geowatch.tasks.fusion.datamodules.kwcoco_dataset module

Defines a torch Dataset for kwcoco video data.

The parameters to each are handled by scriptconfig objects, which prevents us from needing to specify what the available options are in multiple places.

# import liberator # lib = liberator.Liberator() # lib.add_dynamic(KWCocoVideoDataset) # lib.expand([‘geowatch’]) # print(lib.current_sourcecode())

For notes on Spaces, see

~/code/geowatch/docs/development/coding_conventions.rst

CommandLine

xdoctest -m geowatch.tasks.fusion.datamodules.kwcoco_dataset __doc__:0 --show
xdoctest -m geowatch.tasks.fusion.datamodules.kwcoco_dataset __doc__:1 --show

Example

>>> # Basic Data Sampling
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import ndsampler
>>> import kwcoco
>>> import geowatch
>>> coco_dset = geowatch.coerce_kwcoco('vidshapes1', num_frames=10)
>>> sampler = ndsampler.CocoSampler(coco_dset)
>>> self = KWCocoVideoDataset(sampler, time_dims=4, window_dims=(300, 300),
>>>                           channels='r|g|b')
>>> self.disable_augmenter = True
>>> index = self.sample_grid['targets'][self.sample_grid['positives_indexes'][0]]
>>> item = self[index]
>>> # Summarize batch item in text
>>> summary = self.summarize_item(item)
>>> print('item summary: ' + ub.urepr(summary, nl=2))
>>> # Draw batch item
>>> canvas = self.draw_item(item)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(canvas)
>>> kwplot.show_if_requested()

Example

>>> # Demo toy data without augmentation
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import kwcoco
>>> coco_dset = kwcoco.CocoDataset.demo('vidshapes2-multispectral', num_frames=10)
>>> channels = 'B10,B8a|B1,B8'
>>> self = KWCocoVideoDataset(coco_dset, time_dims=4, window_dims=(300, 300),
>>>                           channels=channels,
>>>                           input_space_scale='native',
>>>                           output_space_scale=None,
>>>                           window_space_scale=1.2,
>>>                           augment_space_shift_rate=0.5,
>>>                           use_grid_negatives=False,
>>>                           use_grid_positives=False,
>>>                           use_centered_positives=True,
>>>                           absolute_weighting=True,
>>>                           time_sampling='uniform',
>>>                           time_kernel='-1year,0,1month,1year',
>>>                           modality_dropout=0.5,
>>>                           channel_dropout=0.5,
>>>                           temporal_dropout=0.7,
>>>                           temporal_dropout_rate=1.0)
>>> # Add weights to annots
>>> annots = self.sampler.dset.annots()
>>> annots.set('weight', 2 + np.random.rand(len(annots)) * 10)
>>> self.disable_augmenter = False
>>> index = self.sample_grid['targets'][self.sample_grid['positives_indexes'][3]]
>>> item = self[index]
>>> summary = self.summarize_item(item)
>>> print('item summary: ' + ub.urepr(summary, nl=3))
>>> canvas = self.draw_item(item, overlay_on_image=0, rescale=0, max_dim=1024)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(canvas)
>>> kwplot.show_if_requested()

Example

>>> # Demo toy data with augmentation
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import kwcoco
>>> coco_dset = kwcoco.CocoDataset.demo('vidshapes2-multispectral', num_frames=10)
>>> channels = 'B10,B8a|B1,B8'
>>> self = KWCocoVideoDataset(coco_dset, time_dims=3, window_dims=(300, 300),
>>>                           channels=channels,
>>>                           input_space_scale='native',
>>>                           output_space_scale=None,
>>>                           window_space_scale=1.2,
>>>                           time_sampling='soft2+distribute',
>>>                           time_kernel='-1y,0,1y',
>>>                           modality_dropout=0.5,
>>>                           temporal_dropout=0.5)
>>> assert not self.disable_augmenter
>>> index = self.sample_grid['targets'][self.sample_grid['positives_indexes'][3]]
>>> item = self[index]
>>> assert item['target']['allow_augment']
>>> print('item summary: ' + ub.urepr(self.summarize_item(item), nl=3))
>>> canvas = self.draw_item(item, overlay_on_image=0, rescale=0)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(canvas)
>>> kwplot.show_if_requested()

Known Issues

• [ ] FIXME: sensorchan codes should exclude non-specified sensors immediately before temporal sampling. Currently temporal sampling is given everything. E.g. (L8,S2):red|green|blue should not allow WV to be included in sampling.

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.KWCocoVideoDatasetConfig(*args, **kwargs)

Bases: DataConfig

This is the configuration for a single dataset that could be used for train, test, or validation.

In the future this might be convertable to, or handled by omegaconfig

The core spacetime parameters are:

• window_space_scale

• input_space_scale

• output_space_scale

• time_steps

• time_sampling

• chip_dims / window_space_dims

Valid options: []

Parameters:

• *args – positional arguments for this data config

• **kwargs – keyword arguments for this data config

default = {'absolute_weighting': <Value(False)>, 'augment_space_rot': <Value(True)>, 'augment_space_shift_rate': <Value(0.9)>, 'augment_space_xflip': <Value(True)>, 'augment_space_yflip': <Value(True)>, 'augment_time_resample_rate': <Value(0.8)>, 'balance_areas': <Value(False)>, 'balance_options': <Value(None)>, 'channel_dropout': <Value(0.0)>, 'channel_dropout_rate': <Value(0.0)>, 'channels': <Value(None)>, 'chip_dims': <Value(128)>, 'chip_overlap': <Value(0.0)>, 'default_class_behavior': <Value('background')>, 'dist_weights': <Value(0)>, 'downweight_nan_regions': <Value(True)>, 'dynamic_fixed_resolution': <Value(None)>, 'exclude_sensors': <Value(None)>, 'failed_sample_policy': <Value('warn')>, 'fixed_resolution': <Value(None)>, 'force_bad_frames': <Value(False)>, 'ignore_dilate': <Value(0)>, 'include_sensors': <Value(None)>, 'input_space_scale': <Value(None)>, 'mask_low_quality': <Value(False)>, 'mask_nan_bands': <Value('')>, 'mask_samecolor_bands': <Value('red')>, 'mask_samecolor_method': <Value(None)>, 'mask_samecolor_values': <Value(0)>, 'max_epoch_length': <Value(None)>, 'min_spacetime_weight': <Value(0.9)>, 'modality_dropout': <Value(0.0)>, 'modality_dropout_rate': <Value(0.0)>, 'neg_to_pos_ratio': <Value(1.0)>, 'normalize_perframe': <Value(False)>, 'normalize_peritem': <Value(None)>, 'num_balance_trees': <Value(16)>, 'observable_threshold': <Value(0.0)>, 'output_space_scale': <Value(None)>, 'prenormalize_inputs': <Value(None)>, 'quality_threshold': <Value(0.0)>, 'resample_invalid_frames': <Value(3)>, 'reseed_fit_random_generators': <Value(True)>, 'sampler_backend': <Value(None)>, 'sampler_workdir': <Value(None)>, 'sampler_workers': <Value('avail/2')>, 'select_images': <Value(None)>, 'select_videos': <Value(None)>, 'set_cover_algo': <Value(None)>, 'temporal_dropout': <Value(0.0)>, 'temporal_dropout_rate': <Value(1.0)>, 'time_kernel': <Value(None)>, 'time_sampling': <Value('contiguous')>, 'time_span': <Value(None)>, 'time_steps': <Value(2)>, 'upweight_centers': <Value(True)>, 'upweight_time': <Value(None)>, 'use_centered_positives': <Value(False)>, 'use_cloudmask': <Value(None)>, 'use_grid_cache': <Value(True)>, 'use_grid_negatives': <Value(True)>, 'use_grid_positives': <Value(True)>, 'use_grid_valid_regions': <Value(True)>, 'weight_dilate': <Value(0)>, 'window_space_scale': <Value(None)>}

normalize()

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.TruthMixin

Bases: object

Methods related to drawing truth rasters / training objectives

ComamndLine:

LINE_PROFILE=1 xdoctest -m geowatch.tasks.fusion.datamodules.kwcoco_dataset TruthMixin:0 –bench

Example

>>> # xdoctest: +REQUIRES(--bench)
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import KWCocoVideoDataset
>>> import ndsampler
>>> import geowatch
>>> coco_dset = geowatch.coerce_kwcoco('vidshapes2', num_frames=10)
>>> sampler = ndsampler.CocoSampler(coco_dset)
>>> self = KWCocoVideoDataset(sampler, mode="fit", time_dims=4, window_dims=(196, 196),
>>>                           channels='r|g|b', neg_to_pos_ratio=0)
>>> for index in ub.ProgIter(range(1000)):
>>>     self.getitem(index)

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.GetItemMixin

Bases: TruthMixin

This mixin defines what is needed for the getitem method.

getitem(index)

This is just the same thing as __getitem__ but it raises an error when it fails, which is handled by __getitem__.

Parameters:

index (int | Dict) – index or target

Returns:

Dict

CommandLine

LINE_PROFILE=1 xdoctest -m geowatch.tasks.fusion.datamodules.kwcoco_dataset GetItemMixin.getitem

Example

>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import kwcoco
>>> import geowatch
>>> coco_dset = geowatch.coerce_kwcoco('geowatch-msi-dates-geodata-heatmap', num_frames=5, image_size=(128, 128), num_videos=1)
>>> # Remove two annotations to test new time weights
>>> aids = coco_dset.images().take([0]).annots[0].lookup('id')
>>> coco_dset.remove_annotations(aids)
>>> #
>>> # Each sensor uses all of its own channels
>>> channels = 'auto'
>>> self = KWCocoVideoDataset(coco_dset, time_dims=5,
>>>                           window_resolution='0.09GSD',
>>>                           input_resolution='0.09GSD',
>>>                           window_dims=(128, 128),
>>>                           channels=channels,
>>>                           balance_areas=True,
>>>                           weight_dilate=3,
>>>                           normalize_perframe=False)
>>> self.disable_augmenter = True
>>> # Pretend that some external object has given us information about desired class weights
>>> from geowatch.tasks.fusion.methods import watch_module_mixins
>>> dataset_stats = self.cached_dataset_stats()
>>> from geowatch.tasks.fusion.methods.network_modules import _class_weights_from_freq
>>> class_keys = dataset_stats['class_freq']
>>> total_freq = np.array(list(dataset_stats['class_freq'].values()))
>>> class_importance_weights = _class_weights_from_freq(total_freq)
>>> catname_to_weight = ub.dzip(class_keys, class_importance_weights)
>>> catname_to_weight['star'] = 2.0
>>> self.catname_to_weight = catname_to_weight
>>> #
>>> index = 0
>>> index = target = self.sample_grid['targets'][self.sample_grid['positives_indexes'][4]]
>>> item = self[index]
>>> # xdoctest: +REQUIRES(--show)
>>> canvas = self.draw_item(item)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(canvas)
>>> kwplot.show_if_requested()

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.IntrospectMixin

Bases: object

Methods for introspection / visualization of data

draw_item(item, item_output=None, combinable_extra=None, max_channels=5, max_dim=224, norm_over_time='auto', overlay_on_image=False, draw_weights=True, rescale='auto', classes=None, show_summary_text=True, **kwargs)

Visualize an item produced by this DataSet.

Each channel will be a row, and each column will be a timestep.

Parameters:

• item (Dict) – An item returned from the torch Dataset.

• overlay_on_image (bool) – if True, the truth and prediction is drawn on top of an image, otherwise it is drawn on a black image.

• max_dim (int) – max dimension to resize each grid cell to.

• max_channels (int) – maximum number of channel rows to draw

• item_output (Dict) – Special task keys that we know how to plot. These should be some sort of binary or class prediction from the network. I’m not sure how best to pass the details of how they should be interpreted.

  Known keys:

  change_probs saliency_probs class_probs pred_ltrb

• classes (kwcoco.CategoryTree | None) – Classes any “class_probs” in the ‘item_output’ dictionary corresponds to. If unspecified uses the classes from the datamodule.

• show_summary_text (bool) – if True, draw additional summary debug information. Defaults to True.

• **kwargs – additional arguments to BatchVisualizationBuilder.

Note

The self.requested_tasks controls the task labels returned by getitem, and hence what can be visualized here.

Example

>>> # Basic Data Sampling with lots of small objects
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import geowatch
>>> anchors = np.array([[0.1, 0.1]])
>>> size = (96, 96)
>>> coco_dset = geowatch.coerce_kwcoco('vidshapes1', num_frames=4, num_tracks=40, anchors=anchors, image_size=size)
>>> self = KWCocoVideoDataset(coco_dset, time_dims=4, window_dims=size, default_class_behavior='ignore')
>>> self._notify_about_tasks(predictable_classes=['star', 'eff'])
>>> self.requested_tasks['change'] = False
>>> index = self.sample_grid['targets'][self.sample_grid['positives_indexes'][0]]
>>> item = self[index]
>>> canvas = self.draw_item(item, draw_weights=False)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> label_to_color = {
>>>     node: data['color']
>>>     for node, data in self.predictable_classes.graph.nodes.items()}
>>> label_to_color = ub.sorted_keys(label_to_color)
>>> legend_img = utils._memo_legend(label_to_color)
>>> legend_img = kwimage.imresize(legend_img, scale=4.0)
>>> show_canvas = kwimage.stack_images([canvas, legend_img], axis=1)
>>> kwplot.imshow(show_canvas)
>>> kwplot.show_if_requested()

Example

>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import kwcoco
>>> import kwarray
>>> import rich
>>> coco_dset = kwcoco.CocoDataset.demo('vidshapes2-multispectral', num_frames=5)
>>> channels = 'B10|B8a|B1|B8|B11'
>>> combinable_extra = [['B10', 'B8', 'B8a']]  # special behavior
>>> # combinable_extra = None  # uncomment for raw behavior
>>> mode = 'fit'
>>> mode = 'test'
>>> coco_dset.clear_annotations()
>>> self = KWCocoVideoDataset(coco_dset, mode=mode, time_dims=5, window_dims=(530, 610), channels=channels, balance_areas=True)
>>> #index = len(self) // 4
>>> #index = self.sample_grid['targets'][self.sample_grid['positives_indexes'][5]]
>>> index = self.sample_grid['targets'][0]
>>> # More controlled settings for debug
>>> self.disable_augmenter = True
>>> item = self[index]
>>> item_output = self._build_demo_outputs(item)
>>> rich.print('item summary: ' + ub.urepr(self.summarize_item(item), nl=3))
>>> canvas = self.draw_item(item, item_output, combinable_extra=combinable_extra, overlay_on_image=1)
>>> canvas2 = self.draw_item(item, item_output, combinable_extra=combinable_extra, max_channels=3, overlay_on_image=0)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(canvas, fnum=1, pnum=(1, 2, 1))
>>> kwplot.imshow(canvas2, fnum=1, pnum=(1, 2, 2))
>>> kwplot.show_if_requested()

summarize_item(item, stats=False)

Return debugging stats about the item

Parameters:

• item (dict) – an item returned by __getitem__

• stats (bool) – if True, include statistics on input datas.

Returns:

a summary of the item

Return type:

dict

Example

>>> # xdoctest: +SKIP
>>> from geowatch.tasks.fusion.datamodules import kwcoco_dataset
>>> import kwcoco
>>> import geowatch
>>> coco_dset = geowatch.coerce_kwcoco('vidshapes1', num_frames=10)
>>> self = kwcoco_dataset.KWCocoVideoDataset(
>>>     coco_dset, time_dims=4, window_dims=(300, 300),
>>>     channels='r|g|b')
>>> self.disable_augmenter = True
>>> index = self.sample_grid['targets'][self.sample_grid['positives_indexes'][0]]
>>> item = self[index]
>>> item_summary = self.summarize_item(item, stats=True)
>>> print(f'item_summary = {ub.urepr(item_summary, nl=-1)}')

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.BalanceMixin

Bases: object

Helpers to build the sample grid and balance it

CommandLine

LINE_PROFILE=1 xdoctest -m geowatch.tasks.fusion.datamodules.kwcoco_dataset BalanceMixin:1 --bench

Example

>>> # Test the legacy neg_to_pos_ratio setting (todo: use more general balance_options)
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import KWCocoVideoDataset
>>> import ndsampler
>>> import geowatch
>>> coco_dset = geowatch.coerce_kwcoco('vidshapes2', num_frames=10)
>>> sampler = ndsampler.CocoSampler(coco_dset)
>>> num_samples = 50
>>> neg_to_pos_ratio = 0
>>> self = KWCocoVideoDataset(sampler, mode="fit", time_dims=4, window_dims=(300, 300),
>>>                           channels='r|g|b', neg_to_pos_ratio=neg_to_pos_ratio)
>>> num_targets = len(self.sample_grid['targets'])
>>> positives_indexes = self.sample_grid['positives_indexes']
>>> negatives_indexes = self.sample_grid['negatives_indexes']
>>> print('dataset positive ratio:', len(positives_indexes) / num_targets)
>>> print('dataset negative ratio:', len(negatives_indexes) / num_targets)
>>> print('specified neg_to_pos_ratio:', neg_to_pos_ratio)
>>> sampled_indexes = [self[x]['resolved_index'] for x in range(num_samples)]
>>> num_positives = sum([x in positives_indexes for x in sampled_indexes])
>>> num_negatives = num_samples - num_positives
>>> print('sampled positive ratio:', num_positives / num_samples)
>>> print('sampled negative ratio:', num_negatives / num_samples)
>>> assert all([x in positives_indexes for x in sampled_indexes])
>>> assert num_negatives == 0
>>> assert num_positives > num_negatives
>>> #...
>>> neg_to_pos_ratio = .1
>>> self = KWCocoVideoDataset(sampler, time_dims=4, window_dims=(300, 300),
>>>                           channels='r|g|b', neg_to_pos_ratio=neg_to_pos_ratio)
>>> num_targets = len(self.sample_grid['targets'])
>>> positives_indexes = self.sample_grid['positives_indexes']
>>> negatives_indexes = self.sample_grid['negatives_indexes']
>>> print('dataset positive ratio:', len(positives_indexes) / num_targets)
>>> print('dataset negative ratio:', len(negatives_indexes) / num_targets)
>>> print('specified neg_to_pos_ratio:', neg_to_pos_ratio)
>>> sampled_indexes = [self[x]['resolved_index'] for x in range(num_samples)]
>>> num_positives = sum([x in positives_indexes for x in sampled_indexes])
>>> num_negatives = num_samples - num_positives
>>> print('sampled positive ratio:', num_positives / num_samples)
>>> print('sampled negative ratio:', num_negatives / num_samples)
>>> assert num_negatives > 0
>>> assert num_positives > num_negatives

Example

>>> # xdoctest: +REQUIRES(--bench)
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import KWCocoVideoDataset
>>> import ndsampler
>>> import geowatch
>>> import kwcoco
>>> coco_fpath = '/media/joncrall/flash1/smart_phase3_data/Drop8-Cropped2GSD-V1/data_vali_rawbands_split6_n004_f9b08cce.kwcoco.zip'
>>> coco_fpath = '/media/joncrall/flash1/smart_drop7/Drop7-Cropped2GSD-V2/data_vali_rawbands_split6.kwcoco.zip'
>>> coco_dset = kwcoco.CocoDataset(coco_fpath)
>>> self = KWCocoVideoDataset(coco_dset, mode="fit", time_dims=4, window_dims=(300, 300),
>>>                           channels='red|green|blue', neg_to_pos_ratio=1.0)

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.PreprocessMixin

Bases: object

Methods related to dataset preprocessing

cached_dataset_stats(num=None, num_workers=0, batch_size=2, with_intensity=True, with_class=True)

Compute the normalization stats, and caches them

Todo

• [ ] Does this dataset have access to the workdir?

• [ ] Cacher needs to depend on config of this dataset

compute_dataset_stats(num=None, num_workers=0, batch_size=2, with_intensity=True, with_class=True, with_vidid=True)

Parameters:

num (int | None) – number of input items to compute stats for

CommandLine

xdoctest -m geowatch.tasks.fusion.datamodules.kwcoco_dataset KWCocoVideoDataset.compute_dataset_stats:2

Example

>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import kwcoco
>>> dct_dset = coco_dset = kwcoco.CocoDataset.demo('vidshapes2-multispectral', num_frames=3)
>>> self = KWCocoVideoDataset(dct_dset, time_dims=2, window_dims=(256, 256), channels='auto')
>>> self.compute_dataset_stats(num_workers=2)

Example

>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import kwcoco
>>> coco_dset = kwcoco.CocoDataset.demo('vidshapes2')
>>> self = KWCocoVideoDataset(coco_dset, time_dims=2, window_dims=(256, 256), channels='auto')
>>> stats = self.compute_dataset_stats()
>>> assert stats['class_freq']['star'] > 0 or stats['class_freq']['superstar'] > 0 or stats['class_freq']['eff'] > 0
>>> #assert stats['class_freq']['background'] > 0

Example

>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import geowatch
>>> from geowatch.tasks.fusion import datamodules
>>> num = 1
>>> datamodule = datamodules.KWCocoVideoDataModule(
>>>     train_dataset='vidshapes-geowatch', window_dims=64, time_steps=3,
>>>     num_workers=0, batch_size=3, channels='auto',
>>>     normalize_inputs=num)
>>> datamodule.setup('fit')
>>> self = datamodule.torch_datasets['train']
>>> coco_dset = self.sampler.dset
>>> print({c.get('sensor_coarse') for c in coco_dset.images().coco_images})
>>> print({c.channels.spec for c in coco_dset.images().coco_images})
>>> num_workers = 0
>>> batch_size = 6
>>> s = (self.compute_dataset_stats(num=num))
>>> print('s = {}'.format(ub.urepr(s, nl=3)))
>>> stats1 = self.compute_dataset_stats(num=num, with_intensity=False)
>>> stats2 = self.compute_dataset_stats(num=num, with_class=False)
>>> stats3 = self.compute_dataset_stats(num=num, with_class=False, with_intensity=False)

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.MiscMixin

Bases: object

TODO: better groups

reseed()

Reinitialize the random number generator

Todo

HELP WANTED: Lack of determenism likely comes from this module and the order it gives data to predict. It would be very nice if we could fix that.

property coco_dset

make_loader(subset=None, batch_size=1, num_workers=0, shuffle=False, pin_memory=False)

Use this to make the dataloader so we ensure that we have the right worker init function.

Parameters:

subset (None | Dataset) – if specified, the loader is made for this dataset instead of self.

Example

>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import kwcoco
>>> coco_dset = kwcoco.CocoDataset.demo('vidshapes2-multispectral', num_frames=5)
>>> self = KWCocoVideoDataset(coco_dset, time_dims=3, window_dims=(530, 610), channels='auto')
>>> loader = self.make_loader(batch_size=2)
>>> batch = next(iter(loader))

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.BackwardCompatMixin

Bases: object

Backwards compatability for modified properties. (These may eventually be deprecated).

property new_sample_grid

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.KWCocoVideoDataset(sampler, mode='fit', test_with_annot_info=False, autobuild=True, **kwargs)

Bases: Dataset, GetItemMixin, BalanceMixin, PreprocessMixin, IntrospectMixin, MiscMixin, SpacetimeAugmentMixin, BackwardCompatMixin, SMARTDataMixin

Accepted keyword arguments are specified in KWCocoVideoDatasetConfig

Example

>>> # Native Data Sampling
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import ndsampler
>>> import kwcoco
>>> import geowatch
>>> coco_dset = geowatch.coerce_kwcoco('geowatch-multisensor-msi', geodata=True)
>>> print({c.get('sensor_coarse') for c in coco_dset.images().coco_images})
>>> print({c.channels.spec for c in coco_dset.images().coco_images})
>>> sampler = ndsampler.CocoSampler(coco_dset)
>>> self = KWCocoVideoDataset(sampler, time_dims=4, window_dims=(100, 200),
>>>                           input_space_scale='native',
>>>                           window_space_scale='0.05GSD',
>>>                           output_space_scale='native',
>>>                           channels='auto',
>>> )
>>> self.disable_augmenter = True
>>> target = self.sample_grid['targets'][self.sample_grid['positives_indexes'][3]]
>>> item = self[target]
>>> canvas = self.draw_item(item, overlay_on_image=0, rescale=0, max_channels=3)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(canvas)
>>> kwplot.show_if_requested()

Example

>>> # Target GSD Data Sampling
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import ndsampler
>>> import kwcoco
>>> import geowatch
>>> coco_dset = geowatch.coerce_kwcoco('geowatch', geodata=True)
>>> print({c.get('sensor_coarse') for c in coco_dset.images().coco_images})
>>> print({c.channels.spec for c in coco_dset.images().coco_images})
>>> sampler = ndsampler.CocoSampler(coco_dset)
>>> self = KWCocoVideoDataset(sampler, window_dims=(100, 100), time_dims=5,
>>>                           input_space_scale='0.35GSD',
>>>                           window_space_scale='0.7GSD',
>>>                           output_space_scale='0.2GSD',
>>>                           channels='auto',
>>> )
>>> self.disable_augmenter = True
>>> index = self.sample_grid['targets'][self.sample_grid['positives_indexes'][3]]
>>> Box = kwimage.Box
>>> index['space_slice'] = Box.from_slice(index['space_slice']).translate((30, 0)).quantize().to_slice()
>>> item = self[index]
>>> #print('item summary: ' + ub.urepr(self.summarize_item(item), nl=3))
>>> canvas = self.draw_item(item, overlay_on_image=1, rescale=0, max_channels=3)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(canvas)
>>> kwplot.show_if_requested()

Parameters:

• sampler (kwcoco.CocoDataset | ndsampler.CocoSampler) – kwcoco dataset

• mode (str) – fit or predict

• autobuild (bool) – if False, defer potentially expensive initialization. In this case the user must call ._init()

• **kwargs – see KWCocoVideoDatasetConfig for valid options these options will be stored in the .config attribute.

geowatch.tasks.fusion.datamodules.kwcoco_dataset.more_demos()

CommandLine

USE_RTREE=1 DVC_DPATH=1 XDEV_PROFILE=1 xdoctest -m geowatch.tasks.fusion.datamodules.kwcoco_dataset more_demos:0
USE_RTREE=0 DVC_DPATH=1 XDEV_PROFILE=1 xdoctest -m geowatch.tasks.fusion.datamodules.kwcoco_dataset more_demos:0

Example

>>> # xdoctest: +REQUIRES(env:DVC_DPATH)
>>> # Demo with real data
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import geowatch
>>> import kwcoco
>>> dvc_dpath = geowatch.find_dvc_dpath(tags='phase2_data', hardware='auto')
>>> coco_fpath = dvc_dpath / 'Drop6/data_vali_split1.kwcoco.zip'
>>> coco_dset = kwcoco.CocoDataset(coco_fpath)
>>> ##'red|green|blue',
>>> self = KWCocoVideoDataset(
>>>     coco_dset,
>>>     time_dims=7, window_dims=(196, 196),
>>>     window_overlap=0,
>>>     channels="(S2,L8):blue|green|red|nir",
>>>     input_space_scale='3.3GSD',
>>>     window_space_scale='3.3GSD',
>>>     output_space_scale='1GSD',
>>>     prenormalize_inputs=True,
>>>     #normalize_peritem='nir',
>>>     dist_weights=0,
>>>     quality_threshold=0,
>>>     neg_to_pos_ratio=0, time_sampling='soft2',
>>> )
>>> self.requested_tasks['change'] = 1
>>> self.requested_tasks['saliency'] = 1
>>> self.requested_tasks['class'] = 0
>>> self.requested_tasks['boxes'] = 1
>>> index = self.sample_grid['targets'][self.sample_grid['positives_indexes'][3]]
>>> index['allow_augment'] = False
>>> item = self[index]
>>> target = item['target']
>>> #for idx in range(100):
... #    self[idx]
>>> print('item summary: ' + ub.urepr(self.summarize_item(item), nl=3))
>>> # xdoctest: +REQUIRES(--show)
>>> canvas = self.draw_item(item, max_channels=10, overlay_on_image=0, rescale=1)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(canvas, fnum=1)
>>> kwplot.show_if_requested()

Example

>>> # xdoctest: +REQUIRES(env:DVC_DPATH)
>>> # This shows how you can use the dataloader to sample an arbitrary
>>> # spacetime volume.
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import geowatch
>>> import kwcoco
>>> dvc_dpath = geowatch.find_dvc_dpath(tags='phase2_data', hardware='auto')
>>> #coco_fpath = dvc_dpath / 'Drop4-BAS/data_vali.kwcoco.json'
>>> coco_fpath = dvc_dpath / 'Drop6/data_vali_split1.kwcoco.zip'
>>> coco_dset = kwcoco.CocoDataset(coco_fpath)
>>> ##'red|green|blue',
>>> self = KWCocoVideoDataset(
>>>     coco_dset,
>>>     time_dims=7, window_dims=(196, 196),
>>>     window_overlap=0,
>>>     channels="(S2,L8):blue|green|red|nir",
>>>     input_space_scale='3.3GSD',
>>>     window_space_scale='3.3GSD',
>>>     output_space_scale='1GSD',
>>>     #normalize_peritem='nir',
>>>     dist_weights=0,
>>>     quality_threshold=0,
>>>     neg_to_pos_ratio=0, time_sampling='soft2',
>>> )
>>> self.requested_tasks['change'] = 1
>>> self.requested_tasks['saliency'] = 1
>>> self.requested_tasks['class'] = 0
>>> self.requested_tasks['boxes'] = 1
>>> target = {
>>>     'video_id': 3,
>>>     'gids': [529, 555, 607, 697, 719, 730, 768],
>>>     'main_idx': 3,
>>>     'space_slice': (slice(0, 65, None), slice(130, 195, None)),
>>> }
>>> item = self[target]

Example

>>> # xdoctest: +REQUIRES(env:DVC_DPATH)
>>> # Tests the hard negative sampling
>>> from geowatch.tasks.fusion.datamodules.kwcoco_dataset import *  # NOQA
>>> import geowatch
>>> import kwcoco
>>> dvc_dpath = geowatch.find_dvc_dpath(tags='phase2_data', hardware='auto')
>>> coco_fpath = dvc_dpath / 'Drop6-MeanYear10GSD/data.kwcoco.zip'
>>> coco_dset = kwcoco.CocoDataset(coco_fpath)
>>> ##'red|green|blue',
>>> self = KWCocoVideoDataset(
>>>     coco_dset,
>>>     time_dims=5, window_dims=(196, 196),
>>>     window_overlap=0,
>>>     channels="(S2,L8):blue|green|red",
>>>     fixed_resolution='10GSD',
>>>     normalize_peritem=True,
>>>     use_grid_negatives='cleared',
>>>     use_grid_positives=False,
>>>     use_centered_positives= True,
>>>     time_kernel='(-2y,-1y,0,1y,2y)',
>>> )
>>> self.requested_tasks['change'] = 1
>>> self.requested_tasks['saliency'] = 1
>>> self.requested_tasks['class'] = 0
>>> self.requested_tasks['boxes'] = 1

>>> # Check that all of the negative regions are from cleared videos
>>> videos = self.sampler.dset.videos()
>>> vidid_to_cleared = ub.udict(ub.dzip(videos.lookup('id'), videos.lookup('cleared', False)))
>>> assert self.config['use_grid_negatives'] == 'cleared'
>>> positive_idxs = self.sample_grid['positives_indexes']
>>> negative_idxs = self.sample_grid['negatives_indexes']
>>> targets = self.sample_grid['targets']
>>> negative_video_ids = {targets[x]['video_id'] for x in negative_idxs}
>>> positive_video_ids = {targets[x]['video_id'] for x in positive_idxs}
>>> assert all(vidid_to_cleared.subdict(negative_video_ids).values())

>>> index = 0
>>> item = self[index]
>>> target = item['target']
>>> print('item summary: ' + ub.urepr(self.summarize_item(item), nl=3))
>>> # xdoctest: +REQUIRES(--show)
>>> canvas = self.draw_item(item, max_channels=10, overlay_on_image=0, rescale=1)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(canvas, fnum=1)
>>> kwplot.show_if_requested()

geowatch.tasks.fusion.datamodules.kwcoco_dataset.worker_init_fn(worker_id)

exception geowatch.tasks.fusion.datamodules.kwcoco_dataset.FailedSample

Bases: Exception

Used to indicate that a sample should be skipped.

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.Modality(sensor: str, channels: str, domain: str)

Bases: NamedTuple

A modality consists of a domain, a sensor, and a FusedChannelSpec

Create new instance of Modality(sensor, channels, domain)

sensor: str

Alias for field number 0

channels: str

Alias for field number 1

domain: str

Alias for field number 2

class geowatch.tasks.fusion.datamodules.kwcoco_dataset.Domain(sensor: str, channels: str, video_name: str)

Bases: NamedTuple

DO NOT USE. BUT DO NOT REMOVE. NEEDED FOR BACKWARDS COMPAT Use Modality instead.

Create new instance of Domain(sensor, channels, video_name)

sensor: str

Alias for field number 0

channels: str

Alias for field number 1

video_name: str

Alias for field number 2