Update faq/local/index_en.rst #9947
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
translation version 1.0
|
|
abhinavarora
suggested changes
Apr 16, 2018
doc/v2/faq/local/index_en.rst
Outdated
| TBD | ||
| .. contents:: | ||
|
|
||
| 1. Reduce Memory Consuming |
There was a problem hiding this comment.
Consuming -> Consumption
doc/v2/faq/local/index_en.rst
Outdated
| 1. Reduce Memory Consuming | ||
| ------------------- | ||
|
|
||
| The training procedure of neural networks demands dozens gigabytes of host memory or serval gigabytes of device memory, which is a rather memory consuming work. The memory consumed by PaddlePaddle framework mainly includes: |
There was a problem hiding this comment.
dozens gigabytes -> dozens of gigabytes
doc/v2/faq/local/index_en.rst
Outdated
| Reduce DataProvider cache memory | ||
| ++++++++++++++++++++++++++ | ||
|
|
||
| PyDataProvider works under asynchronously mechanism, it loads together with the data fetch and shuffle procedure in host memory: |
There was a problem hiding this comment.
asynchronously -> asynchronous
doc/v2/faq/local/index_en.rst
Outdated
| Data Files -> Host Memory Pool -> PaddlePaddle Training | ||
| } | ||
|
|
||
| Thus the reduction of the DataProvider cache memory can reduce memory occupancy, meanwhile speed up the data loading procedure before training. However, the size of the memory pool can actually effect the granularity of shuffle，which means a shuffle operation is needed before each data file reading process to ensure the randomness of data when try to reduce the size of the memory pool. |
doc/v2/faq/local/index_en.rst
Outdated
|
|
||
| .. literalinclude:: src/reduce_min_pool_size.py | ||
|
|
||
| In such way, the memory consuming can be significantly reduced and hence the training procedure can be accelerated. More details are demonstrated in :ref:`api_pydataprovider2`. |
There was a problem hiding this comment.
In such way -> In this way
There was a problem hiding this comment.
memory consuming -> memory consumption
doc/v2/faq/local/index_en.rst
Outdated
|
|
||
| * Parameters or gradients during training are oversize, which leads to floating overflow during calculation. | ||
| * The model failed to convergence and divert to a big value. | ||
| * Errors in training data leads to parameters converge to a singularity situation. This may also due to the large scale of input data, which contains millions of parameter values, and that will raise float overflow when operating matrix multiplication. |
There was a problem hiding this comment.
Errors in training data leads to parameters converge to a singularity situation.
This sentence does not make any sense. What are you trying to say here?
There was a problem hiding this comment.
also due to -> also be due to
doc/v2/faq/local/index_en.rst
Outdated
|
|
||
| Details can refer to example `machine translation <https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/train.py#L66>`_ 。 | ||
|
|
||
| The main difference of these two methods are: |
There was a problem hiding this comment.
of these two -> between these two
doc/v2/faq/local/index_en.rst
Outdated
|
|
||
| The main difference of these two methods are: | ||
|
|
||
| 1. They both block the gradient, but within different occasion，the former one happens when then :code:`optimzier` updates the network parameters while the latter happens when the back propagation computing of activation functions. |
There was a problem hiding this comment.
but within different occasion. What does this mean?
doc/v2/faq/local/index_en.rst
Outdated
| * Output sequence layer and non sequence layer; | ||
| * Multiple output layers process multiple sequence with different length; | ||
|
|
||
| Such issue can be avoid by calling infer interface and set :code:`flatten_result=False`. Thus, the infer interface returns a python list, in which |
doc/v2/faq/local/index_en.rst
Outdated
| 7. Fetch parameters’ weight and gradient during training | ||
| ----------------------------------------------- | ||
|
|
||
| Under certain situations, know the weights of currently training mini-batch can provide more inceptions of many problems. Their value can be acquired by printing values in :code:`event_handler` (note that to gain such parameters when training on GPU, you should set :code:`paddle.event.EndForwardBackward`). Detailed code is as following: |
abhinavarora
approved these changes
Apr 23, 2018
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
translation version 1.0
fix #8953