Deterministic seeds in debug mode (#3589)

Add support for deterministic seed generation in debug mode when using pipeline seed.

Currently seed generation differ between standard and debug mode, in the future we might standardize this.

Signed-off-by: ksztenderski <ksztenderski@nvidia.com>

Author: ksztenderski

dali/python/nvidia/dali/_debug_mode.py

@@ -134,6 +134,12 @@ def __init__(self, exec_func, **kwargs):
         self._cur_subpipeline_id = -1
         self._exec_func = exec_func
 
+        import numpy as np
+        seed = kwargs.get('seed', -1)
+        if seed < 0:
+            seed  = np.random.randint(0, 2**32)
+        self._seed_generator = np.random.default_rng(seed)
+
     def __enter__(self):
         raise RuntimeError("Currently pipeline in debug mode works only with `pipeline_def` decorator."
                            "Using `with` statement is not supported.")
@@ -259,6 +265,8 @@ def pipe():
                 else:
                     kwargs_preprocessed[key] = value
 
+            if 'seed' not in kwargs_preprocessed and op_wrapper.__name__ != '_arithm_op':
+                kwargs_preprocessed['seed'] = self._seed_generator.integers(0, 2**32)
             res = op_wrapper(*inputs_preprocessed, **kwargs_preprocessed)
 
             return tuple(res) if isinstance(res, list) else res

dali/test/python/test_pipeline_debug.py

@@ -340,3 +340,38 @@ def _test_shape_pipeline(device):
 def test_shape_pipeline():
     for device in ['cpu', 'mixed']:
         yield _test_shape_pipeline, device
+
+
+@pipeline_def(batch_size=8, num_threads=3, device_id=0, seed=47, debug=True)
+def seed_pipeline():
+    coin_flip = fn.random.coin_flip()
+    normal = fn.random.normal()
+    uniform = fn.random.uniform()
+    batch_permutation = fn.batch_permutation()
+    return coin_flip, normal, uniform, batch_permutation
+
+
+def test_seed_generation():
+    pipe1 = seed_pipeline()
+    pipe2 = seed_pipeline()
+    compare_pipelines(pipe1, pipe2, 8, 10)
+
+
+@pipeline_def(batch_size=8, num_threads=3, device_id=0, seed=47, debug=True)
+def seed_rn50_pipeline_base():
+    rng = fn.random.coin_flip(probability=0.5)
+    jpegs, labels = fn.readers.file(
+        file_root=file_root, shard_id=0, num_shards=2, random_shuffle=True)
+    images = fn.decoders.image(jpegs, device='mixed', output_type=types.RGB)
+    resized_images = fn.random_resized_crop(images, device="gpu", size=(224, 224))
+    out_type = types.FLOAT16
+
+    output = fn.crop_mirror_normalize(resized_images.gpu(), mirror=rng, device="gpu", dtype=out_type, crop=(
+        224, 224), mean=[0.485 * 255, 0.456 * 255, 0.406 * 255], std=[0.229 * 255, 0.224 * 255, 0.225 * 255])
+    return rng, jpegs, labels, images, resized_images, output
+
+
+def test_seed_generation_base():
+    pipe1 = seed_rn50_pipeline_base()
+    pipe2 = seed_rn50_pipeline_base()
+    compare_pipelines(pipe1, pipe2, 8, 10)