tensorflow
TypeError: __init__() missing 1 required positional argument: 'distribution'
I tried to use Ray Tune with with tfp.NoUTurn Sampler but I got this error TypeError: __init__() missing 1 required positional argument: 'distribution'. I tried it with HMC and VI and got the same error. It seems tfp can't find Ray tune's values for the specified hyperparameter.
params = {'num_burnin_steps': tune.randint(30, 100)}
def train model(config):
#preceding code here...
num_burnin_steps =config["num_burnin_steps"]
sampler = tfp.mcmc.TransformedTransitionKernel(
tfp.mcmc.NoUTurnSampler(
target_log_prob_fn=pinned_model.unnormalized_log_prob,
step_size=0.1),
bijector=constraining_bijector)
remaining code here...
return {'score': score}
import ray tune and define your parameters
trainable_with_resources = tune.with_resources(train_model, {"cpu": 1})
tuner = tune.Tuner(
trainable_with_resources,
param_space=trial_space,
tune_config=tune.TuneConfig(num_samples=10)
)
results = tuner.fit()`
Tensorflow probability version: 0.19.0 Ray Tune version: 2.3.1
What is in constraining_bijector? Consider using tfp.experimental.mcmc.windowed_adaptive_nuts(..) instead. It's not clear how to further debug this without a stack trace or more code.
Brian Patton | Software Engineer | @.***
On Tue, Apr 11, 2023 at 4:41 AM bumie @.***> wrote:
I tried to use Ray Tune with with tfp.NoUTurn Sampler but I got this error TypeError: init() missing 1 required positional argument: 'distribution'. I tried it with HMC and VI and got the same error. It seems tfp can't find Ray tune's values for the specified hyperparameter.
params = {'num_burnin_steps': tune.randint(30, 100)}
def train model(config): preceding code here...
num_burnin_steps =config["num_burnin_steps"] sampler = tfp.mcmc.TransformedTransitionKernel( tfp.mcmc.NoUTurnSampler( target_log_prob_fn=pinned_model.unnormalized_log_prob, step_size=0.1), bijector=constraining_bijector)
remaining code here...
return {'score': score} import ray tune
trainable_with_resources = tune.with_resources(train_model, {"cpu": 1}) tuner = tune.Tuner( trainable_with_resources, param_space=trial_space, tune_config=tune.TuneConfig(num_samples=10) ) results = tuner.fit()
— Reply to this email directly, view it on GitHub https://github.com/tensorflow/probability/issues/1711, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFJFSIZAFJ4JHEBFSPXTVMDXAUKKLANCNFSM6AAAAAAWZ6TBSA . You are receiving this because you are subscribed to this thread.Message ID: @.***>
Hi,
Thank you for the quick response.
Here's the full code
def build_model(observed_time_series, data):
smooth_seasonal = sts.SmoothSeasonal(
period=365.25, # Annual period
observed_time_series= observed_time_series,
frequency_multipliers=[1, 2, 3],
name='smooth_seasonal'
)
autoregressive = tfp.sts.Autoregressive(
order=7,
observed_time_series=observed_time_series,
coefficients_prior=None,
level_scale_prior=None,
initial_state_prior=None,
name='autoregressive')
model_components = [
smooth_seasonal,
autoregressive,
]
model = tfp.sts.Sum(model_components,
observed_time_series=observed_time_series)
return model
sales_model = build_model(sales_y_data, sales_data)
positive_bijector = tfb.Softplus()
approximate_unconstrained_rates = positive_bijector.inverse(
tf.convert_to_tensor(sales_y_data) + 0.01)
def sts_with_poisson_likelihood_model():
param_vals = []
for param in sales_model.parameters:
param_val = yield param.prior
param_vals.append(param_val)
unconstrained_rate = yield sales_model.make_state_space_model(
14, param_vals)
rate = positive_bijector.forward(unconstrained_rate[..., 0])
observed_counts = yield tfd.Poisson(rate, name='observed_counts')
model = tfd.JointDistributionCoroutineAutoBatched(sts_with_poisson_likelihood_model)
def train_model(config):
start_time = time.time()
num_burnin_steps =config["num_burnin_steps"]
sampler = tfp.mcmc.TransformedTransitionKernel(
tfp.mcmc.NoUTurnSampler(
target_log_prob_fn=pinned_model.unnormalized_log_prob,
step_size=0.1),
bijector=constraining_bijector)
adaptive_sampler = tfp.mcmc.DualAveragingStepSizeAdaptation(
inner_kernel=sampler,
num_adaptation_steps=int(0.8 * num_burnin_steps),
target_accept_prob=0.75)
initial_state = constraining_bijector.forward(
type(pinned_model.event_shape)(
*(tf.random.normal(part_shape)
for part_shape in constraining_bijector.inverse_event_shape(
pinned_model.event_shape))))
num_results = 14
@tf.function(autograph=False, jit_compile=True)
def do_sampling():
return tfp.mcmc.sample_chain(
kernel=adaptive_sampler,
current_state=initial_state,
num_results=num_results,
num_burnin_steps=num_burnin_steps,
trace_fn=None)
t0 = time.time()
samples = do_sampling()
t1 = time.time()
return {"time": t1-t0}
# Define trial parameters
trial_space = {
"num_burnin_steps": tune.randint(30, 100),
}
trainable_with_resources = tune.with_resources(train_model, {"cpu": 1})
tuner = tune.Tuner(
trainable_with_resources,
param_space=trial_space,
tune_config=tune.TuneConfig(num_samples=10)
)
results = tuner.fit()
Looks OK to me (apart from the raytune part, which I dropped). Suggests the problem is actually in tune. constraining_bijector was still undefined, so I defined it. https://colab.research.google.com/gist/brianwa84/3c0c6859b07607416380a1e83be5e430/untitled47.ipynb
Brian Patton | Software Engineer | @.***
On Wed, Apr 12, 2023 at 9:11 AM bumie @.***> wrote:
Hi,
Thank you for the quick response.
Here's the full code
def build_model(observed_time_series, data):
smooth_seasonal = sts.SmoothSeasonal( period=365.25, # Annual period observed_time_series= observed_time_series, frequency_multipliers=[1, 2, 3], name='smooth_seasonal' ) autoregressive = tfp.sts.Autoregressive( order=7, observed_time_series=observed_time_series, coefficients_prior=None, level_scale_prior=None, initial_state_prior=None, name='autoregressive') model_components = [ smooth_seasonal, autoregressive, ] model = tfp.sts.Sum(model_components, observed_time_series=observed_time_series) return modelsales_model = build_model(sales_y_data, sales_data)
positive_bijector = tfb.Softplus() approximate_unconstrained_rates = positive_bijector.inverse( tf.convert_to_tensor(sales_y_data) + 0.01)
def sts_with_poisson_likelihood_model():
param_vals = [] for param in sales_model.parameters: param_val = yield param.prior param_vals.append(param_val) unconstrained_rate = yield sales_model.make_state_space_model( 14, param_vals) rate = positive_bijector.forward(unconstrained_rate[..., 0]) observed_counts = yield tfd.Poisson(rate, name='observed_counts')model = tfd.JointDistributionCoroutineAutoBatched(sts_with_poisson_likelihood_model)
def train_model(config):
start_time = time.time() num_burnin_steps =config["num_burnin_steps"] sampler = tfp.mcmc.TransformedTransitionKernel( tfp.mcmc.NoUTurnSampler( target_log_prob_fn=pinned_model.unnormalized_log_prob, step_size=0.1), bijector=constraining_bijector) adaptive_sampler = tfp.mcmc.DualAveragingStepSizeAdaptation( inner_kernel=sampler, num_adaptation_steps=int(0.8 * num_burnin_steps), target_accept_prob=0.75) initial_state = constraining_bijector.forward( type(pinned_model.event_shape)( *(tf.random.normal(part_shape) for part_shape in constraining_bijector.inverse_event_shape( pinned_model.event_shape)))) num_results = 14 @tf.function(autograph=False, jit_compile=True) def do_sampling(): return tfp.mcmc.sample_chain( kernel=adaptive_sampler, current_state=initial_state, num_results=num_results, num_burnin_steps=num_burnin_steps, trace_fn=None) t0 = time.time() samples = do_sampling() t1 = time.time() return {"time": t1-t0}Define trial parameters
trial_space = { "num_burnin_steps": tune.randint(30, 100), } trainable_with_resources = tune.with_resources(train_model, {"cpu": 1}) tuner = tune.Tuner( trainable_with_resources, param_space=trial_space, tune_config=tune.TuneConfig(num_samples=10) ) results = tuner.fit()
— Reply to this email directly, view it on GitHub https://github.com/tensorflow/probability/issues/1711#issuecomment-1505253989, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFJFSIZ3M7IENBCQ73U37DTXA2SYFANCNFSM6AAAAAAWZ6TBSA . You are receiving this because you commented.Message ID: @.***>
Hi,
It seems the error happens when I try to use distributed training with tfp. I replaced RayTune with Optuna, ran the model on a Dask Cluster with 2 GPUs and 2 CPUs, and got the same error. I tried to investigate the issue with distributed training by following an example implementation of Sharded Distribution with Jax but got stuck trying to share the training data across devices. How do I resolve this?
Here's the link to the Jax Implementation https://www.tensorflow.org/probability/examples/TensorFlow_Probability_on_JAX
Yes. Stack trace with RayTune and Optuna as the optimizer algorithm
I updated the notebook you sent me with the code.
https://colab.research.google.com/gist/bumie-e/4af2d03e9c446d2de374ebb03ccd41dd/untitled47.ipynb
/usr/local/lib/python3.9/dist-packages/ray/tune/search/optuna/optuna_search.py:685: FutureWarning: IntUniformDistribution has been deprecated in v3.0.0. This feature will be removed in v6.0.0. See https://github.com/optuna/optuna/releases/tag/v3.0.0. Use :class:`~optuna.distributions.IntDistribution` instead.
return ot.distributions.IntUniformDistribution(
[I 2023-04-20 15:35:06,736] A new study created in memory with name: optuna
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
[<ipython-input-6-7f2989bce142>](https://localhost:8080/#) in <cell line: 15>()
13 )
14 )
---> 15 results = tuner.fit()
14 frames
[/usr/local/lib/python3.9/dist-packages/ray/tune/tuner.py](https://localhost:8080/#) in fit(self)
290 if not self._is_ray_client:
291 try:
--> 292 return self._local_tuner.fit()
293 except TuneError as e:
294 raise TuneError(
[/usr/local/lib/python3.9/dist-packages/ray/tune/impl/tuner_internal.py](https://localhost:8080/#) in fit(self)
453 if not self._is_restored:
454 param_space = copy.deepcopy(self._param_space)
--> 455 analysis = self._fit_internal(trainable, param_space)
456 else:
457 analysis = self._fit_resume(trainable)
[/usr/local/lib/python3.9/dist-packages/ray/tune/impl/tuner_internal.py](https://localhost:8080/#) in _fit_internal(self, trainable, param_space)
570 **self._tuner_kwargs,
571 }
--> 572 analysis = run(
573 **args,
574 )
[/usr/local/lib/python3.9/dist-packages/ray/tune/tune.py](https://localhost:8080/#) in run(run_or_experiment, name, metric, mode, stop, time_budget_s, config, resources_per_trial, num_samples, local_dir, search_alg, scheduler, keep_checkpoints_num, checkpoint_score_attr, checkpoint_freq, checkpoint_at_end, verbose, progress_reporter, log_to_file, trial_name_creator, trial_dirname_creator, chdir_to_trial_dir, sync_config, export_formats, max_failures, fail_fast, restore, server_port, resume, reuse_actors, raise_on_failed_trial, callbacks, max_concurrent_trials, trial_executor, _experiment_checkpoint_dir, _remote, _remote_string_queue)
754 )
755 while not runner.is_finished() and not experiment_interrupted_event.is_set():
--> 756 runner.step()
757 if has_verbosity(Verbosity.V1_EXPERIMENT):
758 _report_progress(runner, progress_reporter)
[/usr/local/lib/python3.9/dist-packages/ray/tune/execution/trial_runner.py](https://localhost:8080/#) in step(self)
951 )
952
--> 953 next_trial = self._update_trial_queue_and_get_next_trial()
954 if next_trial:
955 logger.debug(f"Got new trial to run: {next_trial}")
[/usr/local/lib/python3.9/dist-packages/ray/tune/execution/trial_runner.py](https://localhost:8080/#) in _update_trial_queue_and_get_next_trial(self)
887 # Create pending trials until it fails.
888 while num_pending_trials < self._max_pending_trials:
--> 889 if not self._update_trial_queue(blocking=wait_for_trial):
890 break
891 wait_for_trial = False # wait at most one trial
[/usr/local/lib/python3.9/dist-packages/ray/tune/execution/trial_runner.py](https://localhost:8080/#) in _update_trial_queue(self, blocking, timeout)
1473 Boolean indicating if a new trial was created or not.
1474 """
-> 1475 trial = self._search_alg.next_trial()
1476 if blocking and not trial:
1477 start = time.time()
[/usr/local/lib/python3.9/dist-packages/ray/tune/search/search_generator.py](https://localhost:8080/#) in next_trial(self)
98 """
99 if not self.is_finished():
--> 100 return self.create_trial_if_possible(
101 self._experiment.spec, self._experiment.dir_name
102 )
[/usr/local/lib/python3.9/dist-packages/ray/tune/search/search_generator.py](https://localhost:8080/#) in create_trial_if_possible(self, experiment_spec, output_path)
123 self._counter += 1
124 tag = "{0}_{1}".format(str(self._counter), format_vars(flattened_config))
--> 125 trial = _create_trial_from_spec(
126 spec,
127 output_path,
[/usr/local/lib/python3.9/dist-packages/ray/tune/experiment/config_parser.py](https://localhost:8080/#) in _create_trial_from_spec(spec, output_path, parser, **trial_kwargs)
227 checkpoint_config = spec.get("checkpoint_config", CheckpointConfig())
228
--> 229 return Trial(
230 # Submitting trial via server in py2.7 creates Unicode, which does not
231 # convert to string in a straightforward manner.
[/usr/local/lib/python3.9/dist-packages/ray/tune/experiment/trial.py](https://localhost:8080/#) in __init__(self, trainable_name, config, trial_id, local_dir, evaluated_params, experiment_tag, resources, placement_group_factory, stopping_criterion, experiment_dir_name, sync_config, checkpoint_config, export_formats, restore_path, trial_name_creator, trial_dirname_creator, log_to_file, max_failures, stub, _setup_default_resource)
300 self.experiment_tag = experiment_tag
301 self.location = _Location()
--> 302 trainable_cls = self.get_trainable_cls()
303 if trainable_cls and _setup_default_resource:
304 default_resources = trainable_cls.default_resource_request(self.config)
[/usr/local/lib/python3.9/dist-packages/ray/tune/experiment/trial.py](https://localhost:8080/#) in get_trainable_cls(self)
868 if self.stub:
869 return None
--> 870 return get_trainable_cls(self.trainable_name)
871
872 def get_trial_checkpoints(self) -> List[_TrackedCheckpoint]:
[/usr/local/lib/python3.9/dist-packages/ray/tune/registry.py](https://localhost:8080/#) in get_trainable_cls(trainable_name)
44 def get_trainable_cls(trainable_name):
45 validate_trainable(trainable_name)
---> 46 return _global_registry.get(TRAINABLE_CLASS, trainable_name)
47
48
[/usr/local/lib/python3.9/dist-packages/ray/tune/registry.py](https://localhost:8080/#) in get(self, category, key)
213 "Registry value for {}/{} doesn't exist.".format(category, key)
214 )
--> 215 return pickle.loads(value)
216 else:
217 return pickle.loads(self._to_flush[(category, key)])
[/usr/local/lib/python3.9/dist-packages/tensorflow_probability/python/distributions/sample.py](https://localhost:8080/#) in __new__(cls, *args, **kwargs)
368
369 if not isinstance(distribution, tf.__internal__.CompositeTensor):
--> 370 return _Sample(*args, **kwargs)
371 return super(Sample, cls).__new__(cls)
372
TypeError: __init__() missing 1 required positional argument: 'distribution'