optimizer.minimize(avg_cost) 无法反向传播

[xubin1994]

代码数据读取正常，网络框架正常，但整体拼接在一起时使用optimizer.minimize(avg_cost) 报错，似乎是说无法反向传播 训练部分代码如下：

from dataloader import *
from IresNet import IresNet
from loss import model_loss
if __name__ == '__main__':
    #main_program = fluid.default_main_program()
    datapath = '/data/home/xubin/dataset/dataset/GANet_dataset'
    file_name = 'sceneflow_train.txt'
    max_disp = 192
    left_input = fluid.layers.data(name='left_input', shape=[3, 384, 768], dtype='float32')
    right_input = fluid.layers.data(name='right_input', shape=[3, 384, 768], dtype='float32')
    left_disp_input = fluid.layers.data(name='left_disp_input', shape=[1, 384, 768], dtype='float32')
    predict_final, r_res2_predict, r_res1_predict, r_res0=IresNet(left_input,right_input)
    avg_cost = model_loss(predict_final, r_res2_predict, r_res1_predict, r_res0,left_disp_input,max_disp)
    optimizer = fluid.optimizer.AdamOptimizer(learning_rate=0.001)
    opt = optimizer.minimize(avg_cost)
    place = fluid.CPUPlace()
    exe = fluid.Executor(place=place)
    exe.run(program=fluid.default_startup_program())
    train_reader = paddle.batch(reader=paddle.reader.shuffle(reader=train_r(datapath,file_name), buf_size=3000),
                                batch_size=1)
    feeder = fluid.DataFeeder(place=place, feed_list=[left_input, right_input,left_disp_input])
    for pass_id in range(2):
        for batch_id, data in enumerate(train_reader()):
            train_cost, train_acc = exe.run(program=fluid.default_main_program(),
                                            feed=feeder.feed(data),
                                            fetch_list=[avg_cost])
    print(avg_cost)
`
报错log如下：
Traceback (most recent call last):
  File "xubin_train.py", line 15, in <module>
    opt = optimizer.minimize(avg_cost)
  File "</usr/local/lib/python3.6/dist-packages/decorator.py:decorator-gen-20>", line 2, in minimize
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/wrapped_decorator.py", line 25, in __impl__
    return wrapped_func(*args, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/dygraph/base.py", line 87, in __impl__
    return func(*args, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/optimizer.py", line 609, in minimize
    no_grad_set=no_grad_set)
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/optimizer.py", line 494, in backward
    no_grad_set, callbacks)
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/backward.py", line 706, in append_backward
    _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map)
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/backward.py", line 518, in _append_backward_vars_
    op_desc.infer_shape(block.desc)
IndexError: _Map_base::at

请问这种是出现了什么问题？

[xubin1994]

重新贴一遍训练代码，上面的格式有点问题

from dataloader import *
from IresNet import IresNet
from loss import model_loss
if __name__ == '__main__':
    #main_program = fluid.default_main_program()
    datapath = '/data/home/xubin/dataset/dataset/GANet_dataset'
    file_name = 'sceneflow_train.txt'
    max_disp = 192
    left_input = fluid.layers.data(name='left_input', shape=[3, 384, 768], dtype='float32')
    right_input = fluid.layers.data(name='right_input', shape=[3, 384, 768], dtype='float32')
    left_disp_input = fluid.layers.data(name='left_disp_input', shape=[1, 384, 768], dtype='float32')
    predict_final, r_res2_predict, r_res1_predict, r_res0=IresNet(left_input,right_input)
    avg_cost = model_loss(predict_final, r_res2_predict, r_res1_predict, r_res0,left_disp_input,max_disp)
    optimizer = fluid.optimizer.AdamOptimizer(learning_rate=0.001)
    opt = optimizer.minimize(avg_cost)
    place = fluid.CPUPlace()
    exe = fluid.Executor(place=place)
    exe.run(program=fluid.default_startup_program())
    train_reader = paddle.batch(reader=paddle.reader.shuffle(reader=train_r(datapath,file_name), buf_size=3000),
                                batch_size=1)
    feeder = fluid.DataFeeder(place=place, feed_list=[left_input, right_input,left_disp_input])
    for pass_id in range(2):
        for batch_id, data in enumerate(train_reader()):
            train_cost, train_acc = exe.run(program=fluid.default_main_program(),
                                            feed=feeder.feed(data),
                                            fetch_list=[avg_cost])
    print(avg_cost)

[heavengate]

model_loss的代码麻烦贴一下

[xubin1994]

model_loss 的代码麻烦贴一下

loss代码：

import paddle
import paddle.fluid as fluid
def loss_func(predict,disp_gt,maxdisp):
    mask1 = (disp_gt > 0)
    mask2 = (disp_gt < maxdisp)
    mask = paddle.fluid.layers.cast(mask1, 'float32') * paddle.fluid.layers.cast(mask2, 'float32')
    #mask.stop_gradient =True
    sum_mask = paddle.fluid.layers.reduce_sum(mask)
    predict_valid = predict * mask
    disp_gt_valid = disp_gt * mask
    loss_sum = paddle.fluid.layers.reduce_sum(paddle.fluid.layers.abs(predict_valid - disp_gt_valid))
    loss_mean = loss_sum/(sum_mask+1e-6)
    return loss_mean
def model_loss(predict_final, r_res2_predict, r_res1_predict, r_res0, disp_gt,maxdisp):
    all_losses = []
    #loss 0
    loss0 = loss_func(fluid.layers.relu(predict_final), disp_gt, maxdisp)
    all_losses.append(loss0)
    #loss 1
    loss1 = loss_func(fluid.layers.relu(predict_final + r_res0),disp_gt,maxdisp)
    all_losses.append(loss1)
    #loss2
    gt2 = paddle.fluid.layers.image_resize(disp_gt, scale=0.25, resample='BILINEAR',
                                    align_corners=False, align_mode=1)/4.0
    predict_final_2 = paddle.fluid.layers.image_resize(predict_final, scale=0.25, resample='BILINEAR',
                                    align_corners=False, align_mode=1)/4.0
    loss2 = loss_func(fluid.layers.relu(predict_final_2 + r_res2_predict), gt2, maxdisp//4)
    all_losses.append(0.2 * loss2)
    #loss3
    gt3 = paddle.fluid.layers.image_resize(disp_gt, scale=0.5, resample='BILINEAR',
                                    align_corners=False, align_mode=1)/2.0
    predict_final_3 = paddle.fluid.layers.image_resize(predict_final, scale=0.5, resample='BILINEAR',
                                    align_corners=False, align_mode=1)/2.0
    loss3 = loss_func(fluid.layers.relu(predict_final_3 + r_res1_predict), gt3, maxdisp // 2)
    all_losses.append(0.2 * loss3)
    return sum(all_losses)

[heavengate]

maxdisp, mask_sum这些设置stop_gradient=True试下

[xubin1994]

maxdisp, mask_sum 这些设置 stop_gradient=True 试下

设置了： mask.stop_gradient =True mask_sum.stop_gradient =True 还是报同样的错误，‘maxdisp’是一个我外部传入的一个整数，int类型，不存在stop_gradient属性的

[heavengate]

"/usr/local/lib/python3.6/dist-packages/paddle/fluid/backward.py", line 518 上一行把op_desc.type() print出来看下

[xubin1994]

op_desc.type() 按你的要求，插入代码如下： 打印信息如下： test_op fill_constant test_op elementwise_add_grad test_op elementwise_add_grad test_op elementwise_add_grad test_op elementwise_add_grad test_op elementwise_mul_grad test_op elementwise_div_grad test_op elementwise_add_grad test_op reduce_sum_grad test_op abs_grad test_op elementwise_sub_grad test_op elementwise_mul_grad test_op elementwise_mul_grad test_op reduce_sum_grad test_op sum test_op elementwise_mul_grad test_op cast test_op cast test_op relu_grad test_op elementwise_add_grad test_op elementwise_div_grad test_op bilinear_interp_grad test_op elementwise_div_grad test_op elementwise_mul_grad test_op elementwise_div_grad test_op elementwise_add_grad test_op reduce_sum_grad test_op abs_grad test_op elementwise_sub_grad test_op elementwise_mul_grad test_op elementwise_mul_grad test_op reduce_sum_grad test_op sum test_op elementwise_mul_grad test_op cast test_op cast test_op relu_grad test_op elementwise_add_grad test_op elementwise_div_grad test_op bilinear_interp_grad test_op elementwise_div_grad test_op elementwise_div_grad test_op elementwise_add_grad test_op reduce_sum_grad test_op abs_grad test_op elementwise_sub_grad test_op elementwise_mul_grad test_op elementwise_mul_grad test_op reduce_sum_grad test_op sum test_op elementwise_mul_grad test_op cast test_op cast test_op relu_grad test_op elementwise_add_grad test_op elementwise_div_grad test_op elementwise_add_grad test_op reduce_sum_grad test_op abs_grad test_op elementwise_sub_grad test_op elementwise_mul_grad test_op elementwise_mul_grad test_op reduce_sum_grad test_op sum test_op elementwise_mul_grad test_op cast test_op cast test_op relu_grad test_op conv2d_grad test_op conv2d_grad test_op concat_grad test_op leaky_relu_grad test_op conv2d_transpose_grad test_op conv2d_transpose_grad test_op sum test_op conv2d_grad test_op sum test_op conv2d_grad test_op concat_grad test_op leaky_relu_grad test_op conv2d_transpose_grad test_op conv2d_transpose_grad test_op sum test_op conv2d_grad test_op sum test_op leaky_relu_grad test_op conv2d_grad test_op leaky_relu_grad test_op conv2d_grad test_op sum test_op leaky_relu_grad test_op conv2d_grad test_op concat_grad test_op concat_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op pad_grad test_op slice_grad test_op unsqueeze2_grad test_op reduce_mean_grad test_op elementwise_mul_grad test_op sum test_op leaky_relu_grad test_op conv2d_grad test_op sum test_op leaky_relu_grad test_op conv2d_grad test_op leaky_relu_grad test_op conv2d_grad test_op sum test_op leaky_relu_grad test_op conv2d_grad test_op concat_grad test_op abs_grad test_op elementwise_sub_grad test_op transpose2_grad test_op reshape2_grad test_op elementwise_add_grad test_op elementwise_mul_grad test_op elementwise_mul_grad test_op unsqueeze2_grad test_op elementwise_sub_grad test_op unsqueeze2_grad test_op elementwise_sub_grad test_op cast test_op cast test_op gather_grad test_op gather_grad test_op cast

[heavengate]

确认下你代码里面哪里连续使用的gather和cast OP

[xubin1994]

确认下你代码里面哪里连续使用的 gather 和 cast OP gather连续使用: im_flat = paddle.fluid.layers.cast(im_flat, 'float32') idx_a = paddle.fluid.layers.cast(idx_a, 'int32') idx_b = paddle.fluid.layers.cast(idx_b, 'int32') Ia = paddle.fluid.layers.gather(im_flat, idx_a) Ib = paddle.fluid.layers.gather(im_flat, idx_b) cast连续使用的地方有很多，不能连续使用吗？

[heavengate]

可以连续用，报错报在这个OP上，涉及到这两个OP的代码麻烦贴下，我确认下

[xubin1994]

可以连续用，报错报在这个 OP 上，涉及到这两个 OP 的代码麻烦贴下，我确认下

关于cast: 第一部分： B = fluid.layers.slice(fluid.layers.shape(fea1), starts=[0], ends=[1], axes=[0]) B = fluid.layers.cast(x=B, dtype='int32') 第二部分： x0_f = paddle.fluid.layers.cast(x0, 'float32') x1_f = paddle.fluid.layers.cast(x1, 'float32') 第三部分： batch1 = fluid.layers.slice(fluid.layers.shape(T), starts=[0], ends=[1], axes=[0]) batch1 = fluid.layers.cast(x=batch1, dtype='int32')

第四部分： grid = fluid.layers.cast(x=grid, dtype='int32') grid = paddle.fluid.layers.expand(grid, [batch1,1]) print(grid.shape)

最后一个维度应该是height*width

    grid = paddle.fluid.layers.reshape(grid, [batch1, 2, height * width])
    grid = fluid.layers.cast(x=grid, dtype='float32')

cast和gather主要集中在下面这个类上：

class SpatialTransformer(object):
    """
    Spatial Transformer Network
    """

    def __init__(self):
        pass

    def __call__(self, U, T, data_format='channel_first', **kwargs):
        if (data_format == 'channel_first'):
            U = paddle.fluid.layers.transpose(U, (0, 2, 3, 1))
            T = paddle.fluid.layers.transpose(T, (0, 2, 3, 1))
        elif (data_format != 'channel_last'):
            raise ValueError('data_format only can be one of channel_last or channel_first')
        batch = U.shape[0]
        height = U.shape[1]
        width = U.shape[2]
        channels = U.shape[3]

        # batch = paddle.fluid.layers.shape(U)[0]

        # process T to [-1.0, 1.0]
        batch1 = fluid.layers.slice(fluid.layers.shape(T), starts=[0], ends=[1], axes=[0])
        batch1 = fluid.layers.cast(x=batch1, dtype='int32')
        # print('batch1',batch1)
        d = paddle.fluid.layers.reshape(T, [-1, 1, height * width])
        # print(d)
        print(d.shape)
        zeros = paddle.fluid.layers.zeros_like(d)
        d = paddle.fluid.layers.concat([d, zeros], axis=1)
        print(d.shape)
        # #grid的shape为[2,540*960]
        grid = self.__meshgrid(height, width)
        grid = paddle.fluid.layers.reshape(grid, [-1])
        grid = paddle.fluid.layers.unsqueeze(grid,[0])
        grid = fluid.layers.cast(x=grid, dtype='int32')
        grid = paddle.fluid.layers.expand(grid, [batch1,1])
        print(grid.shape)
        # # 最后一个维度应该是height*width
        grid = paddle.fluid.layers.reshape(grid, [batch1, 2, height * width])
        grid = fluid.layers.cast(x=grid, dtype='float32')
        T_g = paddle.fluid.layers.elementwise_add(grid, d)  # target grid
        #T_g = grid
        x_s = T_g[:, 0, :]
        y_s = T_g[:, 1, :]
        # # x_s = paddle.fluid.layers.slice(T_g, [0, 0, 0], [-1, 1, -1])
        # # y_s = paddle.fluid.layers.slice(T_g, [0, 1, 0], [-1, 1, -1])
        x_s_flat = paddle.fluid.layers.reshape(x_s, [-1])
        y_s_flat = paddle.fluid.layers.reshape(y_s, [-1])
        #
        input_transformed = self.__interpolate(U, x_s_flat, y_s_flat)
        # # 四舍五入取整之后再进行类型转换
        # # input_transformed = tf.round(input_transformed)
        # # input_transformed = tf.cast(tf.floor(input_transformed1), U.dtype)
        output = paddle.fluid.layers.reshape(input_transformed,
                                             [batch1, height, width, channels])
        # # output = paddle.fluid.layers.cast(paddle.fluid.layers.reshape(input_transformed,
        # #                                                               paddle.fluid.layers.stack(
        # #                                                                   [batch, height, width, channels])),'float32')
        if (data_format == 'channel_first'):
             output = paddle.fluid.layers.transpose(output, (0, 3, 1, 2))
        return output
    #网格代码正确运行
    def __meshgrid(self, height, width):
        x_t = paddle.fluid.layers.matmul(paddle.fluid.layers.ones(shape=[height, 1], dtype='float32'),
                                         paddle.fluid.layers.transpose(paddle.fluid.layers.unsqueeze(
                                             paddle.fluid.layers.cast(paddle.fluid.layers.range(0, width, 1, 'float32'),
                                                                      'float32'), [1]), [1, 0]))
        y_t = paddle.fluid.layers.matmul(
            paddle.fluid.layers.unsqueeze(
                paddle.fluid.layers.cast(paddle.fluid.layers.range(0, height, 1, 'float32'), 'float32'), [1]),
            paddle.fluid.layers.ones(shape=[1, width], dtype='float32'))
        x_t_flat = paddle.fluid.layers.reshape(x_t, (1, -1))
        y_t_flat = paddle.fluid.layers.reshape(y_t, (1, -1))
        grid = paddle.fluid.layers.concat([x_t_flat, y_t_flat], 0)
        return grid
    #base = self.__repeat(paddle.fluid.layers.range(0, num_batch, 1, 'int32') * dim1, height * width)
    def __repeat(self, x, n_repeats):
        #rep[1,height*width]
        rep = paddle.fluid.layers.transpose(
                paddle.fluid.layers.ones(shape=[n_repeats,1], dtype='float32'), [1, 0])
        x = paddle.fluid.layers.cast(x, 'float32')
        print('rep',rep.shape)
        x = paddle.fluid.layers.matmul(paddle.fluid.layers.reshape(x, (-1, 1)), rep)
        return paddle.fluid.layers.reshape(x, [-1])

    def __interpolate(self, im, x, y):
        # constants
        # num_batch = paddle.fluid.layers.shape(im)[0]
        # height = paddle.fluid.layers.shape(im)[1]
        # width = paddle.fluid.layers.shape(im)[2]
        # channels = paddle.fluid.layers.shape(im)[-1]
        #num_batch = im.shape[0]
        num_batch = fluid.layers.slice(fluid.layers.shape(im), starts=[0], ends=[1], axes=[0])
        num_batch = fluid.layers.cast(x=num_batch, dtype='int32')
        height = im.shape[1]
        width = im.shape[2]
        channels = im.shape[-1]

        x = paddle.fluid.layers.cast(x, 'float32')
        y = paddle.fluid.layers.cast(y, 'int32')
        zero = paddle.fluid.layers.zeros([], 'int32')
        max_x = float(im.shape[2] - 1)

        # max_x = paddle.fluid.layers.cast(im.shape[2] - 1, 'float32')
        # scale indices from [-1, 1] to [0, width/height]
        # x = (x + 1.0) * (width_f) / 2.0
        # y = (y + 1.0) * (height_f) / 2.0

        # do sampling
        x0 = paddle.fluid.layers.cast(paddle.fluid.layers.floor(x), 'float32')
        x1 = x0 + 1

        x0 = paddle.fluid.layers.clip(x0, 0.0, max_x)
        x1 = paddle.fluid.layers.clip(x1, 0.0, max_x)

        dim2 = width
        dim1 = width * height
        base = self.__repeat(paddle.fluid.layers.range(0, num_batch, 1, 'int32') * dim1, height * width)
        base_y = base + y * dim2
        #idx_b - idx_a = 1
        idx_a = base_y + x0
        idx_b = base_y + x1

        # use indices to lookup pixels in the
        # flat image and restore channels dim
        im_flat = paddle.fluid.layers.reshape(im, [-1, channels])
        im_flat = paddle.fluid.layers.cast(im_flat, 'float32')
        idx_a = paddle.fluid.layers.cast(idx_a, 'int32')
        idx_b = paddle.fluid.layers.cast(idx_b, 'int32')
        Ia = paddle.fluid.layers.gather(im_flat, idx_a)
        Ib = paddle.fluid.layers.gather(im_flat, idx_b)
        print('Ia',Ia.shape)
        # and finally calculate interpolated values
        x0_f = paddle.fluid.layers.cast(x0, 'float32')
        x1_f = paddle.fluid.layers.cast(x1, 'float32')
        print('x1_f', x1_f.shape)
        wa = paddle.fluid.layers.unsqueeze((x1_f - x), [1])
        wb = paddle.fluid.layers.unsqueeze((x - x0_f), [1])
        # wa = paddle.fluid.layers.expand(wa,[1,channels])
        # wb = paddle.fluid.layers.expand(wb, [1, channels])
        # print('wa', wa.shape)
        # output = wa * Ia + wb * Ib
        output = paddle.fluid.layers.elementwise_mul(Ia, wa) + paddle.fluid.layers.elementwise_mul(Ib, wb)
        #output = Ib
        return output

[heavengate]

        idx_a = paddle.fluid.layers.cast(idx_a, 'int32')
        idx_a.stop_gradient = True
        idx_b = paddle.fluid.layers.cast(idx_b, 'int32')
        idx_b.stop_gradient = True
        Ia = paddle.fluid.layers.gather(im_flat, idx_a)
        Ib = paddle.fluid.layers.gather(im_flat, idx_b)

试下

[xubin1994]

        idx_a = paddle.fluid.layers.cast(idx_a, 'int32')
        idx_a.stop_gradient = True
        idx_b = paddle.fluid.layers.cast(idx_b, 'int32')
        idx_b.stop_gradient = True
        Ia = paddle.fluid.layers.gather(im_flat, idx_a)
        Ib = paddle.fluid.layers.gather(im_flat, idx_b)

试下

更新一下刚才的回复，我发现在windows下和linux下跑同样的代码结果不一样，linux下的log更新了，报错如下：

Traceback (most recent call last):
  File "xubin_train.py", line 26, in <module>
    fetch_list=[avg_cost])
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/executor.py", line 657, in run
    use_program_cache=use_program_cache)
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/executor.py", line 755, in _run
    exe.run(program.desc, scope, 0, True, True, fetch_var_name)
paddle.fluid.core_avx.EnforceNotMet: Invoke operator expand_grad error.
Python Call stacks:
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/framework.py", line 1774, in append_op
    attrs=kwargs.get("attrs", None))
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/layer_helper.py", line 43, in append_op
    return self.main_program.current_block().append_op(*args, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/layers/nn.py", line 9518, in expand
    type='expand', inputs=inputs, outputs={'Out': out}, attrs=attrs)
  File "/data/home/xubin/data/Paddle-IresNet/stn_paddlepaddle.py", line 85, in __call__
    grid = paddle.fluid.layers.expand(grid, [batch1,1])
  File "/data/home/xubin/data/Paddle-IresNet/submodule.py", line 222, in DRS_net
    w_up_1b2b = transformer(up_1b2b, minus_predict_final)
  File "/data/home/xubin/data/Paddle-IresNet/IresNet.py", line 6, in IresNet
    r_res2_predict, r_res1_predict, r_res0 = DRS_net(conv1a, conv1b, up_1a2a, up_1b2b, predict_final)
  File "xubin_train.py", line 12, in <module>
    predict_final, r_res2_predict, r_res1_predict, r_res0=IresNet(left_input,right_input)
C++ Call stacks:
op expand_grad does not have kernel for data_type[int]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN] at [/paddle/paddle/fluid/framework/operator.cc:979]
PaddlePaddle Call Stacks:
0       0x7f9e7b30d278p void paddle::platform::EnforceNotMet::Init<std::string>(std::string, char const*, int) + 360
1       0x7f9e7b30d5c7p paddle::platform::EnforceNotMet::EnforceNotMet(std::string const&, char const*, int) + 87
2       0x7f9e7d3aed09p paddle::framework::OperatorWithKernel::ChooseKernel(paddle::framework::RuntimeContext const&, paddle::framework::Scope const&, boost::variant<paddle::platform::CUDAPlace, paddle::platform::CPUPlace, paddle::platform::CUDAPinnedPlace, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_> const&) const + 1289
3       0x7f9e7d3b0a58p paddle::framework::OperatorWithKernel::RunImpl(paddle::framework::Scope const&, boost::variant<paddle::platform::CUDAPlace, paddle::platform::CPUPlace, paddle::platform::CUDAPinnedPlace, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_> const&, paddle::framework::RuntimeContext*) const + 728
4       0x7f9e7d3b0cd1p paddle::framework::OperatorWithKernel::RunImpl(paddle::framework::Scope const&, boost::variant<paddle::platform::CUDAPlace, paddle::platform::CPUPlace, paddle::platform::CUDAPinnedPlace, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_> const&) const + 529
5       0x7f9e7d3ab863p paddle::framework::OperatorBase::Run(paddle::framework::Scope const&, boost::variant<paddle::platform::CUDAPlace, paddle::platform::CPUPlace, paddle::platform::CUDAPinnedPlace, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_> const&) + 307
6       0x7f9e7b497dfep paddle::framework::Executor::RunPreparedContext(paddle::framework::ExecutorPrepareContext*, paddle::framework::Scope*, bool, bool, bool) + 398
7       0x7f9e7b49ae9fp paddle::framework::Executor::Run(paddle::framework::ProgramDesc const&, paddle::framework::Scope*, int, bool, bool, std::vector<std::string, std::allocator<std::string> > const&, bool) + 143
8       0x7f9e7b2fe1fdp
9       0x7f9e7b33fd56p
10            0x56204cp _PyCFunction_FastCallDict + 860
11            0x4f88bap
12            0x4f98c7p _PyEval_EvalFrameDefault + 1127
13            0x4f6128p
14            0x4f7d60p
15            0x4f876dp
16            0x4fa6c0p _PyEval_EvalFrameDefault + 4704
17            0x4f6128p
18            0x4f7d60p
19            0x4f876dp
20            0x4fa6c0p _PyEval_EvalFrameDefault + 4704
21            0x4f6128p
22            0x4f9023p PyEval_EvalCode + 35
23            0x6415b2p
24            0x64166ap PyRun_FileExFlags + 154
25            0x643730p PyRun_SimpleFileExFlags + 400
26            0x62b26ep Py_Main + 1438
27            0x4b4cb0p main + 224
28      0x7f9f1ac9fb97p __libc_start_main + 231
29            0x5bdf6ap _start + 42

[heavengate]

看报错是expand 不支持int32，可以排查下是否有int32的expand操作

[xubin1994]

看报错是 expand 不支持 int32，可以排查下是否有 int32 的 expand 操作 有一个expand 的int32操作： grid = fluid.layers.cast(x=grid, dtype='int32') grid = paddle.fluid.layers.expand(grid, [batch1,1]) 但是我去掉int32的转换，它又给我报错说应该是要int的，日志如下： Python Call stacks: File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/framework.py", line 1774, in append_op attrs=kwargs.get("attrs", None)) File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/layer_helper.py", line 43, in append_op return self.main_program.current_block().append_op(*args, **kwargs) File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/layers/nn.py", line 9518, in expand type='expand', inputs=inputs, outputs={'Out': out}, attrs=attrs) File "/data/home/xubin/data/Paddle-IresNet/stn_paddlepaddle.py", line 86, in call grid = paddle.fluid.layers.expand(grid, [batch1,1]) File "/data/home/xubin/data/Paddle-IresNet/submodule.py", line 222, in DRS_net w_up_1b2b = transformer(up_1b2b, minus_predict_final) File "/data/home/xubin/data/Paddle-IresNet/IresNet.py", line 6, in IresNet r_res2_predict, r_res1_predict, r_res0 = DRS_net(conv1a, conv1b, up_1a2a, up_1b2b, predict_final) File "xubin_train.py", line 12, in predict_final, r_res2_predict, r_res1_predict, r_res0=IresNet(left_input,right_input) C++ Call stacks: Tensor holds the wrong type, it holds float, but desires to be int at [/paddle/paddle/fluid/framework/tensor_impl.h:30]

[heavengate]

看报错是 expand 不支持 int32，可以排查下是否有 int32 的 expand 操作 有一个expand 的int32操作： grid = fluid.layers.cast(x=grid, dtype='int32') grid = paddle.fluid.layers.expand(grid, [batch1,1]) 但是我去掉int32的转换，它又给我报错说应该是要int的，日志如下： Python Call stacks: File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/framework.py", line 1774, in append_op attrs=kwargs.get("attrs", None)) File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/layer_helper.py", line 43, in append_op return self.main_program.current_block().append_op(*args, kwargs) File "/usr/local/lib/python3.6/dist-packages/paddle/fluid/layers/nn.py", line 9518, in expand type='expand', inputs=inputs, outputs={'Out': out}, attrs=attrs) File "/data/home/xubin/data/Paddle-IresNet/stn_paddlepaddle.py", line 86, in call** grid = paddle.fluid.layers.expand(grid, [batch1,1]) File "/data/home/xubin/data/Paddle-IresNet/submodule.py", line 222, in DRS_net w_up_1b2b = transformer(up_1b2b, minus_predict_final) File "/data/home/xubin/data/Paddle-IresNet/IresNet.py", line 6, in IresNet r_res2_predict, r_res1_predict, r_res0 = DRS_net(conv1a, conv1b, up_1a2a, up_1b2b, predict_final) File "xubin_train.py", line 12, in predict_final, r_res2_predict, r_res1_predict, r_res0=IresNet(left_input,right_input) C++ Call stacks: Tensor holds the wrong type, it holds float, but desires to be int at [/paddle/paddle/fluid/framework/tensor_impl.h:30]

确认了下，expand_grad没有int的kernel，如果你这个grid是不需要梯度的，可以设置一下stop_gradient，如果需要梯度，建议把cast放到expand后面

[xubin1994]

grid

搞定，的确加个grid.stop_gradient就解决了。把cast放到expand后面应该是行不通的，我前面提到了这里不在前面加cast就会报另一个错。

[heavengate]

grid

搞定，的确加个grid.stop_gradient就解决了。把cast放到expand后面应该是行不通的，我前面提到了这里不在前面加cast就会报另一个错。把cast和expand换个顺序

不是不加cast，是cast之前先expand，因为expand_grad不支持int，所以在float上先expand，expand完之后再cast成int32， cast和expand换个顺序

[xubin1994]

grid

搞定，的确加个 grid.stop_gradient 就解决了。把 cast 放到 expand 后面应该是行不通的，我前面提到了这里不在前面加 cast 就会报另一个错。把 cast 和 expand 换个顺序

不是不加 cast，是 cast 之前先 expand，因为 expand_grad 不支持 int，所以在 float 上先 expand，expand 完之后再 cast 成 int32， cast 和 expand 换个顺序

我的意思是说，按照下面的操作，报错会提示在expand上，

grid = fluid.layers.cast(x=grid, dtype='int32')

grid = paddle.fluid.layers.expand(grid, [batch1,1])

grid = fluid.layers.cast(x=grid, dtype='int32')

log:()更详细的报错在今早第一次回复上 C++ Call stacks: Tensor holds the wrong type, it holds float, but desires to be int at [/paddle/paddle/fluid/framework/tensor_impl.h:30]

[heavengate]

grid

搞定，的确加个 grid.stop_gradient 就解决了。把 cast 放到 expand 后面应该是行不通的，我前面提到了这里不在前面加 cast 就会报另一个错。把 cast 和 expand 换个顺序

不是不加 cast，是 cast 之前先 expand，因为 expand_grad 不支持 int，所以在 float 上先 expand，expand 完之后再 cast 成 int32， cast 和 expand 换个顺序

我的意思是说，按照下面的操作，报错会提示在expand上，

grid = fluid.layers.cast(x=grid, dtype='int32')

grid = paddle.fluid.layers.expand(grid, [batch1,1])

grid = fluid.layers.cast(x=grid, dtype='int32')

log:()更详细的报错在今早第一次回复上 C++ Call stacks: Tensor holds the wrong type, it holds float, but desires to be int at [/paddle/paddle/fluid/framework/tensor_impl.h:30]

好的，那还是加stop_gradient吧

[paddle-bot-old[bot]]

Since you haven\'t replied for more than a year, we have closed this issue/pr. If the problem is not solved or there is a follow-up one, please reopen it at any time and we will continue to follow up. 由于您超过一年未回复，我们将关闭这个issue/pr。 若问题未解决或有后续问题，请随时重新打开，我们会继续跟进。