The big refactor #67

patricksnape · 2015-08-13T13:45:44Z

OK, so this is the big refactoring PR. I rebased this version from the refactor branch so that we have a consistent history. Some of the commits might be a little borked from this, but it won't matter much. The end result of that final merge is that this branch is now 100% consistent with the refactor branch.

I want to get this in to master ASAP so that people can start to play with it. Given that everything at least appears to work, I think that this is OK. It also has the advantage of building master conda builds that we can share with people.

The major missing components are:

Testing
Documentation
Fitting results refactoring
New algorithms

However, the major components of the SDM, CLM, AAM, ATM and LK algorithms are now all fairly consistent and appear to be working.

Brief Tour Of Changes

New CLMs and AAMs. Adds all of @jalabort work on new AAM and CLM implementations. There are now many new algorithms and features such as priors for AAMs. CLMs in particular have had a lot of work done to them including fancy FFT convolution.
Goodbye builders. The base classes are now builders themselves. They no longer have train methods and instead build in the constructor. This was necessary for the next update...
Incremental models. SDM, ATM, AAM and CLM can all be trained incrementally. This is great for building large models, which also enables...
Batched building. You can now build models using a generator, exhausting a finite number of inputs at a time. Where possible, we have attempted to perform an exact incremental update. However, this is not always possible for methods like the SDM - where small batch sizes will likely cause divergence.
More sensible initialization schemes. @jalabort updated the way in which perturbations are generated, to a much more sensible scheme that is more easily parametrizable.

Conflicts: menpofit/aam/base.py menpofit/aam/builder.py menpofit/aam/fitter.py menpofit/builder.py menpofit/checks.py menpofit/fitter.py menpofit/fittingresult.py menpofit/sdm/fitter.py

Interfaces are now concrete rather than passing a class.

This makes the objects much less heavyweight

No more exceptions from iterations or error types.

Because we just deepcopy the shape model, we were throwing the components away too early. Now we build all the shape models, then truncate them at the end.

Also, add HolisticAAM as AAM alias

@jalabort

After speaking to @jalabort - he found this more useful.

Working on refactoring the ATMs

This fix is like not modelling the noise at all - its a reasonable fix but we should think carefully about it.

Handle images and shapes

Helps visualize just work at the moment

Now ATMs act like AAMs - including the new interface usage so that the algorithms just take a single interface. A lot of the code is similar to the AAM - but not identical!

Just use the correct ordering for scales etc. Also, use _prepare_image rather than _prepare_template to make sure everything is consistent

Since we use Patch all the time in SDM etc, we should probably use it for the 'parts' AAM which is more similar to SDM. Since 'Patch' was being used merely for an AAM where the appearance model has a disjoint mask applied - we change the name to 'Masked'

AAM/ATM/LK/SDM Refactoring

Now everyone method has essentially the same _train method. Then, this method calls a _train_batch method that is different for each method and actually performs training.

Consistency for training

Wasn't passing parameters up to super class properly.

When holistic feature is no_op, don't copy images or even bother running through the loop, just skip it.

Includes non-verbose bug for CLMs

This renames the fitting methods and adds a new method that can fit from bounding boxes.

This is consistent with the previous renaming of parts aam to patch aam

Now everyone method has essentially the same _train method. Then, this method calls a _train_batch method that is different for each method and actually performs training.

Wasn't passing parameters up to super class properly.

When holistic feature is no_op, don't copy images or even bother running through the loop, just skip it.

Includes non-verbose bug for CLMs

This renames the fitting methods and adds a new method that can fit from bounding boxes.

This is consistent with the previous renaming of parts aam to patch aam

This is consistent with Menpo's extract_patches method and with CLM

features -> holistic_features normalize_parts -> patch_normalisation normalise_callable -> patch_normalisation patch_features -> patch_features At the moment, only SDMs have patch_features due to the lack of distinction on menpo between image and vector features.

Consistent with other renaming

Conflicts: menpofit/aam/algorithm/sd.py menpofit/checks.py menpofit/fitter.py

jabooth · 2015-08-13T13:49:11Z

Fantastic work guys.

As you've already been doing code reviews on the refactor branch I see no reason why we would delay getting this in to master - as you've said once there it will be built with conda and easier to test internally before release.

+1 after CI builds succeed.

nontas · 2015-08-13T13:56:44Z

Awesome!! +1

Empty strings are no longer allowed as colours - use None instead.

Important to stop appveyor from hangin

Fixes WIndows builds.

Removal of *Widget classes

This helps when you have different number of features per patch, otherwise, you get an object array back. (SDM)

patricksnape · 2015-08-26T14:05:15Z

Noone is saying anything about this... so! MERGING

The big refactor

patricksnape and others added 30 commits August 4, 2015 14:30

Merge branch 'refactor' into aam_fixes

45486aa

Conflicts: menpofit/aam/base.py menpofit/aam/builder.py menpofit/aam/fitter.py menpofit/builder.py menpofit/checks.py menpofit/fitter.py menpofit/fittingresult.py menpofit/sdm/fitter.py

Fix incremental Gauss-Newton

afc4c43

Refactoring AAM to match SD-AAM

b0fefcc

Interfaces are now concrete rather than passing a class.

Refactor results to take a transform rather than algorithm

b9a4cd9

This makes the objects much less heavyweight

Fix menpofit widgets

dc67955

No more exceptions from iterations or error types.

Nasty bug about truncating shape model

6c25a53

Because we just deepcopy the shape model, we were throwing the components away too early. Now we build all the shape models, then truncate them at the end.

Fix __str__ for AAMs

ee9e58b

Also, add HolisticAAM as AAM alias

Change results back to taking algorithm

d887081

After speaking to @jalabort - he found this more useful.

Partial commit - moving home due to tube strike

b4fe608

Working on refactoring the ATMs

Remove xrange - for Python 3 compat

537b6df

Incorrect call to super - fixed

15706b6

Fix noise variance bug

ca0f3f7

This fix is like not modelling the noise at all - its a reasonable fix but we should think carefully about it.

Refactor trilist check method

a95f861

Handle images and shapes

Remove mention of downscale etc from visualize

03c3216

Helps visualize just work at the moment

Remove unused methods

f5b4217

Fix bugs using wrong index in AAM

d782cf8

Update ATM to work with AAM layout

5640d60

Now ATMs act like AAMs - including the new interface usage so that the algorithms just take a single interface. A lot of the code is similar to the AAM - but not identical!

Refactor LK to match new scheme

665ef0a

Just use the correct ordering for scales etc. Also, use _prepare_image rather than _prepare_template to make sure everything is consistent

Update travis and appveyor

b548310

Fix a bunch of relative import errors

13394e1

Merge pull request #60 from menpo/aam_fixes

9eaf568

AAM/ATM/LK/SDM Refactoring

Refactor training to be more consistent

5a01c06

Now everyone method has essentially the same _train method. Then, this method calls a _train_batch method that is different for each method and actually performs training.

Move _train to top in CLM - consistent with AAM etc

5ab5f67

Merge pull request #62 from menpo/cleaning

0b7e0b7

Consistency for training

Fix silly bug in PatchAAM

d642ab1

Wasn't passing parameters up to super class properly.

Small inhancement for no_op

95c965d

When holistic feature is no_op, don't copy images or even bother running through the loop, just skip it.

Fix printing a bit

554ccd1

Includes non-verbose bug for CLMs

Change fitting methods to fit_from_shape/bb

343190e

This renames the fitting methods and adds a new method that can fit from bounding boxes.

Rename Parts interfaces to patch

497e527

This is consistent with the previous renaming of parts aam to patch aam

patricksnape added 12 commits August 13, 2015 14:25

Fix a bunch of relative import errors

0353e72

Refactor training to be more consistent

d53a9c1

Now everyone method has essentially the same _train method. Then, this method calls a _train_batch method that is different for each method and actually performs training.

Move _train to top in CLM - consistent with AAM etc

15b34c2

Fix silly bug in PatchAAM

b0a02a2

Wasn't passing parameters up to super class properly.

Small inhancement for no_op

ff07836

When holistic feature is no_op, don't copy images or even bother running through the loop, just skip it.

Fix printing a bit

d592ddb

Includes non-verbose bug for CLMs

Change fitting methods to fit_from_shape/bb

4ea9c52

This renames the fitting methods and adds a new method that can fit from bounding boxes.

Rename Parts interfaces to patch

00b28cc

This is consistent with the previous renaming of parts aam to patch aam

Universally rename patch_shape to patch_size

fe6cdec

This is consistent with Menpo's extract_patches method and with CLM

Rename partsinterface to patchinterface

bc57cbd

Consistent with other renaming

Merge branch 'refactor' into refactor_merge

b838476

Conflicts: menpofit/aam/algorithm/sd.py menpofit/checks.py menpofit/fitter.py

jabooth added the in progress label Aug 13, 2015

jalabort mentioned this pull request Aug 13, 2015

AAMs problems in Refactor branch #68

Open

patricksnape added 9 commits August 18, 2015 13:15

Fixes for Jupyter 4.0

5330aaa

Empty strings are no longer allowed as colours - use None instead.

Update to v0.4.4 condaci

2a57367

Important to stop appveyor from hangin

Update condaci

7d0806a

Fixes WIndows builds.

Upgrade to latest condaci

cc3509d

More fixes for Jupyter 4.0

f86830c

Removal of *Widget classes

Move len(scales) to AFTER it has been checked

1f406b5

ipywidgets rather than IPython (jupyter 4)

4b127f7

Fix passing single scale

c6fea14

Change asarray to stacking

1a7a2f4

This helps when you have different number of features per patch, otherwise, you get an object array back. (SDM)

patricksnape pushed a commit that referenced this pull request Aug 26, 2015

Merge pull request #67 from menpo/refactor_merge

aa056af

The big refactor

patricksnape merged commit aa056af into master Aug 26, 2015

jabooth removed the in progress label Aug 26, 2015

patricksnape deleted the refactor_merge branch August 26, 2015 14:05

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

The big refactor #67

The big refactor #67

patricksnape commented Aug 13, 2015

jabooth commented Aug 13, 2015

nontas commented Aug 13, 2015

patricksnape commented Aug 26, 2015

The big refactor #67

The big refactor #67

Conversation

patricksnape commented Aug 13, 2015

Brief Tour Of Changes

jabooth commented Aug 13, 2015

nontas commented Aug 13, 2015

patricksnape commented Aug 26, 2015