如何将此函数转换为生成器？

Question

我有一个函数（patch_generator），它需要两个图像（作为numpy数组）并生成给window_size的补丁。 该功能可以正常工作，但是它会同时生成所有补丁。 我想批量生成图像补丁。

def patch_generator(x, y, window_size):
    # x and y are numpy arrays with shape of: (bands, height, width)
    # generates image patch of shape (patches, size, size, bands)
    # and a ground truth patch of shape (patches,)
    index_patch = 0
    x = x.reshape((x.shape[1], x.shape[2], x.shape[0]))
    y = y.reshape((y.shape[1], y.shape[2], y.shape[0]))
    if window_size % 2 == 0:
        margin = int(window_size/2)
    else:
        margin = int((window_size - 1) / 2)
    x_zeros = pad_zeros(x, margin)
    x_patch = np.zeros((x.shape[0]*x.shape[1], window_size, window_size, x.shape[2]))
    y_patch = np.zeros((x.shape[0]*x.shape[1]))
    row_range = range(margin, x_zeros.shape[0] - margin)
    col_range = range(margin, x_zeros.shape[1] - margin)
    for r in product(row_range, col_range):
        if window_size % 2 == 0:
            patch = x_zeros[r[0] - margin:r[0] + margin,
                            r[1] - margin:r[1] + margin]
        else:
            patch = x_zeros[r[0] - margin:r[0]+margin+1,
                            r[1] - margin:r[1]+margin+1]
        x_patch[index_patch, :, :, :] = patch
        y_patch[index_patch] = y[r[0]-margin, r[1]-margin]
        index_patch += 1
    del margin, x_zeros, row_range, col_range, patch
    return x_patch.astype(np.float16), y_patch.astype(np.float16)

我希望通过包含一个附加的函数参数“ batch_size”将同一函数转换为生成器。 这应该每次都生成batch_size图像补丁。

def pad_zeros(x, margin=2):
    # x is a numpy array with shape of: (width, height, bands)
    new_x = np.zeros((x.shape[0]+2*margin, x.shape[1]+2*margin, x.shape[2]))
    new_x[margin:x.shape[0]+margin, margin:x.shape[1]+margin, :] = x
    return new_x

例如，

X = np.array([[0, 2, 3, 4, 5, 6, 7 , 8],
              [1, 1, 1, 1, 1, 1, 1 , 1],
              [2, 2, 3, 4, 0, 6, 5 , 4],
              [0, 2, 3, 4, 5, 6, 7 , 8],
              [3, 2, 1, 3, 0, 9, 1 , 0],
              [1, 1, 3, 4, 5, 7, 6 , 8],
              [0, 3, 3, 4, 5, 6, 1 , 0]).reshape(1, 7, 8)

Y = np.array([[0, 0, 0, 0, 0, 0, 0 , 0],
              [1, 1, 1, 1, 1, 1, 1 , 1],
              [0, 0, 0, 0, 0, 0, 0 , 0],
              [0, 0, 0, 0, 0, 0, 0 , 0],
              [0, 0, 1, 0, 0, 0, 1 , 0],
              [1, 1, 0, 0, 0, 0, 0 , 0],
              [0, 0, 0, 0, 0, 0, 1 , 0]).reshape(1, 7, 8)

Answer 1

也许是这样的话？ 基本上要生成一个发电机，您需要使用关键字yield而不是return。 有关示例，请参见此博客 。

def patch_generator(x, y, window_size, batch_size=64):
    # x and y are numpy arrays with shape of: (bands, height, width)
    # generates image patch of shape (patches, size, size, bands)
    # and a ground truth patch of shape (patches,)

    x = x.reshape((x.shape[1], x.shape[2], x.shape[0]))
    y = y.reshape((y.shape[1], y.shape[2], y.shape[0]))
    if window_size % 2 == 0:
        margin = int(window_size/2)
    else:
        margin = int((window_size - 1) / 2)
    x_zeros = pad_zeros(x, margin)

    row_range = range(margin, x_zeros.shape[0] - margin)
    col_range = range(margin, x_zeros.shape[1] - margin)

    # Prepare vectors with batch_size entries
    x_patch = np.zeros((batch_size, window_size, window_size, x.shape[2]))
    y_patch = np.zeros(batch_size)

    for index_patch,r in enumerate(product(row_range, col_range)):

        if window_size % 2 == 0:
            patch = x_zeros[r[0] - margin:r[0] + margin,
                            r[1] - margin:r[1] + margin]
        else:
            patch = x_zeros[r[0] - margin:r[0]+margin+1,
                            r[1] - margin:r[1]+margin+1]

        # Overwrite x_patch and y_patch at each new batch
        x_patch[index_patch % batch_size, :, :, :] = patch
        y_patch[index_patch % batch_size] = y[r[0]-margin, r[1]-margin]

        if ((index_patch+1) % batch_size == 0) or (index_patch==len(row_range)*len(col_range)-1):
            # When we reach batch_size, return the result
            yield (x_patch.astype(np.float16),
                   y_patch.astype(np.float16))

而对于测试（ patch_fn是您的原始函数， patch_generator是修改后的生成器版本）：

x = np.array([[0, 2, 3, 4, 5, 6, 7 , 8],
              [1, 1, 1, 1, 1, 1, 1 , 1],
              [2, 2, 3, 4, 0, 6, 5 , 4],
              [0, 2, 3, 4, 5, 6, 7 , 8],
              [3, 2, 1, 3, 0, 9, 1 , 0],
              [1, 1, 3, 4, 5, 7, 6 , 8],
              [0, 3, 3, 4, 5, 6, 1 , 0]]).reshape(1, 7, 8)

y = np.array([[0, 0, 0, 0, 0, 0, 0 , 0],
              [1, 1, 1, 1, 1, 1, 1 , 1],
              [0, 0, 0, 0, 0, 0, 0 , 0],
              [0, 0, 0, 0, 0, 0, 0 , 0],
              [0, 0, 1, 0, 0, 0, 1 , 0],
              [1, 1, 0, 0, 0, 0, 0 , 0],
              [0, 0, 0, 0, 0, 0, 1 , 0]]).reshape(1, 7, 8)

window_size = 3
batch_size = 2

res_w_fn = patch_fn(x,y,window_size)
res_w_gen = patch_generator(x,y,window_size,batch_size)

for i,(x,y) in enumerate(res_w_gen):
    x_t,y_t = (res_w_fn[0][i*batch_size:(i+1)*batch_size,:,:,:],
               res_w_fn[1][i*batch_size:(i+1)*batch_size])
    if np.sum(x!=x_t)==0 and np.sum(y!=y_t)==0:
        print("Batch #{} - Success".format(i))
    else:
        print("Batch #{} - Fail".format(i))