caffe SSD 添加新层时出现的问题

caffe代码，在迭代１００００次的时候需要进行ｔｅｓｔ，但是test的时候遇见问题。

I0512 14:40:29.685868 15163 upgrade_proto.cpp:77] Attempting to upgrade batch norm layers using deprecated params: snapshot_iter_10000.caffemodel

I0512 14:40:29.685925 15163 upgrade_proto.cpp:80] Successfully upgraded batch norm layers using deprecated params.

I0512 14:40:29.710816 15163 sgd_solver.cpp:356] SGDSolver: restoring history

I0512 14:40:29.788755 15163 caffe.cpp:251] Starting Optimization

I0512 14:40:29.788800 15163 solver.cpp:294] Solving VGG_VOC0712_SSD_300x300_test_train

I0512 14:40:29.788805 15163 solver.cpp:295] Learning Rate Policy: multistep

I0512 14:40:29.799832 15163 solver.cpp:433] Iteration 10000, Testing net (#0)

I0512 14:40:29.817718 15163 net.cpp:693] Ignoring source layer mbox_loss

F0512 14:40:32.193830 15163 solver.cpp:464] Check failed: result[j]->width() == 5 (150 vs. 5)

*** Check failure stack trace: ***

    @     0x7f82cfdf75cd  google::LogMessage::Fail()

    @     0x7f82cfdf9433  google::LogMessage::SendToLog()

    @     0x7f82cfdf715b  google::LogMessage::Flush()

    @     0x7f82cfdf9e1e  google::LogMessageFatal::~LogMessageFatal()

    @     0x7f82d04aef2c  caffe::Solver<>::TestDetection()

    @     0x7f82d04afe39  caffe::Solver<>::TestAll()

    @     0x7f82d04aff3c  caffe::Solver<>::Step()

    @     0x7f82d04b0abe  caffe::Solver<>::Solve()

    @           0x40bcf4  train()

    @           0x4077c8  main

    @     0x7f82ce58e830  __libc_start_main

    @           0x408099  _start

    @              (nil)  (unknown)

Aborted (core dumped)

注：遇见问题，如果抛出错误，首先要去读源码。例如这个，要去solver.cpp中，看抛出的问题在什么地方。

看不懂代码也是悲伤！

ＳＳＤ的solver源代码　为什么在test的时候，对每一次获得的blob 的ｗｉｄｔｈ要和5进行对比呢？

＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊solver源代码分割线＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊

    Dtype iter_loss;

    const vector<Blob<Dtype>*>& result = test_net->Forward(&iter_loss);

    if (param_.test_compute_loss()) {

      loss += iter_loss;

    }

    for (int j = 0; j < result.size(); ++j) {

      CHECK_EQ(result[j]->width(), 5);

      const Dtype* result_vec = result[j]->cpu_data();

      int num_det = result[j]->height();

      for (int k = 0; k < num_det; ++k) {

        int item_id = static_cast<int>(result_vec[k * 5]);

        int label = static_cast<int>(result_vec[k * 5 + 1]);

        if (item_id == -1) {

          // Special row of storing number of positives for a label.

          if (all_num_pos[j].find(label) == all_num_pos[j].end()) {

            all_num_pos[j][label] = static_cast<int>(result_vec[k * 5 + 2]);

          } else {

            all_num_pos[j][label] += static_cast<int>(result_vec[k * 5 + 2]);

          }

        } else {

          // Normal row storing detection status.

          float score = result_vec[k * 5 + 2];

          int tp = static_cast<int>(result_vec[k * 5 + 3]);

          int fp = static_cast<int>(result_vec[k * 5 + 4]);

          if (tp == 0 && fp == 0) {

            // Ignore such case. It happens when a detection bbox is matched to

            // a difficult gt bbox and we don't evaluate on difficult gt bbox.

            continue;

          }

          all_true_pos[j][label].push_back(std::make_pair(score, tp));

          all_false_pos[j][label].push_back(std::make_pair(score, fp));

        }

      }

    }

  }

***********************************2017/5/13*********************************

去ｇｉｔ的主页上找到了一个人遇见了同样的问题，添加新的层会遇见同样的问题

原问题：
Thanks to Wei for the great work, and sharing it with everybody.

I need to get TEST loss during training, so that I can get some idea about how well the network is doing in generalization. In the python script I tried to add MultiBoxLoss layer to the test net the same way train net did, however it give me error at test
time:

solver.cpp:464 check failed: result[j]->width() ==5 (1 vs 5)

I am not familiar with the internals of Caffe, could you give me a pointer on proper way of enabling test loss in the Python script?

Your help is greatly appreciated.

作者回答：

I think you have to add a silence layer after the loss layer.

在python的接口的代码
in the python script,when you create the test net, after CreateMultiBoxHead, add these lines

name = "test_loss"
mbox_layers.append(net.label)
net[name] = L.MultiBoxLoss(*mbox_layers, multibox_loss_param=multibox_loss_param,
loss_param=loss_param, include=dict(phase=caffe_pb2.Phase.Value('TEST')),
propagate_down=[True, True, False, False])
name="silence"
net[name] = L.Silence(net['test_loss'],ntop=0)

also, add this to the solver_param:

'test_compute_loss':True,