!2912 Update ssim and add msssim

Merge pull request !2912 from fanglei/msssim

mindspore/nn/layer/image.py

@@ -21,9 +21,13 @@ from mindspore.ops import functional as F
 from mindspore.ops.primitive import constexpr
 from mindspore._checkparam import Validator as validator
 from mindspore._checkparam import Rel
+from .conv import Conv2d
+from .container import CellList
+from .pooling import AvgPool2d
+from .activation import ReLU
 from ..cell import Cell
 
-__all__ = ['ImageGradients', 'SSIM', 'PSNR', 'CentralCrop']
+__all__ = ['ImageGradients', 'SSIM', 'MSSSIM', 'PSNR', 'CentralCrop']
 
 class ImageGradients(Cell):
     r"""
@@ -83,21 +87,6 @@ def _convert_img_dtype_to_float32(img, max_val):
         ret = ret * scale
     return ret
 
-
-@constexpr
-def _gauss_kernel_helper(filter_size):
-    """gauss kernel helper"""
-    filter_size = F.scalar_cast(filter_size, mstype.int32)
-    coords = ()
-    for i in range(filter_size):
-        i_cast = F.scalar_cast(i, mstype.float32)
-        offset = F.scalar_cast(filter_size-1, mstype.float32)/2.0
-        element = i_cast-offset
-        coords = coords+(element,)
-    g = np.square(coords).astype(np.float32)
-    g = Tensor(g)
-    return filter_size, g
-
 @constexpr
 def _check_input_4d(input_shape, param_name, func_name):
     if len(input_shape) != 4:
@@ -110,9 +99,65 @@ def _check_input_filter_size(input_shape, param_name, filter_size, func_name):
     validator.check(param_name + " shape[2]", input_shape[2], "filter_size", filter_size, Rel.GE, func_name)
     validator.check(param_name + " shape[3]", input_shape[3], "filter_size", filter_size, Rel.GE, func_name)
 
-@constexpr
-def _check_input_dtype(input_dtype, param_name, allow_dtypes, cls_name):
-    validator.check_type_name(param_name, input_dtype, allow_dtypes, cls_name)
+def _conv2d(in_channels, out_channels, kernel_size, weight, stride=1, padding=0):
+    return Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride,
+                  weight_init=weight, padding=padding, pad_mode="valid")
+
+def _create_window(size, sigma):
+    x_data, y_data = np.mgrid[-size // 2 + 1:size // 2 + 1, -size // 2 + 1:size // 2 + 1]
+    x_data = np.expand_dims(x_data, axis=-1).astype(np.float32)
+    x_data = np.expand_dims(x_data, axis=-1) ** 2
+    y_data = np.expand_dims(y_data, axis=-1).astype(np.float32)
+    y_data = np.expand_dims(y_data, axis=-1) ** 2
+    sigma = 2 * sigma ** 2
+    g = np.exp(-(x_data + y_data) / sigma)
+    return np.transpose(g / np.sum(g), (2, 3, 0, 1))
+
+def _split_img(x):
+    _, c, _, _ = F.shape(x)
+    img_split = P.Split(1, c)
+    output = img_split(x)
+    return output, c
+
+def _compute_per_channel_loss(c1, c2, img1, img2, conv):
+    """computes ssim index between img1 and img2 per single channel"""
+    dot_img = img1 * img2
+    mu1 = conv(img1)
+    mu2 = conv(img2)
+    mu1_sq = mu1 * mu1
+    mu2_sq = mu2 * mu2
+    mu1_mu2 = mu1 * mu2
+    sigma1_tmp = conv(img1 * img1)
+    sigma1_sq = sigma1_tmp - mu1_sq
+    sigma2_tmp = conv(img2 * img2)
+    sigma2_sq = sigma2_tmp - mu2_sq
+    sigma12_tmp = conv(dot_img)
+    sigma12 = sigma12_tmp - mu1_mu2
+    a = (2 * mu1_mu2 + c1)
+    b = (mu1_sq + mu2_sq + c1)
+    v1 = 2 * sigma12 + c2
+    v2 = sigma1_sq + sigma2_sq + c2
+    ssim = (a * v1) / (b * v2)
+    cs = v1 / v2
+    return ssim, cs
+
+def _compute_multi_channel_loss(c1, c2, img1, img2, conv, concat, mean):
+    """computes ssim index between img1 and img2 per color channel"""
+    split_img1, c = _split_img(img1)
+    split_img2, _ = _split_img(img2)
+    multi_ssim = ()
+    multi_cs = ()
+    for i in range(c):
+        ssim_per_channel, cs_per_channel = _compute_per_channel_loss(c1, c2, split_img1[i], split_img2[i], conv)
+        multi_ssim += (ssim_per_channel,)
+        multi_cs += (cs_per_channel,)
+
+    multi_ssim = concat(multi_ssim)
+    multi_cs = concat(multi_cs)
+
+    ssim = mean(multi_ssim, (2, 3))
+    cs = mean(multi_cs, (2, 3))
+    return ssim, cs
 
 class SSIM(Cell):
     r"""
@@ -157,67 +202,126 @@ class SSIM(Cell):
         self.max_val = max_val
         self.filter_size = validator.check_integer('filter_size', filter_size, 1, Rel.GE, self.cls_name)
         self.filter_sigma = validator.check_float_positive('filter_sigma', filter_sigma, self.cls_name)
-        validator.check_value_type('k1', k1, [float], self.cls_name)
-        self.k1 = validator.check_number_range('k1', k1, 0.0, 1.0, Rel.INC_NEITHER, self.cls_name)
-        validator.check_value_type('k2', k2, [float], self.cls_name)
-        self.k2 = validator.check_number_range('k2', k2, 0.0, 1.0, Rel.INC_NEITHER, self.cls_name)
-        self.mean = P.DepthwiseConv2dNative(channel_multiplier=1, kernel_size=filter_size)
+        self.k1 = validator.check_value_type('k1', k1, [float], self.cls_name)
+        self.k2 = validator.check_value_type('k2', k2, [float], self.cls_name)
+        window = _create_window(filter_size, filter_sigma)
+        self.conv = _conv2d(1, 1, filter_size, Tensor(window))
+        self.conv.weight.requires_grad = False
+        self.reduce_mean = P.ReduceMean()
+        self.concat = P.Concat(axis=1)
 
     def construct(self, img1, img2):
-        _check_input_dtype(F.dtype(img1), "img1", [mstype.float32, mstype.float16], self.cls_name)
         _check_input_filter_size(F.shape(img1), "img1", self.filter_size, self.cls_name)
         P.SameTypeShape()(img1, img2)
         max_val = _convert_img_dtype_to_float32(self.max_val, self.max_val)
         img1 = _convert_img_dtype_to_float32(img1, self.max_val)
         img2 = _convert_img_dtype_to_float32(img2, self.max_val)
 
-        kernel = self._fspecial_gauss(self.filter_size, self.filter_sigma)
-        kernel = P.Tile()(kernel, (1, P.Shape()(img1)[1], 1, 1))
+        c1 = (self.k1 * max_val) ** 2
+        c2 = (self.k2 * max_val) ** 2
+
+        ssim_ave_channel, _ = _compute_multi_channel_loss(c1, c2, img1, img2, self.conv, self.concat, self.reduce_mean)
+        loss = self.reduce_mean(ssim_ave_channel, -1)
+
+        return loss
+
+def _downsample(img1, img2, op):
+    a = op(img1)
+    b = op(img2)
+    return a, b
+
+class MSSSIM(Cell):
+    r"""
+    Returns MS-SSIM index between img1 and img2.
+
+    Its implementation is based on Wang, Zhou, Eero P. Simoncelli, and Alan C. Bovik. `Multiscale structural similarity
+    for image quality assessment <https://ieeexplore.ieee.org/document/1292216>`_.
+    Signals, Systems and Computers, 2004.
 
-        mean_ssim = self._calculate_mean_ssim(img1, img2, kernel, max_val, self.k1, self.k2)
+    .. math::
 
-        return mean_ssim
+        l(x,y)&=\frac{2\mu_x\mu_y+C_1}{\mu_x^2+\mu_y^2+C_1}, C_1=(K_1L)^2.\\
+        c(x,y)&=\frac{2\sigma_x\sigma_y+C_2}{\sigma_x^2+\sigma_y^2+C_2}, C_2=(K_2L)^2.\\
+        s(x,y)&=\frac{\sigma_{xy}+C_3}{\sigma_x\sigma_y+C_3}, C_3=C_2/2.\\
+        MSSSIM(x,y)&=l^alpha_M*{\prod_{1\leq j\leq M} (c^beta_j*s^gamma_j)}.
 
-    def _calculate_mean_ssim(self, x, y, kernel, max_val, k1, k2):
-        """calculate mean ssim"""
-        c1 = (k1 * max_val) * (k1 * max_val)
-        c2 = (k2 * max_val) * (k2 * max_val)
+    Args:
+        max_val (Union[int, float]): The dynamic range of the pixel values (255 for 8-bit grayscale images).
+          Default: 1.0.
+        power_factors (Union[tuple, list]): Iterable of weights for each of the scales.
+          Default: (0.0448, 0.2856, 0.3001, 0.2363, 0.1333). Default values obtained by Wang et al.
+        filter_size (int): The size of the Gaussian filter. Default: 11.
+        filter_sigma (float): The standard deviation of Gaussian kernel. Default: 1.5.
+        k1 (float): The constant used to generate c1 in the luminance comparison function. Default: 0.01.
+        k2 (float): The constant used to generate c2 in the contrast comparison function. Default: 0.03.
 
-        # SSIM luminance formula
-        # (2 * mean_{x} * mean_{y} + c1) / (mean_{x}**2 + mean_{y}**2 + c1)
-        mean_x = self.mean(x, kernel)
-        mean_y = self.mean(y, kernel)
-        square_sum = F.square(mean_x)+F.square(mean_y)
-        luminance = (2*mean_x*mean_y+c1)/(square_sum+c1)
+    Inputs:
+        - **img1** (Tensor) - The first image batch with format 'NCHW'. It should be the same shape and dtype as img2.
+        - **img2** (Tensor) - The second image batch with format 'NCHW'. It should be the same shape and dtype as img1.
 
-        # SSIM contrast*structure formula (when c3 = c2/2)
-        # (2 * conv_{xy} + c2) / (conv_{xx} + conv_{yy} + c2), equals to
-        # (2 * (mean_{xy} - mean_{x}*mean_{y}) + c2) / (mean_{xx}-mean_{x}**2 + mean_{yy}-mean_{y}**2 + c2)
-        mean_xy = self.mean(x*y, kernel)
-        mean_square_add = self.mean(F.square(x)+F.square(y), kernel)
+    Outputs:
+        Tensor, has the same dtype as img1. It is a 1-D tensor with shape N, where N is the batch num of img1.
 
-        cs = (2*(mean_xy-mean_x*mean_y)+c2)/(mean_square_add-square_sum+c2)
+    Examples:
+        >>> net = nn.MSSSIM(power_factors=(0.033, 0.033, 0.033))
+        >>> img1 = Tensor(np.random.random((1,3,128,128)))
+        >>> img2 = Tensor(np.random.random((1,3,128,128)))
+        >>> msssim = net(img1, img2)
+    """
+    def __init__(self, max_val=1.0, power_factors=(0.0448, 0.2856, 0.3001, 0.2363, 0.1333), filter_size=11,
+                 filter_sigma=1.5, k1=0.01, k2=0.03):
+        super(MSSSIM, self).__init__()
+        validator.check_value_type('max_val', max_val, [int, float], self.cls_name)
+        validator.check_number('max_val', max_val, 0.0, Rel.GT, self.cls_name)
+        self.max_val = max_val
+        validator.check_value_type('power_factors', power_factors, [tuple, list], self.cls_name)
+        self.filter_size = validator.check_integer('filter_size', filter_size, 1, Rel.GE, self.cls_name)
+        self.filter_sigma = validator.check_float_positive('filter_sigma', filter_sigma, self.cls_name)
+        self.k1 = validator.check_value_type('k1', k1, [float], self.cls_name)
+        self.k2 = validator.check_value_type('k2', k2, [float], self.cls_name)
+        window = _create_window(filter_size, filter_sigma)
+        self.level = len(power_factors)
+        self.conv = []
+        for i in range(self.level):
+            self.conv.append(_conv2d(1, 1, filter_size, Tensor(window)))
+            self.conv[i].weight.requires_grad = False
+        self.multi_convs_list = CellList(self.conv)
+        self.weight_tensor = Tensor(power_factors, mstype.float32)
+        self.avg_pool = AvgPool2d(kernel_size=2, stride=2, pad_mode='valid')
+        self.relu = ReLU()
+        self.reduce_mean = P.ReduceMean()
+        self.prod = P.ReduceProd()
+        self.pow = P.Pow()
+        self.pack = P.Pack(axis=-1)
+        self.concat = P.Concat(axis=1)
 
-        # SSIM formula
-        # luminance * cs
-        ssim = luminance*cs
+    def construct(self, img1, img2):
+        _check_input_4d(F.shape(img1), "img1", self.cls_name)
+        _check_input_4d(F.shape(img2), "img2", self.cls_name)
+        P.SameTypeShape()(img1, img2)
+        max_val = _convert_img_dtype_to_float32(self.max_val, self.max_val)
+        img1 = _convert_img_dtype_to_float32(img1, self.max_val)
+        img2 = _convert_img_dtype_to_float32(img2, self.max_val)
 
-        mean_ssim = P.ReduceMean()(ssim, (-3, -2, -1))
+        c1 = (self.k1 * max_val) ** 2
+        c2 = (self.k2 * max_val) ** 2
 
-        return mean_ssim
+        sim = ()
+        mcs = ()
 
-    def _fspecial_gauss(self, filter_size, filter_sigma):
-        """get gauss kernel"""
-        filter_size, g = _gauss_kernel_helper(filter_size)
+        for i in range(self.level):
+            sim, cs = _compute_multi_channel_loss(c1, c2, img1, img2,
+                                                  self.multi_convs_list[i], self.concat, self.reduce_mean)
+            mcs += (self.relu(cs),)
+            img1, img2 = _downsample(img1, img2, self.avg_pool)
 
-        square_sigma_scale = -0.5/(filter_sigma * filter_sigma)
-        g = g*square_sigma_scale
-        g = F.reshape(g, (1, -1))+F.reshape(g, (-1, 1))
-        g = F.reshape(g, (1, -1))
-        g = P.Softmax()(g)
-        ret = F.reshape(g, (1, 1, filter_size, filter_size))
-        return ret
+        mcs = mcs[0:-1:1]
+        mcs_and_ssim = self.pack(mcs + (self.relu(sim),))
+        mcs_and_ssim = self.pow(mcs_and_ssim, self.weight_tensor)
+        ms_ssim = self.prod(mcs_and_ssim, -1)
+        loss = self.reduce_mean(ms_ssim, -1)
 
+        return loss
 
 class PSNR(Cell):
     r"""

tests/ut/python/nn/test_msssim.py

@@ -0,0 +1,135 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""
+test msssim
+"""
+import numpy as np
+import pytest
+
+import mindspore.common.dtype as mstype
+import mindspore.nn as nn
+from mindspore import Tensor
+from mindspore.common.api import _executor
+
+_MSSSIM_WEIGHTS = (0.0448, 0.2856, 0.3001, 0.2363, 0.1333)
+
+class MSSSIMNet(nn.Cell):
+    def __init__(self, max_val=1.0, power_factors=_MSSSIM_WEIGHTS, filter_size=11, filter_sigma=1.5, k1=0.01, k2=0.03):
+        super(MSSSIMNet, self).__init__()
+        self.net = nn.MSSSIM(max_val, power_factors, filter_size, filter_sigma, k1, k2)
+
+    def construct(self, img1, img2):
+        return self.net(img1, img2)
+
+
+def test_compile():
+    factors = (0.033, 0.033, 0.033)
+    net = MSSSIMNet(power_factors=factors)
+    img1 = Tensor(np.random.random((8, 3, 128, 128)))
+    img2 = Tensor(np.random.random((8, 3, 128, 128)))
+    _executor.compile(net, img1, img2)
+
+
+def test_compile_grayscale():
+    max_val = 255
+    factors = (0.033, 0.033, 0.033)
+    net = MSSSIMNet(max_val=max_val, power_factors=factors)
+    img1 = Tensor(np.random.randint(0, 256, (8, 3, 128, 128), np.uint8))
+    img2 = Tensor(np.random.randint(0, 256, (8, 3, 128, 128), np.uint8))
+    _executor.compile(net, img1, img2)
+
+
+def test_msssim_max_val_negative():
+    max_val = -1
+    with pytest.raises(ValueError):
+        _ = MSSSIMNet(max_val)
+
+
+def test_msssim_max_val_bool():
+    max_val = True
+    with pytest.raises(TypeError):
+        _ = MSSSIMNet(max_val)
+
+
+def test_msssim_max_val_zero():
+    max_val = 0
+    with pytest.raises(ValueError):
+        _ = MSSSIMNet(max_val)
+
+
+def test_msssim_power_factors_set():
+    with pytest.raises(TypeError):
+        _ = MSSSIMNet(power_factors={0.033, 0.033, 0.033})
+
+
+def test_msssim_filter_size_float():
+    with pytest.raises(TypeError):
+        _ = MSSSIMNet(filter_size=1.1)
+
+
+def test_msssim_filter_size_zero():
+    with pytest.raises(ValueError):
+        _ = MSSSIMNet(filter_size=0)
+
+
+def test_msssim_filter_sigma_zero():
+    with pytest.raises(ValueError):
+        _ = MSSSIMNet(filter_sigma=0.0)
+
+
+def test_msssim_filter_sigma_negative():
+    with pytest.raises(ValueError):
+        _ = MSSSIMNet(filter_sigma=-0.1)
+
+
+def test_msssim_different_shape():
+    shape_1 = (8, 3, 128, 128)
+    shape_2 = (8, 3, 256, 256)
+    factors = (0.033, 0.033, 0.033)
+    img1 = Tensor(np.random.random(shape_1))
+    img2 = Tensor(np.random.random(shape_2))
+    net = MSSSIMNet(power_factors=factors)
+    with pytest.raises(ValueError):
+        _executor.compile(net, img1, img2)
+
+
+def test_msssim_different_dtype():
+    dtype_1 = mstype.float32
+    dtype_2 = mstype.float16
+    factors = (0.033, 0.033, 0.033)
+    img1 = Tensor(np.random.random((8, 3, 128, 128)), dtype=dtype_1)
+    img2 = Tensor(np.random.random((8, 3, 128, 128)), dtype=dtype_2)
+    net = MSSSIMNet(power_factors=factors)
+    with pytest.raises(TypeError):
+        _executor.compile(net, img1, img2)
+
+
+def test_msssim_invalid_5d_input():
+    shape_1 = (8, 3, 128, 128)
+    shape_2 = (8, 3, 256, 256)
+    invalid_shape = (8, 3, 128, 128, 1)
+    factors = (0.033, 0.033, 0.033)
+    img1 = Tensor(np.random.random(shape_1))
+    invalid_img1 = Tensor(np.random.random(invalid_shape))
+    img2 = Tensor(np.random.random(shape_2))
+    invalid_img2 = Tensor(np.random.random(invalid_shape))
+
+    net = MSSSIMNet(power_factors=factors)
+    with pytest.raises(ValueError):
+        _executor.compile(net, invalid_img1, img2)
+    with pytest.raises(ValueError):
+        _executor.compile(net, img1, invalid_img2)
+    with pytest.raises(ValueError):
+        _executor.compile(net, invalid_img1, invalid_img2)

tests/ut/python/nn/test_ssim.py

@@ -78,26 +78,6 @@ def test_ssim_filter_sigma_negative():
         _ = SSIMNet(filter_sigma=-0.1)
 
 
-def test_ssim_k1_k2_wrong_value():
-    with pytest.raises(ValueError):
-        _ = SSIMNet(k1=1.1)
-    with pytest.raises(ValueError):
-        _ = SSIMNet(k1=1.0)
-    with pytest.raises(ValueError):
-        _ = SSIMNet(k1=0.0)
-    with pytest.raises(ValueError):
-        _ = SSIMNet(k1=-1.0)
-
-    with pytest.raises(ValueError):
-        _ = SSIMNet(k2=1.1)
-    with pytest.raises(ValueError):
-        _ = SSIMNet(k2=1.0)
-    with pytest.raises(ValueError):
-        _ = SSIMNet(k2=0.0)
-    with pytest.raises(ValueError):
-        _ = SSIMNet(k2=-1.0)
-
-
 def test_ssim_different_shape():
     shape_1 = (8, 3, 16, 16)
     shape_2 = (8, 3, 8, 8)