Preprocessing: faster build vocab + multiple weighted datasets (OpenN…

…MT#1413)

* handle multiple training corpora and enable weighting
* move fields vocab building logic in function
* fix device handling MultipleDatasetIterator
* fix multi/yield_raw_batch parameter DatasetLazyIter
* update FAQ.md
* add -pool_factor option
* reduce pool_factor for travis runs

.travis.yml

@@ -31,7 +31,7 @@ script:
   # test nmt preprocessing
   - python preprocess.py -train_src data/src-train.txt -train_tgt data/tgt-train.txt -valid_src data/src-val.txt -valid_tgt data/tgt-val.txt -save_data /tmp/data -src_vocab_size 1000 -tgt_vocab_size 1000 && rm -rf /tmp/data*.pt
   # test im2text preprocessing
-  - python preprocess.py -data_type img -shard_size 3 -src_dir /tmp/im2text/images -train_src /tmp/im2text/src-train.txt -train_tgt /tmp/im2text/tgt-train.txt -valid_src /tmp/im2text/src-val.txt -valid_tgt /tmp/im2text/tgt-val.txt -save_data /tmp/im2text/data && rm -rf /tmp/im2text/data*.pt
+  - python preprocess.py -data_type img -shard_size 100 -src_dir /tmp/im2text/images -train_src /tmp/im2text/src-train.txt -train_tgt /tmp/im2text/tgt-train.txt -valid_src /tmp/im2text/src-val.txt -valid_tgt /tmp/im2text/tgt-val.txt -save_data /tmp/im2text/data && rm -rf /tmp/im2text/data*.pt
   # test speech2text preprocessing
   - python preprocess.py -data_type audio -shard_size 300 -src_dir /tmp/speech/an4_dataset -train_src /tmp/speech/src-train.txt -train_tgt /tmp/speech/tgt-train.txt -valid_src /tmp/speech/src-val.txt -valid_tgt /tmp/speech/tgt-val.txt -save_data /tmp/speech/data && rm -rf /tmp/speech/data*.pt
   # test nmt translation
@@ -43,14 +43,14 @@ script:
   # test speech2text translation
   - head /tmp/speech/src-val.txt > /tmp/speech/src-val-head.txt; head /tmp/speech/tgt-val.txt > /tmp/speech/tgt-val-head.txt; python translate.py -data_type audio -src_dir /tmp/speech/an4_dataset -model /tmp/test_model_speech.pt -src /tmp/speech/src-val-head.txt -tgt /tmp/speech/tgt-val-head.txt -verbose -out /tmp/speech/trans; diff /tmp/speech/tgt-val-head.txt /tmp/speech/trans
   # test nmt preprocessing and training
-  - head data/src-val.txt > /tmp/src-val.txt; head data/tgt-val.txt > /tmp/tgt-val.txt; python preprocess.py -train_src /tmp/src-val.txt -train_tgt /tmp/tgt-val.txt -valid_src /tmp/src-val.txt -valid_tgt /tmp/tgt-val.txt -save_data /tmp/q -src_vocab_size 1000 -tgt_vocab_size 1000; python train.py -data /tmp/q -rnn_size 2 -batch_size 10 -word_vec_size 5 -report_every 5 -rnn_size 10 -train_steps 10 && rm -rf /tmp/q*.pt
+  - head -500 data/src-val.txt > /tmp/src-val.txt; head -500 data/tgt-val.txt > /tmp/tgt-val.txt; python preprocess.py -train_src /tmp/src-val.txt -train_tgt /tmp/tgt-val.txt -valid_src /tmp/src-val.txt -valid_tgt /tmp/tgt-val.txt -save_data /tmp/q -src_vocab_size 1000 -tgt_vocab_size 1000; python train.py -data /tmp/q -rnn_size 2 -batch_size 2 -word_vec_size 5 -report_every 5 -rnn_size 10 -train_steps 10 && rm -rf /tmp/q*.pt
   # test nmt preprocessing w/ sharding and training w/copy
-  - head data/src-val.txt > /tmp/src-val.txt; head data/tgt-val.txt > /tmp/tgt-val.txt; python preprocess.py -train_src /tmp/src-val.txt -train_tgt /tmp/tgt-val.txt -valid_src /tmp/src-val.txt -valid_tgt /tmp/tgt-val.txt -shard_size 1 -dynamic_dict -save_data /tmp/q -src_vocab_size 1000 -tgt_vocab_size 1000; python train.py -data /tmp/q -rnn_size 2 -batch_size 10 -word_vec_size 5 -report_every 5 -rnn_size 10 -copy_attn -train_steps 10 && rm -rf /tmp/q*.pt
+  - head -50 data/src-val.txt > /tmp/src-val.txt; head -50 data/tgt-val.txt > /tmp/tgt-val.txt; python preprocess.py -train_src /tmp/src-val.txt -train_tgt /tmp/tgt-val.txt -valid_src /tmp/src-val.txt -valid_tgt /tmp/tgt-val.txt -shard_size 25 -dynamic_dict -save_data /tmp/q -src_vocab_size 1000 -tgt_vocab_size 1000; python train.py -data /tmp/q -rnn_size 2 -batch_size 2 -word_vec_size 5 -report_every 5 -rnn_size 10 -copy_attn -train_steps 10 -pool_factor 10 && rm -rf /tmp/q*.pt
 
   # test im2text preprocessing and training
-  - head /tmp/im2text/src-val.txt > /tmp/im2text/src-val-head.txt; head /tmp/im2text/tgt-val.txt > /tmp/im2text/tgt-val-head.txt; python preprocess.py -data_type img -src_dir /tmp/im2text/images -train_src /tmp/im2text/src-val-head.txt -train_tgt /tmp/im2text/tgt-val-head.txt -valid_src /tmp/im2text/src-val-head.txt -valid_tgt /tmp/im2text/tgt-val-head.txt -save_data /tmp/im2text/q; python train.py -model_type img -data /tmp/im2text/q -rnn_size 2 -batch_size 10 -word_vec_size 5 -report_every 5 -rnn_size 10 -train_steps 10 && rm -rf /tmp/im2text/q*.pt
+  - head -50 /tmp/im2text/src-val.txt > /tmp/im2text/src-val-head.txt; head -50 /tmp/im2text/tgt-val.txt > /tmp/im2text/tgt-val-head.txt; python preprocess.py -data_type img -src_dir /tmp/im2text/images -train_src /tmp/im2text/src-val-head.txt -train_tgt /tmp/im2text/tgt-val-head.txt -valid_src /tmp/im2text/src-val-head.txt -valid_tgt /tmp/im2text/tgt-val-head.txt -save_data /tmp/im2text/q -tgt_seq_length 100; python train.py -model_type img -data /tmp/im2text/q -rnn_size 2 -batch_size 2 -word_vec_size 5 -report_every 5 -rnn_size 10 -train_steps 10 -pool_factor 10 && rm -rf /tmp/im2text/q*.pt
   # test speech2text preprocessing and training
-  - head /tmp/speech/src-val.txt > /tmp/speech/src-val-head.txt; head /tmp/speech/tgt-val.txt > /tmp/speech/tgt-val-head.txt; python preprocess.py -data_type audio -src_dir /tmp/speech/an4_dataset -train_src /tmp/speech/src-val-head.txt -train_tgt /tmp/speech/tgt-val-head.txt -valid_src /tmp/speech/src-val-head.txt -valid_tgt /tmp/speech/tgt-val-head.txt -save_data /tmp/speech/q; python train.py -model_type audio -data /tmp/speech/q -rnn_size 2 -batch_size 10 -word_vec_size 5 -report_every 5 -rnn_size 10 -train_steps 10 && rm -rf /tmp/speech/q*.pt
+  - head -100 /tmp/speech/src-val.txt > /tmp/speech/src-val-head.txt; head -100 /tmp/speech/tgt-val.txt > /tmp/speech/tgt-val-head.txt; python preprocess.py -data_type audio -src_dir /tmp/speech/an4_dataset -train_src /tmp/speech/src-val-head.txt -train_tgt /tmp/speech/tgt-val-head.txt -valid_src /tmp/speech/src-val-head.txt -valid_tgt /tmp/speech/tgt-val-head.txt -save_data /tmp/speech/q; python train.py -model_type audio -data /tmp/speech/q -rnn_size 2 -batch_size 2 -word_vec_size 5 -report_every 5 -rnn_size 10 -train_steps 10 -pool_factor 10 && rm -rf /tmp/speech/q*.pt
   # test nmt translation
   - python translate.py -model onmt/tests/test_model2.pt  -src  data/morph/src.valid  -verbose -batch_size 10 -beam_size 10 -tgt data/morph/tgt.valid -out /tmp/trans; diff  data/morph/tgt.valid /tmp/trans
   # test nmt translation with random sampling

docs/source/FAQ.md

@@ -2,9 +2,9 @@
 
 ## How do I use Pretrained embeddings (e.g. GloVe)?
 
-Using vocabularies from OpenNMT-py preprocessing outputs, `embeddings_to_torch.py` to generate encoder and decoder embeddings initialized with GloVe’s values.
+Using vocabularies from OpenNMT-py preprocessing outputs, `embeddings_to_torch.py` to generate encoder and decoder embeddings initialized with GloVe's values.
 
-the script is a slightly modified version of ylhsieh’s one2.
+the script is a slightly modified version of ylhsieh's one2.
 
 Usage:
 
@@ -105,4 +105,34 @@ if you use a regular network card (1 Gbps) then we suggest to use a higher accum
 You can specify several models in the translate.py command line: -model model1_seed1 model2_seed2
 Bear in mind that your models must share the same traget vocabulary.
 
+## How can I weight different corpora at training?
 
+### Preprocessing
+
+We introduced `-train_ids` which is a list of IDs that will be given to the preprocessed shards.
+
+E.g. we have two corpora : `parallel.en` and  `parallel.de` + `from_backtranslation.en` `from_backtranslation.de`, we can pass the following in the `preprocess.py` command:
+```
+...
+-train_src parallel.en from_backtranslation.en \
+-train_tgt parallel.de from_backtranslation.de \
+-train_ids A B \
+-save_data my_data \
+...
+```
+and it will dump `my_data.train_A.X.pt` based on `parallel.en`//`parallel.de` and `my_data.train_B.X.pt` based on `from_backtranslation.en`//`from_backtranslation.de`.
+
+### Training
+
+We introduced `-data_ids` based on the same principle as above, as well as `-data_weights`, which is the list of the weight each corpus should have.
+E.g.
+```
+...
+-data my_data \
+-data_ids A B \
+-data_weights 1 7 \
+...
+```
+will mean that we'll look for `my_data.train_A.*.pt` and `my_data.train_B.*.pt`, and that when building batches, we'll take 1 example from corpus A, then 7 examples from corpus B, and so on.
+
+**Warning**: This means that we'll load as many shards as we have `-data_ids`, in order to produce batches containing data from every corpus. It may be a good idea to reduce the `-shard_size` at preprocessing.