def _jit_compute_loss_fn(predict_fn, loss_fn, n_devices, jit=True):
"""Returns a (JIT-compiled) function that computes the loss for one step."""
if n_devices == 1: # TODO(lukaszkaiser): remove branch when not needed.
def single_compute_loss(opt_state, batch, state, rng):
rng, subrng = jax_random.split(rng[0])
loss_val, state = loss_fn(opt_state[0], batch, predict_fn, state,
rng)
return loss_val, state, [subrng]
return backend.jit(single_compute_loss) if jit else single_compute_loss
# Else, for n_devices > 1:
@functools.partial(backend.pmap, axis_name='batch')
def mapped_compute_loss(opt_state, batch, state, rng):
"""This is a multi-device version of the update function above."""
# We assume all tensors have the first dimension = n_devices.
rng, subrng = jax_random.split(rng)
loss_val, state = loss_fn(opt_state[0], batch, predict_fn, state, rng)
return loss_val, state, subrng
def compute_loss(opt_state, batch, state, rng):
return mapped_compute_loss(opt_state,
_reshape_by_device(batch, n_devices), state,
rng)
return compute_loss
def __init__(self,
model,
batch_size,
observation_space,
action_space,
reward_range,
discrete_rewards,
history_stream,
output_dir,
model_predict_kwargs=None):
"""Initializes the env.
Args:
model: Trax model.
batch_size: (int) Number of simulated environments run in parallel.
observation_space: (gym.Space) Observation space.
action_space: (gym.Space) Action space.
reward_range: (tuple) Pair (min_reward, max_reward).
discrete_rewards: (bool) Whether to discretize the rewards.
history_stream: Iterator yielding batches of initial input data for the
model. The format is implementation-specific.
output_dir: (str) Output dir.
model_predict_kwargs: (dict) Additional model keyword arguments for
inference. Useful when different config is needed for training and
inference, e.g. train with memory efficient attention and predict with
the regular one.
"""
self._model = model
if model_predict_kwargs is None:
model_predict_kwargs = {}
model_predict = self._model(mode='predict', **model_predict_kwargs)
# NOTE: can set non-default PRNG key by: model_predict._set_rng(...)
def predict_with_state(*args, **kwargs):
output = model_predict(*args, **kwargs)
return (output, model_predict.state)
self._model_predict = backend.jit(predict_with_state)
self._model_initialize = model_predict.initialize_once
self._observation_space = observation_space
self._action_space = action_space
self._reward_range = reward_range
self._output_dir = output_dir
self._predict_fn = None
self._rng = None
self._model_state = None
self._history_stream = None
# Call the super's ctor. It will use some of the member fields, so we call
# it in the end.
super(SimulatedEnvProblem, self).__init__(
batch_size=batch_size,
discrete_rewards=discrete_rewards,
history_stream=history_stream,
)
self.seed()
def _jit_update_fn(predict_fn, loss_fn, optimizer, n_devices, jit=True):
"""Returns a (JIT-compiled) function that computes updates for one step."""
model_and_loss = tl.Serial(predict_fn, loss_fn)
# Gradients are always wrt. the first argument, so putting weights first.
def model_and_loss_call(weights, batch, state, rng):
res = model_and_loss(batch, weights=weights, state=state, rng=rng)
return res, model_and_loss.state
if n_devices == 1: # TODO(lukaszkaiser): remove branch when not needed.
def single_update(i, opt_state, batch, state, rng):
weights, slots, opt_params = opt_state
rng, subrng = jax_random.split(rng[0])
grad_fn = backend.grad(model_and_loss_call, has_aux=True)
grads, state = grad_fn(weights, batch, state, rng)
return optimizer.tree_update(i, grads, weights, slots,
opt_params), state, [subrng]
return backend.jit(single_update) if jit else single_update
# Else, for n_devices > 1:
@functools.partial(backend.pmap, axis_name='batch')
def mapped_update(i, opt_state, batch, state, rng):
"""This is a multi-device version of the update function above."""
# We assume all tensors have the first dimension = n_devices.
weights, slots, opt_params = opt_state
rng, subrng = jax_random.split(rng)
grad_fn = backend.grad(model_and_loss_call, has_aux=True)
grads, state = grad_fn(weights, batch, state, rng)
# We do a psum(1.0) here instead of `n_devices` since `n_devices` is just
# the number of devices on this host machine, however psum goes over all
# devices of all hosts (ex: a TPU pod) and we need to be averaging over all
# of them.
grads = jax.tree_util.tree_map(
lambda g: backend.psum(g, 'batch') / backend.psum(1.0, 'batch'),
grads)
return optimizer.tree_update(i, grads, weights, slots,
opt_params), state, subrng
def update(i, opt_state, batch, state, rng):
return mapped_update(np.repeat(i, n_devices), opt_state, batch, state,
rng)
return update
def _jit_update_fn(predict_fn, loss_fn, optimizer, n_devices, jit=True):
"""Returns a (JIT-compiled) function that computes updates for one step."""
model_and_loss = layers.Serial(predict_fn, loss_fn)
# Gradients are always wrt. the first argument, so putting params first.
def model_and_loss_call(params, batch, state, rng):
res = model_and_loss(batch, params=params, state=state, rng=rng)
return res, model_and_loss.state
if n_devices == 1: # TODO(lukaszkaiser): remove branch when not needed.
def single_update(i, opt_state, batch, state, rng):
params, slots, opt_params = opt_state
rng, subrng = jax_random.split(rng[0])
grad_fn = backend.grad(model_and_loss_call, has_aux=True)
grads, state = grad_fn(params, batch, state, rng)
return optimizer.tree_update(i, grads, params, slots,
opt_params), state, [subrng]
return backend.jit(single_update) if jit else single_update
# Else, for n_devices > 1:
@functools.partial(backend.pmap, axis_name='batch')
def mapped_update(i, opt_state, batch, state, rng):
"""This is a multi-device version of the update function above."""
# We assume all tensors have the first dimension = n_devices.
params, slots, opt_params = opt_state
rng, subrng = jax_random.split(rng)
grad_fn = backend.grad(model_and_loss_call, has_aux=True)
grads, state = grad_fn(params, batch, state, rng)
grads = jax.tree_util.tree_map(lambda g: backend.psum(g, 'batch'),
grads)
return optimizer.tree_update(i, grads, params, slots,
opt_params), state, subrng
def update(i, opt_state, batch, state, rng):
return mapped_update(np.repeat(i, n_devices), opt_state, batch, state,
rng)
return update
def __init__(self,
model,
loss_fn,
optimizer,
lr_schedule,
inputs,
output_dir=None,
random_seed=None,
n_devices=None,
save_steps=None,
should_save_checkpoints=True,
should_write_summaries=True,
has_weights=False,
nontrainable_param_map=None,
mask_id=None,
metrics=None):
if backend.get_name() == 'jax':
self._host_id = jax.host_id()
self._host_count = jax.host_count()
else:
self._host_id = 0
self._host_count = 1
self._is_chief = (self._host_id == 0)
if save_steps is None:
save_steps = []
self._save_steps = save_steps
self._should_save_checkpoints = should_save_checkpoints
self._should_write_summaries = should_write_summaries
self._has_weights = has_weights
self._mask_id = mask_id
self._metrics_dict = _METRICS if metrics is None else metrics
loss_fn = loss_fn(has_weights=has_weights, mask_id=mask_id)
device_count = backend.device_count()
n_devices = n_devices or device_count
# TODO(lukaszkaiser): remove this restriction when possible.
if n_devices != device_count and backend.get_name() == 'jax':
raise ValueError(
'JAX cannot work yet with n_devices != all devices: '
'%d != %d' % (n_devices, device_count))
self._n_devices = n_devices
# Simple differential seeding of RNG across hosts by host_id and time.
if random_seed is None and self._host_count > 1:
_, random_seed = divmod(
int(time.time() * 1e6) + int(self._host_id * 1e6), 2**32)
rng = get_random_number_generator_and_set_seed(random_seed)
inputs = inputs(n_devices)
self._inputs = inputs
# Initialize the learning rate to a dummy value. It will be set in reset().
opt = optimizer(learning_rate=0.0)
# Setup the model.
model_train = model(mode='train')
model_predict_eval = model(mode='eval')
# Setup state.
rng, init_rng = jax_random.split(rng)
self._rngs = np.stack(jax_random.split(rng, n_devices))
first_shape = inputs.input_shape[0]
# If the inputs are a tuple/list, add [None] (batch) to each element.
if isinstance(first_shape, (list, tuple)):
model_input_shape = tuple(
tuple([None] + list(shape)) for shape in inputs.input_shape)
model_target_shape = tuple(
tuple([None] + list(shape)) for shape in inputs.target_shape)
else: # Otherwise just add [None] to the input shape.
model_input_shape = tuple([None] + list(inputs.input_shape))
model_target_shape = tuple([None] + list(inputs.target_shape))
# Change all None to 1 in input and target shape.
model_input_shape = backend.nested_map(lambda x: x or 1,
model_input_shape)
model_target_shape = backend.nested_map(lambda x: x or 1,
model_target_shape)
def new_opt_state_and_model_state(input_shape, input_dtype,
target_shape, target_dtype, rng):
"""Returns optimizer and model states suitable for training a model."""
# Combine inputs and targets on the stack.
if not isinstance(input_dtype, (list, tuple)):
input_dtype = [input_dtype]
input_shape = [input_shape]
if not isinstance(target_dtype, (list, tuple)):
target_dtype = [target_dtype]
target_shape = [target_shape]
dtypes = list(input_dtype) + list(target_dtype)
shapes = list(input_shape) + list(target_shape)
if self._has_weights:
shapes += list(target_shape)
dtypes += [np.float32 for _ in target_dtype]
input_signature = tuple(
ShapeDtype(s, d) for (s, d) in zip(shapes, dtypes))
# We need to create a new model instance and not reuse `model_train` here,
# because `m.initialize` puts cached parameter values in `m` and hence the
# next call of `m.initialize` will give wrong results.
m = tl.Serial(model(mode='train'), loss_fn)
m._set_rng_recursive(rng) # pylint: disable=protected-access
weights, state = m.init(input_signature)
(slots, opt_params) = opt.tree_init(weights)
return (OptState(weights, slots, opt_params), state)
if _is_jit_init():
# JIT parameter initialization to avoid memory fragmentation
new_opt_state_and_model_state = backend.jit(
new_opt_state_and_model_state, static_argnums=(0, 1, 2, 3))
self._new_opt_state_and_model_state = (
lambda: new_opt_state_and_model_state( # pylint: disable=g-long-lambda
model_input_shape, self._inputs.input_dtype,
model_target_shape, self._inputs.target_dtype, init_rng))
# Arrange and initialize metrics layers.
self._metrics = list(sorted(self._metrics_dict.keys()))
metrics_layers = [
self._metrics_dict[m](has_weights=self._has_weights,
mask_id=self._mask_id) for m in self._metrics
]
metrics_in_parallel = tl.Branch(*metrics_layers)
# TODO(lukaszkaiser): clean this up once layer API stabilizes.
# For now, we need to initialize metric layers somehow, so here we go.
# We assume that they do not have any parameters, so this is a dummy.
dummy_shapes = ((1, 2), (1, ),
(1, )) if self._has_weights else ((1, 2), (1, ))
dummy_signature = tuple(ShapeDtype(s) for s in dummy_shapes)
metrics_in_parallel._set_rng_recursive(init_rng) # pylint: disable=protected-access
m_weights, m_state = metrics_in_parallel.init(dummy_signature)
self._metrics_weights = self._for_n_devices(m_weights)
self._metrics_state = self._for_n_devices(m_state)
# Jit model_predict and update so they're fast.
self._jit_eval = _jit_predict_fn(model_predict_eval,
metrics_in_parallel, n_devices)
self._jit_update_fn = _jit_update_fn(model_train, loss_fn, opt,
n_devices)
self._model_train = model_train
self._model_predict_eval = model_predict_eval
self._loss_fn = loss_fn
# TODO(pkozakowski): "Learning rate schedules" are currently able to control
# control all optimizer parameters and model state, so let's rename them
# accordingly.
self._lr_schedule = lr_schedule
if nontrainable_param_map is None:
nontrainable_param_map = {}
self._nontrainable_param_map = nontrainable_param_map
# Those fields will be set in reset().
self._output_dir = None
self._train_sw = None
self._eval_sw = None
self._history = None
self._lr_fn = None
self._opt_state = None
self._step = None
self._model_state = None
if output_dir is not None:
self.reset(output_dir)
