justdata 1.2.1__tar.gz → 1.3.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (123) hide show
  1. {justdata-1.2.1 → justdata-1.3.0}/PKG-INFO +29 -5
  2. {justdata-1.2.1 → justdata-1.3.0}/README.md +28 -4
  3. {justdata-1.2.1 → justdata-1.3.0}/pyproject.toml +1 -1
  4. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/__init__.py +1 -1
  5. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/execution.py +5 -0
  6. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/geometry.py +143 -27
  7. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/pipelines.py +2 -2
  8. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/stages.py +2 -2
  9. {justdata-1.2.1 → justdata-1.3.0}/LICENSE +0 -0
  10. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/__init__.py +0 -0
  11. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/_dcase.py +0 -0
  12. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/_manifests.py +0 -0
  13. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/_random.py +0 -0
  14. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/_signal.py +0 -0
  15. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/adapters.py +0 -0
  16. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/augment/__init__.py +0 -0
  17. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/augment/_common.py +0 -0
  18. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/augment/batch.py +0 -0
  19. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/augment/configs.py +0 -0
  20. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/augment/patchout.py +0 -0
  21. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/augment/spectrogram.py +0 -0
  22. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/augment/waveform.py +0 -0
  23. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/batching.py +0 -0
  24. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/channel.py +0 -0
  25. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/compat/__init__.py +0 -0
  26. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/compat/ast.py +0 -0
  27. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/compat/ced.py +0 -0
  28. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/compat/efficientat.py +0 -0
  29. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/compat/panns.py +0 -0
  30. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/compat/passt.py +0 -0
  31. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/configs.py +0 -0
  32. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/corruptions/__init__.py +0 -0
  33. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/corruptions/codec.py +0 -0
  34. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/corruptions/datasets.py +0 -0
  35. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/corruptions/device.py +0 -0
  36. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/corruptions/dynamics.py +0 -0
  37. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/corruptions/filters.py +0 -0
  38. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/corruptions/noise.py +0 -0
  39. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/corruptions/registry.py +0 -0
  40. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/corruptions/reverb.py +0 -0
  41. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/corruptions/timefreq.py +0 -0
  42. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/datasets.py +0 -0
  43. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/dcase2025.py +0 -0
  44. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/decoding.py +0 -0
  45. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/eval_views.py +0 -0
  46. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/evals.py +0 -0
  47. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/filters.py +0 -0
  48. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/frontends/__init__.py +0 -0
  49. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/frontends/ast.py +0 -0
  50. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/frontends/kaldi.py +0 -0
  51. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/frontends/log.py +0 -0
  52. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/frontends/mel.py +0 -0
  53. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/frontends/mfcc.py +0 -0
  54. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/frontends/pcen.py +0 -0
  55. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/frontends/stft.py +0 -0
  56. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/jax.py +0 -0
  57. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/labels.py +0 -0
  58. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/layouts.py +0 -0
  59. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/metadata.py +0 -0
  60. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/normalization.py +0 -0
  61. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/pipelines.py +0 -0
  62. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/postprocessing.py +0 -0
  63. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/preprocessing.py +0 -0
  64. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/presets.py +0 -0
  65. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/registry.py +0 -0
  66. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/resampling.py +0 -0
  67. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/schema.py +0 -0
  68. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/segment.py +0 -0
  69. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/sources.py +0 -0
  70. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/stages.py +0 -0
  71. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/stats.py +0 -0
  72. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/acoustic/tasks.py +0 -0
  73. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/audio/__init__.py +0 -0
  74. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/__init__.py +0 -0
  75. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/_records.py +0 -0
  76. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/adapters.py +0 -0
  77. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/augmentations.py +0 -0
  78. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/cache.py +0 -0
  79. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/config_resolution.py +0 -0
  80. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/executed_config.py +0 -0
  81. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/filters.py +0 -0
  82. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/finalization.py +0 -0
  83. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/inventory.py +0 -0
  84. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/label_mixing.py +0 -0
  85. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/loader.py +0 -0
  86. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/metadata.py +0 -0
  87. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/presets.py +0 -0
  88. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/registry.py +0 -0
  89. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/replay.py +0 -0
  90. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/sources.py +0 -0
  91. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/core/stats.py +0 -0
  92. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/corruptions.py +0 -0
  93. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/py.typed +0 -0
  94. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/__init__.py +0 -0
  95. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/augmentations/__init__.py +0 -0
  96. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/augmentations/auto.py +0 -0
  97. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/augmentations/color.py +0 -0
  98. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/augmentations/composed.py +0 -0
  99. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/augmentations/geometric.py +0 -0
  100. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/augmentations/mixing.py +0 -0
  101. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/augmentations/registry.py +0 -0
  102. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/corruptions/__init__.py +0 -0
  103. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/corruptions/blur.py +0 -0
  104. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/corruptions/digital.py +0 -0
  105. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/corruptions/noise.py +0 -0
  106. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/corruptions/registry.py +0 -0
  107. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/corruptions/standard.py +0 -0
  108. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/corruptions/weather.py +0 -0
  109. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/datasets.py +0 -0
  110. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/encodings/__init__.py +0 -0
  111. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/encodings/heatmaps.py +0 -0
  112. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/encodings/panoptic_targets.py +0 -0
  113. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/encodings/semantic_targets.py +0 -0
  114. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/minic.py +0 -0
  115. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/presets.py +0 -0
  116. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/sources.py +0 -0
  117. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/tasks/__init__.py +0 -0
  118. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/tasks/classification.py +0 -0
  119. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/tasks/depth.py +0 -0
  120. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/tasks/panoptic.py +0 -0
  121. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/tasks/segmentation.py +0 -0
  122. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/transforms.py +0 -0
  123. {justdata-1.2.1 → justdata-1.3.0}/src/justdata/vision/utils.py +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: justdata
3
- Version: 1.2.1
3
+ Version: 1.3.0
4
4
  Summary: A TensorFlow-native data pipeline library with a modality-neutral core and first-class computer vision and acoustic recipes
5
5
  Author: clementpoiret
6
6
  Author-email: clementpoiret <clement@linux.com>
@@ -546,7 +546,9 @@ For semantic segmentation with replayable geometry, select `vision/segmentation`
546
546
  `geometry_kwargs`. It supports paired aspect-preserving training resize and crops, longer-side evaluation caps,
547
547
  independent RGB/ignore padding, per-pixel validity, and original-coordinate score restoration. See
548
548
  [the dense geometry contract](docs/vision.md#12-replayable-dense-segmentation) for configuration, record fields, replay,
549
- and batching requirements.
549
+ and batching requirements. For ordered image/mask file lists without a registered source,
550
+ `justdata.vision.geometry.dense_views_from_files` builds the same views with explicit per-example seeds in a parallel,
551
+ order-preserving `tf.data` map.
550
552
 
551
553
  For instance-aware labels, select `vision/panoptic_segmentation` with a segment-ID map and category table. It preserves
552
554
  thing instances, merges stuff by category, and excludes crowd pixels from panoptic supervision. The
@@ -1094,14 +1096,36 @@ Replay prefetches after skipping committed batches and defaults to one batch.
1094
1096
  When another framework uses the accelerator, configure TensorFlow before creating tensors or datasets:
1095
1097
 
1096
1098
  ```python
1099
+ import jax
1100
+
1101
+ jax.devices() # initialize JAX's GPU backend before TensorFlow is imported
1102
+
1097
1103
  from justdata.core import configure_tensorflow_cpu
1098
1104
 
1099
1105
  device_report = configure_tensorflow_cpu(intra_op_threads=1, inter_op_threads=1)
1100
1106
  assert device_report["logical_gpus"] == []
1101
1107
  ```
1102
1108
 
1103
- The report includes physical and logical device names and process thread settings. The helper raises if the runtime has
1104
- initialized too early for the requested change. Core, vision, acoustic, and audio imports permit this setup.
1109
+ The report includes physical and logical device names and process thread settings. The helper raises if the TensorFlow
1110
+ runtime has initialized too early for the requested change. Core, vision, acoustic, and audio imports do not initialize
1111
+ it, so the helper can run after them.
1112
+
1113
+ Those imports do import TensorFlow. On Linux, the CUDA-enabled `tensorflow` wheel then exposes stubs for its CUDA
1114
+ libraries (CUDA 12 as of TensorFlow 2.21) to the whole process. A framework built for another CUDA version that
1115
+ initializes its GPU backend afterwards can bind to those stubs: JAX with its CUDA 13 plugin reports an outdated cuBLAS
1116
+ and falls back to the CPU. Hiding TensorFlow's GPUs does not prevent this. Either initialize the other framework's
1117
+ backend before the first TensorFlow import, as above (`import jax` alone is not enough), or install `tensorflow-cpu`,
1118
+ which exports no CUDA symbols, in place of `tensorflow`. With uv, on Linux x86_64:
1119
+
1120
+ ```toml
1121
+ [project]
1122
+ dependencies = ["justdata", "tensorflow-cpu"]
1123
+
1124
+ [tool.uv]
1125
+ exclude-dependencies = ["tensorflow"]
1126
+ ```
1127
+
1128
+ Set `JAX_PLATFORMS=cuda` to make a JAX fallback to the CPU an error rather than a warning.
1105
1129
 
1106
1130
  `CachePolicy` enables identity-checked file caches on the existing preprocessing and model-input cache stages. The cache
1107
1131
  path is a directory whose parent exists. The input identity must cover ordered content. `load_inventory` binds its
@@ -1306,7 +1330,7 @@ package-support jobs pass. Build artifacts are retained for one week. Tag pipeli
1306
1330
  to complete a release.
1307
1331
 
1308
1332
  Before tagging, update `pyproject.toml`, `src/justdata/__init__.py`, and `CHANGELOG.md`, then run `uv lock`. Push a tag
1309
- matching the package version, with an optional `v` prefix (for example, `v1.2.1` or `1.2.1`). Publishing rejects tags
1333
+ matching the package version, with an optional `v` prefix (for example, `v1.3.0` or `1.3.0`). Publishing rejects tags
1310
1334
  that disagree with either version declaration. Branch pipelines build artifacts without publishing.
1311
1335
 
1312
1336
  Configure a [PyPI GitLab trusted publisher](https://docs.pypi.org/trusted-publishers/adding-a-publisher/) for the
@@ -503,7 +503,9 @@ For semantic segmentation with replayable geometry, select `vision/segmentation`
503
503
  `geometry_kwargs`. It supports paired aspect-preserving training resize and crops, longer-side evaluation caps,
504
504
  independent RGB/ignore padding, per-pixel validity, and original-coordinate score restoration. See
505
505
  [the dense geometry contract](docs/vision.md#12-replayable-dense-segmentation) for configuration, record fields, replay,
506
- and batching requirements.
506
+ and batching requirements. For ordered image/mask file lists without a registered source,
507
+ `justdata.vision.geometry.dense_views_from_files` builds the same views with explicit per-example seeds in a parallel,
508
+ order-preserving `tf.data` map.
507
509
 
508
510
  For instance-aware labels, select `vision/panoptic_segmentation` with a segment-ID map and category table. It preserves
509
511
  thing instances, merges stuff by category, and excludes crowd pixels from panoptic supervision. The
@@ -1051,14 +1053,36 @@ Replay prefetches after skipping committed batches and defaults to one batch.
1051
1053
  When another framework uses the accelerator, configure TensorFlow before creating tensors or datasets:
1052
1054
 
1053
1055
  ```python
1056
+ import jax
1057
+
1058
+ jax.devices() # initialize JAX's GPU backend before TensorFlow is imported
1059
+
1054
1060
  from justdata.core import configure_tensorflow_cpu
1055
1061
 
1056
1062
  device_report = configure_tensorflow_cpu(intra_op_threads=1, inter_op_threads=1)
1057
1063
  assert device_report["logical_gpus"] == []
1058
1064
  ```
1059
1065
 
1060
- The report includes physical and logical device names and process thread settings. The helper raises if the runtime has
1061
- initialized too early for the requested change. Core, vision, acoustic, and audio imports permit this setup.
1066
+ The report includes physical and logical device names and process thread settings. The helper raises if the TensorFlow
1067
+ runtime has initialized too early for the requested change. Core, vision, acoustic, and audio imports do not initialize
1068
+ it, so the helper can run after them.
1069
+
1070
+ Those imports do import TensorFlow. On Linux, the CUDA-enabled `tensorflow` wheel then exposes stubs for its CUDA
1071
+ libraries (CUDA 12 as of TensorFlow 2.21) to the whole process. A framework built for another CUDA version that
1072
+ initializes its GPU backend afterwards can bind to those stubs: JAX with its CUDA 13 plugin reports an outdated cuBLAS
1073
+ and falls back to the CPU. Hiding TensorFlow's GPUs does not prevent this. Either initialize the other framework's
1074
+ backend before the first TensorFlow import, as above (`import jax` alone is not enough), or install `tensorflow-cpu`,
1075
+ which exports no CUDA symbols, in place of `tensorflow`. With uv, on Linux x86_64:
1076
+
1077
+ ```toml
1078
+ [project]
1079
+ dependencies = ["justdata", "tensorflow-cpu"]
1080
+
1081
+ [tool.uv]
1082
+ exclude-dependencies = ["tensorflow"]
1083
+ ```
1084
+
1085
+ Set `JAX_PLATFORMS=cuda` to make a JAX fallback to the CPU an error rather than a warning.
1062
1086
 
1063
1087
  `CachePolicy` enables identity-checked file caches on the existing preprocessing and model-input cache stages. The cache
1064
1088
  path is a directory whose parent exists. The input identity must cover ordered content. `load_inventory` binds its
@@ -1263,7 +1287,7 @@ package-support jobs pass. Build artifacts are retained for one week. Tag pipeli
1263
1287
  to complete a release.
1264
1288
 
1265
1289
  Before tagging, update `pyproject.toml`, `src/justdata/__init__.py`, and `CHANGELOG.md`, then run `uv lock`. Push a tag
1266
- matching the package version, with an optional `v` prefix (for example, `v1.2.1` or `1.2.1`). Publishing rejects tags
1290
+ matching the package version, with an optional `v` prefix (for example, `v1.3.0` or `1.3.0`). Publishing rejects tags
1267
1291
  that disagree with either version declaration. Branch pipelines build artifacts without publishing.
1268
1292
 
1269
1293
  Configure a [PyPI GitLab trusted publisher](https://docs.pypi.org/trusted-publishers/adding-a-publisher/) for the
@@ -1,6 +1,6 @@
1
1
  [project]
2
2
  name = "justdata"
3
- version = "1.2.1"
3
+ version = "1.3.0"
4
4
  description = "A TensorFlow-native data pipeline library with a modality-neutral core and first-class computer vision and acoustic recipes"
5
5
  readme = "README.md"
6
6
  license = "MIT"
@@ -1,3 +1,3 @@
1
- __version__ = "1.2.1"
1
+ __version__ = "1.3.0"
2
2
 
3
3
  __all__ = ["__version__"]
@@ -14,6 +14,11 @@ def configure_tensorflow_cpu(
14
14
 
15
15
  Call before constructing tensors, datasets, or invoking other TensorFlow
16
16
  operations. A runtime that was initialized too early raises RuntimeError.
17
+
18
+ Hiding GPUs does not isolate TensorFlow's CUDA libraries. With the
19
+ CUDA-enabled ``tensorflow`` wheel, another framework must initialize its
20
+ GPU backend before TensorFlow is first imported; ``tensorflow-cpu`` has no
21
+ such ordering requirement.
17
22
  """
18
23
  for name, value in (
19
24
  ("intra_op_threads", intra_op_threads),
@@ -1,11 +1,14 @@
1
1
  """Versioned, paired dense geometry in HWC coordinates.
2
2
 
3
3
  Records are dictionaries of numeric tensors, suitable for tf.data, NumPy storage,
4
- and replay without a random generator. Version 1 uses half-pixel resize centers,
4
+ and replay without a random generator. Both versions use half-pixel resize centers,
5
5
  bilinear RGB sampling, nearest categorical sampling, and bottom/right padding.
6
+ Version 1 resamples RGB over the whole resized frame before cropping; version 2,
7
+ used by new records, resamples only a crop that lies inside the resized frame.
6
8
  """
7
9
 
8
10
  import math
11
+ import os
9
12
  from dataclasses import asdict, dataclass
10
13
 
11
14
  import tensorflow as tf
@@ -138,19 +141,32 @@ def _bottom_right(amount):
138
141
  return tf.stack([0, amount[0], 0, amount[1]])
139
142
 
140
143
 
144
+ def _canonical_seed(seed):
145
+ """Widen an int32 or int64 stateless seed of shape [2] to int64.
146
+
147
+ ``stateless_split`` draws in the seed dtype, so equal int32 and int64 seed
148
+ values would otherwise produce different views.
149
+ """
150
+ seed = tf.ensure_shape(tf.convert_to_tensor(seed), [2])
151
+ if seed.dtype not in (tf.int32, tf.int64):
152
+ raise TypeError("seed must have int32 or int64 dtype")
153
+ return tf.cast(seed, tf.int64)
154
+
155
+
141
156
  def sample_dense_geometry(
142
157
  original_size, config: DenseGeometryConfig, *, is_training, seed=None
143
158
  ):
144
- """Create a version-1 record; training requires an int32/int64 seed of shape [2]."""
159
+ """Create a version-2 record; training requires an int32/int64 seed of shape [2].
160
+
161
+ Seeds are widened to int64 first, so equal values select the same view
162
+ whatever their dtype.
163
+ """
145
164
  original = tf.ensure_shape(tf.convert_to_tensor(original_size, tf.int32), [2])
146
165
  tf.debugging.assert_positive(original, message="original_size must be positive")
147
166
  if is_training:
148
167
  if seed is None:
149
168
  raise ValueError("training geometry requires a stateless seed")
150
- seed = tf.ensure_shape(tf.convert_to_tensor(seed), [2])
151
- if seed.dtype not in (tf.int32, tf.int64):
152
- raise TypeError("seed must have int32 or int64 dtype")
153
- seeds = tf.random.experimental.stateless_split(seed, 4)
169
+ seeds = tf.random.experimental.stateless_split(_canonical_seed(seed), 4)
154
170
  if config.train_resize_range is not None:
155
171
  low, high = config.train_resize_range
156
172
  target = tf.random.stateless_uniform(
@@ -196,7 +212,7 @@ def sample_dense_geometry(
196
212
  after = (config.patch_size - cropped % config.patch_size) % config.patch_size
197
213
  input_size = cropped + after
198
214
  return {
199
- "version": tf.constant(1, tf.int32),
215
+ "version": tf.constant(2, tf.int32),
200
216
  "original_size": original,
201
217
  "resized_size": resized,
202
218
  "crop_box": tf.stack([top, left, cropped[0], cropped[1]]),
@@ -207,7 +223,7 @@ def sample_dense_geometry(
207
223
  "model_input_size": input_size,
208
224
  "patch_size": tf.constant(config.patch_size, tf.int32),
209
225
  "grid_size": input_size // config.patch_size,
210
- # Numeric enum values are part of version 1, documented in docs/vision.md.
226
+ # Numeric enum values are shared by versions 1 and 2, see docs/vision.md.
211
227
  "image_interpolation": tf.constant(1, tf.int32),
212
228
  "mask_interpolation": tf.constant(0, tf.int32),
213
229
  "alignment": tf.constant(1, tf.int32),
@@ -224,7 +240,7 @@ def sample_dense_geometry(
224
240
  def sample_panoptic_geometry(
225
241
  original_size, config: PanopticGeometryConfig, *, void_value, is_training, seed=None
226
242
  ):
227
- """Sample version-1 geometry with a panoptic label contract."""
243
+ """Sample version-2 geometry with a panoptic label contract."""
228
244
  record = sample_dense_geometry(
229
245
  original_size, _spatial_config(config), is_training=is_training, seed=seed
230
246
  )
@@ -238,8 +254,8 @@ def _check_record(record):
238
254
  record = tf.nest.map_structure(tf.convert_to_tensor, record)
239
255
  record["version"] = tf.ensure_shape(record["version"], [])
240
256
  tf.debugging.assert_equal(
241
- record["version"],
242
- 1,
257
+ (record["version"] == 1) | (record["version"] == 2),
258
+ True,
243
259
  message="unsupported geometry version",
244
260
  )
245
261
  for key, value in (
@@ -307,12 +323,13 @@ def validate_dense_sample(sample, class_values, ignore_value):
307
323
  return result
308
324
 
309
325
 
310
- def _nearest(tensor, size):
326
+ def _nearest(tensor, size, start, stop):
327
+ """Sample rows and columns ``[start, stop)`` of ``tensor`` resized to ``size``."""
311
328
  # Gather directly so categorical int64 IDs never pass through float32.
312
329
  indices = []
313
330
  for axis in (0, 1):
314
331
  length = tf.shape(tensor)[axis]
315
- centers = tf.cast(tf.range(size[axis]), tf.float64) + 0.5
332
+ centers = tf.cast(tf.range(start[axis], stop[axis]), tf.float64) + 0.5
316
333
  scale = tf.cast(length, tf.float64) / tf.cast(size[axis], tf.float64)
317
334
  nearest = tf.cast(tf.floor(centers * scale), tf.int32)
318
335
  indices.append(tf.minimum(nearest, length - 1))
@@ -379,36 +396,71 @@ def replay_dense_geometry(sample, record):
379
396
  return result | {"pixel_valid_mask": valid}
380
397
 
381
398
 
399
+ def _resample_rgb(image, record, start, stop):
400
+ """Bilinearly resample rows and columns ``[start, stop)`` of the resized frame."""
401
+ scale = tf.cast(record["resized_size"], tf.float32) / tf.cast(
402
+ tf.shape(image)[:2], tf.float32
403
+ )
404
+
405
+ def triangle(antialias):
406
+ # The op behind antialiased tf.image.resize. Translating by -start
407
+ # computes only the window; over the whole frame it matches that resize.
408
+ return lambda: tf.raw_ops.ScaleAndTranslate(
409
+ images=image[None],
410
+ size=stop - start,
411
+ scale=scale,
412
+ translation=-tf.cast(start, tf.float32),
413
+ kernel_type="triangle",
414
+ antialias=antialias,
415
+ )[0]
416
+
417
+ return tf.cond(
418
+ record["image_antialias"],
419
+ triangle(True),
420
+ lambda: tf.cond(
421
+ record["version"] == 1,
422
+ lambda: _bilinear_scores(image, record["resized_size"]),
423
+ triangle(False),
424
+ ),
425
+ )
426
+
427
+
382
428
  def _replay_spatial(sample, record):
383
429
  tf.debugging.assert_equal(
384
430
  tf.shape(sample["image"])[:2],
385
431
  record["original_size"],
386
432
  message="replay requires original sample dimensions",
387
433
  )
388
- image = tf.cast(sample["image"], tf.float32)
389
- image = tf.cond(
390
- record["image_antialias"],
391
- lambda: tf.image.resize(
392
- image, record["resized_size"], method="bilinear", antialias=True
393
- ),
394
- lambda: tf.image.resize(
395
- image, record["resized_size"], method="bilinear", antialias=False
396
- ),
434
+ resized = record["resized_size"]
435
+ origin, size = record["crop_box"][:2], record["crop_box"][2:]
436
+ before = record["pre_padding"][::2]
437
+ # Version 2 resamples only a crop inside the resized frame. Axes whose crop
438
+ # reaches padding, and all version-1 axes, keep the frame to pad from it.
439
+ window = (
440
+ (record["version"] == 2)
441
+ & (origin >= before)
442
+ & (origin + size <= before + resized)
397
443
  )
444
+ start = tf.where(window, origin - before, 0)
445
+ stop = tf.where(window, start + size, resized)
446
+ pre_padding = tf.where(tf.repeat(window, 2), 0, record["pre_padding"])
447
+ offset = tf.where(window, 0, origin)
398
448
 
399
449
  def crop_flip(tensor):
400
- top, left, height, width = tf.unstack(record["crop_box"])
450
+ top, left = tf.unstack(offset)
451
+ height, width = tf.unstack(size)
401
452
  tensor = tensor[top : top + height, left : left + width]
402
453
  return tf.cond(
403
454
  record["horizontal_flip"], lambda: tf.reverse(tensor, [1]), lambda: tensor
404
455
  )
405
456
 
406
- image = _pad_rgb(image, record["pre_padding"], record)
457
+ image = _resample_rgb(tf.cast(sample["image"], tf.float32), record, start, stop)
458
+ image = _pad_rgb(image, pre_padding, record)
407
459
  image = _pad_rgb(crop_flip(image), record["post_padding"], record)
408
460
 
409
461
  def categorical(tensor, fill):
410
- tensor = _nearest(tensor, record["resized_size"])
411
- tensor = _constant_pad(tensor, record["pre_padding"], fill)
462
+ tensor = _nearest(tensor, resized, start, stop)
463
+ tensor = _constant_pad(tensor, pre_padding, fill)
412
464
  return _constant_pad(crop_flip(tensor), record["post_padding"], fill)
413
465
 
414
466
  support = categorical(tf.ones(record["original_size"], tf.bool), False)
@@ -416,6 +468,69 @@ def _replay_spatial(sample, record):
416
468
  return result, categorical
417
469
 
418
470
 
471
+ _PNG_SIGNATURE_AND_HEADER_CHUNK = b"\x89PNG\r\n\x1a\n\x00\x00\x00\rIHDR"
472
+
473
+
474
+ def _read_mask_png(path):
475
+ contents = tf.io.read_file(path)
476
+ # decode_png rescales 16-bit samples, maps palettes through RGB, and also
477
+ # decodes JPEG data; each would silently rewrite class IDs.
478
+ tf.debugging.assert_equal(
479
+ (tf.strings.substr(contents, 0, 16) == _PNG_SIGNATURE_AND_HEADER_CHUNK)
480
+ & (tf.strings.substr(contents, 24, 2) == b"\x08\x00"),
481
+ True,
482
+ message="mask files must be 8-bit grayscale PNG",
483
+ )
484
+ return tf.io.decode_png(contents, channels=1)
485
+
486
+
487
+ def dense_views_from_files(
488
+ image_paths,
489
+ config: DenseGeometryConfig,
490
+ *,
491
+ mask_paths=None,
492
+ seeds=None,
493
+ num_parallel_calls=tf.data.AUTOTUNE,
494
+ ):
495
+ """Replay one dense view per image file, in input order, in parallel.
496
+
497
+ Images decode to RGB uint8 and optional masks from 8-bit grayscale PNG.
498
+ Element ``i`` equals ``replay_dense_geometry`` of the decoded sample with
499
+ the training record for ``seeds[i]`` (``[N, 2]`` int32 or int64), or with
500
+ the evaluation record when ``seeds`` is None. Set threading options on the
501
+ returned dataset to bound its CPU use.
502
+ """
503
+ columns = {
504
+ "image": tf.constant([os.fspath(path) for path in image_paths], tf.string)
505
+ }
506
+ if mask_paths is not None:
507
+ columns["mask"] = tf.constant(
508
+ [os.fspath(path) for path in mask_paths], tf.string
509
+ )
510
+ if seeds is not None:
511
+ columns["seed"] = seeds
512
+
513
+ def view(column):
514
+ sample = {
515
+ "image": tf.io.decode_image(
516
+ tf.io.read_file(column["image"]), channels=3, expand_animations=False
517
+ )
518
+ }
519
+ if "mask" in column:
520
+ sample["mask"] = _read_mask_png(column["mask"])
521
+ record = sample_dense_geometry(
522
+ tf.shape(sample["image"])[:2],
523
+ config,
524
+ is_training=seeds is not None,
525
+ seed=column.get("seed"),
526
+ )
527
+ return replay_dense_geometry(sample, record)
528
+
529
+ return tf.data.Dataset.from_tensor_slices(columns).map(
530
+ view, num_parallel_calls=num_parallel_calls, deterministic=True
531
+ )
532
+
533
+
419
534
  def _visible_segments(mask, segments, max_segments):
420
535
  """Keep segments visible in a replayed view and recompute area and bbox."""
421
536
  flat = tf.reshape(mask, [-1])
@@ -464,7 +579,7 @@ def _visible_segments(mask, segments, max_segments):
464
579
  def replay_panoptic_geometry(
465
580
  sample, record, *, class_values, thing_class_values, void_value, max_segments
466
581
  ):
467
- """Replay a version-1 record, including integer map and crowd validity."""
582
+ """Replay a record of either version, including integer map and crowd validity."""
468
583
  record = _check_record(record)
469
584
  tf.debugging.assert_equal(
470
585
  tf.size(record["class_values"]),
@@ -573,6 +688,7 @@ __all__ = [
573
688
  "sample_panoptic_geometry",
574
689
  "validate_dense_sample",
575
690
  "replay_dense_geometry",
691
+ "dense_views_from_files",
576
692
  "replay_panoptic_geometry",
577
693
  "restore_dense_scores",
578
694
  "restore_dense_predictions",
@@ -105,7 +105,7 @@ def _validate_recorded_postprocess(post, flags):
105
105
 
106
106
 
107
107
  def _recorded_geometry_contract(geometry, is_training):
108
- """Describe the version-1 geometry shared by semantic and panoptic views."""
108
+ """Describe the version-2 geometry shared by semantic and panoptic views."""
109
109
  if not is_training:
110
110
  resize_policy = "long_side_cap"
111
111
  elif geometry["train_scale_range"] is not None:
@@ -113,7 +113,7 @@ def _recorded_geometry_contract(geometry, is_training):
113
113
  else:
114
114
  resize_policy = "uniform_integer_short_side"
115
115
  return geometry | {
116
- "record_version": 1,
116
+ "record_version": 2,
117
117
  "resize_policy": resize_policy,
118
118
  "image_interpolation": "bilinear",
119
119
  "mask_interpolation": "nearest",
@@ -5,7 +5,7 @@ import tensorflow as tf
5
5
 
6
6
  from justdata.vision.augmentations.color import color_jitter, photometric_distortion
7
7
  from justdata.vision.augmentations.registry import get_crop_strategy
8
- from justdata.vision.geometry import _train_input_size
8
+ from justdata.vision.geometry import _canonical_seed, _train_input_size
9
9
  from justdata.vision.transforms import nhwc_to_nchw, normalize, resize_image
10
10
 
11
11
 
@@ -270,7 +270,7 @@ def recorded_view_augmentations(
270
270
 
271
271
  def augmentations(sample, seed):
272
272
  geometry_seed, color_seed = tf.unstack(
273
- tf.random.experimental.stateless_split(seed, 2)
273
+ tf.random.experimental.stateless_split(_canonical_seed(seed), 2)
274
274
  )
275
275
  if photometric_kwargs is not None:
276
276
  sample = sample | {
File without changes
File without changes