Python version degrades

[Kalyan Guntupalli]

while using this conda install -y -c caffe2 "pytorch-caffe2-cuda${TMP_CUDA_VERSION}.0-cudnn7"
python version degrades to 2.7.15 and hence is not able to export model using https://github.com/pytorch/translate#exporting-a-model-with-onnx

Can this conversion be done on a CPU
Does pytorch1.0 does not need many of the above steps and facilitate the conversion process..

[Liezl]

export should be doable on CPU. Note that none of our ONNX tests require GPU: https://github.com/pytorch/translate/blob/master/pytorch_translate/test/test_onnx.py

whereas training DOES require GPU: https://github.com/pytorch/translate/blob/master/pytorch_translate/test/test_train.py

Tests which require GPU have this decorator @unittest.skipIf(torch.cuda.device_count() < 1, "No GPU available for test.")

[Kalyan Guntupalli]

the build fails at make 2>&1 | tee MAKE_OUT
with the log ::

Scanning dependencies of target translation_decoder
[ 16%] Building CXX object CMakeFiles/translation_decoder.dir/Decoder.cpp.o
In file included from /home/local/usr/miniconda3/envs/pytrans/include/caffe2/core/logging.h:12:0,
from /home/local/usr/miniconda3/envs/pytrans/include/caffe2/core/init.h:6,
from /home/local/usr/translate/pytorch_translate/cpp/Decoder.cpp:32:
/home/local/usr/miniconda3/envs/pytrans/include/caffe2/proto/caffe2.pb.h:17:2: error: #error This file was generated by an older version of protoc which is
#error This file was generated by an older version of protoc which is
^

Please help @liezl200

According to this BVLC/caffe#5645 uninstall libprotobuf in conda installed, solves it, but the pytorch+caffe2 also gets uninstalled with it .

These are exactly the steps I follow :: (I just need to export a fairseq trained translation model to onnx on CPU)

git clone https://github.com/pytorch/translate.git
pushd translate

conda install -y -c caffe2 pytorch-caffe2
conda install -y numpy==1.14 --no-deps
export CONDA_PATH="$(dirname $(which conda))/.."

git clone --recursive https://github.com/onnx/onnx.git
yes | pip install ./onnx 2>&1 | tee ONNX_OUT

pip uninstall -y pytorch-translate
python3 setup.py build develop
pushd pytorch_translate/cpp

mkdir build && pushd build
cmake
-DCMAKE_PREFIX_PATH="${CONDA_PATH}/usr/local"
-DCMAKE_INSTALL_PREFIX="${CONDA_PATH}" ..
2>&1 | tee CMAKE_OUT

[Kalyan Guntupalli]

Is there average attention, shared embeddings, knowledge distillation under works as the Marian-NMT people do? @liezl200

[jhcross]

@gvskalyan, knowledge distillation is indeed in the works. Note also that shared embeddings are already an option for the transformer architecture [1]. We don't have immediate plans to add average attention, but we suggest looking at the hybrid architecture [2], which in the same spirit (faster inference for transformer models).

[1]

translate/pytorch_translate/transformer.py

Lines 159 to 171 in 7752ad7

 parser.add_argument( 
     "--share-decoder-input-output-embed", 
     default=False, 
     action="store_true", 
     help="share decoder input and output embeddings", 
 ) 
 parser.add_argument( 
     "--share-all-embeddings", 
     default=False, 
     action="store_true", 
     help="share encoder, decoder and output embeddings" 
     " (requires shared dictionary and embed dim)", 
 ) 

[2]

translate/pytorch_translate/hybrid_transformer_rnn.py

Line 23 in 7752ad7

 class HybridTransformerRNNModel(FairseqModel): 

[Kalyan Guntupalli]

I am still unable to make build due to the above error.

[FYVictor93]

I am still unable to make build due to the above error.

I had the same problem. Python 3.6 degrades to 2.7.

[Kalyan Guntupalli]

@FuKaiYin You could try training the model in translate itself and use the docker image to export the model that might it!