[OpenVINO backend] support categorical

Author: Mohamed-Ashraf273

keras/src/backend/openvino/excluded_concrete_tests.txt

@@ -188,3 +188,58 @@ CoreOpsDtypeTest::test_convert_to_tensor12
 CoreOpsDtypeTest::test_convert_to_tensor14
 CoreOpsDtypeTest::test_convert_to_tensor25
 CoreOpsDtypeTest::test_convert_to_tensor37
+RandomCorrectnessTest::test_beta0
+RandomCorrectnessTest::test_beta1
+RandomCorrectnessTest::test_beta2
+RandomCorrectnessTest::test_binomial0
+RandomCorrectnessTest::test_binomial1
+RandomCorrectnessTest::test_binomial2
+RandomCorrectnessTest::test_dropout
+RandomCorrectnessTest::test_dropout_noise_shape
+RandomCorrectnessTest::test_gamma0
+RandomCorrectnessTest::test_gamma1
+RandomCorrectnessTest::test_gamma2
+RandomCorrectnessTest::test_randint0
+RandomCorrectnessTest::test_randint1
+RandomCorrectnessTest::test_randint2
+RandomCorrectnessTest::test_randint3
+RandomCorrectnessTest::test_randint4
+RandomCorrectnessTest::test_shuffle
+RandomCorrectnessTest::test_truncated_normal0
+RandomCorrectnessTest::test_truncated_normal1
+RandomCorrectnessTest::test_truncated_normal2
+RandomCorrectnessTest::test_truncated_normal3
+RandomCorrectnessTest::test_truncated_normal4
+RandomCorrectnessTest::test_truncated_normal5
+RandomCorrectnessTest::test_uniform0
+RandomCorrectnessTest::test_uniform1
+RandomCorrectnessTest::test_uniform2
+RandomCorrectnessTest::test_uniform3
+RandomCorrectnessTest::test_uniform4
+RandomBehaviorTest::test_beta_tf_data_compatibility
+RandomDTypeTest::test_beta_bfloat16
+RandomDTypeTest::test_beta_float16
+RandomDTypeTest::test_beta_float32
+RandomDTypeTest::test_beta_float64
+RandomDTypeTest::test_binomial_bfloat16
+RandomDTypeTest::test_binomial_float16
+RandomDTypeTest::test_binomial_float32
+RandomDTypeTest::test_binomial_float64
+RandomDTypeTest::test_dropout_bfloat16
+RandomDTypeTest::test_dropout_float16
+RandomDTypeTest::test_dropout_float32
+RandomDTypeTest::test_dropout_float64
+RandomDTypeTest::test_gamma_bfloat16
+RandomDTypeTest::test_gamma_float16
+RandomDTypeTest::test_gamma_float32
+RandomDTypeTest::test_gamma_float64
+RandomDTypeTest::test_normal_bfloat16
+RandomDTypeTest::test_randint_int16
+RandomDTypeTest::test_randint_int32
+RandomDTypeTest::test_randint_int64
+RandomDTypeTest::test_randint_int8
+RandomDTypeTest::test_randint_uint16
+RandomDTypeTest::test_randint_uint32
+RandomDTypeTest::test_randint_uint8
+RandomDTypeTest::test_truncated_normal_bfloat16
+RandomDTypeTest::test_uniform_bfloat16

keras/src/backend/openvino/excluded_tests.txt

@@ -33,7 +33,7 @@ keras/src/ops/math_test.py
 keras/src/ops/nn_test.py
 keras/src/optimizers
 keras/src/quantizers
-keras/src/random
+keras/src/random/seed_generator_test.py
 keras/src/regularizers
 keras/src/saving
 keras/src/trainers

keras/src/backend/openvino/random.py

@@ -39,9 +39,36 @@ def uniform(shape, minval=0.0, maxval=1.0, dtype=None, seed=None):
 
 
 def categorical(logits, num_samples, dtype="int64", seed=None):
-    raise NotImplementedError(
-        "`categorical` is not supported with openvino backend"
+    if isinstance(logits, OpenVINOKerasTensor):
+        logits = convert_to_numpy(logits)
+    assert isinstance(logits, np.ndarray), (
+        "logits must be a numpy array or an OpenVINOKerasTensor, "
+        "got: {}".format(type(logits))
     )
+    # Compute probabilities
+    probs = np.exp(logits - np.max(logits, axis=-1, keepdims=True))
+    probs = probs / np.sum(probs, axis=-1, keepdims=True)
+    rng = np.random.default_rng(draw_seed(seed).data)
+    # Batched sampling
+    if probs.ndim == 1:
+        samples = rng.choice(
+            logits.shape[-1],
+            size=(num_samples,),
+            p=probs,
+            replace=True,
+        )
+        samples = samples[None, :]  # Add batch dim for consistency
+    else:
+        samples = np.empty((probs.shape[0], num_samples), dtype=dtype)
+        for i in range(probs.shape[0]):
+            samples[i] = rng.choice(
+                logits.shape[-1],
+                size=(num_samples,),
+                p=probs[i],
+                replace=True,
+            )
+    samples = samples.astype(dtype)
+    return OpenVINOKerasTensor(ov_opset.constant(samples).output(0))
 
 
 def randint(shape, minval, maxval, dtype="int32", seed=None):