Question

我在使用自己的损失函数训练模型时遇到错误。

我想要实现的是：

loss function

我的实施是：

train.prototxt

layer {
  name: 'similarity_loss'
  type: 'Python'
  bottom: 'fc7'
  bottom: 'labels'
  propagate_down: 1
  propagate_down: 0
  top: 'loss_similarity'
  loss_weight: 1
  python_param {
    module: 'roi_similarity_measure.roi_similarity_loss_layer'
    layer: 'RoISimilarityLossLayer'
  }
}

损失层

import caffe
import numpy as np
class RoISimilarityLossLayer(caffe.Layer):
    def setup(self, bottom, top):
        # check input pair
        if len(bottom) != 2:
            raise Exception("Need two inputs to compute distance - fc7 and labels.")
        # set textual probability matrix for 20 classes
        self._num_classes = 20 + 1
        self._num_rois = (bottom[1].data.shape)[0]
        self._probs = np.ndarray(shape=(self._num_classes, self._num_classes), dtype=float, order='C')
    def reshape(self, bottom, top):
        pass
    def forward(self, bottom, top):
        self._similarity = np.zeros((128, 128), dtype=np.float)
        labels = bottom[1].data.astype(np.int)
        for i in xrange(self._num_rois):
            for j in xrange(self._num_rois):
                self._similarity[i][j] = self._probs[labels[i]][labels[j]]
        fc7 = bottom[0].data
        fc7_euclidean = 0
        for i in xrange(self._num_rois):
            for j in xrange(self._num_rois):
                fc7_euclidean += ((fc7[i] - fc7[j]) ** 2) * self._similarity[i][j]
        top[0].data[...] = fc7_euclidean
        top[0].diff[...] = np.zeros((128, 4096), dtype=np.float)
    def backward(self, top, propagate_down, bottom):
        for num in xrange(len(propagate_down)):
            if not propagate_down[num]:
                continue
            D = np.zeros((128, 128), dtype=np.float)
            for i in xrange(self._num_rois):
                D[i][i] = self._similarity[i].sum()
                L = D - self._similarity
            # diff in this layer is calculated by : 2 * (L + transpose L)*fc7
            top[0].diff[...] = 2*(L+L.transpose()).dot(bottom[0].data)

错误

I0320 09:01:25.520689 22184 layer_factory.hpp:77] Creating layer similarity_loss
I0320 09:01:25.521440 22184 net.cpp:84] Creating Layer similarity_loss
I0320 09:01:25.521452 22184 net.cpp:406] similarity_loss <- fc7_drop7_0_split_2
I0320 09:01:25.521471 22184 net.cpp:406] similarity_loss <- labels_roi-data_1_split_1
I0320 09:01:25.521478 22184 net.cpp:380] similarity_loss -> loss_similarity
F0320 09:01:25.521548 22184 blob.cpp:145] Check failed: diff_ 
*** Check failure stack trace: ***
./experiments/scripts/faster_rcnn_end2end.sh: line 57: 22184 Aborted (core dumped) ./tools/train_net.py --gpu ${GPU_ID} --solver models/${PT_DIR}/${NET}/faster_rcnn_end2end/solver.prototxt --weights data/imagenet_models/${NET}.v2.caffemodel --imdb ${TRAIN_IMDB} --iters ${ITERS} --cfg experiments/cfgs/faster_rcnn_end2end.yml ${EXTRA_ARGS}

我是caffe的新手，这对我来说很困惑。谢谢你的帮助。

自定义咖啡损失图层时出错

0 个答案: