Pascal Lamblin

> Well, I'm not sure about the 'without to/form converters' part, but torch [already supports protobuf](https://github.com/szagoruyko/loadcaffe). As does caffe. mxnet https://github.com/dmlc/mxnet/blob/master/tools/caffe_converter/README.md tensorflow https://github.com/ethereon/caffe-tensorflow [sklearn-theano](https://github.com/sklearn-theano/sklearn-theano) also has a [converter](https://github.com/sklearn-theano/sklearn-theano/blob/master/sklearn_theano/feature_extraction/caffe/caffemodel.py). ## Pascal

Such a solution (credit mostly goes to @dwf, actually) has also been implemented recently in [Blocks](https://github.com/bartvm/blocks) by @bartvm, if you want to have a look.

We use [`tf.image.decode_image`](https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/io/decode_image) passing explicitly `channels=3`, see [here](https://github.com/google-research/meta-dataset/blob/main/meta_dataset/data/decoder.py#L35). [Then](https://github.com/google-research/meta-dataset/blob/main/meta_dataset/data/decoder.py#L106-L111) we resize with [`tf.image.resize_images`](https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/image/resize_images), and map values to [-1, 1].

This is really surprising, as I would expect just reading and writing the data back would take quite some time, just because of the bandwidth needed to read / write...

Sorry if I did not answer your question. It is normal that about 400 instances "skipping file ..." is printed during the generation of tfrecords, they correspond to images that...

https://github.com/tensorflow/tensorflow/issues/41532#issuecomment-759075803 suggests that TF may use more than the numbers of available threads, and suggests things to check. You could try using `ulimit -u`, as explained [here](https://rudametw.github.io/blog/posts/2014.04.10/not-enough-threads.html) (in another context)...

We do not currently have anything explicitly handling semi-supervised classification. It should be relatively straightforward to add unlabeled examples in the data pipeline, but none of our learners or cost...

Hello, I don't think it is a typo, but I can double-check what happens if I relaunch the experiment. In a previous version the article (https://arxiv.org/abs/1903.03096v1, https://github.com/google-research/meta-dataset/tree/arxiv_v1), using a resnet...

That is right, we are just using the one nearest neighbour. We initially envisioned _k_ as a hyper-parameter that might be tuned, but we did not end up tuning it...

Thanks for the contribution! I would prefer to keep the minimum of files in the root of the repo, do you think there would be a way of making the...