mindspore.dataset.TFRecordDataset
- class mindspore.dataset.TFRecordDataset(dataset_files, schema=None, columns_list=None, num_samples=None, num_parallel_workers=None, shuffle=Shuffle.GLOBAL, num_shards=None, shard_id=None, shard_equal_rows=False, cache=None, compression_type=None)[source]
A source dataset that reads and parses datasets stored on disk in TFData format.
The columns of generated dataset depend on the source TFRecord files.
Note
'TFRecordDataset' is not support on Windows platform yet.
- Parameters
dataset_files (Union[str, list[str]]) – String or list of files to be read or glob strings to search for a pattern of files. The list will be sorted in lexicographical order.
schema (Union[str, Schema], optional) – Data format policy, which specifies the data types and shapes of the data column to be read. Both JSON file path and objects constructed by
mindspore.dataset.Schema
are acceptable. Default:None
.columns_list (list[str], optional) – List of columns to be read. Default:
None
, read all columns.num_samples (int, optional) –
The number of samples (rows) to be included in the dataset. Default:
None
. When num_shards and shard_id are specified, it will be interpreted as number of rows per shard. Processing priority for num_samples is as the following:If specify num_samples with value > 0, read num_samples samples.
If no num_samples and specify numRows(parsed from schema) with value > 0, read numRows samples.
If no num_samples and no schema, read the full dataset.
num_parallel_workers (int, optional) – Number of worker threads to read the data. Default:
None
, will use global default workers(8), it can be set bymindspore.dataset.config.set_num_parallel_workers()
.shuffle (Union[bool, Shuffle], optional) –
Perform reshuffling of the data every epoch. Default:
Shuffle.GLOBAL
. Bool type and Shuffle enum are both supported to pass in. If shuffle isFalse
, no shuffling will be performed. If shuffle isTrue
, perform global shuffle. There are three levels of shuffling, desired shuffle enum defined bymindspore.dataset.Shuffle
.Shuffle.GLOBAL
: Shuffle both the files and samples, same as setting shuffle toTrue
.Shuffle.FILES
: Shuffle files only.
num_shards (int, optional) – Number of shards that the dataset will be divided into. Default:
None
. When this argument is specified, num_samples reflects the maximum sample number per shard. Used in data parallel training .shard_id (int, optional) – The shard ID within num_shards . Default:
None
. This argument can only be specified when num_shards is also specified.shard_equal_rows (bool, optional) – Get equal rows for all shards. Default:
False
. If shard_equal_rows is False, the number of rows of each shard may not be equal, and may lead to a failure in distributed training. When the number of samples per TFRecord file are not equal, it is suggested to set it toTrue
. This argument should only be specified when num_shards is also specified. When compression_type is notNone
, and num_samples or numRows (parsed from schema ) is provided, shard_equal_rows will be implied asTrue
.cache (DatasetCache, optional) – Use tensor caching service to speed up dataset processing. More details: Single-Node Data Cache . Default:
None
, which means no cache is used.compression_type (str, optional) – The type of compression used for all files, must be either
''
,'GZIP'
, or'ZLIB'
. Default:None
, as in empty string. It is highly recommended to provide num_samples or numRows (parsed from schema) when compression_type is"GZIP"
or"ZLIB"
to avoid performance degradation caused by multiple decompressions of the same file to obtain the file size.
- Raises
ValueError – If dataset_files are not valid or do not exist.
ValueError – If num_parallel_workers exceeds the max thread numbers.
RuntimeError – If num_shards is specified but shard_id is None.
RuntimeError – If shard_id is specified but num_shards is None.
ValueError – If shard_id is not in range of [0, num_shards ).
ValueError – If compression_type is not
''
,'GZIP'
or'ZLIB'
.ValueError – If compression_type is provided, but the number of dataset files < num_shards .
ValueError – If num_samples < 0.
Examples
>>> import mindspore.dataset as ds >>> from mindspore import dtype as mstype >>> >>> tfrecord_dataset_dir = ["/path/to/tfrecord_dataset_file"] # contains 1 or multiple TFRecord files >>> tfrecord_schema_file = "/path/to/tfrecord_schema_file" >>> >>> # 1) Get all rows from tfrecord_dataset_dir with no explicit schema. >>> # The meta-data in the first row will be used as a schema. >>> dataset = ds.TFRecordDataset(dataset_files=tfrecord_dataset_dir) >>> >>> # 2) Get all rows from tfrecord_dataset_dir with user-defined schema. >>> schema = ds.Schema() >>> schema.add_column(name='col_1d', de_type=mstype.int64, shape=[2]) >>> dataset = ds.TFRecordDataset(dataset_files=tfrecord_dataset_dir, schema=schema) >>> >>> # 3) Get all rows from tfrecord_dataset_dir with the schema file. >>> dataset = ds.TFRecordDataset(dataset_files=tfrecord_dataset_dir, schema=tfrecord_schema_file)
Pre-processing Operation
Apply a function in this dataset. |
|
Concatenate the dataset objects in the input list. |
|
Filter dataset by prediction. |
|
Map func to each row in dataset and flatten the result. |
|
Apply each operation in operations to this dataset. |
|
The specified columns will be selected from the dataset and passed into the pipeline with the order specified. |
|
Rename the columns in input datasets. |
|
Repeat this dataset count times. |
|
Reset the dataset for next epoch. |
|
Save the dynamic data processed by the dataset pipeline in common dataset format. |
|
Shuffle the dataset by creating a cache with the size of buffer_size . |
|
Skip the first N elements of this dataset. |
|
Split the dataset into smaller, non-overlapping datasets. |
|
Take the first specified number of samples from the dataset. |
|
Zip the datasets in the sense of input tuple of datasets. |
Batch
Combine batch_size number of consecutive rows into batch which apply per_batch_map to the samples first. |
|
Bucket elements according to their lengths. |
|
Combine batch_size number of consecutive rows into batch which apply pad_info to the samples first. |
Iterator
Create an iterator over the dataset that yields samples of type dict, while the key is the column name and the value is the data. |
|
Create an iterator over the dataset that yields samples of type list, whose elements are the data for each column. |
Attribute
Return the size of batch. |
|
Get the mapping dictionary from category names to category indexes. |
|
Return the names of the columns in dataset. |
|
Return the number of batches in an epoch. |
|
Get the replication times in RepeatDataset. |
|
Get the column index, which represents the corresponding relationship between the data column order and the network when using the sink mode. |
|
Get the number of classes in a dataset. |
|
Get the shapes of output data. |
|
Get the types of output data. |
Apply Sampler
Add a child sampler for the current dataset. |
|
Replace the last child sampler of the current dataset, remaining the parent sampler unchanged. |
Others
Release a blocking condition and trigger callback with given data. |
|
Add a blocking condition to the input Dataset and a synchronize action will be applied. |
|
Serialize a pipeline into JSON string and dump into file if filename is provided. |