[automl] Memory Aware Config Tuning (#1257)

ludwig/automl/auto_tune_config.py

@@ -0,0 +1,164 @@
+import copy
+from collections import OrderedDict
+
+import psutil
+import ray
+
+try:
+    import GPUtil
+except ImportError:
+    raise ImportError(
+        ' ray is not installed. '
+        'In order to use auto_train please run '
+        'pip install ludwig[ray]'
+    )
+
+from ludwig.api import LudwigModel
+from ludwig.automl.utils import get_available_resources
+from ludwig.data.preprocessing import preprocess_for_training
+from ludwig.features.feature_registries import update_config_with_metadata
+from ludwig.utils.defaults import merge_with_defaults
+from ludwig.constants import COMBINER, HYPEROPT, BATCH_SIZE, TRAINING, TYPE, PREPROCESSING, SPACE
+
+# maps variable search space that can be modified to minimum permissible value for the range
+RANKED_MODIFIABLE_PARAM_LIST = {
+    'tabnet': OrderedDict({
+        'training.batch_size': 32,
+        'combiner.size': 8,
+        'combiner.output_size': 8,
+    }),
+    'concat': OrderedDict({
+        'training.batch_size': 32,
+        'combiner.fc_size': 64,
+        'combiner.num_fc_layers': 1,
+
+    }),
+    'tabtransformer': OrderedDict({
+        'training.batch_size': 32,
+        'combiner.num_heads:': 4,
+        'combiner.output_size': 8,
+        'combiner.num_layers': 4,
+        'combiner.num_fc_layers': 1,
+    }),
+}
+
+
+BYTES_PER_MiB = 1048576
+
+
+def get_trainingset_metadata(config, dataset):
+    (_, _, _, training_set_metadata) = preprocess_for_training(
+        config,
+        dataset=dataset,
+        preprocessing_params=config[PREPROCESSING])
+    return training_set_metadata
+
+
+def get_machine_memory():
+
+    if ray.is_initialized():  # using ray cluster
+        @ray.remote(num_gpus=1)
+        def get_remote_gpu():
+            gpus = GPUtil.getGPUs()
+            total_mem_mb = gpus[0].memory_total
+            return total_mem_mb * BYTES_PER_MiB
+
+        @ray.remote(num_cpus=1)
+        def get_remote_cpu():
+            total_mem = psutil.virtual_memory().total
+            return total_mem
+
+        resources = get_available_resources()  # check if cluster has GPUS
+
+        if resources['gpu'] > 0:
+            machine_mem = ray.get(get_remote_gpu.remote())
+        else:
+            machine_mem = ray.get(get_remote_cpu.remote())
+    else:  # not using ray cluster
+        if GPUtil.getGPUs():
+            machine_mem = GPUtil.getGPUs()[0].memory_total * BYTES_PER_MiB
+        else:
+            machine_mem = psutil.virtual_memory().total
+
+    return machine_mem
+
+
+def compute_memory_usage(config, training_set_metadata) -> int:
+    update_config_with_metadata(config, training_set_metadata)
+    lm = LudwigModel.create_model(config)
+    lm.get_connected_model()
+    model_tensors = lm.collect_weights()
+    total_size = 0
+    batch_size = config[TRAINING][BATCH_SIZE]
+    for tnsr in model_tensors:
+        total_size += tnsr[1].numpy().size * batch_size
+    total_bytes = total_size * 32  # assumes 32-bit precision
+    return total_bytes
+
+
+def sub_new_params(config: dict, new_param_vals: dict):
+    new_config = copy.deepcopy(config)
+    for param, val in new_param_vals.items():
+        config_section = param.split(".")[0]
+        param_name = param.split(".")[1]
+        new_config[config_section][param_name] = val
+    return new_config
+
+
+def get_new_params(current_param_values, hyperparam_search_space, params_to_modify):
+    for param, _ in params_to_modify.items():
+        if hyperparam_search_space[param][SPACE] == "choice":
+            current_param_values[param] = hyperparam_search_space[param]['categories'][-1]
+        else:
+            current_param_values[param] = hyperparam_search_space[param]['upper']
+    return current_param_values
+
+
+def memory_tune_config(config, dataset):
+    fits_in_memory = False
+    raw_config = merge_with_defaults(config)
+    training_set_metadata = get_trainingset_metadata(raw_config, dataset)
+    modified_hyperparam_search_space = copy.deepcopy(
+        raw_config[HYPEROPT]['parameters'])
+    params_to_modify = RANKED_MODIFIABLE_PARAM_LIST[raw_config[COMBINER][TYPE]]
+    param_list = list(params_to_modify.keys())
+    current_param_values = {}
+    max_memory = get_machine_memory()
+
+    while param_list is not None:
+        # compute memory utilization
+        current_param_values = get_new_params(
+            current_param_values, modified_hyperparam_search_space, params_to_modify)
+        temp_config = sub_new_params(raw_config, current_param_values)
+        if compute_memory_usage(temp_config, training_set_metadata) < max_memory:
+            fits_in_memory = True
+            break
+        # check if we have exhausted tuning of current param (e.g. we can no longer reduce the param value)
+        param, min_value = param_list[0],  params_to_modify[param_list[0]]
+
+        if param in modified_hyperparam_search_space.keys():
+            param_space = modified_hyperparam_search_space[param]["space"]
+            if param_space == "choice":
+                if len(modified_hyperparam_search_space[param]['categories']) > 2 and \
+                        modified_hyperparam_search_space[param]['categories'][-2] > min_value:
+                    modified_hyperparam_search_space[param][
+                        'categories'] = modified_hyperparam_search_space[param]['categories'][:-1]
+                else:
+                    param_list.pop(0)  # exhausted reduction of this parameter
+            else:
+                # reduce by 10%
+                upper_bound, lower_bound = modified_hyperparam_search_space[param][
+                    "upper"], modified_hyperparam_search_space[param]["lower"]
+                reduction_val = (upper_bound - lower_bound) * 0.1
+                new_upper_bound = upper_bound - reduction_val
+                if (new_upper_bound) > lower_bound and new_upper_bound > min_value:
+                    modified_hyperparam_search_space[param]["upper"] = new_upper_bound
+                else:
+                    param_list.pop(0)  # exhausted reduction of this parameter
+        else:
+            param_list.pop(0)  # param not in hyperopt search space
+
+    modified_config = copy.deepcopy(config)
+
+    modified_config[HYPEROPT]["parameters"] = modified_hyperparam_search_space
+    return modified_config, fits_in_memory

ludwig/automl/automl.py

@@ -16,6 +16,7 @@
 
 from ludwig.api import LudwigModel
 from ludwig.automl.base_config import _create_default_config, DatasetInfo
+from ludwig.automl.auto_tune_config import memory_tune_config
 from ludwig.automl.utils import _ray_init
 from ludwig.constants import COMBINER, TYPE
 from ludwig.hyperopt.run import hyperopt
@@ -61,6 +62,7 @@ def auto_train(
     target: str,
     time_limit_s: Union[int, float],
     output_directory: str = OUTPUT_DIR,
+    tune_for_memory: bool = False,
     **kwargs
 ) -> AutoTrainResults:
     """
@@ -81,7 +83,8 @@ def auto_train(
     # Returns
     :return: (AutoTrainResults) results containing hyperopt experiments and best model
     """
-    config = create_auto_config(dataset, target, time_limit_s)
+    config = create_auto_config(
+        dataset, target, time_limit_s, tune_for_memory, **kwargs)
     return train_with_config(
         dataset,
         config,
@@ -94,6 +97,7 @@ def create_auto_config(
     dataset: Union[str, pd.DataFrame, dd.core.DataFrame, DatasetInfo],
     target: str,
     time_limit_s: Union[int, float],
+    tune_for_memory: bool,
 ) -> dict:
     """
     Returns an auto-generated Ludwig config with the intent of training
@@ -111,6 +115,13 @@ def create_auto_config(
     """
     default_configs = _create_default_config(dataset, target, time_limit_s)
     model_config = _model_select(default_configs)
+    if tune_for_memory:
+        if ray.is_initialized():
+            model_config, _ = ray.get(ray.remote(num_cpus=1)(
+                memory_tune_config
+            ).remote(model_config, dataset))
+        else:
+            model_config, _ = memory_tune_config(model_config, dataset)
     return model_config
 
 

requirements_ray.txt

@@ -1,3 +1,5 @@
 ray[default,tune]
 pickle5
 tensorboardX<2.3
+GPUtil
+