Python Dataset Reader (#2414)

* Add skeleton for new python data reader

* Implement basic functionality

* Fix initialization for distconv

* Add support for labels

* Add python library supporting classes

* clang format

* Raise exception if rank/io parts not set

* Rename to python dataset

* Add optional module dir argument to add to path

* Add unit tests

* Simplify naming

* Add cosmoflow example and reader helper

* Update release notes

* Save dataset pickle in work dir

* Overhaul new data reader to support prefetching multiple samples/batches

* Fix worker index calculation

* clang-format

* Clarify proto comments

* Throw error if file fails to open

* Add docstrings and type hints

* Update CosmoFlow example and enable parallel IO

* Add basic sample size checking, remove label reconstruction, general clean up

* Switch to multiprocessing pool

* Implement response shuffling for distconv

* fix typo

Co-authored-by: Tal Ben-Nun <[email protected]>

---------

Co-authored-by: Tal Ben-Nun <[email protected]>

Author: fiedorowicz1

Diff for: ReleaseNotes.txt

@@ -19,6 +19,8 @@ Experiments & Applications:
 Internal features:
 
 I/O & data ingestion:
+ - Added a new python dataset reader for simple, flexible, and distconv-supported
+   python data readers.
 
 Build system:
 

Diff for: applications/physics/cosmology/cosmoflow/cosmoflow_dataset.py

@@ -0,0 +1,31 @@
+import numpy as np
+from glob import glob
+from lbann.util.data import Sample, SampleDims, Dataset, DistConvDataset
+import h5py as h5
+import os
+
+    
+class CosmoFlowDataset(DistConvDataset):
+    def __init__(self, data_dir, input_width, num_secrets):
+        self.data_dir = data_dir
+        self.input_width = input_width
+        self.num_secrets = num_secrets
+        self.samples = glob(os.path.join(data_dir, '*.hdf5'))
+        self.samples.sort()
+    
+    def __len__(self):
+        return len(self.samples)
+    
+    def __getitem__(self, index) -> Sample:
+        data = h5.File(self.samples[index], 'r')
+        slice_width = self.input_width // self.num_io_partitions
+        slice_ind = self.rank % self.num_io_partitions
+        full = data['full'][:,
+                            slice_ind*slice_width:(slice_ind+1)*slice_width,
+                            :self.input_width,
+                            :self.input_width].astype(np.float32)
+        par = data['unitPar'][:].astype(np.float32)
+        return Sample(sample=np.ascontiguousarray(full), response=par)
+    
+    def get_sample_dims(self):
+        return SampleDims(sample=[4, self.input_width, self.input_width, self.input_width], response=self.num_secrets)

Diff for: applications/physics/cosmology/cosmoflow/train_cosmoflow.py

@@ -10,7 +10,21 @@
 import lbann.contrib.args
 import lbann.contrib.launcher
 from lbann.core.util import get_parallel_strategy_args
+import lbann.util.data
 import os
+from cosmoflow_dataset import CosmoFlowDataset
+
+def create_python_dataset_reader(args):
+    """Create a python dataset reader for CosmoFlow."""
+
+    readers = []
+    for role in ['train', 'val', 'test']:
+        role_dir = getattr(args, f'{role}_dir')
+        dataset = CosmoFlowDataset(role_dir, args.input_width, args.num_secrets)
+        reader = lbann.util.data.construct_python_dataset_reader(dataset, role=role)
+        readers.append(reader)
+
+    return lbann.reader_pb2.DataReader(reader=readers)
 
 def create_cosmoflow_data_reader(
         train_path, val_path, test_path, num_responses):
@@ -133,6 +147,9 @@ def create_synthetic_data_reader(input_width: int, num_responses: int) -> Any:
     parser.add_argument(
         '--synthetic', action='store_true',
         help='Use synthetic data')
+    parser.add_argument(
+        '--python-dataset', action='store_true',
+        help='Use python dataset reader')
     parser.add_argument(
         '--no-datastore', action='store_true',
         help='Disable the data store')
@@ -220,22 +237,26 @@ def create_synthetic_data_reader(input_width: int, num_responses: int) -> Any:
     # optimizer.learn_rate *= 1e-2
 
     # Setup data reader
+    serialize_io = False
     if args.synthetic:
         data_reader = create_synthetic_data_reader(
             args.input_width, args.num_secrets)
+    elif args.python_dataset:
+        data_reader = create_python_dataset_reader(args)
     else:
         data_reader = create_cosmoflow_data_reader(
             args.train_dir,
             args.val_dir,
             args.test_dir,
             num_responses=args.num_secrets)
+        serialize_io = True
 
     # Setup trainer
     random_seed_arg = {'random_seed': args.random_seed} \
         if args.random_seed is not None else {}
     trainer = lbann.Trainer(
         mini_batch_size=args.mini_batch_size,
-        serialize_io=True,
+        serialize_io=serialize_io,
         **random_seed_arg)
 
     # Runtime parameters/arguments

Diff for: ci_test/unit_tests/test_unit_datareader_python_dataset.py

@@ -0,0 +1,137 @@
+import os
+import os.path
+import sys
+import numpy as np
+from lbann.util.data import Dataset, Sample, SampleDims, construct_python_dataset_reader
+
+# Bamboo utilities
+current_file = os.path.realpath(__file__)
+current_dir = os.path.dirname(current_file)
+sys.path.insert(0, os.path.join(os.path.dirname(current_dir), 'common_python'))
+import tools
+
+# ==============================================
+# Objects for Python dataset data reader
+# ==============================================
+# Note: The Python dataset data reader loads the dataset constructed below.
+
+# Data
+class TestDataset(Dataset):
+    def __init__(self):
+        np.random.seed(20240109)
+        self.num_samples = 29
+        self.sample_size = 7
+        self.samples = np.random.normal(size=(self.num_samples,self.sample_size)).astype(np.float32)
+    
+    def __len__(self):
+        return self.num_samples
+    
+    def __getitem__(self, index):
+        return Sample(sample=self.samples[index,:])
+    
+    def get_sample_dims(self):
+        return SampleDims(sample=[self.sample_size])
+
+test_dataset = TestDataset()
+
+# ==============================================
+# Setup LBANN experiment
+# ==============================================
+
+def setup_experiment(lbann, weekly):
+    """Construct LBANN experiment.
+
+    Args:
+        lbann (module): Module for LBANN Python frontend
+
+    """
+    mini_batch_size = len(test_dataset) // 4
+    trainer = lbann.Trainer(mini_batch_size)
+    model = construct_model(lbann)
+    data_reader = construct_data_reader(lbann)
+    optimizer = lbann.NoOptimizer()
+    return trainer, model, data_reader, optimizer, None # Don't request any specific number of nodes
+
+def construct_model(lbann):
+    """Construct LBANN model.
+
+    Args:
+        lbann (module): Module for LBANN Python frontend
+
+    """
+
+    # Layer graph
+    x = lbann.Input(data_field='samples')
+    y = lbann.L2Norm2(x)
+    layers = list(lbann.traverse_layer_graph(x))
+    metric = lbann.Metric(y, name='obj')
+    callbacks = []
+
+    # Compute expected value with NumPy
+    vals = []
+    for i in range(len(test_dataset)):
+        x = test_dataset[i].sample.astype(np.float64)
+        y = tools.numpy_l2norm2(x)
+        vals.append(y)
+    val = np.mean(vals)
+    tol = 8 * val * np.finfo(np.float32).eps
+    callbacks.append(lbann.CallbackCheckMetric(
+        metric=metric.name,
+        lower_bound=val-tol,
+        upper_bound=val+tol,
+        error_on_failure=True,
+        execution_modes='test'))
+
+    # Construct model
+    num_epochs = 0
+    return lbann.Model(num_epochs,
+                       layers=layers,
+                       metrics=[metric],
+                       callbacks=callbacks)
+
+def construct_data_reader(lbann):
+    """Construct Protobuf message for Python dataset data reader.
+
+    The Python data reader will import the current Python file to
+    access the sample access functions.
+
+    Args:
+        lbann (module): Module for LBANN Python frontend
+
+    """
+
+    dataset_path = os.path.join(work_dir, 'dataset.pkl')
+    
+    # Note: The training data reader should be removed when
+    # https://github.com/LLNL/lbann/issues/1098 is resolved.
+    message = lbann.reader_pb2.DataReader()
+    message.reader.extend([
+        construct_python_dataset_reader(
+            test_dataset,
+            dataset_path,
+            'train',
+            shuffle=False
+        )
+    ])
+    message.reader.extend([
+        construct_python_dataset_reader(
+            test_dataset,
+            dataset_path,
+            'test',
+            shuffle=False
+        )
+    ])
+    return message
+
+# ==============================================
+# Setup PyTest
+# ==============================================
+
+work_dir = os.path.join(os.path.dirname(__file__),
+                        'experiments',
+                        os.path.basename(__file__).split('.py')[0])
+os.makedirs(work_dir, exist_ok=True)
+
+# Create test functions that can interact with PyTest
+for _test_func in tools.create_tests(setup_experiment, __file__, work_dir=work_dir):
+    globals()[_test_func.__name__] = _test_func