Inference example and inference-framework unit-test for NMT model #8314
Conversation
… bc_machinetranslation
… bc_machinetranslation
… bc_machinetranslation
… bc_machinetranslation
|
|
||
| #include <gtest/gtest.h> | ||
| #include "gflags/gflags.h" | ||
| #include "test_helper.h" |
There was a problem hiding this comment.
Please use full path name with #include. Here it should be
#include "paddle/fluid/inference/tests/book/test_helper.h"
generic.cmake should fail if it is not full-path inclusion; I will figure out why it doesn't.
…bc_machinetranslation
… bc_machinetranslation
… bc_machinetranslation
|
|
||
| #ifdef PADDLE_WITH_CUDA | ||
| LOG(INFO) << "Beam search isn't supported on gpu yet"; | ||
| /* |
There was a problem hiding this comment.
Please delete code which are no longer needed.
| @@ -29,7 +29,7 @@ | |||
| max_length = 8 | |||
| topk_size = 50 | |||
| trg_dic_size = 10000 | |||
| beam_size = 2 | |||
| beam_size = 1 | |||
There was a problem hiding this comment.
Why beam_size =1 is better than 2?
There was a problem hiding this comment.
I think I just changed it for debugging purposes. Thanks for pointing it out, will fix.
| set_tests_properties(test_inference_machine_translation_train | ||
| PROPERTIES DEPENDS test_machine_translation) | ||
| set_tests_properties(test_inference_machine_translation_decode | ||
| PROPERTIES DEPENDS test_machine_translation) |
There was a problem hiding this comment.
Please try to use inference_test here, and try to simplify the unittest.
There was a problem hiding this comment.
I couldn't get both the tests to be expressed with the inference_test function. So I added explicit statements. Do you suggest modifying the inference_test function to accommodate this?
|
|
||
| // Setup input sequence of 5 words | ||
| paddle::framework::LoDTensor input_sequence; | ||
| std::vector<int64_t> inp_data = {100, 50, 8, 94, 122}; |
There was a problem hiding this comment.
Does this data have a specific meaning? Or can we use random data?
There was a problem hiding this comment.
This is a random sequence of 5 words (The ids need to be within a range). I will make it clearer.
|
Since there is an issue with the |
|
Closing the PR |
Addresses second part of #7999
Background
The fluid example of NMT model has 2 parts:
train_main()anddecode_main(). As discussed with @jacquesqiao , the current implementation of the book chapter (as of Feb 8, 2018), is a bit disjoint. More specifically, thedecode_mainshould ideally use the parameters post training (that is, after callingtrain_main()). However, currently thetrain_main()demonstrates the working example of dynamic RNN training, and currentdecode_main()uses a random initialization of the weights in the encoder-decoder and demonstrates correctness of beam search. Hence, both of the modules are pretty independent.Current progress (updated)
In order to add inference example for the "inference project", one option is to add "infer" methods for each of the two
train_main()anddecode_main(). This is what is done in this PR.train_main, we have to provide both source as well as target sequence, even though ideally we don't have access to the target (This is the same issue as in PR Add Inference example and unit test for rnn_encoder_decoder #8176). This works.decode_main, we add the function in Python and it works.