Hello,

According to HuggingFace Transformers documentation website (https://huggingface.co/transformers/model_doc/gpt2.html#gpt2doubleheadsmodel), under the GPT2DoubleHeadsModel, it defines the output lm_prediction_scores as the following:

lm_prediction_scores: torch.FloatTensor of shape (batch_size, num_choices, sequence_length, config.vocab_size)

To me this doesn't make sense. Shouldn

With the latest version of scipy.misc, scipy.misc.toimage is no longer available. To load and save an image as png we now have to use PIL, breaking tensorboard image summary.

Here is how I fixed the bug:
1./ At the end of main.py, log a uint8 image
logger.image_summary(tag, (images * 255).astype(np.uint8), step+1)
2./ In Logger class, package image as bytes with the PIL library (mode="L

From here:

A particularity of the SV2TTS framework is that all models can be trained
separately and on distinct datasets. For the encoder, one seeks to have a model
that is robust to noise and able to capture the many characteristics of the human
voice. Therefore, a large corpus of many different speakers wou

I had download pretrained model by 'bash scripts/download_cyclegan_model.sh cityscapes_label2photo',
and run 'test.py' to generate photos,
and run 'scripts/eval_cityscapes/evaluate.py',
finally the result as following is far less than the paper result

the paper result is following:

Judging by the logic in https://github.com/horovod/horovod/blob/38e91bee84efbb5b563a4928027a75dc3974633b/setup.py#L1369 it is clear, that before installing Horovod one needs to install the underlying framework(s) (TensorFlow, PyTorch, ...).

This is not mentioned in the installation instructions which made me think, I can install Horovod and then any framework I like (or switch between them) and

I trained RetinaNet with Resnet 34 backbone on COCO.

I would like to add it to the model Zoo.

It looks like weights for other networks are located at https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/

How can I upload weights there?

Forgive me, this is purely out of curiosity. If I had a document embedding that is created from the average of all word embeddings created by BERT. Could I do the inverse and take that ending embedding and generate a word or even a sentence from it?

#3602 switched our docs from restructured text to markdown, which is a big improvement. However, there are some left over traces of rst formatting in the docstrings. It would be great if we could comb through these and update them.

• commands
• common
• data
• interpret
• models
• modules
• nn
• predictors
• tools
• training

How to help

Support for storing large tensor values in external files was introduced in #678, but AFAICT is undocumented.

This is a pretty important feature, functionally, but it's also important for end users who may not realise that they need to move around more than just the *.onnx file.

I would suggest it should be documented in IR.md, and perhaps there are other locations from which it could be s

Hi, Is there any pretrained BART model for Japanese? If not, could you please explain the procedure to train new BART model for Japanese data from scratch?

Traceback (most recent call last):
File "main.py", line 234, in
fire.Fire()
File "/home/zhangqilong/anaconda3/lib/python3.6/site-packages/fire/core.py", line 127, in Fire
component_trace = _Fire(component, args, context, name)
File "/home/zhangqilong/anaconda3/lib/python3.6/site-packages/fire/core.py", line 366, in _Fire
component, remaining_args)
File "/home/zh

❓ Questions & Help

when i run the following example:
examples/mnist_voxel_grid.py

it stoped and warning :

python3.7/site-packages/torch_geometric/nn/conv/spline_conv.py:104: UserWarning: We do not recommend using the non-optimized CPU version of SplineConv. If possible, please convert your data to the GPU.
warnings.warn('We do not recommend using the non-optimized CPU '

how can i fi

The last pooling layer of SENet should be change from self.avg_pool = nn.AvgPool2d(7, stride=1) to self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) so that the model can take different input size.

When I projecting an embedding with different labels, for example:

writer.add_embedding(same_embedding, labels_str_two,
                             tag=f'labels_str_two')
writer.add_embedding(same_embedding, labels_str_one, tag='labels_str_one')

I got two different pictures, just like these two pictures. So why relatively distances between points are different when projecting

Similar to the tutorial on custom losses in SVI, we should have a tutorial on implementing custom MCMC kernels using the new MCMC API. Something simple like SGLD seems like a good starting point.

is it Grid Search can solve CASH problems with NNI , it seems that it is usually used for hyper-parameters optimization, have you guys have finished some revision for Grid Search for solving CASH problems.

about Cash problems can refer to ：microsoft/nni#1178

Describe the bug
FixedPrecision tensor has become a catch-all for a myraid of functionality which have nothing to do with fixing precision. At present it seems to include:

0. The ability to model Fixed Precision Numbers using underlying Integer representations
1. BigInt functionality allowing for large precision
2. Convolutional functionality
3. many references to specific subclass type

anaconda3/envs/py27t04/include/python2.7 -c _roi_crop.c -o ./_roi_crop.o
In file included from /home/wangzhonghao/anaconda3/envs/py27t04/lib/python2.7/site-packages/torch/utils/ffi/../../lib/include/THC/THC.h:4:0,
from _roi_crop.c:493:
/home/wangzhonghao/anaconda3/envs/py27t04/lib/python2.7/site-packages/torch/utils/ffi/../../lib/include/THC/THCGeneral.h:12:18: fatal error: cud