Skip to content

Commit

Permalink
update prototype dataset README (#5777)
Browse files Browse the repository at this point in the history 
* update prototype dataset README

* fix header level

* Apply suggestions from code review

Co-authored-by: Nicolas Hug <contact@nicolas-hug.com>

Co-authored-by: Nicolas Hug <contact@nicolas-hug.com>
  • Loading branch information
@pmeier @NicolasHug
pmeier and NicolasHug authored Apr 7, 2022
1 parent 4c73a5e commit cd36d06
Showing 1 changed file with 169 additions and 66 deletions.
235 changes: 169 additions & 66 deletions torchvision/prototype/datasets/_builtin/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -12,51 +12,66 @@ Finally, `from torchvision.prototype import datasets` is implied below.

Before we start with the actual implementation, you should create a module in `torchvision/prototype/datasets/_builtin`
that hints at the dataset you are going to add. For example `caltech.py` for `caltech101` and `caltech256`. In that
module create a class that inherits from `datasets.utils.Dataset` and overwrites at minimum three methods that will be
discussed in detail below:
module create a class that inherits from `datasets.utils.Dataset` and overwrites four methods that will be discussed in
detail below:

```python
from typing import Any, Dict, List
import pathlib
from typing import Any, BinaryIO, Dict, List, Tuple, Union

from torchdata.datapipes.iter import IterDataPipe
from torchvision.prototype.datasets.utils import Dataset, DatasetInfo, DatasetConfig, OnlineResource
from torchvision.prototype.datasets.utils import Dataset, OnlineResource

from .._api import register_dataset, register_info

NAME = "my-dataset"

@register_info(NAME)
def _info() -> Dict[str, Any]:
return dict(
...
)

@register_dataset(NAME)
class MyDataset(Dataset):
def _make_info(self) -> DatasetInfo:
def __init__(self, root: Union[str, pathlib.Path], *, ..., skip_integrity_check: bool = False) -> None:
...
super().__init__(root, skip_integrity_check=skip_integrity_check)

def resources(self, config: DatasetConfig) -> List[OnlineResource]:
def _resources(self) -> List[OnlineResource]:
...

def _make_datapipe(
self, resource_dps: List[IterDataPipe], *, config: DatasetConfig,
) -> IterDataPipe[Dict[str, Any]]:
def _datapipe(self, resource_dps: List[IterDataPipe[Tuple[str, BinaryIO]]]) -> IterDataPipe[Dict[str, Any]]:
...

def __len__(self) -> int:
...
```

### `_make_info(self)`
In addition to the dataset, you also need to implement an `_info()` function that takes no arguments and returns a
dictionary of static information. The most common use case is to provide human-readable categories.
[See below](#how-do-i-handle-a-dataset-that-defines-many-categories) how to handle cases with many categories.

The `DatasetInfo` carries static information about the dataset. There are two required fields:
Finally, both the dataset class and the info function need to be registered on the API with the respective decorators.
With that they are loadable through `datasets.load("my-dataset")` and `datasets.info("my-dataset")`, respectively.

- `name`: Name of the dataset. This will be used to load the dataset with `datasets.load(name)`. Should only contain
lowercase characters.
### `__init__(self, root, *, ..., skip_integrity_check = False)`

There are more optional parameters that can be passed:
Constructor of the dataset that will be called when the dataset is instantiated. In addition to the parameters of the
base class, it can take arbitrary keyword-only parameters with defaults. The checking of these parameters as well as
setting them as instance attributes has to happen before the call of `super().__init__(...)`, because that will invoke
the other methods, which possibly depend on the parameters. All instance attributes must be private, i.e. prefixed with
an underscore.

- `dependencies`: Collection of third-party dependencies that are needed to load the dataset, e.g. `("scipy",)`. Their
availability will be automatically checked if a user tries to load the dataset. Within the implementation, import
these packages lazily to avoid missing dependencies at import time.
- `categories`: Sequence of human-readable category names for each label. The index of each category has to match the
corresponding label returned in the dataset samples.
[See below](#how-do-i-handle-a-dataset-that-defines-many-categories) how to handle cases with many categories.
- `valid_options`: Configures valid options that can be passed to the dataset. It should be `Dict[str, Sequence[Any]]`.
The options are accessible through the `config` namespace in the other two functions. First value of the sequence is
taken as default if the user passes no option to `torchvision.prototype.datasets.load()`.
If the implementation of the dataset depends on third-party packages, pass them as a collection of strings to the base
class constructor, e.g. `super().__init__(..., dependencies=("scipy",))`. Their availability will be automatically
checked if a user tries to load the dataset. Within the implementation of the dataset, import these packages lazily to
avoid missing dependencies at import time.

## `resources(self, config)`
### `_resources(self)`

Returns `List[datasets.utils.OnlineResource]` of all the files that need to be present locally before the dataset with a
specific `config` can be build. The download will happen automatically.
Returns `List[datasets.utils.OnlineResource]` of all the files that need to be present locally before the dataset can be
build. The download will happen automatically.

Currently, the following `OnlineResource`'s are supported:

Expand All @@ -81,7 +96,7 @@ def sha256sum(path, chunk_size=1024 * 1024):
print(checksum.hexdigest())
```

### `_make_datapipe(resource_dps, *, config)`
### `_datapipe(self, resource_dps)`

This method is the heart of the dataset, where we transform the raw data into a usable form. A major difference compared
to the current stable datasets is that everything is performed through `IterDataPipe`'s. From the perspective of someone
Expand All @@ -99,60 +114,112 @@ All of them can be imported `from torchdata.datapipes.iter`. In addition, use `f
needs extra arguments. If the provided `IterDataPipe`'s are not sufficient for the use case, it is also not complicated
to add one. See the MNIST or CelebA datasets for example.

`make_datapipe()` receives `resource_dps`, which is a list of datapipes that has a 1-to-1 correspondence with the return
value of `resources()`. In case of archives with regular suffixes (`.tar`, `.zip`, ...), the datapipe will contain
tuples comprised of the path and the handle for every file in the archive. Otherwise the datapipe will only contain one
`_datapipe()` receives `resource_dps`, which is a list of datapipes that has a 1-to-1 correspondence with the return
value of `_resources()`. In case of archives with regular suffixes (`.tar`, `.zip`, ...), the datapipe will contain
tuples comprised of the path and the handle for every file in the archive. Otherwise, the datapipe will only contain one
of such tuples for the file specified by the resource.

Since the datapipes are iterable in nature, some datapipes feature an in-memory buffer, e.g. `IterKeyZipper` and
`Grouper`. There are two issues with that: 1. If not used carefully, this can easily overflow the host memory, since
most datasets will not fit in completely. 2. This can lead to unnecessarily long warm-up times when data is buffered
that is only needed at runtime.
`Grouper`. There are two issues with that:

1. If not used carefully, this can easily overflow the host memory, since most datasets will not fit in completely.
2. This can lead to unnecessarily long warm-up times when data is buffered that is only needed at runtime.

Thus, all buffered datapipes should be used as early as possible, e.g. zipping two datapipes of file handles rather than
trying to zip already loaded images.

There are two special datapipes that are not used through their class, but through the functions `hint_shuffling` and
`hint_sharding`. As the name implies they only hint part in the datapipe graph where shuffling and sharding should take
place, but are no-ops by default. They can be imported from `torchvision.prototype.datasets.utils._internal` and are
required in each dataset. `hint_shuffling` has to be placed before `hint_sharding`.
`hint_sharding`. As the name implies they only hint at a location in the datapipe graph where shuffling and sharding
should take place, but are no-ops by default. They can be imported from `torchvision.prototype.datasets.utils._internal`
and are required in each dataset. `hint_shuffling` has to be placed before `hint_sharding`.

Finally, each item in the final datapipe should be a dictionary with `str` keys. There is no standardization of the
names (yet!).

### `__len__`

This returns an integer denoting the number of samples that can be drawn from the dataset. Please use
[underscores](https://peps.python.org/pep-0515/) after every three digits starting from the right to enhance the
readability. For example, `1_281_167` vs. `1281167`.

If there are only two different numbers, a simple `if` / `else` is fine:

```py
def __len__(self):
return 12_345 if self._split == "train" else 6_789
```

If there are more options, using a dictionary usually is the most readable option:

```py
def __len__(self):
return {
"train": 3,
"val": 2,
"test": 1,
}[self._split]
```

If the number of samples depends on more than one parameter, you can use tuples as dictionary keys:

```py
def __len__(self):
return {
("train", "bar"): 4,
("train", "baz"): 3,
("test", "bar"): 2,
("test", "baz"): 1,
}[(self._split, self._foo)]
```

The length of the datapipe is only an annotation for subsequent processing of the datapipe and not needed during the
development process. Since it is an `@abstractmethod` you still have to implement it from the start. The canonical way
is to define a dummy method like

```py
def __len__(self):
return 1
```

and only fill it with the correct data if the implementation is otherwise finished.
[See below](#how-do-i-compute-the-number-of-samples) for a possible way to compute the number of samples.

## Tests

To test the dataset implementation, you usually don't need to add any tests, but need to provide a mock-up of the data.
This mock-up should resemble the original data as close as necessary, while containing only few examples.

To do this, add a new function in [`test/builtin_dataset_mocks.py`](../../../../test/builtin_dataset_mocks.py) with the
same name as you have defined in `_make_config()` (if the name includes hyphens `-`, replace them with underscores `_`)
and decorate it with `@register_mock`:
same name as you have used in `@register_info` and `@register_dataset`. This function is called "mock data function".
Decorate it with `@register_mock(configs=[dict(...), ...])`. Each dictionary denotes one configuration that the dataset
will be loaded with, e.g. `datasets.load("my-dataset", **config)`. For the most common case of a product of all options,
you can use the `combinations_grid()` helper function, e.g.
`configs=combinations_grid(split=("train", "test"), foo=("bar", "baz"))`.

In case the name of the dataset includes hyphens `-`, replace them with underscores `_` in the function name and pass
the `name` parameter to `@register_mock`

```py
# this is defined in torchvision/prototype/datasets/_builtin
@register_dataset("my-dataset")
class MyDataset(Dataset):
def _make_info(self) -> DatasetInfo:
return DatasetInfo(
"my-dataset",
...
)

@register_mock
def my_dataset(info, root, config):
...

@register_mock(name="my-dataset", configs=...)
def my_dataset(root, config):
...
```

The function receives three arguments:
The mock data function receives two arguments:

- `info`: The return value of `_make_info()`.
- `root`: A [`pathlib.Path`](https://docs.python.org/3/library/pathlib.html#pathlib.Path) of a folder, in which the data
needs to be placed.
- `config`: The configuration to generate the data for. This is the same value that `_make_datapipe()` receives.
- `config`: The configuration to generate the data for. This is one of the dictionaries defined in
`@register_mock(configs=...)`

The function should generate all files that are needed for the current `config`. Each file should be complete, e.g. if
the dataset only has a single archive that contains multiple splits, you need to generate all regardless of the current
`config`. Although this seems odd at first, this is important. Consider the following original data setup:
the dataset only has a single archive that contains multiple splits, you need to generate the full archive regardless of
the current `config`. Although this seems odd at first, this is important. Consider the following original data setup:

```
root
Expand All @@ -167,9 +234,8 @@ root
For map-style datasets (like the one currently in `torchvision.datasets`), one explicitly selects the files they want to
load. For example, something like `(root / split).iterdir()` works fine even if only the specific split folder is
present. With iterable-style datasets though, we get something like `root.iterdir()` from `resource_dps` in
`_make_datapipe()` and need to manually `Filter` it to only keep the files we want. If we would only generate the data
for the current `config`, the test would also pass if the dataset is missing the filtering, but would fail on the real
data.
`_datapipe()` and need to manually `Filter` it to only keep the files we want. If we would only generate the data for
the current `config`, the test would also pass if the dataset is missing the filtering, but would fail on the real data.

For datasets that are ported from the old API, we already have some mock data in
[`test/test_datasets.py`](../../../../test/test_datasets.py). You can find the test case corresponding test case there
Expand All @@ -178,8 +244,6 @@ and have a look at the `inject_fake_data` function. There are a few differences
- `tmp_dir` corresponds to `root`, but is a `str` rather than a
[`pathlib.Path`](https://docs.python.org/3/library/pathlib.html#pathlib.Path). Thus, you often see something like
`folder = pathlib.Path(tmp_dir)`. This is not needed.
- Although both parameters are called `config`, the value in the new tests is a namespace. Thus, please use `config.foo`
over `config["foo"]` to enhance readability.
- The data generated by `inject_fake_data` was supposed to be in an extracted state. This is no longer the case for the
new mock-ups. Thus, you need to use helper functions like `make_zip` or `make_tar` to actually generate the files
specified in the dataset.
Expand All @@ -196,17 +260,17 @@ Finally, you can run the tests with `pytest test/test_prototype_builtin_datasets

### How do I start?

Get the skeleton of your dataset class ready with all 3 methods. For `_make_datapipe()`, you can just do
Get the skeleton of your dataset class ready with all 4 methods. For `_datapipe()`, you can just do
`return resources_dp[0]` to get started. Then import the dataset class in
`torchvision/prototype/datasets/_builtin/__init__.py`: this will automatically register the dataset and it will be
`torchvision/prototype/datasets/_builtin/__init__.py`: this will automatically register the dataset, and it will be
instantiable via `datasets.load("mydataset")`. On a separate script, try something like

```py
from torchvision.prototype import datasets

dataset = datasets.load("mydataset")
for sample in dataset:
print(sample) # this is the content of an item in datapipe returned by _make_datapipe()
print(sample) # this is the content of an item in datapipe returned by _datapipe()
break
# Or you can also inspect the sample in a debugger
```
Expand All @@ -217,15 +281,24 @@ datapipes and return the appropriate dictionary format.

### How do I handle a dataset that defines many categories?

As a rule of thumb, `datasets.utils.DatasetInfo(..., categories=)` should only be set directly for ten categories or
fewer. If more categories are needed, you can add a `$NAME.categories` file to the `_builtin` folder in which each line
specifies a category. If `$NAME` matches the name of the dataset (which it definitively should!) it will be
automatically loaded if `categories=` is not set.
As a rule of thumb, `categories` in the info dictionary should only be set manually for ten categories or fewer. If more
categories are needed, you can add a `$NAME.categories` file to the `_builtin` folder in which each line specifies a
category. To load such a file, use the `from torchvision.prototype.datasets.utils._internal import read_categories_file`
function and pass it `$NAME`.

In case the categories can be generated from the dataset files, e.g. the dataset follows an image folder approach where
each folder denotes the name of the category, the dataset can overwrite the `_generate_categories` method. It gets
passed the `root` path to the resources, but they have to be manually loaded, e.g.
`self.resources(config)[0].load(root)`. The method should return a sequence of strings representing the category names.
each folder denotes the name of the category, the dataset can overwrite the `_generate_categories` method. The method
should return a sequence of strings representing the category names. In the method body, you'll have to manually load
the resources, e.g.

```py
resources = self._resources()
dp = resources[0].load(self._root)
```

Note that it is not necessary here to keep a datapipe until the final step. Stick with datapipes as long as it makes
sense and afterwards materialize the data with `next(iter(dp))` or `list(dp)` and proceed with that.

To generate the `$NAME.categories` file, run `python -m torchvision.prototype.datasets.generate_category_files $NAME`.

### What if a resource file forms an I/O bottleneck?
Expand All @@ -235,3 +308,33 @@ the performance hit becomes significant, the archives can still be preprocessed.
`preprocess` parameter that can be a `Callable[[pathlib.Path], pathlib.Path]` where the input points to the file to be
preprocessed and the return value should be the result of the preprocessing to load. For convenience, `preprocess` also
accepts `"decompress"` and `"extract"` to handle these common scenarios.

### How do I compute the number of samples?

Unless the authors of the dataset published the exact numbers (even in this case we should check), there is no other way
than to iterate over the dataset and count the number of samples:

```py
import itertools
from torchvision.prototype import datasets


def combinations_grid(**kwargs):
return [dict(zip(kwargs.keys(), values)) for values in itertools.product(*kwargs.values())]


# If you have implemented the mock data function for the dataset tests, you can simply copy-paste from there
configs = combinations_grid(split=("train", "test"), foo=("bar", "baz"))

for config in configs:
dataset = datasets.load("my-dataset", **config)

num_samples = 0
for _ in dataset:
num_samples += 1

print(", ".join(f"{key}={value}" for key, value in config.items()), num_samples)
```

To speed this up, it is useful to temporarily comment out all unnecessary I/O, such as loading of images or annotation
files.

0 comments on commit cd36d06

Please sign in to comment.