Add features and labels view by vinx13 · Pull Request #4352 · shogun-toolbox/shogun

vinx13 · 2018-07-03T15:58:03Z

Continue #3970 , but features and labels view return raw pointer now instead of Some because that will cause covariant type problem, and Some is not available in SWIG

karlnapf · 2018-07-04T09:38:19Z


+CFeatures* CFeatures::view(const SGVector<index_t>& subset)
+{
+	auto feats_view = this->duplicate();


i would prefer a copy ctor here

this is a virtual function that calls copy ctor of subclasses

ah sorry I see now, totally justified!

karlnapf · 2018-07-04T09:39:58Z

+	return feats_view;
+}
+
+CFeatures* CFeatures::view(const std::vector<index_t>& subset)


this doesnt belong into CFeatures imo, this is conversion code between std::vector and SGVector and therefore should sit in SGVector
features->view(SGVector<float64_t>(my_std_vector))

or even done implicitly (if possible)

karlnapf · 2018-07-04T09:40:33Z


+CLabels* CBinaryLabels::duplicate() const
+{
+	return new CBinaryLabels(*this);


why have this method why not just use the copy ctor?

karlnapf · 2018-07-04T09:41:08Z

+{
+	auto labels_view = this->duplicate();
+
+	auto sg_subset = SGVector<index_t>(subset.size());


same comment as above

karlnapf · 2018-07-04T09:42:17Z

-		// apply subset
-		if (m_subset_frac!=1.0)
-			apply_subset(feats,interf);
+		const auto result = get_subset(feats, interf);


why all these changes? shouldnt we focus on the view first?

this is taken from michele's pr, we need small refactor on this to deploy views

ok,
I technically would prefer first adding the view and then have a second Pr to deploy it.
But I think others see that differently, so ok leave it in :)

karlnapf · 2018-07-04T09:42:59Z

+    CRegressionLabels* inter_f, CLabels* labs)
 {
-	REQUIRE(m_labels,"training labels not set!\n")
-	SGVector<float64_t> labels=(dynamic_cast<CDenseLabels*>(m_labels))->get_labels();


pls avoid such whitespace changes in PRs.
They should be in a sep PR. It just pollutes the diff (i.e. hard to review, takes longer)

karlnapf · 2018-07-04T09:43:32Z

 }

-void CStochasticGBMachine::apply_subset(CDenseFeatures<float64_t>* f, CLabels* interf)
+std::tuple<Some<CDenseFeatures<float64_t>>, Some<CRegressionLabels>,


can you explain the reasonaing here?

since view create a new instance, using smart pointers prevent ref/unref

doesnt this create the same problem with Some?

actually no, this function is only used internally, so we know the static type of return value based on types of args

karlnapf · 2018-07-04T09:44:24Z

 	lab[4]=0;

-	CDenseFeatures<float64_t>* feats=new CDenseFeatures<float64_t>(data);
+	auto feats = some<CDenseFeatures<float64_t>>(data);


again, I would prefer those changes to happen in a separate minimal PR

karlnapf · 2018-07-05T12:28:41Z

 	return feats_view;
 }

-CFeatures* CFeatures::view(const std::vector<index_t>& subset)


can you explain ?
We need to use the base class interface in things like xvalidation ..

ah sorry nevermind I didnt see the arugment was the std

instead: yes ! :)

karlnapf · 2018-07-05T12:32:21Z

+CFeatures* CFeatures::view(const SGVector<index_t>& subset)
+{
+	auto feats_view = this->duplicate();
+	feats_view->add_subset(subset.clone());


i wonder, do we want to clone the subset?
I think actually we shouldnt
And also I think that add_subset should accept a const vector

but maybe you have a different opinion?

agree, we don't need clone here.
add_subset should accept a const vector, and we need to clone the vector in ctor of CSubset

karlnapf · 2018-07-05T12:45:26Z

+	ASSERT_EQ(labels_subset->get_num_labels(), subset.vlen);
+	for (auto i : range(subset.vlen))
+	{
+		EXPECT_EQ(labels_subset->get_int_label(i), labels_true[subset[i]]);


maybe we can actually systematically test all access methods of labels and features somehow?
I would also assert that the view's data pointer is the same as the original one (i.e. no copy happened)

we don't have a way to test different features/labels types, and different access methods now.
and we can't check view's data pointer because that's a private member, calling get_labels will create a copy if subset is present

karlnapf · 2018-07-25T10:43:03Z

+		static_assert(
+		    std::is_base_of<CFeatures, T>::value ||
+		        std::is_base_of<CLabels, T>::value,
+		    "Only CFeatures and CLabels are viewable.");


I am not a fan of error messages that contain things that might change in the future, i.e. a new class is made viewable.
I would just state: Class is not viewable. The compiler error will provide the T and also the static assert which will tell the caller what is viewable and what is not. Or?

karlnapf · 2018-07-25T10:43:26Z

+{
+
+	template <class T>
+	T* view(T* viewable, const SGVector<index_t>& subset)


could we make viewable const?

karlnapf · 2018-07-25T10:44:45Z

+{
+
+	template <class T>
+	T* view(T* viewable, const SGVector<index_t>& subset)


is this used anywhere?
How are we going to do this without multiple inheritance?
Maybe a mixin approach would be better, see IterativeMachine

@karlnapf i'm not clear, could you explain the multiple inheritance problem?
currently we expect T to be either features or labels, but we can also use viewable mixin in features and labels

@karlnapf you mean about the duplciate method or what exactly do you mean by multiple inheritance?

I just realised that I had a mistake in my thinking. This is just a global templated function so all good, nevermind!
I like this idea of having the method with the static assert!

k... note that needs extra love in case of SWIG interfaces... although it's a good question whether we actually wanna expose view mechanism for swig?

I think no.

vigsterkr · 2018-07-24T13:53:48Z

-	REQUIRE(m_labels,"training labels not set!\n")
-	SGVector<float64_t> labels=(dynamic_cast<CDenseLabels*>(m_labels))->get_labels();
+	SGVector<float64_t> labels =
+	    (dynamic_cast<CDenseLabels*>(labs))->get_labels();


since there's already a change here we could just port using as<CDenseLabels>

vigsterkr · 2018-07-24T13:59:28Z

 		SGVector<int32_t> subset(train_indices.data(),train_indices.size(),false);
-		data->add_subset(subset);
-		m_labels->add_subset(subset);
+		auto dense_labels = m_labels->as<CDenseLabels>();


as agreed no need for this, drop it...

vigsterkr · 2018-07-25T11:39:27Z

+{
+
+	template <class T>
+	T* view(T* viewable, const SGVector<index_t>& subset)


@karlnapf you mean about the duplciate method or what exactly do you mean by multiple inheritance?

vinx13 · 2018-07-25T12:47:43Z

+	}
+
+	template <class T>
+	T* view(Some<T> viewable, const SGVector<index_t>& subset)


@karlnapf @vigsterkr maybe we can just return Some<T> here?

Yes please! If you pass Some you should get Some

vigsterkr · 2018-07-25T13:21:38Z

@vinx13 yep cool! this way the c++ code became much much cleaner!

karlnapf · 2018-07-25T17:54:47Z

+	 * @return new viewable instance
+	 */
+	template <class T>
+	T* view(T* viewable, const SGVector<index_t>& subset)


I still vote for const T*

karlnapf · 2018-07-25T17:55:02Z

+	 * @return new viewable instance
+	 */
+	template <class T>
+	Some<T> view(Some<T> viewable, const SGVector<index_t>& subset)


vigsterkr

rebase with latest develop and use override where necessary. this way the CIs wont fail :)

vigsterkr · 2018-07-26T00:08:14Z

 	 */
 	virtual ELabelType get_label_type() const;

+	virtual CLabels* duplicate() const;


use override

vigsterkr · 2018-07-26T00:08:24Z

+		 *
+		 * @return labels object
+		 */
+		virtual CLabels* duplicate() const


use override

karlnapf reviewed Jul 4, 2018

View reviewed changes

karlnapf reviewed Jul 5, 2018

View reviewed changes

vigsterkr force-pushed the feature/transformers branch 2 times, most recently from 77a9f85 to 8711ea6 Compare July 10, 2018 15:12

vinx13 force-pushed the feature/features_view branch 3 times, most recently from d1a175a to ffb3fcc Compare July 12, 2018 06:20

vigsterkr force-pushed the feature/transformers branch from a241ab1 to dbfd69e Compare July 12, 2018 16:02

vinx13 changed the base branch from feature/transformers to develop July 12, 2018 16:38

vinx13 force-pushed the feature/features_view branch 4 times, most recently from 47a55bb to 83884f1 Compare July 15, 2018 06:40

vinx13 force-pushed the feature/features_view branch from 83884f1 to f444cec Compare July 24, 2018 13:33

karlnapf reviewed Jul 25, 2018

View reviewed changes

vigsterkr reviewed Jul 25, 2018

View reviewed changes

vinx13 commented Jul 25, 2018

View reviewed changes

karlnapf reviewed Jul 25, 2018

View reviewed changes

vigsterkr reviewed Jul 26, 2018

View reviewed changes

vinx13 force-pushed the feature/features_view branch from c6ca257 to 9c75248 Compare July 26, 2018 01:13

vinx13 added 13 commits July 26, 2018 15:50

Add features and labels view

319a7d3

Use features view instead of subset

3c3eeee

Add DenseFeatures view test

08f267f

Drop std::vector view api

b7b1f14

Add labels view test

b7765e8

Make subset api accept const vector

7d4ae6b

Don't clone subset when creating features view

0f8e17c

Add view as template function

10d7374

Replace method view with global template function

67c476f

Update assert message

fda53fe

Make arg of view const

67605c3

Replace dynamic cast with as

eea455d

Add override

ec57067

vinx13 force-pushed the feature/features_view branch from 9c75248 to ec57067 Compare July 26, 2018 07:52

Fix ref counting

d98b926

vigsterkr merged commit 2ae291e into shogun-toolbox:develop Jul 26, 2018

ktiefe pushed a commit to ktiefe/shogun that referenced this pull request Jul 30, 2019

Add features and labels view (shogun-toolbox#4352)

420435c

samdbrice added the Tag: GSoC label Sep 6, 2019

Uh oh!

Uh oh!

Conversation

vinx13 commented Jul 3, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

vinx13 Jul 5, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

vinx13 Jul 5, 2018 •

edited

Loading