File post-processing¶

Utilities for processing files generated during training/evaluation.

Exported memories¶

This library provides an implementation for quickly loading exported memory files to inspect their contents in future post-processing steps. See the section Library utilities (skrl.utils module) for example usage.

Usage¶

from skrl.utils import postprocessing


# assuming there is a directory called "memories" with Torch files in it
memory_iterator = postprocessing.MemoryFileIterator("memories/*.pt")
for filename, data in memory_iterator:
    filename: str  # basename of the current file
    data: dict  # keys: tensor names, values: tensors with shape (memory_size, num_envs, data_size)

    # example of simple usage:
    # print the filenames of all memories and their tensor shapes
    print("\nfilename:", filename)
    print("  |-- observations:", data["observations"].shape)
    print("  |-- actions:", data["actions"].shape)
    print("  |-- rewards:", data["rewards"].shape)
    print("  |-- next_observations:", data["next_observations"].shape)
    print("  |-- terminated:", data["terminated"].shape)
    print("  |-- truncated:", data["truncated"].shape)

from skrl.utils import postprocessing


# assuming there is a directory called "memories" with NumPy files in it
memory_iterator = postprocessing.MemoryFileIterator("memories/*.npz")
for filename, data in memory_iterator:
    filename: str  # basename of the current file
    data: dict  # keys: array names, values: arrays with shape (memory_size, num_envs, data_size)

    # example of simple usage:
    # print the filenames of all memories and their array shapes
    print("\nfilename:", filename)
    print("  |-- observations:", data["observations"].shape)
    print("  |-- actions:", data["actions"].shape)
    print("  |-- rewards:", data["rewards"].shape)
    print("  |-- next_observations:", data["next_observations"].shape)
    print("  |-- terminated:", data["terminated"].shape)
    print("  |-- truncated:", data["truncated"].shape)

from skrl.utils import postprocessing


# assuming there is a directory called "memories" with CSV files in it
memory_iterator = postprocessing.MemoryFileIterator("memories/*.csv")
for filename, data in memory_iterator:
    filename: str  # basename of the current file
    data: dict  # keys: list names, values: lists with length (memory_size * num_envs) of sub-lists with length (data_size)

    # example of simple usage:
    # print the filenames of all memories and their list lengths
    print("\nfilename:", filename)
    print("  |-- observations:", data["observations"].shape)
    print("  |-- actions:", data["actions"].shape)
    print("  |-- rewards:", data["rewards"].shape)
    print("  |-- next_observations:", data["next_observations"].shape)
    print("  |-- terminated:", data["terminated"].shape)
    print("  |-- truncated:", data["truncated"].shape)

API¶

class skrl.utils.postprocessing.MemoryFileIterator(pathname: str)[source]¶

Bases: object

Python iterator for loading data from exported memories.

The iterator will load the next memory file in the list of path names. The output of the iterator is a tuple of the filename and the memory data where the memory data is a dictionary of torch.Tensor (PyTorch), numpy.ndarray (NumPy) or lists (CSV) depending on the format and the keys of the dictionary are the names of the variables.

Supported formats:

PyTorch (pt)
NumPy (npz)
Comma-separated values (csv)

Expected output shapes:

PyTorch: (memory_size, num_envs, data_size)
NumPy: (memory_size, num_envs, data_size)
Comma-separated values: (memory_size * num_envs, data_size)

Parameters:: pathname – String containing a path specification for the exported memories. Python glob function is used to find all files matching the path specification.

__iter__() → MemoryFileIterator[source]¶: Return self to make iterable.

__next__() → tuple[str, dict[str, Any]][source]¶

Return next batch.

Returns:: Tuple of file name and data.

Methods:

`_format_csv`()	Load CSV file from file.
`_format_numpy`()	Load NumPy array from file.
`_format_torch`()	Load PyTorch tensor from file.

_format_csv() → tuple[str, dict[str, Any]][source]¶

Load CSV file from file.

Returns:: Tuple of file name and data.

_format_numpy() → tuple[str, dict[str, Any]][source]¶

Load NumPy array from file.

Returns:: Tuple of file name and data.

_format_torch() → tuple[str, dict[str, Any]][source]¶

Load PyTorch tensor from file.

Returns:: Tuple of file name and data.

TensorBoard files¶

This library provides an implementation for quickly loading TensorBoard files to inspect their contents in future post-processing steps. See the section Library utilities (skrl.utils module) for example usage.

Requirements¶

This utility requires the TensorFlow package to be installed to load and parse TensorBoard files:

pip install tensorflow

Usage¶

from skrl.utils import postprocessing


# assuming there is a directory called "runs" with experiments and TensorBoard files in it
tensorboard_iterator = postprocessing.TensorboardFileIterator(
    "runs/*/events.out.tfevents.*", tags=["Reward / Total reward (mean)"]
)
for dirname, data in tensorboard_iterator:
    dirname: str  # path of the directory (experiment name) containing the TensorBoard file
    data: dict  # keys: tags, values: lists of [step, value] pairs

    # example of simple usage:
    # print the directory name and the value length for the "Reward / Total reward (mean)" tag
    print("\ndirname:", dirname)
    for tag, values in data.items():
        print("  |-- tag:", tag)
        print("  |   |-- value length:", len(values))

API¶

class skrl.utils.postprocessing.TensorboardFileIterator(pathname: str, tags: str | list[str])[source]¶

Bases: object

Python iterator for loading data from TensorBoard files.

The iterator will load the next TensorBoard file in the list of path names. The iterator’s output is a tuple of the directory name and the TensorBoard variables selected by the tags. The TensorBoard data is returned as a dictionary with the tag as the key and a list of steps and values as the value.

Parameters:

pathname – String containing a path specification for the TensorBoard files. Python glob function is used to find all files matching the path specification.
tags – String or list of strings containing the tags of the variables to load.

__iter__() → TensorboardFileIterator[source]¶: Return self to make iterable.

__next__() → tuple[str, dict[str, Any]][source]¶

Return next batch.

Returns:: Tuple of directory name and data.