Training a network in python

[tdomhan]

I would be nice to be able to train a network directly within python. That is directly invoke caffe on a numpy array instead of writing it out to a leveldb file first. I know that you can pass numpy arrays for pretrained networks, but I'm not aware that training is already possible.

[kloudkl]

A script converting numpy array into leveldb is the easiest solution.
Both @Yangqing' pyleveldb and Russell Power's py-leveldb are good enough for this task.
The conversion can be done in two steps.

// 1. Convert the original data format into datum
bool ReadImageToDatum(const string& filename, const int label,
    const int height, const int width, Datum* datum) {
  cv::Mat cv_img;
  if (height > 0 && width > 0) {
    cv::Mat cv_img_origin = cv::imread(filename, CV_LOAD_IMAGE_COLOR);
    cv::resize(cv_img_origin, cv_img, cv::Size(height, width));
  } else {
    cv_img = cv::imread(filename, CV_LOAD_IMAGE_COLOR);
  }
  if (!cv_img.data) {
    LOG(ERROR) << "Could not open or find file " << filename;
    return false;
  }
  if (height > 0 && width > 0) {

  }
  datum->set_channels(3);
  datum->set_height(cv_img.rows);
  datum->set_width(cv_img.cols);
  datum->set_label(label);
  datum->clear_data();
  datum->clear_float_data();
  string* datum_string = datum->mutable_data();
  for (int c = 0; c < 3; ++c) {
    for (int h = 0; h < cv_img.rows; ++h) {
      for (int w = 0; w < cv_img.cols; ++w) {
        datum_string->push_back(static_cast<char>(cv_img.at<cv::Vec3b>(h, w)[c]));
      }
    }
  }
  return true;
}

// 2. Save the datum
leveldb::WriteBatch* batch = new leveldb::WriteBatch();
datum.SerializeToString(&value);
batch->Put(string(key_cstr), value);

[tdomhan]

Converting numpy arrays to leveldb is of course a first workaround. Writing out a file everytime I want to train a model is too cumbersome in the long run though, because I want to parametrize preprocessing, so it's not a one-off job. But anyway thanks for the pointers.

[sergeyk]

Great proposal, and we would appreciate a PR for this!
In the meantime, take a look at #147.

[tdomhan]

Thanks for pointing me to the HDF5DataLayer. I'll create a PR as soon as I'll have some time to work on this.

[shelhamer]

#286 takes steps to solving this, and will follow-up with numpy array / in-memory inputs for training.