HDF5_DATA + MEMORY_DATA refuse loudly to transform

[shelhamer]

These layers are meant for generic inputs so they do not do transformations. As transform_param was pulled out of data layers' proto definitions for reuse, these layers silently ignored transformation parameters until now instead of signalling their refusal.

Transforming could be added in a future PR. Better to make sure users know the current state first.

See #1718.

[DmitryUlyanov]

I can create PR with transforms for memory_data, do you need that ?

[shelhamer]

@DmitryUlyanov if extending MemoryDataLayer::Reset() to transform is useful to you then you could send a PR -- however we've always thought of this data layer as taking input arrays exactly as they are. Note that it does transform when adding vectors of Datum or cv::Mat as-is.

@longjon thoughts?

[DmitryUlyanov]

@shelhamer well, I don't actually think It's a good way to transform things when doing reset (or Add...Vector) because during training you obviously want to have some random transformations every time. So I did it in Forward_cpu and didn't feel net started to work any slower on my 40x40 images with batch_size = 300, I did it like that (this version is outdated a bit):

void MemoryDataLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top) {
  CHECK(data_) << "MemoryDataLayer needs to be initalized by calling Reset";

  if (this->layer_param_.has_transform_param()) {

    // Copy to blob
    Blob<Dtype> b;
    b.Reshape(batch_size_, channels_, height_, width_);
    b.set_cpu_data(data_ + pos_ * size_);

    // transform
    this->data_transformer_.Transform(&b, top[0]);

    // Copy labels
    top[1]->set_cpu_data(labels_ + pos_);

    } else {
    top[0]->set_cpu_data(data_ + pos_ * size_);
    top[1]->set_cpu_data(labels_ + pos_);
  }
  pos_ = (pos_ + batch_size_) % n_;
  has_new_data_ = false;
}

And anyway user can omit transform_params in config so memory_data will take input arrays exactly as they are with no overhead at all.
So if this way is satisfactory I will send a PR.

[rohitgirdhar]

Even with MemoryDataLayer, I use the prototxt to specify what transform parameters to use (in AddMatVector function). However the check added in this PR fails, and the program aborts. Is there another way to add DataTransfomer parameters through the code? (whatever examples I've seen take it from the prototxt file).

[shelhamer]

This check was accidentally too stringent for MemoryDataLayer. @DmitryUlyanov your approach is better, so please send a PR for transforming memory data as outlined.

(While transforming during forward isn't ideal it's the best solution at present.)