Merge branch 'main' into eddie-changes

TCSo · web-flow · commit df2013b4048f · 2024-04-15T20:33:50.000-07:00
diff --git a/README.md b/README.md
@@ -20,7 +20,7 @@ import fogx as fox
 
 # 🦊 Dataset Creation 
 # from distributed dataset storage 
-dataset = fox.Dataset(load_from = ["/tmp/rtx", "s3://fox_stroage/"])  
+dataset = fox.Dataset(load_from = ["/tmp/rtx", "s3://fox_storage/"])  
 
 # 🦊 Data collection: 
 # create a new trajectory
diff --git a/docs/Reference.md b/docs/Reference.md
@@ -0,0 +1,9 @@
+# API Reference
+
+## Dataset
+::: fog_x.dataset.Dataset
+
+-------
+
+## Episode
+::: fog_x.episode.Episode
diff --git a/docs/index.md b/docs/index.md
@@ -1,17 +1,17 @@
-# Welcome to MkDocs
+# 🦊 Fog-X Documentation
 
-For full documentation visit [mkdocs.org](https://www.mkdocs.org).
+**Fog-X is an efficient and scalable data collection and management framework for robotics learning.**
+Supports datasets from [Open-X-Embodiment](https://robotics-transformer-x.github.io/) and 🤗[HuggingFace](https://huggingface.co/).
+Fog-X considers both speed 🚀 and memory efficiency 📈 with active metadata and lazily-loaded trajectory data. It supports flexible and distributed dataset partitioning. 
 
-## Commands
+## Installation
 
-* `mkdocs new [dir-name]` - Create a new project.
-* `mkdocs serve` - Start the live-reloading docs server.
-* `mkdocs build` - Build the documentation site.
-* `mkdocs -h` - Print help message and exit.
+```bash
+pip install fogx
+```
 
-## Project layout
+## Usage
 
-    mkdocs.yml    # The configuration file.
-    docs/
-        index.md  # The documentation homepage.
-        ...       # Other markdown pages, images and other files.
+See [Usage Guide](./usage.md) for an overview of how to use Fog-X.
+
+You can also view [working examples on GitHub](https://github.com/KeplerC/fog_x/tree/main/examples).
diff --git a/docs/usage.md b/docs/usage.md
@@ -0,0 +1,140 @@
+# Usage Guide
+The code examples below will assume the following import:
+```py
+import fog_x as fox 
+```
+
+## Definitions
+- **episode**: one robot trajectory or action, consisting of multiple step data.
+- **step data**: data representing a snapshot of the robot action at a certain time
+- **metadata**: information that is consistent across a certain episode, e.g. the language instruction associated with the robot action, the name of the person collecting the data, or any other tags/labels.
+
+## The Fog-X Dataset
+To start, create a `Dataset` object. Any data that is collected, loaded, or exported
+will be saved to the provided path.
+There can be existing Fog-X data located at the path as well, so that you can continue
+right where you left off.
+```py
+dataset = fox.Dataset(name="my_fogx_dataset", path="/local/path/my_fogx_dataset")  
+```
+
+## Collecting Robot Data
+
+```py
+# create a new trajectory
+episode = dataset.new_episode()
+
+# run robot and collect data
+while robot_is_running:
+    # at each step, add data to the episode
+    episode.add(feature = "arm_view", value = "image1.jpg")
+
+# Automatically time-aligns and saves the trajectory
+episode.close()
+```
+
+## Exporting Data
+By default, the exported data will be located under the `/export` directory within
+the initialized `dataset.path`.
+Currently, the supported data formats are `rtx`, `open-x`, and `rlds`.
+
+```py
+# Export and share the dataset as standard Open-X-Embodiment format
+dataset.export(desired_episodes, format="rtx")
+```
+
+### PyTorch
+```py
+import torch 
+
+metadata = dataset.get_episode_info()
+metadata = metadata.filter(metadata["feature1"] == "value1")
+pytorch_ds = dataset.pytorch_dataset_builder(metadata=metadata)
+
+# get samples from the dataset
+for data in torch.utils.data.DataLoader(
+    pytorch_ds,
+    batch_size=2,
+    collate_fn=lambda x: x,
+    sampler=torch.utils.data.RandomSampler(pytorch_ds),
+):
+    print(data)
+```
+
+### HuggingFace
+WIP: Currently there is limited support for HuggingFace.
+
+```py
+huggingface_ds = dataset.get_as_huggingface_dataset()
+```
+
+
+## Loading Data from Existing Datasets
+
+### RT-X / Tensorflow Datasets
+Load any RT-X robotics data available at [Tensorflow Datasets](https://www.tensorflow.org/datasets/catalog/).
+You can also find a preview of all the RT-X datasets [here](https://dibyaghosh.com/rtx_viz/).
+
+When loading the episodes, you can optionally specify `additional_metadata` to be associated with it.
+You can also load a specific portion of train or test data with the `split` parameter. See the [Tensorflow Split API](https://www.tensorflow.org/datasets/splits) for specifics.
+
+```py
+# load all berkeley_autolab_ur5 data
+dataset.load_rtx_episodes(name="berkeley_autolab_ur5")
+
+# load 75% of berkeley_autolab_ur5 train data labeled with my_label as train`
+dataset.load_rtx_episodes(
+    name="berkeley_autolab_ur5",
+    split="train[:75%]",
+    additional_metadata={"my_label": "train1"}
+)
+```
+
+## Data Management
+
+### Episode Metadata
+You can retrieve episode-level information (metadata) using `dataset.get_episode_info()`.
+This is a [pandas DataFrame](https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.html),
+meaning you have access to pandas data management methods including `filter`, `map`, `aggregate`, `groupby`, etc.
+After processing the metadata, you can then use the metadata to obtain your
+desired episodes with `dataset.read_by(desired_metadata)`.
+
+```py
+# Retrieve episode-level data as a pandas DataFrame
+episode_info = dataset.get_episode_info()
+
+# Use pandas DataFrame filter to select episodes
+desired_episode_metadata = episode_info.filter(episode_info["natural_language_instruction"] == "open door")
+
+# Obtain the actual episodes containing their step data
+episodes = dataset.read_by(desired_episode_metadata)
+```
+
+### Step Data
+Step data is stored as a [Polars LazyFrame](https://docs.pola.rs/py-polars/html/reference/lazyframe/index.html).
+Lazy loading with Polars results in speedups of 10 to 100 times compared to pandas.
+```py
+# Retrieve Fog-X dataset as a Polars LazyFrame
+step_data = dataset.get_step_data()
+
+# select only the episode_id and natural_language_instruction
+lazy_id_to_language = step_data.select("episode_id", "natural_language_instruction")
+
+# the frame is lazily evaluated at memory when we call collect(). returns a Polars DataFrame
+id_to_language = lazy_id_to_language.collect() 
+
+# drop rows with duplicate natural_language_instruction to see unique instructions
+id_to_language.unique(subset=["natural_language_instruction"], maintain_order=True)
+```
+
+Polars also allows chaining methods:
+```py
+# Same as above example, but chained
+id_to_language = (
+    dataset.get_step_data()
+        .select("episode_id", "natural_language_instruction")
+        .collect()
+        .unique(subset=["natural_language_instruction"], maintain_order=True)
+)
+```
+
diff --git a/fog_x/dataset.py b/fog_x/dataset.py
@@ -4,6 +4,8 @@
 from typing import Any, Dict, List, Optional, Tuple
 
 import numpy as np
+import polars
+import pandas
 
 from fog_x.database import (
     DatabaseConnector,
@@ -36,6 +38,22 @@ def __init__(
         step_data_connector: DatabaseConnector = None,
         storage: Optional[str] = None,
     ) -> None:
+        """
+
+        Args:
+            name (str): Name of this dataset. Used as the directory name when exporting.
+            path (str): Required. Local path of where this dataset should be stored.
+            features (optional Dict[str, FeatureType]): Description of `param1`.
+            enable_feature_inference (bool): enable inferring additional FeatureTypes
+
+        Example:
+        ```
+        >>> dataset = fog_x.Dataset('my_dataset', path='~/fog_x/my_dataset`)
+        ```
+
+        TODO:
+            * is replace_existing actually used anywhere?
+        """
         self.name = name
         path = os.path.expanduser(path)
         self.path = path
@@ -55,23 +73,24 @@ def __init__(
             if not os.path.exists(f"{path}/{name}"):
                 os.makedirs(f"{path}/{name}")
             step_data_connector = LazyFrameConnector(f"{path}/{name}")
-        self.db_manager = DatabaseManager(
-            episode_info_connector, step_data_connector
-        )
+        self.db_manager = DatabaseManager(episode_info_connector, step_data_connector)
         self.db_manager.initialize_dataset(self.name, features)
 
         self.storage = storage
         self.obs_keys = []
         self.act_keys = []
         self.step_keys = []
 
-    def new_episode(
-        self, metadata: Optional[Dict[str, Any]] = None
-    ) -> Episode:
+    def new_episode(self, metadata: Optional[Dict[str, Any]] = None) -> Episode:
         """
         Create a new episode / trajectory.
-        TODO #1: support multiple processes writing to the same episode
-        TODO #2: close the previous episode if not closed
+
+        Returns:
+            Episode
+
+        TODO:
+            * support multiple processes writing to the same episode
+            * close the previous episode if not closed
         """
         return Episode(
             metadata=metadata,
@@ -113,6 +132,10 @@ def export(
     ) -> None:
         """
         Export the dataset.
+
+        Args:
+            export_path (optional str): location of exported data. Uses dataset.path/export by default.
+            format (str): Supported formats are `rtx`, `open-x`, and `rlds`.
         """
         if format == "rtx" or format == "open-x" or format == "rlds":
             self.export_rtx(export_path, max_episodes_per_file, version, obs_keys, act_keys, step_keys)
@@ -274,7 +297,18 @@ def load_rtx_episodes(
         additional_metadata: Optional[Dict[str, Any]] = None,
     ):
         """
-        Load the dataset.
+        Load robot data from Tensorflow Datasets.
+
+        Args:
+            name (str): Name of RT-X episodes, which can be found at [Tensorflow Datasets](https://www.tensorflow.org/datasets/catalog) under the Robotics category
+            split (optional str): the portion of data to load, see [Tensorflow Split API](https://www.tensorflow.org/datasets/splits)
+            additional_metadata (optional Dict[str, Any]): additional metadata to be associated with the loaded episodes
+
+        Example:
+            ```
+            >>> dataset.load_rtx_episodes(name="berkeley_autolab_ur5)
+            >>> dataset.load_rtx_episodes(name="berkeley_autolab_ur5", split="train[:10]", additional_metadata={"data_collector": "Alice", "custom_tag": "sample"})
+            ```
         """
 
         # this is only required if rtx format is used
@@ -334,26 +368,36 @@ def load_rtx_episodes(
                         fog_episode.add(
                             feature=str(k),
                             value=v.numpy(),
-                            feature_type=FeatureType(
-                                tf_feature_spec=data_type[k]
-                            ),
+                            feature_type=FeatureType(tf_feature_spec=data_type[k]),
                         )
                         self.step_keys.append(k)
             fog_episode.close()
 
-    def get_episode_info(self):
+    def get_episode_info(self) -> pandas.DataFrame:
         """
-        Return the metadata as pandas dataframe.
+        Returns:
+            metadata of all episodes as `pandas.DataFrame`
         """
         return self.db_manager.get_episode_info_table()
 
-    def get_step_data(self):
+    def get_step_data(self) -> polars.LazyFrame:
         """
-        Return the all step data as lazy dataframe.
+        Returns:
+            step data of all episodes
         """
         return self.db_manager.get_step_table_all()
 
-    def get_step_data_by_episode_ids(self, episode_ids: List[int], as_lazy_frame = True):
+    def get_step_data_by_episode_ids(
+        self, episode_ids: List[int], as_lazy_frame=True
+    ) -> List[polars.LazyFrame] | List[polars.DataFrame]:
+        """
+        Args:
+            episode_ids (List[int]): list of episode ids
+            as_lazy_frame (bool): whether to return polars.LazyFrame or polars.DataFrame
+
+        Returns:
+            step data of each episode
+        """
         episodes = []
         for episode_id in episode_ids:
             if episode_id == None:
@@ -363,8 +407,17 @@ def get_step_data_by_episode_ids(self, episode_ids: List[int], as_lazy_frame = T
             else:
                 episodes.append(self.db_manager.get_step_table(episode_id).collect())
         return episodes
-    
-    def read_by(self, episode_info: Any = None):
+
+    def read_by(self, episode_info: Any = None) -> List[polars.LazyFrame]:
+        """
+        To be used with `Dataset.get_episode_info`.
+
+        Args:
+            episode_info (pandas.DataFrame): episode metadata information to determine which episodes to read
+
+        Returns:
+            episodes filtered by `episode_info`
+        """
         episode_ids = list(episode_info["episode_id"])
         logger.info(f"Reading episodes as order: {episode_ids}")
         episodes = []
@@ -384,6 +437,11 @@ def get_episodes_from_metadata(self, metadata: Any = None):
         return episodes
 
     def pytorch_dataset_builder(self, metadata=None, **kwargs):
+        """
+        Used for loading current dataset as a PyTorch dataset.
+        To be used with `torch.utils.data.DataLoader`.
+        """
+
         import torch
         from torch.utils.data import Dataset
         episodes = self.get_episodes_from_metadata(metadata)
@@ -394,17 +452,24 @@ def pytorch_dataset_builder(self, metadata=None, **kwargs):
         return pytorch_dataset
 
     def get_as_huggingface_dataset(self):
+        """
+        Load current dataset as a HuggingFace dataset.
+
+        TODO:
+            * currently the support for huggingg face dataset is limited.
+                it only shows its capability of easily returning a hf dataset
+            * add features from the episode metadata
+            * allow selecting episodes based on queries.
+                doing so requires creating a new copy of the dataset on disk
+        """
         import datasets
 
-        # TODO: currently the support for huggingg face dataset is limited 
-        # it only shows its capability of easily returning a hf dataset 
-        # TODO #1: add features from the episode metadata 
-        # TODO #2: allow selecting episodes based on queries 
-        # doing so requires creating a new copy of the dataset on disk 
         dataset_path = self.path + "/" + self.name
-        parquet_files = [os.path.join(dataset_path, f) for f in os.listdir(dataset_path)]
+        parquet_files = [
+            os.path.join(dataset_path, f) for f in os.listdir(dataset_path)
+        ]
 
-        hf_dataset = datasets.load_dataset('parquet', data_files=parquet_files)
+        hf_dataset = datasets.load_dataset("parquet", data_files=parquet_files)
         return hf_dataset
     
 class PyTorchDataset(Dataset):
diff --git a/fog_x/episode.py b/fog_x/episode.py
diff --git a/mkdocs.yml b/mkdocs.yml
