Memory overflow using scale_by_radam

[HGangloff]

Hi,

I have my RAM getting used up to overflow when I use scale_by_radam gradient transform or equivalently optax.radam without JIT compiling the code. The problem appears on CPU and GPU but does not appear when I use JIT compilation. The problem does not seem to exist with optax.adam.

Here is a MWE derived from optax quick start tutorial:

import random
from typing import Tuple
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '' # uncomment to force CPU

import optax
import jax.numpy as jnp
import jax
import numpy as np

BATCH_SIZE = 500
NUM_TRAIN_STEPS = 10000
RAW_TRAINING_DATA = np.random.randint(255, size=(NUM_TRAIN_STEPS, BATCH_SIZE, 1))

TRAINING_DATA = np.unpackbits(RAW_TRAINING_DATA.astype(np.uint8), axis=-1)
LABELS = jax.nn.one_hot(RAW_TRAINING_DATA % 2, 2).astype(jnp.float32).reshape(NUM_TRAIN_STEPS, BATCH_SIZE, 2)

initial_params = {
    'hidden': jax.random.normal(shape=[8, 200], key=jax.random.PRNGKey(0)),
    'hidden2': jax.random.normal(shape=[200, 100], key=jax.random.PRNGKey(0)),
    'output': jax.random.normal(shape=[100, 2], key=jax.random.PRNGKey(1)),
}


def net(x: jnp.ndarray, params: jnp.ndarray) -> jnp.ndarray:
  x = jnp.dot(x, params['hidden'])
  x = jax.nn.relu(x)
  x = jnp.dot(x, params['hidden2'])
  x = jax.nn.relu(x)
  x = jnp.dot(x, params['output'])
  return x


def loss(params: optax.Params, batch: jnp.ndarray, labels: jnp.ndarray) -> jnp.ndarray:
  y_hat = net(batch, params)

  # optax also provides a number of common loss functions.
  loss_value = optax.sigmoid_binary_cross_entropy(y_hat, labels).sum(axis=-1)

  return loss_value.mean()

def fit(params: optax.Params, optimizer: optax.GradientTransformation) -> optax.Params:
  opt_state = optimizer.init(params)

  #@jax.jit
  def step(params, opt_state, batch, labels):
    loss_value, grads = jax.value_and_grad(loss)(params, batch, labels)
    updates, opt_state = optimizer.update(grads, opt_state, params)
    params = optax.apply_updates(params, updates)
    return params, opt_state, loss_value

  for i, (batch, labels) in enumerate(zip(TRAINING_DATA, LABELS)):
    params, opt_state, loss_value = step(params, opt_state, batch, labels)
    if i % 100 == 0:
      print(f'step {i}, loss: {loss_value}')

  return params

# Finally, we can fit our parametrized function using the Adam optimizer
# provided by optax.
optimizer = optax.radam(learning_rate=1e-2)
params = fit(initial_params, optimizer)

Of course this example is simple enough and does not saturate the RAM before a long time but this issue is really problematic in another particular research project.

The problem seems to be linked with this computation specific to RAdam: https://github.com/deepmind/optax/blob/fc5de3d3951c4dfd87513c6426e30baf505d89ae/optax/_src/transform.py#L685C7-L685C7. But I do not know how to investigate further.

Thanks for your feedback.

[itstalmeez]

Hi HGangloff, Prioritize JIT Compilation:

Compile your code using jax.jit whenever possible to benefit from JAX's optimizations and potentially avoid the RAM issue. Investigate RAdam Implementation:

Explore the RAdam implementation in Optax: https://github.com/deepmind/optax/blob/fc5de3d3951c4dfd87513c6426e30baf505d89ae/optax/_src/transform.py#L685C7-L685C7 Focus on areas that might create large temporary arrays or perform memory-intensive operations. Consider profiling memory usage to pinpoint specific lines or functions causing excessive consumption. Experiment with Alternative Optimizers:

If RAdam's performance is crucial for your research, consider: Modifying RAdam's implementation to reduce memory footprint (if feasible). Exploring alternative optimizers like Yogi, which share similarities with RAdam but might have different memory characteristics. Report to Optax Maintainers:

Share your findings and code examples with the Optax maintainers to bring attention to the issue and potentially contribute to a fix. Additional Considerations:

Memory Profiling: Use tools like jax.profiler or external profilers to track memory usage and identify bottlenecks. Batch Size Adjustment: Experiment with smaller batch sizes to reduce memory requirements per step. Hardware Constraints: Consider available RAM and potential hardware limitations. I'm ready to assist further if you have more questions or require additional guidance. I'll be waiting for your positive response!!!

[carlosgmartin]

Since #969 has been closed as completed, I think this issue can be closed.

[vroulet]

Right, thank you for catching this @carlosgmartin ! @HGangloff feel free to reopen if the matter was not solved but I think it should be good.