_and_uint8

import time import jax import jax.numpy as jnp import mediapy import optax import torchvision from flax import nnx from jax import Array from cax.core import ComplexSystem from cax.core.perceive import ConvPerceive from cax.core.update import ResidualUpdate from cax.utils.render import clip_and_uint8

Configuration¶

In [4]:

seed = 0

spatial_dims = (28, 28)
channel_size = 64
perception_size = 64
update_hidden_layer_sizes = (128, 128)
cell_dropout_rate = 0.5

num_steps = 64
batch_size = 8
learning_rate = 1e-3

key = jax.random.key(seed)
rngs = nnx.Rngs(seed)

seed = 0 spatial_dims = (28, 28) channel_size = 64 perception_size = 64 update_hidden_layer_sizes = (128, 128) cell_dropout_rate = 0.5 num_steps = 64 batch_size = 8 learning_rate = 1e-3 key = jax.random.key(seed) rngs = nnx.Rngs(seed)

Dataset¶

In [5]:

# Load MNIST dataset
ds_train = torchvision.datasets.MNIST(root="./data", train=True, download=True)
ds_test = torchvision.datasets.MNIST(root="./data", train=False, download=True)

# Convert to jax.Array
y_train = jnp.array([y.resize(spatial_dims) for y, _ in ds_train])[..., None] / 255
y_test = jnp.array([y.resize(spatial_dims) for y, _ in ds_test])[..., None] / 255

# Visualize
mediapy.show_images(y_train[:8].repeat(4, axis=-3).repeat(4, axis=-2))

# Load MNIST dataset ds_train = torchvision.datasets.MNIST(root="./data", train=True, download=True) ds_test = torchvision.datasets.MNIST(root="./data", train=False, download=True) # Convert to jax.Array y_train = jnp.array([y.resize(spatial_dims) for y, _ in ds_train])[..., None] / 255 y_test = jnp.array([y.resize(spatial_dims) for y, _ in ds_test])[..., None] / 255 # Visualize mediapy.show_images(y_train[:8].repeat(4, axis=-3).repeat(4, axis=-2))

Instantiate system¶

In [6]:

class RRCNN(ComplexSystem):
    """Recurrent Residual Convolutional Neural Network class."""

    def __init__(self, *, rngs: nnx.Rngs):
        """Initialize RRCNN.

        Args:
                rngs: rng key.

        """
        self.perceive = ConvPerceive(
            channel_size=channel_size,
            perception_size=perception_size,
            rngs=rngs,
        )
        self.update = ResidualUpdate(
            num_spatial_dims=len(spatial_dims),
            channel_size=channel_size,
            perception_size=perception_size,
            hidden_layer_sizes=update_hidden_layer_sizes,
            cell_dropout_rate=cell_dropout_rate,
            zeros_init=True,
            rngs=rngs,
        )

    def _step(self, state: Array, input: Array | None = None) -> Array:
        perception = self.perceive(state)
        next_state = self.update(state, perception, input)

        return next_state

    @nnx.jit
    def render(self, state):
        """Render state to RGB."""
        gray = state[..., -1:]
        rgb = jnp.repeat(gray, 3, axis=-1)

        # Clip values to valid range and convert to uint8
        return clip_and_uint8(rgb)

class RRCNN(ComplexSystem): """Recurrent Residual Convolutional Neural Network class.""" def __init__(self, *, rngs: nnx.Rngs): """Initialize RRCNN. Args: rngs: rng key. """ self.perceive = ConvPerceive( channel_size=channel_size, perception_size=perception_size, rngs=rngs, ) self.update = ResidualUpdate( num_spatial_dims=len(spatial_dims), channel_size=channel_size, perception_size=perception_size, hidden_layer_sizes=update_hidden_layer_sizes, cell_dropout_rate=cell_dropout_rate, zeros_init=True, rngs=rngs, ) def _step(self, state: Array, input: Array | None = None) -> Array: perception = self.perceive(state) next_state = self.update(state, perception, input) return next_state @nnx.jit def render(self, state): """Render state to RGB.""" gray = state[..., -1:] rgb = jnp.repeat(gray, 3, axis=-1) # Clip values to valid range and convert to uint8 return clip_and_uint8(rgb)

In [7]:

cs = RRCNN(rngs=rngs)

cs = RRCNN(rngs=rngs)

In [8]:

params = nnx.state(cs, nnx.Param)
print("Number of params:", sum(x.size for x in jax.tree.leaves(params)))

params = nnx.state(cs, nnx.Param) print("Number of params:", sum(x.size for x in jax.tree.leaves(params)))

Number of params: 69952

Sample initial state¶

In [9]:

def add_noise(image, alpha, key):
    """Add noise to the image with a given alpha value."""
    noise = jax.random.normal(key, image.shape)
    noisy_image = (1 - alpha) * image + alpha * noise
    return jnp.clip(noisy_image, 0.0, 1.0)


def sample_state(key):
    """Sample a state with a randomly sampled image and added noise."""
    state = jnp.zeros(y_train.shape[1:3] + (channel_size,))

    sample_key, alpha_key, noise_key = jax.random.split(key, 3)

    # Sample a target image
    y_idx = jax.random.choice(sample_key, y_train.shape[0])
    y = y_train[y_idx]

    # Add noise
    alpha = jax.random.uniform(alpha_key)
    noisy_y = add_noise(y, alpha, noise_key)

    return state.at[..., -1:].set(noisy_y), y_idx

def add_noise(image, alpha, key): """Add noise to the image with a given alpha value.""" noise = jax.random.normal(key, image.shape) noisy_image = (1 - alpha) * image + alpha * noise return jnp.clip(noisy_image, 0.0, 1.0) def sample_state(key): """Sample a state with a randomly sampled image and added noise.""" state = jnp.zeros(y_train.shape[1:3] + (channel_size,)) sample_key, alpha_key, noise_key = jax.random.split(key, 3) # Sample a target image y_idx = jax.random.choice(sample_key, y_train.shape[0]) y = y_train[y_idx] # Add noise alpha = jax.random.uniform(alpha_key) noisy_y = add_noise(y, alpha, noise_key) return state.at[..., -1:].set(noisy_y), y_idx

Train¶

Optimizer¶

In [10]:

lr_sched = optax.linear_schedule(
    init_value=learning_rate, end_value=0.1 * learning_rate, transition_steps=2_000
)

optimizer = optax.chain(
    optax.clip_by_global_norm(1.0),
    optax.adam(learning_rate=lr_sched),
)

update_params = nnx.All(nnx.Param, nnx.PathContains("update"))
optimizer = nnx.Optimizer(cs, optimizer, wrt=update_params)

lr_sched = optax.linear_schedule( init_value=learning_rate, end_value=0.1 * learning_rate, transition_steps=2_000 ) optimizer = optax.chain( optax.clip_by_global_norm(1.0), optax.adam(learning_rate=lr_sched), ) update_params = nnx.All(nnx.Param, nnx.PathContains("update")) optimizer = nnx.Optimizer(cs, optimizer, wrt=update_params)

Loss¶

In [11]:

def mse(state, y):
    """Mean Squared Error."""
    return jnp.mean(jnp.square(state[..., -1:] - y))

def mse(state, y): """Mean Squared Error.""" return jnp.mean(jnp.square(state[..., -1:] - y))

In [12]:

@nnx.jit
def loss_fn(cs, state, y):
    """Loss function."""
    state_axes = nnx.StateAxes({nnx.RngState: 0, ...: None})
    state = nnx.split_rngs(splits=batch_size)(
        nnx.vmap(
            lambda cs, state: cs(state, num_steps=num_steps),
            in_axes=(state_axes, 0),
        )
    )(cs, state)
    loss = mse(state, y)
    return loss

@nnx.jit def loss_fn(cs, state, y): """Loss function.""" state_axes = nnx.StateAxes({nnx.RngState: 0, ...: None}) state = nnx.split_rngs(splits=batch_size)( nnx.vmap( lambda cs, state: cs(state, num_steps=num_steps), in_axes=(state_axes, 0), ) )(cs, state) loss = mse(state, y) return loss

Train step¶

In [13]:

@nnx.jit
def train_step(cs, optimizer, key):
    """Train step."""
    keys = jax.random.split(key, batch_size)
    current_state, y_idx = jax.vmap(sample_state)(keys)
    y = y_train[y_idx]

    loss, grad = nnx.value_and_grad(loss_fn, argnums=nnx.DiffState(0, update_params))(
        cs, current_state, y
    )
    optimizer.update(cs, grad)

    return loss

@nnx.jit def train_step(cs, optimizer, key): """Train step.""" keys = jax.random.split(key, batch_size) current_state, y_idx = jax.vmap(sample_state)(keys) y = y_train[y_idx] loss, grad = nnx.value_and_grad(loss_fn, argnums=nnx.DiffState(0, update_params))( cs, current_state, y ) optimizer.update(cs, grad) return loss

Main loop¶

In [14]:

num_train_steps = 8_192
print_interval = 128

losses = []
start = time.perf_counter()
for i in range(num_train_steps):
    key, subkey = jax.random.split(key)
    loss = train_step(cs, optimizer, subkey)
    losses.append(loss)

    if i % print_interval == 0 or i == num_train_steps - 1:
        avg_loss = sum(losses[-print_interval:]) / len(losses[-print_interval:])
        elapsed = time.perf_counter() - start
        print(f"Step {i:>4}/{num_train_steps} | {elapsed:6.1f}s | Loss {avg_loss:.3e}")

print(f"✨ Trained for {num_train_steps} steps in {time.perf_counter() - start:.0f}s")

num_train_steps = 8_192 print_interval = 128 losses = [] start = time.perf_counter() for i in range(num_train_steps): key, subkey = jax.random.split(key) loss = train_step(cs, optimizer, subkey) losses.append(loss) if i % print_interval == 0 or i == num_train_steps - 1: avg_loss = sum(losses[-print_interval:]) / len(losses[-print_interval:]) elapsed = time.perf_counter() - start print(f"Step {i:>4}/{num_train_steps} | {elapsed:6.1f}s | Loss {avg_loss:.3e}") print(f"✨ Trained for {num_train_steps} steps in {time.perf_counter() - start:.0f}s")

Step    0/8192 |    6.1s | Loss 1.742e-01

Step  128/8192 |    7.9s | Loss 6.702e-02

Step  256/8192 |    9.5s | Loss 5.388e-02

Step  384/8192 |   11.1s | Loss 5.062e-02

Step  512/8192 |   12.8s | Loss 4.548e-02

Step  640/8192 |   14.4s | Loss 4.144e-02

Step  768/8192 |   16.1s | Loss 4.261e-02

Step  896/8192 |   17.8s | Loss 4.335e-02

Step 1024/8192 |   19.4s | Loss 3.948e-02

Step 1152/8192 |   21.1s | Loss 3.982e-02

Step 1280/8192 |   22.7s | Loss 4.190e-02

Step 1408/8192 |   24.4s | Loss 3.994e-02

Step 1536/8192 |   26.0s | Loss 3.917e-02

Step 1664/8192 |   27.7s | Loss 3.595e-02

Step 1792/8192 |   29.3s | Loss 3.742e-02

Step 1920/8192 |   31.0s | Loss 3.579e-02

Step 2048/8192 |   32.6s | Loss 3.524e-02

Step 2176/8192 |   34.3s | Loss 3.755e-02

Step 2304/8192 |   35.9s | Loss 3.589e-02

Step 2432/8192 |   37.6s | Loss 3.647e-02

Step 2560/8192 |   39.3s | Loss 3.518e-02

Step 2688/8192 |   40.9s | Loss 3.576e-02

Step 2816/8192 |   42.6s | Loss 3.539e-02

Step 2944/8192 |   44.2s | Loss 3.634e-02

Step 3072/8192 |   45.9s | Loss 3.675e-02

Step 3200/8192 |   47.5s | Loss 3.578e-02

Step 3328/8192 |   49.2s | Loss 3.435e-02

Step 3456/8192 |   50.8s | Loss 3.677e-02

Step 3584/8192 |   52.5s | Loss 3.607e-02

Step 3712/8192 |   54.1s | Loss 3.520e-02

Step 3840/8192 |   55.8s | Loss 3.654e-02

Step 3968/8192 |   57.5s | Loss 3.514e-02

Step 4096/8192 |   59.1s | Loss 3.494e-02

Step 4224/8192 |   60.8s | Loss 3.450e-02

Step 4352/8192 |   62.5s | Loss 3.723e-02

Step 4480/8192 |   64.1s | Loss 3.623e-02

Step 4608/8192 |   65.8s | Loss 3.521e-02

Step 4736/8192 |   67.4s | Loss 3.559e-02

Step 4864/8192 |   69.1s | Loss 3.565e-02

Step 4992/8192 |   70.7s | Loss 3.531e-02

Step 5120/8192 |   72.3s | Loss 3.579e-02

Step 5248/8192 |   74.0s | Loss 3.571e-02

Step 5376/8192 |   75.6s | Loss 3.496e-02

Step 5504/8192 |   77.3s | Loss 3.447e-02

Step 5632/8192 |   78.9s | Loss 3.455e-02

Step 5760/8192 |   80.6s | Loss 3.495e-02

Step 5888/8192 |   82.3s | Loss 3.391e-02

Step 6016/8192 |   83.9s | Loss 3.437e-02

Step 6144/8192 |   85.6s | Loss 3.549e-02

Step 6272/8192 |   87.3s | Loss 3.641e-02

Step 6400/8192 |   88.9s | Loss 3.472e-02

Step 6528/8192 |   90.5s | Loss 3.475e-02

Step 6656/8192 |   92.2s | Loss 3.383e-02

Step 6784/8192 |   93.8s | Loss 3.360e-02

Step 6912/8192 |   95.5s | Loss 3.488e-02

Step 7040/8192 |   97.1s | Loss 3.717e-02

Step 7168/8192 |   98.8s | Loss 3.469e-02

Step 7296/8192 |  100.5s | Loss 3.575e-02

Step 7424/8192 |  102.1s | Loss 3.539e-02

Step 7552/8192 |  103.7s | Loss 3.494e-02

Step 7680/8192 |  105.4s | Loss 3.557e-02

Step 7808/8192 |  107.0s | Loss 3.538e-02

Step 7936/8192 |  108.7s | Loss 3.390e-02

Step 8064/8192 |  110.3s | Loss 3.423e-02

Step 8191/8192 |  112.0s | Loss 3.493e-02
✨ Trained for 8192 steps in 112s

Run¶

In [15]:

num_examples = 8

key, subkey = jax.random.split(key)
keys = jax.random.split(subkey, num_examples)
state_init, y_idx = jax.vmap(sample_state)(keys)

state_axes = nnx.StateAxes({nnx.RngState: 0, ...: None})
state_final, states = nnx.split_rngs(splits=num_examples)(
    nnx.vmap(
        lambda cs, state: cs(state, num_steps=2 * num_steps, return_states=True),
        in_axes=(state_axes, 0),
    )
)(cs, state_init)

num_examples = 8 key, subkey = jax.random.split(key) keys = jax.random.split(subkey, num_examples) state_init, y_idx = jax.vmap(sample_state)(keys) state_axes = nnx.StateAxes({nnx.RngState: 0, ...: None}) state_final, states = nnx.split_rngs(splits=num_examples)( nnx.vmap( lambda cs, state: cs(state, num_steps=2 * num_steps, return_states=True), in_axes=(state_axes, 0), ) )(cs, state_init)

Visualize¶

In [16]:

states = jnp.concatenate([state_init[:, None], states], axis=1)
frames = nnx.vmap(
    lambda cs, state: cs.render(state),
    in_axes=(None, 0),
)(cs, states)

mediapy.show_images(y_train[y_idx].repeat(4, axis=-3).repeat(4, axis=-2))
mediapy.show_videos(frames.repeat(4, axis=-3).repeat(4, axis=-2))

states = jnp.concatenate([state_init[:, None], states], axis=1) frames = nnx.vmap( lambda cs, state: cs.render(state), in_axes=(None, 0), )(cs, states) mediapy.show_images(y_train[y_idx].repeat(4, axis=-3).repeat(4, axis=-2)) mediapy.show_videos(frames.repeat(4, axis=-3).repeat(4, axis=-2))