Improve apply_prediction #523

Hakuyume · 2018-02-27T04:05:08Z

renamed imgs, pred_values, gt_values -> in_values, out_values, rest_values
in_values are the first n_input elements of the batch
renamed apply_prediction_to_iterator -> apply_to_iterator

yuyu2172 · 2018-02-27T07:34:17Z

chainercv/utils/iterator/apply_to_batch.py

+
+    >>> from chainer import iterators
+    >>>
+    >>> from chainercv.datasets import VOCDetectionDataset


VOCBBoxDataset. This is my mistake. Sorry.

yuyu2172 · 2018-02-27T07:53:15Z

chainercv/utils/iterator/apply_to_batch.py

+
+
+def apply_to_batch(func, iterator, n_input=1, hook=None):
+    """Apply a function/method to an iterator of batches.


This is not directly related to the change proposed in this PR, but this sentence is incorrect.
We apply a function to batches and not an iterator.
Thus, Apply a function/method to batches of data from an iterator is more precise.

I looked up the usage of apply on the internet.
https://pandas.pydata.org/pandas-docs/stable/generated/pandas.DataFrame.apply.html

yuyu2172 · 2018-02-27T08:02:33Z

chainercv/utils/iterator/apply_to_batch.py

+    or
+    >>> # batch: [(in_val0, in_val1, ...)]
+    or
+    >>> # batch: [(in_val0, in_val1, ..., rest_val0, rest_val1, ...)]


n_input should appear at this part of the documentation.

yuyu2172 · 2018-02-27T08:52:22Z

chainercv/utils/iterator/apply_to_batch.py

+    >>> out_vals = func(in_val0, in_val1, ...)
+    >>> # out_vals: [out_val]
+    or
+    >>> out_vals0, out_vals1, ... = func(in_val0, in_val1)


in_val0, in_val1 --> in_val0, in_val1, ... for consistency

I think these examples are not necessary.

I would like to show the returned values should be ([out_val0], [out_val1], ...) and not [(out_val0, out_val1, ...)]. I feel it unclear without examples.

Ah. I see. So you wanted to emphasize that the function returns tuple of lists.

BTW, shouldn't function take list of values?
So out_vals0, out_vals1, ... = func(in_val0, in_val1) --> out_vals0, out_vals1, ... = func(in_vals0, in_vals1).

You are right. That's my mistake.

yuyu2172 · 2018-02-27T08:56:21Z

chainercv/utils/iterator/apply_to_batch.py

+
+    This function applies a function/method to an iterator of batches.
+    It assumes that the iterator returns a batch of input data or
+    a batch of tuples whose first elements ara input data.


I thought it was better to describe the internals of this function more in the beginning.
I wrote an alternative docstring. What do you think?

This function assumes that the iterator iterates over a collection of tuples that contain inputs to func.
Additionally, the tuples may contain values that are not used by func.
For convenience, we allow the iterator to iterate over a collection of inputs that are not tuple.
t
Conceptually, this function does two operations.
First, it splits the iterator into two iterators. The first iterator contains the inputs to func and the second iterator contains the rest of the data from iterator.
Second, it creates another iterator that iterates over a collection of the outputs of the func.

Here is an illustration of the expected behavior of iterator.
M is the number of arguments for func, which is defined by an argument n_input. N is the number of elements in a tuple from iterator that are not used by func. Note that N can be 0.

>>> batch = next(iterator) >>> # batch: [(in_val0, ..., in_val{M-1}, rest_val0, ..., rest_val{N-1})] or >>> # batch: [in_val]

First, it splits the iterator into two iterators. The first iterator contains the inputs to func and the second iterator contains the rest of the data from iterator.
Second, it creates another iterator that iterates over a collection of the outputs of the func.

This explanation sounds this function returns three iterators. However, it is not correct.

Right. That is my mistake.

OK, I will try to merge your solution into docstring.

yuyu2172 · 2018-03-01T03:13:29Z

chainercv/utils/iterator/apply_to_iterator.py

+
+    >>> in_values, out_values, rest_values = apply_to_iterator(
+    >>>     func, iterator, n_input)
+    >>> # in_values: (iter of in_val0, ..., iter of in_val{n_input})


n_input --> n_input-1

yuyu2172 · 2018-03-01T05:12:09Z

chainercv/utils/iterator/apply_to_iterator.py

+def apply_to_iterator(func, iterator, n_input=1, hook=None):
+    """Apply a function/method to batches from an iterator.
+
+    This function applies a function/method to an iterator of batches.


This function applies a function/method to batches from an iterator of type chainer.Iterator.

This makes clear that we are talking about Chainer's iterator.

This is not only for chainer.Iterator. It can work with other iterators. an iterator of type chainer.Iterator sounds it requires some special features.

OK. Then we should change args doc.
https://github.com/Hakuyume/chainercv/blob/0bd2f6d970aeaca436e6858b0e53661e364c1fd6/chainercv/utils/iterator/apply_to_iterator.py#L75

Thank you for pointing out.

yuyu2172 · 2018-03-01T05:12:26Z

chainercv/utils/iterator/apply_to_iterator.py

+
+    It assumes that the iterator iterates over a collection of tuples
+    that contain inputs to :func:`func`.
+    Additionally, the tuples may contain values that are not used by func.


s/by func/by the funciton

yuyu2172

LGTM except one comment.

yuyu2172 · 2018-03-01T06:28:10Z

chainercv/utils/iterator/apply_to_iterator.py

+    that are not used by :func:`func`.
+    For convenience, we allow the iterator to iterate over a collection of
+    inputs that are not tuple.
+    Here is an illustration of the expected behavior of the iterator.


I think this comment is helpful for Chainer users.

Here is an ....
In short, we require iterator to satisfy the interface of chainer.Iterator.

chainer.Iterator requires many properties, such as batch_size, epoch_detail, etc..
apply_to_iterator does not require these properties.
https://docs.chainer.org/en/stable/reference/generated/chainer.iterators.SerialIterator.html

How about chainer.Iteratror satisfies this requirement or chainer.Iteratror satisfies this behaviour? These are correct.

This behaviour is same as chainer.Iterator is also OK.

This behaviour is the same as chainer.Iterator

This one looks good! (“the” was missing)

Hakuyume · 2018-03-05T10:05:49Z

@yuyu2172

Hakuyume added 3 commits February 27, 2018 11:22

add n_input option

4179b00

update ProgressHook

6ebf6e6

update test

4afd154

Hakuyume changed the title ~~Improve apply_prediction~~ [WIP] Improve apply_prediction Feb 27, 2018

Hakuyume added 4 commits February 27, 2018 13:31

rename and update docs

14d0973

update related evaluator

cdbc657

update related examples

592d30a

Merge branch 'master' into improve-apply-prediction

4db7d8b

Hakuyume changed the title ~~[WIP] Improve apply_prediction~~ Improve apply_prediction Feb 27, 2018

Hakuyume assigned yuyu2172 Feb 27, 2018

Hakuyume added 3 commits February 27, 2018 14:32

for Python2

2a6a4a2

fix

7a57c6c

apply_to_batch -> apply_to_iterator

77acf2e

yuyu2172 reviewed Feb 27, 2018

View reviewed changes

yuyu2172 added enhancement no-compat labels Feb 28, 2018

yuyu2172 added this to the v0.9 milestone Feb 28, 2018

update docs

2d3cb55

yuyu2172 reviewed Mar 1, 2018

View reviewed changes

Hakuyume added 4 commits March 1, 2018 12:15

fix docs

a4e7c2e

n_input - 1

2e052a4

fix tuple init

f920744

fix tuple init

cc91196

yuyu2172 reviewed Mar 1, 2018

View reviewed changes

Hakuyume added 3 commits March 1, 2018 14:28

use :func:

0bd2f6d

do not assume chainer.Iterator

2e449a3

batch -> batches

583ed87

yuyu2172 reviewed Mar 1, 2018

View reviewed changes

add note

09f7ade

yuyu2172 approved these changes Mar 5, 2018

View reviewed changes

yuyu2172 merged commit 4d239ad into chainer:master Mar 5, 2018

Hakuyume deleted the improve-apply-prediction branch March 5, 2018 12:43

naoto0804 mentioned this pull request Oct 18, 2018

NameError: name 'apply_prediction_to_iterator' is not defined naoto0804/cross-domain-detection#5

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Improve apply_prediction #523

Improve apply_prediction #523

Hakuyume commented Feb 27, 2018 •

edited

Loading

yuyu2172 Feb 27, 2018

yuyu2172 Feb 27, 2018

yuyu2172 Feb 27, 2018

yuyu2172 Feb 27, 2018

yuyu2172 Feb 27, 2018

yuyu2172 Feb 27, 2018

Hakuyume Mar 1, 2018 •

edited

Loading

yuyu2172 Mar 1, 2018

Hakuyume Mar 1, 2018

yuyu2172 Feb 27, 2018

Hakuyume Feb 27, 2018

yuyu2172 Feb 27, 2018

Hakuyume Feb 27, 2018

yuyu2172 Mar 1, 2018

yuyu2172 Mar 1, 2018

Hakuyume Mar 1, 2018

yuyu2172 Mar 1, 2018

Hakuyume Mar 1, 2018

yuyu2172 Mar 1, 2018

yuyu2172 left a comment

yuyu2172 Mar 1, 2018

Hakuyume Mar 1, 2018

Hakuyume Mar 1, 2018

Hakuyume Mar 1, 2018

yuyu2172 Mar 1, 2018

Hakuyume commented Mar 5, 2018



		def apply_to_batch(func, iterator, n_input=1, hook=None):
		"""Apply a function/method to an iterator of batches.

Improve apply_prediction #523

Improve apply_prediction #523

Conversation

Hakuyume commented Feb 27, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Hakuyume Mar 1, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

yuyu2172 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Hakuyume commented Mar 5, 2018

Hakuyume commented Feb 27, 2018 •

edited

Loading

Hakuyume Mar 1, 2018 •

edited

Loading