Import a JAX model using JAX2TF

import tensorflow as tf
import numpy as np
import jax
import jax.numpy as jnp
import flax
import optax
import os
from matplotlib import pyplot as plt
from jax.experimental import jax2tf
from threading import Lock # Only used in the visualization utility.
from functools import partial

# Needed for TensorFlow and JAX to coexist in GPU memory.
os.environ['XLA_PYTHON_CLIENT_PREALLOCATE'] = "false"
gpus = tf.config.list_physical_devices('GPU')
if gpus:
  try:
    for gpu in gpus:
      tf.config.experimental.set_memory_growth(gpu, True)
  except RuntimeError as e:
    # Memory growth must be set before GPUs have been initialized.
    print(e)

(x_train, train_labels), (x_test, test_labels) = tf.keras.datasets.mnist.load_data()

train_data = tf.data.Dataset.from_tensor_slices((x_train, train_labels))
train_data = train_data.map(lambda x,y: (tf.expand_dims(tf.cast(x, tf.float32)/255.0, axis=-1),
                                         tf.one_hot(y, depth=10)))

BATCH_SIZE = 256
train_data = train_data.batch(BATCH_SIZE, drop_remainder=True)
train_data = train_data.cache()
train_data = train_data.shuffle(5000, reshuffle_each_iteration=True)

test_data = tf.data.Dataset.from_tensor_slices((x_test, test_labels))
test_data = test_data.map(lambda x,y: (tf.expand_dims(tf.cast(x, tf.float32)/255.0, axis=-1),
                                         tf.one_hot(y, depth=10)))
test_data = test_data.batch(10000)
test_data = test_data.cache()

(one_batch, one_batch_labels) = next(iter(train_data)) # just one batch
(all_test_data, all_test_labels) = next(iter(test_data)) # all in one batch since batch size is 10000

# Training hyperparameters.
JAX_EPOCHS = 3
TF_EPOCHS = 7
STEPS_PER_EPOCH = len(train_labels)//BATCH_SIZE
LEARNING_RATE = 0.01
LEARNING_RATE_EXP_DECAY = 0.6

# The learning rate schedule for JAX (with Optax).
jlr_decay = optax.exponential_decay(LEARNING_RATE, transition_steps=STEPS_PER_EPOCH, decay_rate=LEARNING_RATE_EXP_DECAY, staircase=True)

# THe learning rate schedule for TensorFlow.
tflr_decay = tf.keras.optimizers.schedules.ExponentialDecay(initial_learning_rate=LEARNING_RATE, decay_steps=STEPS_PER_EPOCH, decay_rate=LEARNING_RATE_EXP_DECAY, staircase=True)

class ConvModel(flax.linen.Module):

  @flax.linen.compact
  def __call__(self, x, train):
    x = flax.linen.Conv(features=12, kernel_size=(3,3), padding="SAME", use_bias=False)(x)
    x = flax.linen.BatchNorm(use_running_average=not train, use_scale=False, use_bias=True)(x)
    x = x.reshape((x.shape[0], -1))  # flatten
    x = flax.linen.Dense(features=200, use_bias=True)(x)
    x = flax.linen.BatchNorm(use_running_average=not train, use_scale=False, use_bias=True)(x)
    x = flax.linen.Dropout(rate=0.3, deterministic=not train)(x)
    x = flax.linen.relu(x)
    x = flax.linen.Dense(features=10)(x)
    #x = flax.linen.log_softmax(x)
    return x

  # JAX differentiation requires a function `f(params, other_state, data, labels)` -> `loss` (as a single number).
  # `jax.grad` will differentiate it against the fist argument.
  # The user must split trainable and non-trainable variables into `params` and `other_state`.
  # Must pass a different RNG key each time for the dropout mask to be different.
  def loss(self, params, other_state, rng, data, labels, train):
    logits, batch_stats = self.apply({'params': params, **other_state},
                                     data,
                                     mutable=['batch_stats'],
                                     rngs={'dropout': rng},
                                     train=train)
    # The loss averaged across the batch dimension.
    loss = optax.softmax_cross_entropy(logits, labels).mean()
    return loss, batch_stats

  def predict(self, state, data):
    logits = self.apply(state, data, train=False) # predict and accuracy disable dropout and use accumulated batch norm stats (train=False)
    probabilities = flax.linen.log_softmax(logits)
    return probabilities

  def accuracy(self, state, data, labels):
    probabilities = self.predict(state, data)
    predictions = jnp.argmax(probabilities, axis=-1)
    dense_labels = jnp.argmax(labels, axis=-1)
    accuracy = jnp.equal(predictions, dense_labels).mean()
    return accuracy

# The training step.
@partial(jax.jit, static_argnums=[0]) # this forces jax.jit to recompile for every new model
def train_step(model, state, optimizer_state, rng, data, labels):

  other_state, params = state.pop('params') # differentiate only against 'params' which represents trainable variables
  (loss, batch_stats), grads = jax.value_and_grad(model.loss, has_aux=True)(params, other_state, rng, data, labels, train=True)

  updates, optimizer_state = optimizer.update(grads, optimizer_state)
  params = optax.apply_updates(params, updates)
  new_state = state.copy(add_or_replace={**batch_stats, 'params': params})

  rng, _ = jax.random.split(rng)

  return new_state, optimizer_state, rng, loss

def train(model, state, optimizer_state, train_data, epochs, losses, avg_losses, eval_losses, eval_accuracies):
  p = Progress(STEPS_PER_EPOCH)
  rng = jax.random.PRNGKey(0)
  for epoch in range(epochs):

    # This is where the learning rate schedule state is stored in the optimizer state.
    optimizer_step = optimizer_state[1].count

    # Run an epoch of training.
    for step, (data, labels) in enumerate(train_data):
      p.step(reset=(step==0))
      state, optimizer_state, rng, loss = train_step(model, state, optimizer_state, rng, data.numpy(), labels.numpy())
      losses.append(loss)
    avg_loss = np.mean(losses[-step:])
    avg_losses.append(avg_loss)

    # Run one epoch of evals (10,000 test images in a single batch).
    other_state, params = state.pop('params')
    # Gotcha: must discard modified batch_stats here
    eval_loss, _ = model.loss(params, other_state, rng, all_test_data.numpy(), all_test_labels.numpy(), train=False)
    eval_losses.append(eval_loss)
    eval_accuracy = model.accuracy(state, all_test_data.numpy(), all_test_labels.numpy())
    eval_accuracies.append(eval_accuracy)

    print("\nEpoch", epoch, "train loss:", avg_loss, "eval loss:", eval_loss, "eval accuracy", eval_accuracy, "lr:", jlr_decay(optimizer_step))

  return state, optimizer_state

# The model.
model = ConvModel()
state = model.init({'params':jax.random.PRNGKey(0), 'dropout':jax.random.PRNGKey(0)}, one_batch, train=True) # Flax allows a separate RNG for "dropout"

# The optimizer.
optimizer = optax.adam(learning_rate=jlr_decay) # Gotcha: it does not seem to be possible to pass just a callable as LR, must be an Optax Schedule
optimizer_state = optimizer.init(state['params'])

losses=[]
avg_losses=[]
eval_losses=[]
eval_accuracies=[]

new_state, new_optimizer_state = train(model, state, optimizer_state, train_data, JAX_EPOCHS+TF_EPOCHS, losses, avg_losses, eval_losses, eval_accuracies)

display_train_curves(losses, avg_losses, eval_losses, eval_accuracies, len(eval_losses), STEPS_PER_EPOCH, ignore_first_n=1*STEPS_PER_EPOCH)

model = ConvModel()
state = model.init({'params':jax.random.PRNGKey(0), 'dropout':jax.random.PRNGKey(0)}, one_batch, train=True) # Flax allows a separate RNG for "dropout"

# The optimizer.
optimizer = optax.adam(learning_rate=jlr_decay) # LR must be an Optax LR Schedule
optimizer_state = optimizer.init(state['params'])

losses, avg_losses, eval_losses, eval_accuracies = [], [], [], []

state, optimizer_state = train(model, state, optimizer_state, train_data, JAX_EPOCHS, losses, avg_losses, eval_losses, eval_accuracies)

display_train_curves(losses, avg_losses, eval_losses, eval_accuracies, len(eval_losses), STEPS_PER_EPOCH, ignore_first_n=1*STEPS_PER_EPOCH)

# Test data with a different batch size to test polymorphic shapes.
x, y = next(iter(train_data.unbatch().batch(13)))

m = tf.Module()
# Wrap the JAX state in `tf.Variable` (needed when calling the converted JAX function.
state_vars = tf.nest.map_structure(tf.Variable, state)
# Keep the wrapped state as flat list (needed in TensorFlow fine-tuning).
m.vars = tf.nest.flatten(state_vars)
# Convert the desired JAX function (`model.predict`).
predict_fn = jax2tf.convert(model.predict, polymorphic_shapes=["...", "(b, 28, 28, 1)"])
# Wrap the converted function in `tf.function` with the correct `tf.TensorSpec` (necessary for dynamic shapes to work).
@tf.function(autograph=False, input_signature=[tf.TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32)])
def predict(data):
    return predict_fn(state_vars, data)
m.predict = predict
tf.saved_model.save(m, "./")

# Test the converted function.
print("Converted function predictions:", np.argmax(m.predict(x).numpy(), axis=-1))
# Reload the model.
reloaded_model = tf.saved_model.load("./")
# Test the reloaded converted function (the result should be the same).
print("Reloaded  function predictions:", np.argmax(reloaded_model.predict(x).numpy(), axis=-1))

from collections import abc

def _fix_frozen(d):
  """Changes any mappings (e.g. frozendict) back to dict."""
  if isinstance(d, list):
    return [_fix_frozen(v) for v in d]
  elif isinstance(d, tuple):
    return tuple(_fix_frozen(v) for v in d)
  elif not isinstance(d, abc.Mapping):
    return d
  d = dict(d)
  for k, v in d.items():
    d[k] = _fix_frozen(v)
  return d

class TFModel(tf.Module):
  def __init__(self, state, model):
    super().__init__()

    # Special care needed for the train=True/False parameter in the loss
    @jax.jit
    def loss_with_train_bool(state, rng, data, labels, train):
      other_state, params = state.pop('params')
      loss, batch_stats = jax.lax.cond(train,
                                       lambda state, data, labels: model.loss(params, other_state, rng, data, labels, train=True),
                                       lambda state, data, labels: model.loss(params, other_state, rng, data, labels, train=False),
                                       state, data, labels)
      # must use JAX to split the RNG, therefore, must do it in a @jax.jit function
      new_rng, _ = jax.random.split(rng)
      return loss, batch_stats, new_rng

    self.state_vars = tf.nest.map_structure(tf.Variable, state)
    self.vars = tf.nest.flatten(self.state_vars)
    self.jax_rng = tf.Variable(jax.random.PRNGKey(0))

    self.loss_fn = jax2tf.convert(loss_with_train_bool, polymorphic_shapes=["...", "...", "(b, 28, 28, 1)", "(b, 10)", "..."])
    self.accuracy_fn = jax2tf.convert(model.accuracy, polymorphic_shapes=["...", "(b, 28, 28, 1)", "(b, 10)"])
    self.predict_fn = jax2tf.convert(model.predict, polymorphic_shapes=["...", "(b, 28, 28, 1)"])

  # Must specify TensorSpec manually for variable batch size to work
  @tf.function(autograph=False, input_signature=[tf.TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32)])
  def predict(self, data):
    # Make sure the TfModel.predict function implicitly use self.state_vars and not the JAX state directly
    # otherwise, all model weights would be embedded in the TF graph as constants.
    return self.predict_fn(self.state_vars, data)

  @tf.function(input_signature=[tf.TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32),
                                tf.TensorSpec(shape=(None, 10), dtype=tf.float32)],
               autograph=False)
  def train_loss(self, data, labels):
      loss, batch_stats, new_rng = self.loss_fn(self.state_vars, self.jax_rng, data, labels, True)
      # update batch norm stats
      flat_vars = tf.nest.flatten(self.state_vars['batch_stats'])
      flat_values = tf.nest.flatten(batch_stats['batch_stats'])
      for var, val in zip(flat_vars, flat_values):
        var.assign(val)
      # update RNG
      self.jax_rng.assign(new_rng)
      return loss

  @tf.function(input_signature=[tf.TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32),
                                tf.TensorSpec(shape=(None, 10), dtype=tf.float32)],
               autograph=False)
  def eval_loss(self, data, labels):
      loss, batch_stats, new_rng = self.loss_fn(self.state_vars, self.jax_rng, data, labels, False)
      return loss

  @tf.function(input_signature=[tf.TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32),
                                tf.TensorSpec(shape=(None, 10), dtype=tf.float32)],
               autograph=False)
  def accuracy(self, data, labels):
    return self.accuracy_fn(self.state_vars, data, labels)

# Instantiate the model.
tf_model = TFModel(state, model)

# Save the model.
tf.saved_model.save(tf_model, "./")

reloaded_model = tf.saved_model.load("./")

# Test if it works and that the batch size is indeed variable.
x,y = next(iter(train_data.unbatch().batch(13)))
print(np.argmax(reloaded_model.predict(x).numpy(), axis=-1))
x,y = next(iter(train_data.unbatch().batch(20)))
print(np.argmax(reloaded_model.predict(x).numpy(), axis=-1))

print(reloaded_model.accuracy(one_batch, one_batch_labels))
print(reloaded_model.accuracy(all_test_data, all_test_labels))

optimizer = tf.keras.optimizers.Adam(learning_rate=tflr_decay)

# Set the iteration step for the learning rate to resume from where it left off in JAX.
optimizer.iterations.assign(len(eval_losses)*STEPS_PER_EPOCH)

p = Progress(STEPS_PER_EPOCH)

for epoch in range(JAX_EPOCHS, JAX_EPOCHS+TF_EPOCHS):

  # This is where the learning rate schedule state is stored in the optimizer state.
  optimizer_step = optimizer.iterations

  for step, (data, labels) in enumerate(train_data):
    p.step(reset=(step==0))
    with tf.GradientTape() as tape:
      #loss = reloaded_model.loss(data, labels, True)
      loss = reloaded_model.train_loss(data, labels)
      grads = tape.gradient(loss, reloaded_model.vars)
      optimizer.apply_gradients(zip(grads, reloaded_model.vars))
      losses.append(loss)
  avg_loss = np.mean(losses[-step:])
  avg_losses.append(avg_loss)

  eval_loss = reloaded_model.eval_loss(all_test_data.numpy(), all_test_labels.numpy()).numpy()
  eval_losses.append(eval_loss)
  eval_accuracy = reloaded_model.accuracy(all_test_data.numpy(), all_test_labels.numpy()).numpy()
  eval_accuracies.append(eval_accuracy)

  print("\nEpoch", epoch, "train loss:", avg_loss, "eval loss:", eval_loss, "eval accuracy", eval_accuracy, "lr:", tflr_decay(optimizer.iterations).numpy())

display_train_curves(losses, avg_losses, eval_losses, eval_accuracies, len(eval_losses), STEPS_PER_EPOCH, ignore_first_n=2*STEPS_PER_EPOCH)

# The loss takes a hit when the training restarts, but does not go back to random levels.
# This is likely caused by the optimizer momentum being reinitialized.