Refactor tests.gpuarray to use NumPy Generator

Author: kc611

tests/gpuarray/check_dnn_conv.py

@@ -612,7 +612,6 @@ def get_atol_rtol(self, algo, dtype, precision):
         return None, None
 
     def __init__(self):
-        utt.seed_rng(1234)
         self.dtype_configs = cudnn.get_supported_dtype_configs(
             check_dtype_config_support
         )

tests/gpuarray/test_basic_ops.py

@@ -35,14 +35,13 @@
     TestJoinAndSplit,
     TestReshape,
 )
-from tests.tensor.utils import rand, safe_make_node
+from tests.tensor.utils import random, safe_make_node
 
 
 pygpu = pytest.importorskip("pygpu")
 gpuarray = pygpu.gpuarray
 
-utt.seed_rng()
-rng = np.random.RandomState(seed=utt.fetch_seed())
+rng = np.random.default_rng(seed=utt.fetch_seed())
 
 
 def inplace_func(
@@ -79,7 +78,7 @@ def fake_shared(value, name=None, strict=False, allow_downcast=None, **kwargs):
 
 
 def rand_gpuarray(*shape, **kwargs):
-    r = rng.rand(*shape) * 2 - 1
+    r = rng.random(shape) * 2 - 1
     dtype = kwargs.pop("dtype", aesara.config.floatX)
     cls = kwargs.pop("cls", None)
     if len(kwargs) != 0:
@@ -219,7 +218,7 @@ def test_transfer_cpu_gpu():
     a = fmatrix("a")
     g = GpuArrayType(dtype="float32", broadcastable=(False, False))("g")
 
-    av = np.asarray(rng.rand(5, 4), dtype="float32")
+    av = np.asarray(rng.random((5, 4)), dtype="float32")
     gv = gpuarray.array(av, context=get_context(test_ctx_name))
 
     f = aesara.function([a], GpuFromHost(test_ctx_name)(a))
@@ -236,7 +235,7 @@ def test_transfer_gpu_gpu():
         dtype="float32", broadcastable=(False, False), context_name=test_ctx_name
     )()
 
-    av = np.asarray(rng.rand(5, 4), dtype="float32")
+    av = np.asarray(rng.random((5, 4)), dtype="float32")
     gv = gpuarray.array(av, context=get_context(test_ctx_name))
     mode = mode_with_gpu.excluding(
         "cut_gpua_host_transfers", "local_cut_gpua_host_gpua"
@@ -256,7 +255,7 @@ def test_transfer_strided():
     a = fmatrix("a")
     g = GpuArrayType(dtype="float32", broadcastable=(False, False))("g")
 
-    av = np.asarray(rng.rand(5, 8), dtype="float32")
+    av = np.asarray(rng.random((5, 8)), dtype="float32")
     gv = gpuarray.array(av, context=get_context(test_ctx_name))
 
     av = av[:, ::2]
@@ -283,14 +282,14 @@ def gpu_alloc_expected(x, *shp):
     op=lambda *args: alloc(*args) + 1,
     gpu_op=GpuAlloc(test_ctx_name),
     cases=dict(
-        correct01=(rand(), np.int32(7)),
+        correct01=(random(), np.int32(7)),
         # just gives a DeepCopyOp with possibly wrong results on the CPU
-        # correct01_bcast=(rand(1), np.int32(7)),
-        correct02=(rand(), np.int32(4), np.int32(7)),
-        correct12=(rand(7), np.int32(4), np.int32(7)),
-        correct13=(rand(7), np.int32(2), np.int32(4), np.int32(7)),
-        correct23=(rand(4, 7), np.int32(2), np.int32(4), np.int32(7)),
-        bad_shape12=(rand(7), np.int32(7), np.int32(5)),
+        # correct01_bcast=(random(1), np.int32(7)),
+        correct02=(random(), np.int32(4), np.int32(7)),
+        correct12=(random(7), np.int32(4), np.int32(7)),
+        correct13=(random(7), np.int32(2), np.int32(4), np.int32(7)),
+        correct23=(random(4, 7), np.int32(2), np.int32(4), np.int32(7)),
+        bad_shape12=(random(7), np.int32(7), np.int32(5)),
     ),
 )
 
@@ -357,7 +356,7 @@ def test_shape():
 def test_gpu_contiguous():
     a = fmatrix("a")
     i = iscalar("i")
-    a_val = np.asarray(np.random.rand(4, 5), dtype="float32")
+    a_val = np.asarray(np.random.random(4, 5), dtype="float32")
     # The reshape is needed otherwise we make the subtensor on the CPU
     # to transfer less data.
     f = aesara.function(
@@ -390,7 +389,6 @@ def setup_method(self):
 
 class TestGPUComparison(TestComparison):
     def setup_method(self):
-        utt.seed_rng()
         self.mode = mode_with_gpu
         self.shared = gpuarray_shared_constructor
         self.dtypes = ["float64", "float32"]
@@ -415,8 +413,8 @@ def shared(x, **kwargs):
     def test_gpusplit_opt(self):
         # Test that we move the node to the GPU
         # Also test float16 computation at the same time.
-        rng = np.random.RandomState(seed=utt.fetch_seed())
-        m = self.shared(rng.rand(4, 6).astype("float16"))
+        rng = np.random.default_rng(seed=utt.fetch_seed())
+        m = self.shared(rng.random((4, 6)).astype("float16"))
         o = Split(2)(m, 0, [2, 2])
         assert o[0].dtype == "float16"
         f = aesara.function([], o, mode=self.mode)
@@ -433,9 +431,9 @@ def test_gpusplit_opt(self):
 
 def test_gpujoin_gpualloc():
     a = fmatrix("a")
-    a_val = np.asarray(np.random.rand(4, 5), dtype="float32")
+    a_val = np.asarray(np.random.random(4, 5), dtype="float32")
     b = fmatrix("b")
-    b_val = np.asarray(np.random.rand(3, 5), dtype="float32")
+    b_val = np.asarray(np.random.random(3, 5), dtype="float32")
 
     f = aesara.function(
         [a, b],
@@ -514,9 +512,9 @@ def test_hostfromgpu_shape_i():
     )
     a = fmatrix("a")
     ca = aesara.gpuarray.type.GpuArrayType("float32", (False, False))()
-    av = np.asarray(np.random.rand(5, 4), dtype="float32")
+    av = np.asarray(np.random.random(5, 4), dtype="float32")
     cv = gpuarray.asarray(
-        np.random.rand(5, 4), dtype="float32", context=get_context(test_ctx_name)
+        np.random.random(5, 4), dtype="float32", context=get_context(test_ctx_name)
     )
 
     f = aesara.function([a], GpuFromHost(test_ctx_name)(a), mode=m)
@@ -583,12 +581,11 @@ def check_u(m, k=0):
         assert result.dtype == np.dtype(dtype)
         assert any([isinstance(node.op, GpuTri) for node in f.maker.fgraph.toposort()])
 
-    utt.seed_rng()
-    test_rng = np.random.RandomState(seed=utt.fetch_seed())
+    test_rng = np.random.default_rng(seed=utt.fetch_seed())
 
     for dtype in ["float64", "float32", "float16"]:
         # try a big one
-        m = np.asarray(test_rng.rand(5000, 5000) * 2 - 1, dtype=dtype)
+        m = np.asarray(test_rng.random((5000, 5000)) * 2 - 1, dtype=dtype)
         check_l(m, 0)
         check_l(m, 1)
         check_l(m, -1)
@@ -597,7 +594,7 @@ def check_u(m, k=0):
         check_u(m, 1)
         check_u(m, -1)
 
-        m = np.asarray(test_rng.rand(10, 10) * 2 - 1, dtype=dtype)
+        m = np.asarray(test_rng.random((10, 10)) * 2 - 1, dtype=dtype)
         check_l(m, 0)
         check_l(m, 1)
         check_l(m, -1)
@@ -606,7 +603,7 @@ def check_u(m, k=0):
         check_u(m, 1)
         check_u(m, -1)
 
-        m = np.asarray(test_rng.rand(10, 5) * 2 - 1, dtype=dtype)
+        m = np.asarray(test_rng.random((10, 5)) * 2 - 1, dtype=dtype)
         check_l(m, 0)
         check_l(m, 1)
         check_l(m, -1)

tests/gpuarray/test_blocksparse.py

@@ -17,7 +17,6 @@
 
 class TestBlockSparseGemvAndOuterGPUarray(TestBlockSparseGemvAndOuter):
     def setup_method(self):
-        utt.seed_rng()
         self.mode = mode_with_gpu.excluding("constant_folding")
         self.gemv_op = gpu_sparse_block_gemv
         self.outer_op = gpu_sparse_block_outer

tests/gpuarray/test_dnn.py

@@ -170,7 +170,6 @@ def test_dnn_conv_merge():
 def test_dnn_conv_inplace():
     # This test that we have inplace work correctly even when
     # GpuAllocEmpty get merged together.
-    utt.seed_rng()
 
     img_shp = [2, 5, 6, 8]
     kern_shp = [3, 5, 5, 6]
@@ -312,7 +311,6 @@ def assert_types(conv):
 
 
 def test_pooling():
-    utt.seed_rng()
 
     modes = get_dnn_pool_modes()
 
@@ -420,7 +418,6 @@ def fn(x):
 # This test will be run with different values of 'mode'
 # (see next test below).
 def run_pooling_with_tensor_vars(mode):
-    utt.seed_rng()
 
     x = tensor4()
     ws = aesara.shared(np.array([2, 2], dtype="int32"))
@@ -473,7 +470,6 @@ def test_pooling_with_tensor_vars():
 @pytest.mark.skipif(dnn.version(raises=False) < 3000, reason=dnn.dnn_available.msg)
 def test_pooling3d():
     # 3d pooling requires version 3 or newer.
-    utt.seed_rng()
 
     # We force the FAST_RUN as we don't want the reference to run in DebugMode.
     mode_without_gpu_ref = aesara.compile.mode.get_mode("FAST_RUN").excluding(
@@ -582,7 +578,6 @@ def fn(x):
 
 
 def test_pooling_opt():
-    utt.seed_rng()
 
     # 2D pooling
     x = matrix()
@@ -654,7 +649,6 @@ def test_pooling_opt():
 def test_pooling_opt_arbitrary_dimensions():
     # test if input with an arbitrary number of non-pooling dimensions
     # is correctly reshaped to run on the GPU
-    utt.seed_rng()
 
     modes = get_dnn_pool_modes()
 
@@ -1081,7 +1075,6 @@ def test_dnn_conv_border_mode():
 
 
 def test_dnn_conv_alpha_output_merge():
-    utt.seed_rng()
 
     img = tensor4()
     kern = tensor4()
@@ -1151,7 +1144,6 @@ def test_dnn_conv_alpha_output_merge():
 
 
 def test_dnn_conv_grad():
-    utt.seed_rng()
 
     b = 1
     c = 4
@@ -1261,7 +1253,6 @@ def run_conv_small_batched_vs_multicall(inputs_shape, filters_shape, batch_sub):
 
     batch_size = inputs_shape[0]
 
-    utt.seed_rng()
     inputs_val = np.random.random(inputs_shape).astype("float32")
     filters_val = np.random.random(filters_shape).astype("float32")
     # Scale down the input values to prevent very large absolute errors
@@ -1311,8 +1302,6 @@ def test_batched_conv3d_small():
 
 
 def test_conv3d_fwd():
-    utt.seed_rng()
-
     def run_conv3d_fwd(
         inputs_shape, filters_shape, subsample, dilation, border_mode, conv_mode
     ):
@@ -1378,8 +1367,6 @@ def run_conv3d_fwd(
 
 
 def test_conv3d_bwd():
-    utt.seed_rng()
-
     def run_conv3d_bwd(
         inputs_shape, filters_shape, subsample, dilation, border_mode, conv_mode
     ):
@@ -1819,7 +1806,6 @@ def test_dnn_maxandargmax_opt():
 
 
 def test_dnn_batchnorm_train():
-    utt.seed_rng()
 
     for mode in ("per-activation", "spatial"):
         for vartype in (
@@ -2022,7 +2008,6 @@ def test_dnn_batchnorm_train():
 
 def test_dnn_batchnorm_train_without_running_averages():
     # compile and run batch_normalization_train without running averages
-    utt.seed_rng()
 
     x, scale, bias, dy = (
         tensor4("x"),
@@ -2096,7 +2081,6 @@ def test_dnn_batchnorm_train_without_running_averages():
 def test_without_dnn_batchnorm_train_without_running_averages():
     # compile and run batch_normalization_train without running averages
     # But disable cudnn and make sure it run on the GPU.
-    utt.seed_rng()
 
     x, scale, bias, dy = (
         tensor4("x"),
@@ -2163,7 +2147,6 @@ def test_without_dnn_batchnorm_train_without_running_averages():
 @utt.assertFailure_fast
 def test_dnn_batchnorm_train_inplace():
     # test inplace_running_mean and inplace_running_var
-    utt.seed_rng()
 
     x, scale, bias = tensor4("x"), tensor4("scale"), tensor4("bias")
     data_shape = (5, 10, 30, 25)
@@ -2218,7 +2201,6 @@ def test_dnn_batchnorm_train_inplace():
 
 
 def test_batchnorm_inference():
-    utt.seed_rng()
 
     for mode in ("per-activation", "spatial"):
         for vartype in (
@@ -2344,7 +2326,6 @@ def test_batchnorm_inference():
 @utt.assertFailure_fast
 def test_batchnorm_inference_inplace():
     # test inplace
-    utt.seed_rng()
 
     x, scale, bias, mean, var = (
         tensor4(n) for n in ("x", "scale", "bias", "mean", "var")
@@ -2460,7 +2441,6 @@ def test_dnn_batchnorm_valid_and_invalid_axes():
 
 
 def test_dnn_rnn_gru():
-    utt.seed_rng()
 
     # test params
     input_dim = 32
@@ -2569,7 +2549,6 @@ def funcs(out, params, hy=None):
 
 
 def test_dnn_rnn_gru_bidi():
-    utt.seed_rng()
 
     # test params
     input_dim = 32
@@ -2630,7 +2609,6 @@ def funcs(out, params, hy=None):
 
 
 def test_dnn_rnn_lstm():
-    utt.seed_rng()
 
     # test params
     input_dim = 32
@@ -2716,7 +2694,6 @@ def funcs(out, params):
 
 
 def test_dnn_rnn_lstm_grad_c():
-    utt.seed_rng()
 
     # test params
     input_dim = 32
@@ -2819,7 +2796,6 @@ def __init__(self, *args, **kwargs):
 
 
 def test_dnn_spatialtf():
-    utt.seed_rng()
 
     """
     Spatial Transformer implementation using Aesara from Lasagne
@@ -3023,7 +2999,6 @@ def try_theta_shp(theta_shp):
 
 
 def test_dnn_spatialtf_grad():
-    utt.seed_rng()
 
     inputs = tensor4("inputs")
     theta = tensor3("theta")
@@ -3097,7 +3072,7 @@ class TestDnnConv2DRuntimeAlgorithms:
     ]
 
     def __init__(self):
-        utt.seed_rng()
+
         self.runtime_algorithms = (
             "time_once",
             "guess_once",
@@ -3286,7 +3261,7 @@ class TestDnnConv3DRuntimeAlgorithms(TestDnnConv2DRuntimeAlgorithms):
 def test_conv_guess_once_with_dtypes():
     # This test checks that runtime conv algorithm selection does not raise any exception
     # when consecutive functions with different dtypes and precisions are executed.
-    utt.seed_rng()
+
     inputs_shape = (2, 3, 5, 5)
     filters_shape = (2, 3, 40, 4)
     border_mode = "full"