All probabilities 0 on predict with C++. Works in Python.

[Steph van Schalkwyk]

Hi

I have a net trained on gray scale jpgs. As they are jpgs, Caffe sees them as 3 channel. No problems there with Python.

For C++, I changed the predict.prototxt file as follows:

name: "CaffeNet"

#input: "data"  --- this is needed as per the Python tutorial

#input_dim: 1

#input_dim: 3

#input_dim: 100

#input_dim: 100

layers {           # this layer is added for c++

  name: "data"

  type: MEMORY_DATA

  top: "data"

  top: "label"

  memory_data_param {

     channels: 3

     height: 100

     width: 100

     batch_size: 1

  }

  transform_param {

    crop_size: 100

    mean_file: "/xxx/mean.binaryproto"

    mirror: false

  }

}

.....               # all layers same as train_val.prototxt with all training parameters removed

layers {

  name: "prob"

  type: SOFTMAX

  bottom: "fc8"

  top: "prob"

}

layers {

  name: "argmax"

  type: ARGMAX

  bottom: "prob"

  top: "argmax"

  argmax_param {

    top_k: 100    # I need to see probabilities for all classes. Is this correct?

  }

}

The relevant C++ is as follows:

Net<float> caffe_net(argv[1]);

caffe_net.CopyTrainedLayersFrom(argv[2]);

Datum datum;

boost::shared_ptr<MemoryDataLayer<float>> memory_data_layer;

memory_data_layer = boost::static_pointer_cast<MemoryDataLayer<float> >(caffe_net.layer_by_name("data"));

cv::Mat input_image;

input_image = cv::imread(argv[6]);

cv::resize(input_image, input_image, cv::Size(memory_data_layer->width(), memory_data_layer->height()));

CVMatToDatum(input_image, &datum);

std::vector<Datum> datumVector;

datumVector.push_back(datum);

memory_data_layer->AddDatumVector(datumVector);

float loss;

std::vector<Blob<float>*> results = caffe_net.ForwardPrefilled(&loss);

const float* argmaxs = results[1]->cpu_data();

for (int i = 0; i < results[0]->num(); i++) {

   for (int j = 0; j < results[0]->height(); j++) {

cout << "Image: " << i << " prob: " << argmaxs[i*results[0]->height() + j] << endl;

}

   }

   for (int i = 0; i < results[1]->num(); i++) {

for (int j = 0; j < results[1]->height(); j++) {

cout << "Image: " << i << " class:" << argmaxs[i*results[1]->height() + j] << endl;

}

}

I'm always getting probabilities of 0 or exceedingly small numbers (e.g. prob: -1.23558e+33). 

What am I not seeing? 

Thanks

Steph

[Steph van Schalkwyk]

I am now using this and it does seem to work:

// Get probabilities

const boost::shared_ptr<Blob<float> >& probLayer = caffe_net.blob_by_name("prob");

const float* probs_out = probLayer->cpu_data();

// Get argmax results

const boost::shared_ptr<Blob<float> >& argmaxLayer = caffe_net.blob_by_name("argmax");

The "prob" layer does not always seem to return all the channels that the "fc8" layer propagates to it. Has anyone seen this? I can force it to return all channels by indexing into probs_out but most are 0.

[lavaniac]

Hi. 

Were you able to figure out the issue ? I am facing the same issue when I set Caffe::set_mode(Caffe::GPU). But when I set mode to CPU using Caffe::set_mode(Caffe::CPU), I get reasonable results.

Thanks