def to_dataset(data, batch_size, batch_per_thread, validation_data,
feature_cols, labels_cols, hard_code_batch_size,
sequential_order, shuffle, auto_shard_files):
# todo wrap argument into kwargs
if validation_data:
if isinstance(data, SparkXShards):
assert isinstance(validation_data, SparkXShards), \
"train data and validation data should be both SparkXShards"
if isinstance(data, Dataset):
assert isinstance(validation_data, Dataset), \
"train data and validation data should be both orca.data.tf.Dataset"
if isinstance(data, DataFrame):
assert isinstance(validation_data, DataFrame), \
"train data and validation data should be both Spark DataFrame"
if isinstance(data, tf.data.Dataset):
assert isinstance(validation_data, tf.data.Dataset), \
"train data and validation data should be both tf.data.Dataset"
if isinstance(data, SparkXShards):
dataset = xshards_to_tf_dataset(
data,
batch_size,
batch_per_thread,
validation_data,
hard_code_batch_size=hard_code_batch_size,
sequential_order=sequential_order,
shuffle=shuffle)
elif isinstance(data, Dataset):
dataset = TFDataDataset2(data,
batch_size=batch_size,
batch_per_thread=batch_per_thread,
validation_dataset=validation_data)
elif isinstance(data, DataFrame):
dataset = TFDataset.from_dataframe(data, feature_cols, labels_cols,
batch_size, batch_per_thread,
hard_code_batch_size,
validation_data, sequential_order,
shuffle)
elif is_tf_data_dataset(data):
dataset = TFDataset.from_tf_data_dataset(
data,
batch_size,
batch_per_thread,
hard_code_batch_size,
validation_data,
sequential_order,
shuffle,
auto_shard_files=auto_shard_files)
else:
raise ValueError(
"data must be SparkXShards or orca.data.tf.Dataset or "
"Spark DataFrame or tf.data.Dataset")
return dataset
def fit(self,
x=None,
y=None,
batch_size=None,
epochs=1,
validation_data=None,
distributed=False,
**kwargs
):
"""
Train the model for a fixed num of epochs
Arguments:
:param x: Input data. It could be:
- a TFDataset object
- A Numpy array (or array-like), or a list of arrays
(in case the model has multiple inputs).
- A dict mapping input names to the corresponding array/tensors,
if the model has named inputs.
:param y: Target data. Like the input data `x`,
It should be consistent with `x` (you cannot have Numpy inputs and
tensor targets, or inversely). If `x` is a TFDataset, `y` should
not be specified (since targets will be obtained from `x`).
:param batch_size: Integer or `None`.
Number of samples per gradient update.
If `x` is a TFDataset, you do not need to specify batch_size.
:param epochs: Integer. Number of epochs to train the model.
An epoch is an iteration over the entire `x` and `y`
data provided.
:param validation_data: Data on which to evaluate
the loss and any model metrics at the end of each epoch.
The model will not be trained on this data.
`validation_data` could be:
- tuple `(x_val, y_val)` of Numpy arrays or tensors
:param distributed: Boolean. Whether to do prediction in distributed mode or local mode.
Default is True. In local mode, x must be a Numpy array.
"""
if isinstance(x, TFDataset):
# todo check arguments
assert validation_data is None, "validation_data must be None when " \
"using TFDataset as input, please " \
"use set the validation data in TFDataset"
if not x.has_batch:
raise ValueError("The batch_size of TFDataset must be " +
"specified when used in KerasModel fit.")
self._fit_distributed(x, epochs, **kwargs)
elif distributed:
dataset = TFDataset.from_ndarrays((x, y), val_tensors=validation_data,
batch_size=batch_size)
self._fit_distributed(dataset, epochs, **kwargs)
else:
self.model.fit(x=x,
y=y,
batch_size=batch_size,
epochs=epochs,
validation_data=validation_data,
**kwargs
)
def xshards_to_tf_dataset(data_shard,
batch_size=-1, batch_per_thread=-1,
validation_data_shard=None,
hard_code_batch_size=False,
sequential_order=False,
shuffle=True):
# todo data_shard.head ?
feature_spec, label_spec = data_shard._for_each(get_spec(allow_tuple=True, allow_list=False))\
.first()
feature_spec = [(tf.dtypes.as_dtype(spec[0]), spec[1]) for spec in feature_spec]
label_spec = [(tf.dtypes.as_dtype(spec[0]), spec[1]) for spec in label_spec] \
if label_spec is not None else None
assert batch_size != -1 or batch_per_thread != -1, \
"one of batch_size and batch_per_thread should be specified"
val_rdd = None if validation_data_shard is None \
else validation_data_shard.rdd.flatMap(flatten_xy(allow_tuple=True, allow_list=False))
dataset = TFDataset.from_rdd(data_shard.rdd.flatMap(flatten_xy(allow_tuple=True,
allow_list=False)),
features=feature_spec,
labels=label_spec,
batch_size=batch_size,
batch_per_thread=batch_per_thread,
val_rdd=val_rdd,
hard_code_batch_size=hard_code_batch_size,
sequential_order=sequential_order,
shuffle=shuffle)
return dataset
def predict(self,
x,
batch_per_thread=None,
distributed=False):
"""
Use a model to do prediction.
:param x: Input data. It could be:
- a TFDataset object
- A Numpy array (or array-like), or a list of arrays
(in case the model has multiple inputs).
- A dict mapping input names to the corresponding array/tensors,
if the model has named inputs.
:param batch_per_thread:
The default value is 1.
When distributed is True,the total batch size is batch_per_thread * rdd.getNumPartitions.
When distributed is False the total batch size is batch_per_thread * numOfCores.
:param distributed: Boolean. Whether to do prediction in distributed mode or local mode.
Default is True. In local mode, x must be a Numpy array.
"""
if isinstance(x, TFDataset):
# todo check arguments
if not x.has_batch:
raise ValueError("The batch_per_thread of TFDataset" +
" must be specified when used in KerasModel predict.")
if isinstance(x, TFNdarrayDataset):
x = _standarize_feature_dataset(x, self.model)
return self._predict_distributed(x)
else:
if distributed:
sc = getOrCreateSparkContext()
rdd, types, shapes = _create_rdd_x(x, self.model._feed_input_names, sc)
dataset = TFDataset.from_rdd(rdd,
names=self.model._feed_input_names,
types=types,
shapes=shapes,
batch_per_thread=-1 if batch_per_thread is None
else batch_per_thread)
results = self._predict_distributed(dataset).collect()
output_num = len(self.model.outputs)
if output_num == 1:
return np.stack(results)
else:
predictions = []
for i in range(0, output_num):
predictions.append(np.stack([res[i] for res in results]))
return predictions
else:
return self.model.predict(x=x,
batch_size=batch_per_thread)
def evaluate(self,
x=None,
y=None,
batch_per_thread=None,
distributed=False
):
"""
Evaluate a model on a given dataset
:param x: Input data. It could be:
- a TFDataset object
- A Numpy array (or array-like), or a list of arrays
(in case the model has multiple inputs).
- A dict mapping input names to the corresponding array/tensors,
if the model has named inputs.
:param y: Target data. Like the input data `x`,
It should be consistent with `x` (you cannot have Numpy inputs and
tensor targets, or inversely). If `x` is a TFDataset, `y` should
not be specified (since targets will be obtained from `x`).
:param batch_per_thread:
The default value is 1.
When distributed is True,the total batch size is batch_per_thread * rdd.getNumPartitions.
When distributed is False the total batch size is batch_per_thread * numOfCores.
:param distributed: Boolean. Whether to do prediction in distributed mode or local mode.
Default is True. In local mode, x must be a Numpy array.
"""
if isinstance(x, TFDataset):
if not x.has_batch:
raise ValueError("The batch_per_thread of TFDataset must be " +
"specified when used in KerasModel evaluate.")
if isinstance(x, TFNdarrayDataset):
x = _standarize_feature_label_dataset(x, self.model)
# todo check arguments
check_data_compatible(x, self.model, mode="evaluate")
return self._evaluate_distributed(x)
else:
if distributed:
dataset = TFDataset.from_ndarrays((x, y),
batch_per_thread=-1 if batch_per_thread is None
else batch_per_thread
)
dataset = _standarize_feature_label_dataset(dataset, self.model)
return self._evaluate_distributed(dataset)
else:
results = self.model.evaluate(x=x,
y=y,
batch_size=batch_per_thread)
results = dict(zip(self.metrics_names, results))
return results
def _xshards_to_tf_dataset(data_shard,
batch_size=-1,
batch_per_thread=-1,
validation_data_shard=None):
# todo data_shard.head ?
import numpy as np
def check_data_type_and_to_list(data):
result = {}
assert isinstance(data, dict), "each shard should be an dict"
assert "x" in data, "key x should in each shard"
x = data["x"]
if isinstance(x, np.ndarray):
new_x = [x]
elif isinstance(x, tuple) and all(
[isinstance(xi, np.ndarray) for xi in x]):
new_x = x
else:
raise ValueError(
"value of x should be a ndarray or a tuple of ndarrays")
result["x"] = new_x
if "y" in data:
y = data["y"]
if isinstance(y, np.ndarray):
new_y = [y]
elif isinstance(y, tuple) and all(
[isinstance(yi, np.ndarray) for yi in y]):
new_y = y
else:
raise ValueError(
"value of x should be a ndarray or a tuple of ndarrays")
result["y"] = new_y
return result
def get_spec(data):
data = check_data_type_and_to_list(data)
feature_spec = [(feat.dtype, feat.shape[1:]) for feat in data["x"]]
if "y" in data:
label_spec = [(label.dtype, label.shape[1:])
for label in data["y"]]
else:
label_spec = None
return (feature_spec, label_spec)
(feature_spec, label_spec) = data_shard.rdd.map(get_spec).first()
feature_spec = [(tf.dtypes.as_dtype(spec[0]), spec[1])
for spec in feature_spec]
label_spec = [(tf.dtypes.as_dtype(spec[0]), spec[1]) for spec in label_spec] \
if label_spec is not None else None
assert batch_size != -1 or batch_per_thread != -1, \
"one of batch_size and batch_per_thread should be specified"
# todo this might be very slow
def flatten(data):
data = check_data_type_and_to_list(data)
features = data["x"]
has_label = "y" in data
labels = data["y"] if has_label else None
length = features[0].shape[0]
for i in range(length):
fs = [feat[i] for feat in features]
if has_label:
ls = [l[i] for l in labels]
yield (fs, ls)
else:
yield (fs, )
val_rdd = None if validation_data_shard is None \
else validation_data_shard.rdd.flatMap(flatten)
dataset = TFDataset.from_rdd(data_shard.rdd.flatMap(flatten),
features=feature_spec,
labels=label_spec,
batch_size=batch_size,
batch_per_thread=batch_per_thread,
val_rdd=val_rdd)
return dataset
