def printTensor3d(dataTable,
message="",
nFirstDim=0,
nSecondDim=0,
interval=10):
dims = dataTable.getDimensions()
nRows = int(dims[0])
nCols = int(dims[1])
if nFirstDim != 0:
nFirstDim = min(nRows, nFirstDim)
else:
nFirstDim = nRows
if nSecondDim != 0:
nSecondDim = min(nCols, nSecondDim)
else:
nSecondDim = nCols
block = SubtensorDescriptor()
print(message)
for i in range(nFirstDim):
dataTable.getSubtensor([i], 0, nSecondDim, readOnly, block)
nThirdDim = block.getSize() / nSecondDim
printArray(block.getArray(), int(nThirdDim), int(nSecondDim),
int(nThirdDim), "", interval)
dataTable.releaseSubtensor(block)
def printTensor(dataTable,
message="",
nPrintedRows=0,
nPrintedCols=0,
interval=10):
dims = dataTable.getDimensions()
nRows = int(dims[0])
if nPrintedRows != 0:
nPrintedRows = min(nRows, nPrintedRows)
else:
nPrintedRows = nRows
block = SubtensorDescriptor()
dataTable.getSubtensor([], 0, nPrintedRows, readOnly, block)
nCols = int(block.getSize() / nPrintedRows)
if nPrintedCols != 0:
nPrintedCols = min(nCols, nPrintedCols)
else:
nPrintedCols = nCols
printArray(block.getArray(), int(nPrintedCols), int(nPrintedRows),
int(nCols), message, interval)
dataTable.releaseSubtensor(block)
def _readBatchFromDataset(self, in_file, counter):
"""Reads batch of images coresponding to the current reader position"""
imagesBatchSize = self._imagesInBatch * self._imageChannels * self._imageWidth * self._imageHeight
batchPosition = self._imagesPosition + imagesBatchSize * counter
in_file.seek(batchPosition)
dataBatch = allocateTensor(np.float32, self._imagesInBatch,
self._imageChannels, self._imageHeight,
self._imageWidth)
trainTensorSize = dataBatch.getSize()
batchBlock = SubtensorDescriptor(ntype=np.float32)
dataBatch.getSubtensor([], 0, self._imagesInBatch, writeOnly,
batchBlock)
objectsPtr = batchBlock.getArray()
binary_data_str = in_file.read(trainTensorSize)
objectData = np.array(struct.unpack('B' * trainTensorSize,
binary_data_str),
dtype=np.float32)
for x, i in zip(np.nditer(objectsPtr, op_flags=['readwrite']),
range(len(objectData))):
x[...] = objectData[i]
dataBatch.releaseSubtensor(batchBlock)
return dataBatch
def __init__(self, tensor, read_type=np.float32): if not isinstance(tensor, Tensor): raise ValueError(PYDAAL_NOT_A_TENSOR % type(tensor)) self.tensor = tensor self.block = SubtensorDescriptor(ntype=read_type) self.tensor.getSubtensor([], 0, tensor.getDimensionSize(0), readOnly, self.block)
def findClasses(dataTable):
dims1 = dataTable.getDimensions()
nRows1 = int(dims1[0])
block1 = SubtensorDescriptor()
dataTable.getSubtensor([], 0, nRows1, readOnly, block1)
nCols1 = int(block1.getSize() / nRows1)
dataType = block1.getArray().flatten()
dataType = np.reshape(dataType, (nRows1, nCols1))
classes = np.argmax(dataType, axis=1)
dataTable.releaseSubtensor(block1)
return classes
def predict(self, data, batch_size=None, rebuild=True):
"""Predicts labels based on a prediction model.
Supported notation is ``with net.predict(...) as predictions:``
Args:
data (:obj:`daal.data_management.Tensor` or :obj:`numpy.ndarray`): Prediction data.
batch_size (:obj:`int`): Batch size for processing prediction data.
rebuild (:obj:`bool`): Control parameter to force rebuild of the model.
Returns:
:py:class:`pydaalcontrib.nn.DAALNet`: DAAL network with the evaluated predictions.
Raises:
ValueError: If the provided ``data`` are of the wrong type.
"""
if isinstance(data, np.ndarray):
_data = HomogenTensor(data.copy(), ntype=data.dtype)
elif not isinstance(data, Tensor):
raise ValueError('Data is not of numpy.ndarray or Tensor type!')
if not batch_size or batch_size > _data.getDimensionSize(0):
batch_size = _data.getDimensionSize(0)
if rebuild and self.do_rebuild:
#TODO: refactor set rebuild=False once memory allocation is fixed on prediction in Intel DAAL 2018
parameter = prediction.Parameter()
parameter.batchSize = batch_size
self.do_rebuild = False
rebuild_args = {
'data_dims': [batch_size] + _data.getDimensions()[1:],
'parameter': parameter
}
self.model = self.build_model(self.descriptor,
False,
rebuild=rebuild_args,
**self.build_args)
elif 'train_result' in self.__dict__:
self.model = self.train_result.get(
training.model).getPredictionModel_Float32()
net = prediction.Batch()
net.parameter.batchSize = batch_size
net.input.setModelInput(prediction.model, self.model)
net.input.setTensorInput(prediction.data, _data)
self.predictions = SubtensorDescriptor(ntype=data.dtype)
self.predict_result = net.compute().getResult(prediction.prediction)
self.predict_result.getSubtensor(
[], 0, self.predict_result.getDimensionSize(0), readOnly,
self.predictions)
return self
def getNextSubtensor(inputTensor, startPos, nElements):
dims = inputTensor.getDimensions()
dims[0] = nElements
subtensorBlock = SubtensorDescriptor(ntype=np.float32)
inputTensor.getSubtensor([], startPos, nElements, readOnly, subtensorBlock)
subtensorData = np.array(subtensorBlock.getArray(),
copy=True,
dtype=np.float32)
inputTensor.releaseSubtensor(subtensorBlock)
return HomogenTensor(subtensorData, ntype=np.float32)
class DataReader: """Wrapper class for reading Intel DAAL tensors. Supported notation is ``with DataReader(...) as result:``. Args: tensor (:obj:`daal.data_management.Tensor`): Provided tensor. read_type (:obj:`numpy.dtype`, optional): Numpy type for the result tensor. Raises: ValueError: If provided argument is not a :obj:`daal.data_management.Tensor`. """ def __init__(self, tensor, read_type=np.float32): if not isinstance(tensor, Tensor): raise ValueError(PYDAAL_NOT_A_TENSOR % type(tensor)) self.tensor = tensor self.block = SubtensorDescriptor(ntype=read_type) self.tensor.getSubtensor([], 0, tensor.getDimensionSize(0), readOnly, self.block) def __enter__(self): return self.block.getArray() def __exit__(self, type, value, traceback): self.tensor.releaseSubtensor(self.block)
def printTensors(dataTable1,
dataTable2,
title1="",
title2="",
message="",
nPrintedRows=0,
interval=15):
dims1 = dataTable1.getDimensions()
nRows1 = int(dims1[0])
if nPrintedRows != 0:
nPrintedRows = min(nRows1, nPrintedRows)
else:
nPrintedRows = nRows1
block1 = SubtensorDescriptor()
dataTable1.getSubtensor([], 0, nPrintedRows, readOnly, block1)
nCols1 = int(block1.getSize() / nPrintedRows)
dims2 = dataTable2.getDimensions()
nRows2 = int(dims2[0])
if nPrintedRows != 0:
nPrintedRows = min(nRows2, nPrintedRows)
else:
nPrintedRows = nRows2
block2 = SubtensorDescriptor()
dataTable2.getSubtensor([], 0, nPrintedRows, readOnly, block2)
nCols2 = int(block2.getSize() / nPrintedRows)
dataType1 = block1.getArray().flatten()
dataType2 = block2.getArray().flatten()
print(message)
print("{:<{width}}".format(title1, width=(interval * nCols1)), end='')
print("{:<{width}}".format(title2, width=(interval * nCols2)))
for i in range(nPrintedRows):
for j in range(nCols1):
print("{v:<{width}.0f}".format(v=dataType1[i * nCols1 + j],
width=interval),
end='')
for j in range(nCols2):
print("{:<{width}.3f}".format(dataType2[i * nCols2 + j],
width=int(interval / 2)),
end='')
print()
print()
dataTable1.releaseSubtensor(block1)
dataTable2.releaseSubtensor(block2)
def _readGroundTruthFromDataset(self, in_file, counter):
"""Reads batch of labels coresponding to the current reader position"""
batchLabelsSize = self._imagesInBatch * 4
batchPosition = self._classesPosition + batchLabelsSize * counter
in_file.seek(batchPosition)
groundTruthBatch = allocateTensor(np.intc, self._imagesInBatch, 1)
groundTruthBlock = SubtensorDescriptor(ntype=np.intc)
groundTruthBatch.getSubtensor([], 0, self._imagesInBatch, writeOnly,
groundTruthBlock)
groundTruthPtr = groundTruthBlock.getArray()
for x in np.nditer(groundTruthPtr, op_flags=['readwrite']):
x[...] = int(self._readDWORD(in_file))
groundTruthBatch.releaseSubtensor(groundTruthBlock)
return groundTruthBatch
def update(self, _prediction, _groundTruth):
if not _prediction:
raise RuntimeError("Prediction tensor should not be null")
if not _groundTruth:
raise RuntimeError("GroundTruth tensor should not be null")
dimensions = _prediction.getDimensions()
if len(dimensions) != 2:
raise RuntimeError(
"Predictions tensor should have exactly two dimensions")
rowsNum = dimensions[0]
colsNum = dimensions[1]
if colsNum < ClassificationErrorCounter.MAX_ERROR_RATE_CLASSES:
raise RuntimeError(
"Number of classes in prediction result is not enough to compute error rate"
)
predictionBlock = SubtensorDescriptor(ntype=np.float32)
_prediction.getSubtensor([], 0, dimensions[0], readOnly,
predictionBlock)
predictionRows = predictionBlock.getArray()
groundTruthBlock = SubtensorDescriptor(ntype=np.intc)
_groundTruth.getSubtensor([], 0, dimensions[0], readOnly,
groundTruthBlock)
groundTruthClasses = groundTruthBlock.getArray()
for i in range(rowsNum):
row = predictionRows[i]
topIndices = self.findTopIndices(
row, colsNum,
ClassificationErrorCounter.MAX_ERROR_RATE_CLASSES)
groundTruthClass = groundTruthClasses[0][i]
self._totalObjects += 1
if groundTruthClass in topIndices:
self._top5ClassifiedObjects += 1
if groundTruthClass == topIndices[0]:
self._top1ClassifiedObjects += 1
_prediction.releaseSubtensor(predictionBlock)
_groundTruth.releaseSubtensor(groundTruthBlock)
def printPredictedClasses(_predictionResult, _testingGroundTruth):
_prediction = _predictionResult.getResult(prediction.prediction)
predictionDimensions = _prediction.getDimensions()
predictionBlock = SubtensorDescriptor()
_prediction.getSubtensor([], 0, predictionDimensions[0], readOnly,
predictionBlock)
predictionPtr = predictionBlock.getArray()
testGroundTruthBlock = SubtensorDescriptor(ntype=np.intc)
_testingGroundTruth.getSubtensor([], 0, predictionDimensions[0], readOnly,
testGroundTruthBlock)
testGroundTruthPtr = testGroundTruthBlock.getArray().flatten()
# Print predicted classes
maxPIndex = np.argmax(predictionPtr, axis=1)
for i in range(predictionDimensions[0]):
for p in predictionPtr[i]:
print("{:.4f} ".format(p), end="")
print(" -> {} | {}".format(maxPIndex[i], testGroundTruthPtr[i]))
_prediction.releaseSubtensor(predictionBlock)
_testingGroundTruth.releaseSubtensor(testGroundTruthBlock)
def checkResult(predictionResult, testingGroundTruth, TestDataCount):
pred = predictionResult.getResult(prediction.prediction)
predictionDimensions = pred.getDimensions()
predictionBlock = SubtensorDescriptor()
pred.getSubtensor([], 0, predictionDimensions[0], readOnly,
predictionBlock)
predictionPtr = predictionBlock.getArray()
testGroundTruthBlock = SubtensorDescriptor(ntype=np.intc)
testingGroundTruth.getSubtensor([], 0, predictionDimensions[0], readOnly,
testGroundTruthBlock)
testGroundTruthPtr = testGroundTruthBlock.getArray().flatten()
maxPIndex = 0
trueCount = 0
# validation accuracy finding
maxPIndex = np.argmax(predictionPtr, axis=1)
trueCount = np.sum(maxPIndex == testGroundTruthPtr)
pred.releaseSubtensor(predictionBlock)
testingGroundTruth.releaseSubtensor(testGroundTruthBlock)
return True if trueCount / TestDataCount > 0.9 else False
class DAALNet:
"""Wrapper class for working with :obj:`daal.algorithms.neural_networks` package.
Notes
Default working regime is training, see :obj:`daal.algorithms.neural_networks.training.Batch()`.
Default solver used for training is SGD, see :obj:`daal.algorithms.optimization_solver.sgd.Batch()`.
"""
_daal_net_namespace = dict()
def __init__(self):
#TODO: set do_rebuild=False once memory allocation is fixed on prediction in Intel DAAL 2018
self.do_rebuild = True
self.initializer = None
self.solver = sgd.Batch()
self.net = training.Batch(self.solver)
def with_solver(self, solver):
"""Provides a specific solver for the Intel DAAL net/graph.
Args:
solver (from :obj:`daal.algorithms.optimization_solver` module): Intel DAAL solver.
Returns:
:py:class:`pydaalcontrib.nn.DAALNet`: Intel DAAL network with the provided solver.
"""
self.solver = solver
self.net = training.Batch(self.solver)
return self
def with_initializer(self, initializer):
self.initializer = initializer
return self
def train(self, data, labels, **kw_args):
"""Trains a specific Intel DAAL net/graph based on the provided data and labels.
Args:
data (:obj:`daal.data_management.Tensor` or :obj:`numpy.ndarray`): Training data.
labels (:obj:`daal.data_management.Tensor` or :obj:`numpy.ndarray`): Training labels.
**kwargs: Arbitrary keyword arguments (``batch_size`` and ``learning_rate``).
Returns:
:py:class:`pydaalcontrib.nn.DAALNet`: Trained DAAL network.
Raises:
ValueError: If the provided ``data`` or ``labels`` are of the wrong type or the topology is not set.
"""
if 'topology' not in self.__dict__:
raise ValueError('Topology is not intialized!')
if 'batch_size' in kw_args and 'result' not in self.__dict__:
self.solver.parameter.batchSize = kw_args['batch_size']
if 'learning_rate' in kw_args and 'learningRate' in self.solver.parameter.__swig_getmethods__:
self.solver.parameter.learningRate = get_learning_rate(
kw_args['learning_rate'])
if 'learning_rate' in kw_args and 'learningRateSequence' in self.solver.parameter.__swig_getmethods__:
self.solver.parameter.learningRateSequence = get_learning_rate(
kw_args['learning_rate'])
if isinstance(data, Tensor):
self.data = data
elif isinstance(data, np.ndarray) and data.base is not None:
self.data = HomogenTensor(data.copy(), ntype=data.dtype)
elif isinstance(data, np.ndarray) and data.base is None:
self.data = HomogenTensor(data, ntype=data.dtype)
else:
raise ValueError('Data is not of numpy.ndarray or Tensor type!')
if isinstance(labels, Tensor):
self.labels = labels
elif isinstance(labels, np.ndarray):
if len(labels.shape) == 1:
labels = labels.reshape([-1, 1])
if issubdtype(labels, np.int):
labels = labels.astype(np.intc)
elif not issubdtype(labels, np.float):
labels = labels.astype(np.float)
self.labels = HomogenTensor(labels.copy(), ntype=labels.dtype)
else:
raise ValueError('Labels are not of numpy.ndarray or Tensor type!')
if 'train_result' not in self.__dict__ or self.train_result is None:
dims = self.data.getDimensions()[1:]
dims.insert(0, self.solver.parameter.batchSize)
self.net.initialize(dims, self.topology)
# heuristically define the number of iterations for ``self.solver``
batch_size = np.float(self.solver.parameter.batchSize)
n_iter = np.ceil(self.data.getDimensionSize(0) / batch_size)
self.solver.parameter.nIterations = np.int(n_iter)
# Pass a solver, training data and lables to the algorithm
self.net.parameter.optimizationSolver = self.solver
self.net.input.setInput(training.data, self.data)
self.net.input.setInput(training.groundTruth, self.labels)
# Do an actual compute and store the result
self.train_result = self.net.compute()
self.do_rebuild = False
return self
#TODO: refactor set rebuild=False once memory allocation is fixed on prediction in Intel DAAL 2018
def predict(self, data, batch_size=None, rebuild=True):
"""Predicts labels based on a prediction model.
Supported notation is ``with net.predict(...) as predictions:``
Args:
data (:obj:`daal.data_management.Tensor` or :obj:`numpy.ndarray`): Prediction data.
batch_size (:obj:`int`): Batch size for processing prediction data.
rebuild (:obj:`bool`): Control parameter to force rebuild of the model.
Returns:
:py:class:`pydaalcontrib.nn.DAALNet`: DAAL network with the evaluated predictions.
Raises:
ValueError: If the provided ``data`` are of the wrong type.
"""
if isinstance(data, np.ndarray):
_data = HomogenTensor(data.copy(), ntype=data.dtype)
elif not isinstance(data, Tensor):
raise ValueError('Data is not of numpy.ndarray or Tensor type!')
if not batch_size or batch_size > _data.getDimensionSize(0):
batch_size = _data.getDimensionSize(0)
if rebuild and self.do_rebuild:
#TODO: refactor set rebuild=False once memory allocation is fixed on prediction in Intel DAAL 2018
parameter = prediction.Parameter()
parameter.batchSize = batch_size
self.do_rebuild = False
rebuild_args = {
'data_dims': [batch_size] + _data.getDimensions()[1:],
'parameter': parameter
}
self.model = self.build_model(self.descriptor,
False,
rebuild=rebuild_args,
**self.build_args)
elif 'train_result' in self.__dict__:
self.model = self.train_result.get(
training.model).getPredictionModel_Float32()
net = prediction.Batch()
net.parameter.batchSize = batch_size
net.input.setModelInput(prediction.model, self.model)
net.input.setTensorInput(prediction.data, _data)
self.predictions = SubtensorDescriptor(ntype=data.dtype)
self.predict_result = net.compute().getResult(prediction.prediction)
self.predict_result.getSubtensor(
[], 0, self.predict_result.getDimensionSize(0), readOnly,
self.predictions)
return self
def get_predictions(self):
"""Gets the latest predictions after :py:meth:`predict` was called.
Returns:
:py:obj:`numpy.ndarray`: Evaluated predictions.
"""
if 'predictions' in self.__dict__:
predictions_numpy = self.predictions.getArray()
self.predict_result.releaseSubtensor(self.predictions)
return predictions_numpy
else:
return None
def __enter__(self):
return self.predictions.getArray()
def __exit__(self, type, value, traceback):
self.predict_result.releaseSubtensor(self.predictions)
def allocate_model(self, model, args):
"""Allocates a contiguous memory for the model if 'rebuild' option is specified.
Args:
model (:obj:`daal.algorithms.neural_networks.prediction.Model`): instantiated model.
args (:obj:`dict`): Different args which are passed from :py:func:`build_model`.
Returns:
:obj:`daal.algorithms.neural_networks.prediction.Model`
"""
if 'rebuild' in args:
parameter = args['rebuild']['parameter']
data_dims = args['rebuild']['data_dims']
model.allocate_Float32(data_dims, parameter)
return model
def build_model(self, model, trainable, **kw_args):
"""(re)Builds a specific Intel DAAL model based on the provided descriptor.
Args:
model (:py:class:`pydaalcontrib.model.ModelBase` or :obj:`str`): Instance of a model or a path to the folder/file containing the model (*pydaal.model*) file.
trainable (:obj:`bool`): Flag indicating whether `training` or `prediction` topology to be built.
kw_args (:obj:`dict`): Different keyword args which might be of use in sub-classes.
Returns:
:obj:`daal.algorithms.neural_networks.prediction.Model` or ``None``
"""
if 'model' in self.__dict__:
return self.allocate_model(self.model, kw_args)
if isinstance(model, basestring):
self.descriptor = load_model(model)
else:
self.descriptor = model
self.topology = build_topology(self.descriptor,
trainable,
initializer=self.initializer)
#TODO: replace with training.Model(topology) once fixed
return None if trainable else self.allocate_model(
prediction.Model(self.topology), kw_args)
@dispatch(basestring, namespace=_daal_net_namespace)
def build(self, model_path, trainable=False, **kw_args):
self.model = self.build_model(model_path, trainable, **kw_args)
self.build_args = {'model_path': model_path}
self.build_args.update(kw_args)
return self
@dispatch(Model, namespace=_daal_net_namespace)
def build(self, model, trainable=True, **kw_args):
self.model = self.build_model(model, trainable, **kw_args)
self.build_args = kw_args
return self