Merge pull request #1 from confused666/master

测试接口mint.any、mint.max、mint.mean、mint.median、mint.min

Author: lvyufeng

test/test_any.py

@@ -0,0 +1,122 @@
+import pytest
+import numpy as np
+import mindspore as ms
+from mindspore import mint, Tensor, value_and_grad
+import torch
+
+
+dtype_ms = ms.float32
+dtype_torch = torch.float32
+input_data = [[1,0],[0,0],[1,1]]
+ms_tensor = Tensor(input_data, dtype_ms)
+torch_tensor = torch.tensor(input_data, dtype=dtype_torch)
+
+
+def is_same(input_data=[[1,0],[0,0],[1,1]], shape=None, dtype_ms=ms.float32, dtype_torch=torch.float32, dim=-1, keepdim=False):
+    if shape != None:
+        input_data = np.random.randn(*shape)
+
+    ms_tensor = Tensor(input_data, dtype_ms)
+    torch_tensor = torch.tensor(input_data, dtype=dtype_torch)
+
+    ms_result = mint.any(ms_tensor, dim=dim, keepdim=keepdim).asnumpy()
+    torch_result = torch.any(torch_tensor, dim=dim, keepdim=keepdim).numpy()
+
+    return np.allclose(ms_result, torch_result)
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_any_different_dtypes(mode):
+    """测试random输入不同dtype，对比两个框架的支持度"""
+    ms.set_context(mode=mode)
+    ms_dtypes = [ms.int8, ms.int16, ms.int32, ms.int64, ms.uint8,ms.uint16, ms.uint32, ms.uint64, ms.float16, ms.float32, ms.float64, ms.bfloat16, ms.bool_]
+    torch_dtypes = [torch.int8, torch.int16, torch.int32, torch.int64, torch.uint8, torch.uint16, torch.uint32, torch.uint64, torch.float16, torch.float32, torch.float64, torch.bfloat16, torch.bool]
+    
+    for i in range(len(ms_dtypes)):
+        dtype_ms = ms_dtypes[i]
+        dtype_torch = torch_dtypes[i]
+
+        ms_tensor = Tensor(input_data, dtype_ms)
+        torch_tensor = torch.tensor(input_data, dtype=dtype_torch)
+
+        err = False
+        try:
+            ms_result = mint.any(ms_tensor, dim=1).asnumpy()
+        except Exception as e:
+            err = True
+            print(f"mint.any not supported for {dtype_ms}")
+
+        try:
+            torch_result = torch.any(torch_tensor, dim=1).numpy()
+        except Exception as e:
+            err = True
+            print(f"torch.any not supported for {dtype_torch}")
+
+        if not err:
+            assert np.allclose(ms_result, torch_result)
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_any_random_input_fixed_dtype(mode):
+    """测试固定数据类型下的随机输入值，对比输出误差"""
+    ms.set_context(mode=mode)
+
+    shapes = [[5], [5, 2], [5, 4, 3], [4, 6, 7, 8]]
+    for i in range(len(shapes)):
+        shape = shapes[i]
+        result = is_same(shape=shape)
+        assert result
+    
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_any_different_para(mode):
+    """测试固定shape，固定输入值，不同输入参数（string\bool等类型），两个框架的支持度"""
+    ms.set_context(mode=mode)
+
+    dims = [None, 0, 1]
+    keepdims = [True, False]
+    paras = [(dim, keepdim) for dim in dims for keepdim in keepdims]
+
+    for dim, keepdim in paras:
+        result = is_same(dim=dim, keepdim=keepdim)
+        assert result
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_any_wrong_input(mode):
+    """测试随机混乱输入，报错信息的准确性"""
+    ms.set_context(mode=mode)
+    
+    try:
+        ms_result = mint.any(ms_tensor, dim=1, keepdim=1).asnumpy()
+    except Exception as e:
+        print(f"keepdim 不是 bool 类型报错信息：\n{e}")
+
+    try:
+        ms_result = mint.any(input_data, dim=1).asnumpy()
+    except Exception as e:
+        print(f"input 不是 Tensor 报错信息：\n{e}")
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_any_forward_back(mode):
+    """使用Pytorch和MindSpore, 固定输入和权重, 测试正向推理结果和反向梯度"""
+    ms.set_context(mode=mode)
+    torch_tensor = torch.tensor(input_data, dtype=dtype_torch, requires_grad=True)
+
+    def forward_pt(x):
+        return torch.any(x, dim=1, keepdim=False)
+
+    def forward_ms(x):
+        return mint.any(x, dim=1, keepdim=False)
+
+    grad_fn = value_and_grad(forward_ms)
+    output_ms, gradient_ms = grad_fn(ms_tensor)
+    output_pt = forward_pt(torch_tensor)
+    # output_pt.backward()
+    # gradient_pt = torch_tensor.grad
+    assert np.allclose(output_ms.asnumpy(), output_pt.detach().numpy(), atol=1e-3)
+    # assert np.allclose(gradient_ms.asnumpy(), gradient_pt.numpy(), atol=1e-3)
+    # print(output_ms, gradient_ms)
+    # print(output_pt, gradient_pt)
+        

test/test_max.py

@@ -0,0 +1,120 @@
+import pytest
+import numpy as np
+import mindspore as ms
+from mindspore import mint, Tensor, value_and_grad
+import torch
+
+
+
+dtype_ms = ms.float32
+dtype_torch = torch.float32
+input_data = [[1, 6, 2, 4],[7, 3, 8, 2],[2, 9, 11, 5]]
+ms_tensor = Tensor(input_data, dtype_ms)
+torch_tensor = torch.tensor(input_data, dtype=dtype_torch)
+
+
+def is_same(input_data=[[1,6,2,4],[7,3,8,2],[2, 9, 11, 5]], shape=None, dtype_ms=ms.float32, dtype_torch=torch.float32, dim=-1, keepdim=False):
+    if shape != None:
+        input_data = np.random.randn(*shape)
+
+    ms_tensor = Tensor(input_data, dtype_ms)
+    torch_tensor = torch.tensor(input_data, dtype=dtype_torch)
+
+    if dim == None:
+        ms_result = mint.max(ms_tensor)
+        torch_result = torch.max(torch_tensor)
+        return np.allclose(ms_result.asnumpy(), torch_result.numpy())
+    else:
+        ms_result, ms_index = mint.max(ms_tensor, dim=dim, keepdim=keepdim)
+        torch_result, torch_index = torch.max(torch_tensor, dim=dim, keepdim=keepdim)
+        return np.allclose(ms_result.asnumpy(), torch_result.numpy()) and np.allclose(ms_index.asnumpy(), torch_index.numpy()) 
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_max_different_dtypes(mode):
+    """测试random输入不同dtype，对比两个框架的支持度"""
+    ms.set_context(mode=mode)
+    ms_dtypes = [ms.int8, ms.int16, ms.int32, ms.int64, ms.uint8, ms.uint16, ms.uint32, ms.uint64, ms.float16, ms.float32, ms.float64, ms.bfloat16, ms.bool_]
+    torch_dtypes = [torch.int8, torch.int16, torch.int32, torch.int64, torch.uint8, torch.uint16, torch.uint32, torch.uint64, torch.float16, torch.float32, torch.float64, torch.bfloat16, torch.bool]
+
+    for i in range(len(ms_dtypes)):
+        dtype_ms = ms_dtypes[i]
+        dtype_torch = torch_dtypes[i]
+
+        ms_tensor = Tensor(input_data, dtype_ms)
+        torch_tensor = torch.tensor(input_data, dtype=dtype_torch)
+
+        err = False
+        try:
+            ms_result = mint.max(ms_tensor).asnumpy()
+        except Exception as e:
+            err = True
+            print(f"mint.max not supported for {dtype_ms}")
+            # print(e)
+
+        try:
+            torch_result = torch.max(torch_tensor).numpy()
+        except Exception as e:
+            err = True
+            print(f"torch.max not supported for {dtype_torch}")
+            # print(e)
+
+        if not err:
+            assert np.allclose(ms_result, torch_result)
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_max_random_input_fixed_dtype(mode):
+    """测试固定数据类型下的随机输入值，对比输出误差"""
+    ms.set_context(mode=mode)
+
+    shapes = [[5], [5, 2], [5, 4, 3], [4, 6, 7, 8]]
+    for i in range(len(shapes)):
+        shape = shapes[i]
+        result = is_same(shape=shape)
+        assert result
+    
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_max_different_para(mode):
+    """测试固定shape，固定输入值，不同输入参数（string\bool等类型），两个框架的支持度"""
+    ms.set_context(mode=mode)
+
+    dims = [None, 0, 1]
+    keepdims = [True, False]
+    paras = [(dim, keepdim) for dim in dims for keepdim in keepdims]
+    for dim, keepdim in paras:
+        result = is_same(dim=dim, keepdim=keepdim)
+        assert result
+            
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_max_wrong_input(mode):
+    """测试随机混乱输入，报错信息的准确性"""
+    ms.set_context(mode=mode)
+
+    try:
+        ms_result = mint.max(ms_tensor, dim=None, keepdim=True).asnumpy()
+    except Exception as e:
+        print(f"dim 为 None 且 keepdim 不是 False报错信息：\n{e}")
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_max_forward_back(mode):
+    """使用Pytorch和MindSpore, 固定输入和权重, 测试正向推理结果和反向梯度"""
+    ms.set_context(mode=mode)
+    torch_tensor = torch.tensor(input_data, dtype=dtype_torch, requires_grad=True)
+
+    def forward_pt(x):
+        return torch.max(x)
+    
+    def forward_ms(x):
+        return mint.max(x, dim=None, keepdim=False)
+    
+    grad_fn = value_and_grad(forward_ms)
+    output_ms, gradient_ms = grad_fn(ms_tensor)
+    output_pt = forward_pt(torch_tensor)
+    output_pt.backward()
+    gradient_pt = torch_tensor.grad
+    assert np.allclose(output_ms.asnumpy(), output_pt.detach().numpy(), atol=1e-3)
+    assert np.allclose(gradient_ms.asnumpy(), gradient_pt.numpy(), atol=1e-3)

test/test_mean.py

@@ -0,0 +1,116 @@
+import pytest
+import numpy as np
+import mindspore as ms
+from mindspore import mint, Tensor, value_and_grad
+import torch
+
+
+dtype_ms = ms.float32
+dtype_torch = torch.float32
+input_data = [[1,6,2,4],[7,3,8,2],[2, 9, 11, 5]]
+ms_tensor = Tensor(input_data, dtype_ms)
+torch_tensor = torch.tensor(input_data, dtype=dtype_torch)
+
+
+def is_same(input_data=[[1,6,2,4],[7,3,8,2],[2, 9, 11, 5]], shape=None, dtype_ms=ms.float32, dtype_torch=torch.float32, dim=-1, keepdim=False):
+    if shape != None:
+        input_data = np.random.randn(*shape)
+
+    ms_tensor = Tensor(input_data, dtype_ms)
+    torch_tensor = torch.tensor(input_data, dtype=dtype_torch)
+
+    ms_result = mint.mean(ms_tensor, dim=dim, keepdim=keepdim).asnumpy()
+    torch_result = torch.mean(torch_tensor, dim=dim, keepdim=keepdim).numpy()
+
+    return np.allclose(ms_result, torch_result)
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_mean_different_dtypes(mode):
+    """测试random输入不同dtype，对比两个框架的支持度"""
+    ms.set_context(mode=mode)
+    ms_dtypes = [ms.float16, ms.float32, ms.float64]
+    torch_dtypes = [torch.float16, torch.float32, torch.float64]
+
+    for i in range(len(ms_dtypes)):
+        dtype_ms = ms_dtypes[i]
+        dtype_torch = torch_dtypes[i]
+        ms_tensor = Tensor(input_data, dtype_ms)
+        torch_tensor = torch.tensor(input_data, dtype=dtype_torch)
+
+        err = False
+        try:
+            ms_result = mint.mean(ms_tensor).asnumpy() 
+        except Exception as e:
+            er = True
+            print(f"mint.mean not supported for {dtype_ms}")
+            # print(e)
+
+        try:
+            torch_result = torch.mean(torch_tensor).numpy() 
+        except Exception as e:
+            err = True
+            print(f"torch.mean not supported for {dtype_torch}")
+            # print(e)
+        if not err:
+            assert np.allclose(ms_result, torch_result)
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_mean_random_input_fixed_dtype(mode):
+    """测试固定数据类型下的随机输入值，对比输出误差"""
+    ms.set_context(mode=mode)
+
+    shapes = [[5], [5, 2], [5, 4, 3], [4, 6, 7, 8]]
+    for i in range(len(shapes)):
+        shape = shapes[i]
+        result = is_same(shape=shape)
+        assert result
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_mean_different_para(mode):
+    """测试固定shape，固定输入值，不同输入参数（string\bool等类型），两个框架的支持度"""
+    ms.set_context(mode=mode)
+
+    dims = [None, 0, 1]
+    keepdims = [True, False]
+    paras = [(dim, keepdim) for dim in dims for keepdim in keepdims]
+
+    for dim, keepdim in paras:
+        result = is_same(dim=dim, keepdim=keepdim)
+        assert result
+
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_mean_wrong_input(mode):
+    """测试随机混乱输入，报错信息的准确性"""
+    ms.set_context(mode=mode)
+
+    try:
+        ms_result = mint.mean(ms_tensor, dim=2, keepdim=True).asnumpy()
+    except Exception as e:
+        print(f"dim 超出范围报错信息：\n{e}")
+    
+
+@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
+def test_mean_forward_back(mode):
+    """使用Pytorch和MindSpore, 固定输入和权重, 测试正向推理结果和反向梯度"""
+    ms.set_context(mode=mode)
+
+    torch_tensor = torch.tensor(input_data, dtype=dtype_torch, requires_grad=True)
+
+    def forward_pt(x):
+        return torch.mean(x)
+    
+    def forward_ms(x):
+        return mint.mean(x)
+    
+    grad_fn = value_and_grad(forward_ms)
+    output_ms, gradient_ms = grad_fn(ms_tensor)
+    output_pt = forward_pt(torch_tensor)
+    output_pt.backward()
+    gradient_pt = torch_tensor.grad
+    assert np.allclose(output_ms.asnumpy(), output_pt.detach().numpy(), atol=1e-3)
+    assert np.allclose(gradient_ms.asnumpy(), gradient_pt.numpy(), atol=1e-3)
+