[Refactor] Rename asnumpy -> numpy in NDArray (#8083)

apps/android_rpc/tests/android_rpc_test.py

@@ -76,7 +76,7 @@ def test_rpc_module():
     time_f = f2.time_evaluator(f2.entry_name, dev, number=10)
     cost = time_f(a, b).mean
     print("%g secs/op\n" % cost)
-    np.testing.assert_equal(b.asnumpy(), a.asnumpy() + 1)
+    np.testing.assert_equal(b.numpy(), a.numpy() + 1)
 
     # Compile the Graph for OpenCL target
     if test_opencl:
@@ -99,7 +99,7 @@ def test_rpc_module():
         time_f = f1.time_evaluator(f1.entry_name, dev, number=10)
         cost = time_f(a, b).mean
         print("%g secs/op\n" % cost)
-        np.testing.assert_equal(b.asnumpy(), a.asnumpy() + 1)
+        np.testing.assert_equal(b.numpy(), a.numpy() + 1)
 
     # Compile the Graph for Vulkan target
     if test_vulkan:
@@ -122,7 +122,7 @@ def test_rpc_module():
         time_f = f1.time_evaluator(f1.entry_name, dev, number=10)
         cost = time_f(a, b).mean
         print("%g secs/op\n" % cost)
-        np.testing.assert_equal(b.asnumpy(), a.asnumpy() + 1)
+        np.testing.assert_equal(b.numpy(), a.numpy() + 1)
 
 
 if __name__ == "__main__":

apps/extension/tests/test_ext.py

@@ -45,7 +45,7 @@ def check_llvm():
         a = tvm.nd.array(np.random.uniform(size=n).astype(A.dtype), dev)
         b = tvm.nd.array(np.zeros(n, dtype=B.dtype), dev)
         f(a, b)
-        tvm.testing.assert_allclose(b.asnumpy(), a.asnumpy() + 1)
+        tvm.testing.assert_allclose(b.numpy(), a.numpy() + 1)
 
     check_llvm()
 
@@ -92,7 +92,7 @@ def check_llvm():
         a = tvm.nd.array(np.random.uniform(size=n).astype(A.dtype), dev)
         b = tvm.nd.array(np.zeros(n, dtype=B.dtype), dev)
         f(a, b)
-        tvm.testing.assert_allclose(b.asnumpy(), a.asnumpy() + 1)
+        tvm.testing.assert_allclose(b.numpy(), a.numpy() + 1)
 
     check_llvm()
 

apps/howto_deploy/python_deploy.py

@@ -33,8 +33,8 @@ def verify(mod, fname):
     y = tvm.nd.array(np.zeros(N, dtype=np.float32))
     # Invoke the function
     f(x, y)
-    np_x = x.asnumpy()
-    np_y = y.asnumpy()
+    np_x = x.numpy()
+    np_y = y.numpy()
     # Verify correctness of function
     assert np.all([xi + 1 == yi for xi, yi in zip(np_x, np_y)])
     print("Finish verification...")

apps/ios_rpc/tests/ios_rpc_mobilenet.py

@@ -125,7 +125,7 @@ def run(mod, target):
         m.set_input("data", tvm.nd.array(image, dev))
         m.run()
         tvm_output = m.get_output(0)
-        top1 = np.argmax(tvm_output.asnumpy()[0])
+        top1 = np.argmax(tvm_output.numpy()[0])
         print("TVM prediction top-1:", top1, synset[top1])
 
         # evaluate

apps/ios_rpc/tests/ios_rpc_test.py

@@ -96,7 +96,7 @@ def test_rpc_module():
     time_f = f1.time_evaluator(f1.entry_name, dev, number=10)
     cost = time_f(a, b).mean
     print("%g secs/op" % cost)
-    np.testing.assert_equal(b.asnumpy(), a.asnumpy() + 1)
+    np.testing.assert_equal(b.numpy(), a.numpy() + 1)
     # CPU
     dev = remote.cpu(0)
     f2 = remote.load_module("cpu_lib.dylib")
@@ -106,7 +106,7 @@ def test_rpc_module():
     time_f = f2.time_evaluator(f2.entry_name, dev, number=10)
     cost = time_f(a, b).mean
     print("%g secs/op" % cost)
-    np.testing.assert_equal(b.asnumpy(), a.asnumpy() + 1)
+    np.testing.assert_equal(b.numpy(), a.numpy() + 1)
 
 
 def test_rpc_module_with_upload():
@@ -142,7 +142,7 @@ def test_rpc_module_with_upload():
     time_f = f.time_evaluator(f.entry_name, dev, number=10)
     cost = time_f(a, b).mean
     print("%g secs/op" % cost)
-    np.testing.assert_equal(b.asnumpy(), a.asnumpy() + 1)
+    np.testing.assert_equal(b.numpy(), a.numpy() + 1)
 
 
 if __name__ == "__main__":

apps/topi_recipe/broadcast/test_broadcast_map.py

@@ -69,7 +69,7 @@ def test_broadcast_to(in_shape, out_shape):
     out_nd = tvm.nd.array(np.empty(out_shape).astype(B.dtype), tvm.cuda())
     for _ in range(2):
         fcuda(data_nd, out_nd)
-    tvm.testing.assert_allclose(out_nd.asnumpy(), out_npy)
+    tvm.testing.assert_allclose(out_nd.numpy(), out_npy)
 
 
 def test_broadcast_binary_op(lhs_shape, rhs_shape, typ="add"):
@@ -121,7 +121,7 @@ def test_broadcast_binary_op(lhs_shape, rhs_shape, typ="add"):
     out_nd = tvm.nd.array(np.empty(out_npy.shape).astype(B.dtype), tvm.cuda())
     for _ in range(2):
         fcuda(lhs_nd, rhs_nd, out_nd)
-    tvm.testing.assert_allclose(out_nd.asnumpy(), out_npy)
+    tvm.testing.assert_allclose(out_nd.numpy(), out_npy)
 
 
 if __name__ == "__main__":

apps/topi_recipe/conv/depthwise_conv2d_test.py

@@ -146,11 +146,9 @@ def check_device(device):
                 depthwise_conv2d_scipy[:, c, :, :] * scale_np[c] + shift_np[c]
             )
         relu_scipy = np.maximum(scale_shift_scipy, 0)
-        tvm.testing.assert_allclose(
-            depthwise_conv2d_tvm.asnumpy(), depthwise_conv2d_scipy, rtol=1e-5
-        )
-        tvm.testing.assert_allclose(scale_shift_tvm.asnumpy(), scale_shift_scipy, rtol=1e-5)
-        tvm.testing.assert_allclose(relu_tvm.asnumpy(), relu_scipy, rtol=1e-5)
+        tvm.testing.assert_allclose(depthwise_conv2d_tvm.numpy(), depthwise_conv2d_scipy, rtol=1e-5)
+        tvm.testing.assert_allclose(scale_shift_tvm.numpy(), scale_shift_scipy, rtol=1e-5)
+        tvm.testing.assert_allclose(relu_tvm.numpy(), relu_scipy, rtol=1e-5)
         print("success")
 
     for device in ["cuda", "opencl", "rocm"]:
@@ -253,11 +251,9 @@ def check_device(device):
                 depthwise_conv2d_scipy[:, :, :, c] * scale_np[c] + shift_np[c]
             )
         relu_scipy = np.maximum(scale_shift_scipy, 0)
-        tvm.testing.assert_allclose(
-            depthwise_conv2d_tvm.asnumpy(), depthwise_conv2d_scipy, rtol=1e-5
-        )
-        tvm.testing.assert_allclose(scale_shift_tvm.asnumpy(), scale_shift_scipy, rtol=1e-5)
-        tvm.testing.assert_allclose(relu_tvm.asnumpy(), relu_scipy, rtol=1e-5)
+        tvm.testing.assert_allclose(depthwise_conv2d_tvm.numpy(), depthwise_conv2d_scipy, rtol=1e-5)
+        tvm.testing.assert_allclose(scale_shift_tvm.numpy(), scale_shift_scipy, rtol=1e-5)
+        tvm.testing.assert_allclose(relu_tvm.numpy(), relu_scipy, rtol=1e-5)
         print("success")
 
     for device in ["cuda", "opencl", "rocm"]:

apps/topi_recipe/conv/test_conv2d_hwcn_map.py

@@ -88,10 +88,10 @@ def check_device(device):
         ):
             func1 = tvm.build(s1, [A, W, B], device)
             func1(a, w, b)
-            tvm.testing.assert_allclose(b.asnumpy(), b_np, rtol=1e-5)
+            tvm.testing.assert_allclose(b.numpy(), b_np, rtol=1e-5)
             func2 = tvm.build(s2, [A, W, C], device)
             func2(a, w, c)
-            tvm.testing.assert_allclose(c.asnumpy(), c_np, rtol=1e-5)
+            tvm.testing.assert_allclose(c.numpy(), c_np, rtol=1e-5)
 
     for device in ["cuda", "opencl", "rocm"]:
         check_device(device)

apps/topi_recipe/conv/test_conv_int8_arm.py

@@ -186,9 +186,9 @@ def run_inference(
 
         # Functional check
         if data_dtype == "uint8":
-            np.testing.assert_equal(c_orig.asnumpy(), c_sch.asnumpy())
+            np.testing.assert_equal(c_orig.numpy(), c_sch.numpy())
         else:
-            assert np.allclose(c_orig.asnumpy(), c_sch.asnumpy())
+            assert np.allclose(c_orig.numpy(), c_sch.numpy())
 
         evaluator = func.time_evaluator(func.entry_name, DEV, number=1000)
         LOGGER.debug(tvm.lower(sconv, [data, kernel], simple_mode=True))

apps/topi_recipe/conv/test_conv_int8_intel.py

@@ -172,9 +172,9 @@ def run_inference(
 
         # Functional check
         if data_dtype == "uint8":
-            np.testing.assert_equal(c_orig.asnumpy(), c_sch.asnumpy())
+            np.testing.assert_equal(c_orig.numpy(), c_sch.numpy())
         else:
-            assert np.allclose(c_orig.asnumpy(), c_sch.asnumpy())
+            assert np.allclose(c_orig.numpy(), c_sch.numpy())
 
         evaluator = func.time_evaluator(func.entry_name, DEV, number=1000)
         LOGGER.debug(tvm.lower(sconv, [data, kernel], simple_mode=True))

apps/topi_recipe/gemm/android_gemm_square.py

@@ -51,7 +51,7 @@ def evaluate(func, dev, N, times):
     cost = time_f(a, b, c).mean
     gf = ngflops(N) / cost
     print("%g secs/op, %g GFLOPS" % (cost, gf))
-    np.testing.assert_almost_equal(c.asnumpy(), a_np.dot(b_np), decimal=2)
+    np.testing.assert_almost_equal(c.numpy(), a_np.dot(b_np), decimal=2)
 
 
 def test_gemm_gpu(N, times, bn, num_block, num_thread):

apps/topi_recipe/gemm/cuda_gemm_square.py

@@ -136,7 +136,7 @@ def check_device(device):
         c = tvm.nd.array(np.zeros((n, m), dtype=C.dtype), dev)
         for i in range(2):
             f(a, b, c)
-        tvm.testing.assert_allclose(c.asnumpy(), np.dot(b_np.T, a_np), rtol=1e-5)
+        tvm.testing.assert_allclose(c.numpy(), np.dot(b_np.T, a_np), rtol=1e-5)
 
         num_flops = 2 * nn * nn * nn
         num_runs = 10

apps/topi_recipe/gemm/gemm_int8.py

@@ -171,7 +171,7 @@ def block_size_filter(entity):
     f(a, b, c)
 
     tvm.testing.assert_allclose(
-        c.asnumpy(), np.dot(a_np.astype("int32"), b_np.T.astype("int32")), rtol=1e-5
+        c.numpy(), np.dot(a_np.astype("int32"), b_np.T.astype("int32")), rtol=1e-5
     )
 
     num_ops = 2 * l * m * n

apps/topi_recipe/reduce/test_reduce_map.py

@@ -83,7 +83,7 @@ def test_reduce_map(in_shape, axis, keepdims, type="sum", test_id=0):
 
     for _ in range(2):
         fcuda(data_tvm, out_tvm)
-    tvm.testing.assert_allclose(out_tvm.asnumpy(), out_npy, rtol=4e-4, atol=4e-4)
+    tvm.testing.assert_allclose(out_tvm.numpy(), out_npy, rtol=4e-4, atol=4e-4)
 
 
 if __name__ == "__main__":

apps/topi_recipe/rnn/matexp.py

@@ -160,7 +160,7 @@ def check_device(target):
         print("Time cost=%g" % tgap)
         # correctness
         if not SKIP_CHECK:
-            res_cuda = res_a.asnumpy()
+            res_cuda = res_a.numpy()
             res_cmp = np.ones_like(res_np).astype("float64")
             Whh_np = Whh_np.astype("float64")
             for t in range(1, n_num_step):

docs/deploy/hls.rst

@@ -79,7 +79,7 @@ We use two python scripts for this tutorial.
       c = tvm.nd.array(np.zeros(n, dtype="float32"), dev)
 
       fadd(a, b, c)
-      tvm.testing.assert_allclose(c.asnumpy(), a.asnumpy() + b.asnumpy())
+      tvm.testing.assert_allclose(c.numpy(), a.numpy() + b.numpy())
 
 
 Setup

docs/dev/codebase_walkthrough.rst

@@ -169,7 +169,7 @@ The returned module, which can be thought of as a combination of a compiled func
    b = tvm.nd.array(np.random.uniform(size=n).astype(B.dtype), dev)
    c = tvm.nd.array(np.zeros(n, dtype=C.dtype), dev)
    fadd(a, b, c)
-   output = c.asnumpy()
+   output = c.numpy()
 
 Under the hood, TVM allocates device memory and manages memory transfers automatically. To do that, each backend needs to subclass ``DeviceAPI`` class, defined in ``include/tvm/runtime/device_api.h``, and override memory management methods to use device specific API. For example, the CUDA backend implements ``CUDADeviceAPI`` in ``src/runtime/cuda/cuda_device_api.cc`` to use ``cudaMalloc``, ``cudaMemcpy`` etc.
 

docs/dev/debugger.rst

@@ -146,7 +146,7 @@ How to use Debugger?
     m.set_input(**params)
     # execute
     m.run()
-    tvm_out = m.get_output(0, tvm.nd.empty(out_shape, dtype)).asnumpy()
+    tvm_out = m.get_output(0, tvm.nd.empty(out_shape, dtype)).numpy()
 
 The outputs are dumped to a temporary folder in ``/tmp`` folder or the
 folder specified while creating the runtime.

docs/dev/index.rst

@@ -147,7 +147,7 @@ The main goal of TVM's runtime is to provide a minimal API for loading and execu
     arr: tvm.runtime.NDArray = tvm.nd.array([1, 2, 3], device=tvm.cuda(0))
     fun: tvm.runtime.PackedFunc = mod["addone"]
     fun(a)
-    print(a.asnumpy())
+    print(a.numpy())
 
 
 :py:class:`tvm.runtime.Module` encapsulates the result of compilation. A runtime.Module contains a GetFunction method to obtain PackedFuncs by name.
@@ -172,7 +172,7 @@ The above example only deals with a simple `addone` function. The code snippet b
    # execute the model
    gmod["run"]()
    # get the output
-   result = gmod["get_output"](0).asnumpy()
+   result = gmod["get_output"](0).numpy()
 
 The main take away is that runtime.Module and runtime.PackedFunc are sufficient to encapsulate both operator level programs (such as addone), as well as the end-to-end models.
 

docs/langref/relay_pattern.rst

@@ -446,7 +446,7 @@ with a single batch_norm op:
             beta = node_map[self.beta][0]
             gamma = node_map[self.gamma][0]
             eps = node_map[self.eps][0]
-            return relay.op.nn.batch_norm(x, gamma, beta, mean, var, epsilon = eps.data.asnumpy().item())[0]
+            return relay.op.nn.batch_norm(x, gamma, beta, mean, var, epsilon = eps.data.numpy().item())[0]
 
         # A graph of arithmetic operators that are functional equivalent to batch_norm.
         x = relay.var('x')

python/tvm/auto_scheduler/search_task.py

@@ -234,7 +234,7 @@ def _save_buffer_to_file(buffer_name, buffer_data):
 
     File name will be: {buffer_name}.{buffer_shape}_{buffer_data_type}.npy
     """
-    np_data = buffer_data.asnumpy()
+    np_data = buffer_data.numpy()
 
     buffer_name += "."
     for i in np_data.shape:

python/tvm/contrib/sparse.py

@@ -16,6 +16,7 @@
 # under the License.
 """Tensor and Operation class for computation declaration."""
 # pylint: disable=invalid-name
+import warnings
 import numpy as _np
 from tvm.runtime import ndarray as _nd
 from tvm import te
@@ -81,11 +82,21 @@ def __init__(self, arg1, device=None, shape=None):
         )
 
     def asnumpy(self):
+        """Construct a full matrix and convert it to numpy array. This API will be deprecated
+        in TVM v0.8 release. Please use `numpy` instead."""
+        warnings.warn(
+            "CSRNDArray.asnumpy() will be deprecated in TVM v0.8 release. "
+            "Please use CSRNDArray.numpy() instead.",
+            DeprecationWarning,
+        )
+        return self.numpy()
+
+    def numpy(self):
         """Construct a full matrix and convert it to numpy array."""
         full = _np.zeros(self.shape, self.dtype)
-        ridx = _np.diff(self.indptr.asnumpy())
+        ridx = _np.diff(self.indptr.numpy())
         ridx = _np.hstack((_np.ones((v,), itype) * i for i, v in enumerate(ridx)))
-        full[ridx, self.indices.asnumpy().astype(itype)] = self.data.asnumpy()
+        full[ridx, self.indices.numpy().astype(itype)] = self.data.numpy()
         return full
 
 

python/tvm/contrib/target/coreml.py

@@ -69,7 +69,7 @@ def _convert_softmax(builder, name, inputs, outputs, args, attrs):
 
 
 def _convert_conv2d(builder, name, inputs, outputs, args, attrs):
-    weight = args[1].data.asnumpy()
+    weight = args[1].data.numpy()
     if attrs["kernel_layout"] == "OIHW":
         # convert to 'HWIO'
         weight = weight.transpose([2, 3, 1, 0])
@@ -169,7 +169,7 @@ def visit_constant(self, const):
         self.builder.add_load_constant_nd(
             name=output,
             output_name=output,
-            constant_value=const.data.asnumpy(),
+            constant_value=const.data.numpy(),
             shape=const.data.shape,
         )
         self.buf_idx_ = self.buf_idx_ + 1

python/tvm/contrib/target/onnx.py

@@ -684,7 +684,7 @@ def _get_opsets(self):
         return opsets
 
     def make_model(self):
-        """ Creates the onnx model from the graph """
+        """Creates the onnx model from the graph"""
         onnx_graph = onnx.helper.make_graph(
             self._nodes, self._name, self._inputs, self._outputs, self._initializers
         )
@@ -734,7 +734,7 @@ def _get_node_entry(cls, relay_node, name):
         }
 
     def convert_to_onnx(self, func):
-        """ Traverse Relay graph and generate a ONNX model"""
+        """Traverse Relay graph and generate a ONNX model"""
 
         self.visit(func)
         self._add_output(self._node_dict[self.last_node])
@@ -826,7 +826,7 @@ def _add_params(self, node_entry, idx):
             param_name in self._params
         ), "The parameter {0} is not present" "in params dict provided.".format(param_name)
         value = self._params[param_name]
-        numpy_array = value.asnumpy()
+        numpy_array = value.numpy()
         tensor = numpy_helper.from_array(numpy_array, param_name)
         self._mc.add_initializers([tensor])
         dtype = onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[numpy_array.dtype]

python/tvm/driver/tvmc/runner.py

@@ -42,7 +42,7 @@
 
 @register_parser
 def add_run_parser(subparsers):
-    """ Include parser for 'run' subcommand """
+    """Include parser for 'run' subcommand"""
 
     parser = subparsers.add_parser("run", help="run a compiled module")
     parser.set_defaults(func=drive_run)
@@ -420,6 +420,6 @@ def run_module(
     outputs = {}
     for i in range(num_outputs):
         output_name = "output_{}".format(i)
-        outputs[output_name] = module.get_output(i).asnumpy()
+        outputs[output_name] = module.get_output(i).numpy()
 
     return TVMCResult(outputs, times)

python/tvm/relay/analysis/sparse_conv2d.py

@@ -85,7 +85,7 @@ def process_params(expr, params, block_size, sparsity_threshold, layout):
     weight_names = _search_conv2d_op_weight(expr)
     for name in weight_names:
         name = str(name)
-        w_np = params[name].asnumpy()
+        w_np = params[name].numpy()
         # currently only support conv2d_1*1
         if not (
             (w_np.shape[0] == 1 and w_np.shape[1] == 1)

python/tvm/relay/analysis/sparse_dense.py

@@ -83,7 +83,7 @@ def process_params(expr, params, block_size, sparsity_threshold):
     weight_names = _search_dense_op_weight(expr)
     for name in weight_names:
         name = str(name)
-        w_np = params[name].asnumpy()
+        w_np = params[name].numpy()
         sparsity = 1.0 - (np.count_nonzero(w_np) / w_np.size)
         if sparsity >= sparsity_threshold:
             sparse_weight = sp.bsr_matrix(w_np, blocksize=block_size)