!85 add epsilon to random gaussian to avoid zero divisor

Merge pull request !85 from yangsijia/add-epsilon-to-random-gaussian

tests/common/gen_random.py

@@ -28,6 +28,8 @@ from akg.utils.kernel_exec import get_profiling_mode
 
 RANDOM_SEED_NUM = 20
 PROF_ERROR_CODE = 9999999999
+
+
 def func(size_, miu_=0, sigma_=8, seed_=None):
     """
     Select random func according to RANDOM_FUNC_MODE and randint, calculated by the length of the random_func_list.
@@ -59,7 +61,7 @@ def func(size_, miu_=0, sigma_=8, seed_=None):
 
 
 @func_time_required
-def random_gaussian(size, miu=0, sigma=8, seed=None):
+def random_gaussian(size, miu=0, sigma=8, epsilon=0, seed=None):
     """Generate random array with absolution value obeys gaussian distribution."""
     random_data_disk_path = None
     if os.environ.get("RANDOM_DATA_DISK_PATH") is not None:
@@ -93,7 +95,7 @@ def random_gaussian(size, miu=0, sigma=8, seed=None):
                 numbers.extend(func(size_c, miu, sigma, s))
             ret = np.array(numbers)
         ret = ret.flatten()
-        return ret[:size_c].reshape(size)
+        return ret[:size_c].reshape(size) + epsilon
 
     data_len = functools.reduce(lambda x, y: x * y, size)
     data_pool = np.fromfile(random_data_disk_path)
@@ -107,4 +109,8 @@ def random_gaussian(size, miu=0, sigma=8, seed=None):
         np.random.shuffle(data_copy)
         data_copy_list.append(data_copy)
     data_pool = np.concatenate(tuple(data_copy_list), axis=0)
-    return data_pool[0:data_len].reshape(size)
\ No newline at end of file
+    return data_pool[0:data_len].reshape(size) + epsilon
+
+def gen_epsilon(dtype):
+    """Generate suggested epsilon according to data type."""
+    return 1e-7 if dtype == np.float32 else 1e-3

tests/common/test_run/distr_normal_diag_KLdiv_ad_run.py

@@ -16,7 +16,7 @@ import numpy as np
 from tensorio import compare_tensor
 from akg.utils import kernel_exec as utils
 from test_op.prob_program import distr_normal_diag_KLdiv_ad
-from gen_random import random_gaussian
+from gen_random import random_gaussian, gen_epsilon
 from base import get_rtol_atol
 
 
@@ -40,9 +40,8 @@ def gen_data(dtype, shape):
     support_list = {"float16": np.float16, "float32": np.float32}
 
     m, k = shape
-
     mean  = random_gaussian((m, k), miu=1, sigma=0.1).astype(support_list[dtype])
-    scale = random_gaussian((m, k), miu=1, sigma=0.1).astype(support_list[dtype])
+    scale = random_gaussian((m, k), miu=1, sigma=0.1, epsilon=gen_epsilon(dtype)).astype(support_list[dtype])
     head = random_gaussian((m, ), miu=1, sigma=0.1).astype(support_list[dtype])
 
     output1 = np.full((m, k), 0.0, dtype)