Use random.uniform discrete mode

Signed-off-by: Kamil Tokarski <[email protected]>

dali/python/nvidia/dali/auto_aug/auto_augment.py

@@ -17,7 +17,7 @@
 from nvidia.dali import fn
 from nvidia.dali import types
 from nvidia.dali.auto_aug import augmentations as a
-from nvidia.dali.auto_aug.core.utils import operation_idx_random_choice, select
+from nvidia.dali.auto_aug.core.utils import select
 
 
 class Policy:
@@ -160,13 +160,12 @@ def apply_auto_augment(policy: Policy, samples, seed=None, augment_kwargs=None):
     if not use_signed_magnitudes:
         random_sign = None
     else:
-        random_sign = fn.random.uniform(range=[0, 1], dtype=types.INT32, seed=seed,
-                                        shape=(max_policy_len, ))
+        random_sign = fn.random.uniform(values=[0, 1], seed=seed, shape=(max_policy_len, ))
     should_run = fn.random.uniform(range=[0, 1], shape=(max_policy_len, ), dtype=types.FLOAT)
     op_kwargs = dict(samples=samples, should_run=should_run, random_sign=random_sign,
                      num_magnitude_bins=policy.num_magnitude_bins, **augment_kwargs)
     sub_policies = [apply_sub_policy(sub_policy) for sub_policy in sub_policies]
-    policy_id = operation_idx_random_choice(len(sub_policies), 1, seed)
+    policy_id = fn.random.uniform(values=list(range(len(sub_policies))), seed=seed)
     return select(sub_policies, policy_id, op_kwargs)
 
 

dali/python/nvidia/dali/auto_aug/core/utils.py

@@ -12,8 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from nvidia.dali import fn
-
 try:
     import numpy as np
 except ImportError:
@@ -33,16 +31,6 @@ def remap_bin_idx(bin_idx):
     return np.array([remap_bin_idx(bin_idx) for bin_idx in range(2 * len(magnitudes))])
 
 
-def operation_idx_random_choice(num_total_ops, num_levels=1, rng_seed=42):
-    shape = tuple() if num_levels == 1 else (num_levels, )
-    rng = np.random.default_rng(rng_seed)
-
-    def random_choice(_):
-        return rng.choice(range(num_total_ops), shape)
-
-    return fn.external_source(source=random_choice, batch=False)
-
-
 def split_samples_among_ops(op_range_lo, op_range_hi, ops, selected_op_idx, op_kwargs):
     assert op_range_lo <= op_range_hi
     if op_range_lo == op_range_hi:

dali/python/nvidia/dali/auto_aug/rand_augment.py

@@ -15,7 +15,7 @@
 from nvidia.dali import fn
 from nvidia.dali import types
 from nvidia.dali.auto_aug import augmentations as a
-from nvidia.dali.auto_aug.core.utils import operation_idx_random_choice, select
+from nvidia.dali.auto_aug.core.utils import select
 
 rand_augment_suite = {
     "shear_x": a.shear_x.augmentation((0, 0.3), True),
@@ -152,12 +152,12 @@ def apply_rand_augment(augmentations, samples, n, m, num_magnitude_bins, seed, a
         return samples
     augment_kwargs = augment_kwargs or {}
     use_signed_magnitudes = any(aug.randomly_negate for aug in augmentations)
+    shape = tuple() if n == 1 else (n, )
     if not use_signed_magnitudes:
         random_sign = None
     else:
-        random_sign = fn.random.uniform(range=[0, 1], dtype=types.INT32, seed=seed,
-                                        shape=tuple() if n == 1 else (n, ))
-    op_idx = operation_idx_random_choice(len(augmentations), n, seed)
+        random_sign = fn.random.uniform(values=[0, 1], seed=seed, shape=shape)
+    op_idx = fn.random.uniform(values=list(range(len(augmentations))), seed=seed, shape=shape)
     for level_idx in range(n):
         if not use_signed_magnitudes or n == 1:
             level_random_sign = random_sign

dali/python/nvidia/dali/auto_aug/trivial_augment.py

@@ -15,7 +15,7 @@
 from nvidia.dali import fn
 from nvidia.dali import types
 from nvidia.dali.auto_aug import augmentations as a
-from nvidia.dali.auto_aug.core.utils import operation_idx_random_choice, select
+from nvidia.dali.auto_aug.core.utils import select
 
 trivial_augment_wide_suite = {
     "shear_x": a.shear_x.augmentation((0, 0.99), True),
@@ -129,5 +129,5 @@ def apply_trivial_augment(augmentations, samples, num_magnitude_bins, seed, augm
     op_kwargs = dict(samples=samples, magnitude_bin_idx=magnitude_bin_idx,
                      num_magnitude_bins=num_magnitude_bins, random_sign=random_sign,
                      **augment_kwargs)
-    op_idx = operation_idx_random_choice(len(augmentations), 1, seed)
+    op_idx = fn.random.uniform(values=list(range(len(augmentations))), seed=seed)
     return select(augmentations, op_idx, op_kwargs)