Question

我已经从此library实现了roi_pooling。

以roi_pooling身份运行

conv5_3 = net.end_points['conv5_3']
#implement ROI Pooling
input_boxes=tf.dtypes.cast(input_boxes,tf.int32)
pooled_features = roi_pooling(conv5_3, input_boxes, pool_height=5, pool_width=30)

然后pooled_features的形状未知。

(Pdb) p pooled_features
<tf.Tensor 'RoiPooling:0' shape=<unknown> dtype=float32>

我的网络在roi_pooling之后仍然需要通过另一个识别网络。

识别网是

def recognitionnet(inputs, fatness = 64, dilation = True):
    """
    backbone net of vgg16
    """
    # End_points collect relevant activations for external use.
    end_points = {}
    # Original VGG-16 blocks.
    with slim.arg_scope([slim.conv2d, slim.max_pool2d], padding='SAME'):
        # Block1
        net = slim.repeat(inputs, 2, slim.conv2d, fatness, [3, 3], scope='conv1')
        end_points['conv1_2'] = net
        net = slim.max_pool2d(net, [2, 2], scope='pool1')
        end_points['pool1'] = net
        # Block 2.
        net = slim.repeat(net, 2, slim.conv2d, fatness * 2, [3, 3], scope='conv2')
        end_points['conv2_2'] = net
        net = slim.max_pool2d(net, [2, 2], scope='pool2')
        end_points['pool2'] = net
        # fc7 as conv
        net = slim.conv2d(net, fatness * 16, [1, 1], scope='fc7')
        end_points['fc7'] = net
        #model_summary()
        #from keras.utils.visualize_util import plot
        #plot(model, to_file='model.png')
    return net, end_points;

但是当pooled_features通过识别网的第一层时，我的错误为

TypeError: TypeErro...neType',)

pooled_features通过net = slim.repeat(inputs, 2, slim.conv2d, fatness, [3, 3], scope='conv1')传递时可能会出问题。

我该如何解决这个问题？

Answer 1

我所做的是将两个图分开。第一个图具有roi_poling实现，第二个图具有识别网。

由于识别网位于单独的图中，因此我们可以使用placeholder作为识别网图的输入。

运行roi_pooling层的输出到conv2d

1 个答案: