Reshape single input batches for inputs of varying dimension #1313

shelhamer · 2014-10-17T05:55:55Z

To feed inputs of varying dimension, the DATA layer reshapes its prefetch and top blobs when the batch size is 1. This is useful for models of variable input size, such as fully convolutional models.

By the grace of #594 this is a simple change.

simplify to reshape always (there's no cost)
make compatible with crop thanks to @philkr
test data layer reshaping

sguada · 2014-10-18T04:36:16Z

src/caffe/layers/data_layer.cpp

Why not const Datum& datum = iter_->value;

By they way you forgot to add DecodeDatum(&datum); just in case datum was encoded

The Datum isn't stored with the proper dimensions and it's only set right when decoding? Shouldn't encoding / decoding be isolated to the data, float_data fields?

bhack · 2014-12-22T07:43:25Z

@shelhamer Reading this ticket comments the status is unknown. Was self assigned but there is no evidence of further step needs , no todo list, no bullet points.

longjon · 2014-12-24T01:47:34Z

src/caffe/layers/data_layer.cpp

Does this work if crop_size is nonzero?

longjon · 2014-12-24T01:57:29Z

So I would like to merge this soon. Why not lose Reshape, move the current shaping code from DataLayerSetUp to InternalThreadEntry, and add an unconditional reshape to Forward (in BaseDataLayer?).

Then cropping should work fine (it looks broken, no?), and the patch should be only be adding ~2 net loc. Unless I'm forgetting something.

Reshape single input batches for inputs of varying dimension

shelhamer · 2015-01-07T02:13:26Z

@longjon I'll take care of this once #1568 is reviewed and finished.

bhack · 2015-01-08T09:26:45Z

@longjon Can you check some of the last @sguada comments on our #1416? I think that if we want to let top reshaping a forward business we need to use it in the same way in this two PR.

shelhamer · 2015-02-05T19:54:29Z

@longjon re: #1313 (comment) the DataLayerSetup reshaping is needed if the net is to have any dimensions before a call to Forward, as needed for diagnostic output during initialization and for the data layer tests.

It does seem like this could be simplified so that data layers unconditionally reshape, but we need to decide what to do about Net::Init() and I'd rather defer that.

shelhamer · 2015-02-05T20:28:45Z

~~By the way this is still broken w.r.t. crop.~~ This works and has tests but could use a clearer check to catch the case when batch size > 1 but the source data has different shapes instance by instance. This will fail by a dimension check in the data transformer, but that obscures the issue.

It can wait until a more general data reformation though.

shelhamer · 2015-02-06T03:12:08Z

@longjon DATA + IMAGE_DATA now reshape with tests. Could you review + merge?

longjon · 2015-02-11T02:51:30Z

src/caffe/layers/base_data_layer.cpp

Should this be Reshape(this->prefetch_data_.num()? This is unconditional, not just batch size one, right?

Reshape single input batches for inputs of varying dimension

To feed inputs of varying dimension, the `DATA` and `IMAGE_DATA` layer reshapes its prefetch and top blobs when the batch size is 1. The `BasePrefetchingDataLayer` always reshapes on forward.

Reshape single input batches for inputs of varying dimension

dxj19831029 · 2015-04-10T10:00:36Z

em... I think the python's wrapper need to adjust as well, otherwise the preprocess will always resize the input image.

shelhamer assigned longjon Oct 17, 2014

sergeyk force-pushed the dev branch from 2fb4c97 to 1718903 Compare October 17, 2014 18:44

shelhamer mentioned this pull request Oct 18, 2014

leveldb/lmdb refactoring #1238

Merged

sguada reviewed Oct 18, 2014
View reviewed changes

longjon reviewed Dec 24, 2014
View reviewed changes

src/caffe/layers/data_layer.cpp Outdated

Copy link

Contributor

longjon Dec 24, 2014

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does this work if crop_size is nonzero?

longjon added a commit to longjon/caffe that referenced this pull request Dec 29, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

55cf90d

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Dec 29, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

68a8480

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Dec 29, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

2c6bbe0

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Dec 29, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

7ece6f1

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Dec 30, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

7df5459

Reshape single input batches for inputs of varying dimension

shelhamer added focus enhancement labels Dec 30, 2014

longjon added a commit to longjon/caffe that referenced this pull request Dec 30, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

8ea685e

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Dec 30, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

921632c

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Dec 30, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

eac17d5

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Dec 31, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

9167f2f

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Dec 31, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

6da16f7

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Dec 31, 2014

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

e8d9c12

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Jan 1, 2015

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

9646320

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Jan 2, 2015

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

7a7f319

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Jan 2, 2015

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

5507ab2

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Jan 3, 2015

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

9c31951

Reshape single input batches for inputs of varying dimension

longjon added a commit to longjon/caffe that referenced this pull request Jan 3, 2015

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

3f2f7fa

Reshape single input batches for inputs of varying dimension

bhack mentioned this pull request Jan 7, 2015

Added opencv vector<Mat> to memory data layer with tests #1416

Closed

shelhamer force-pushed the reshape-data-layer branch 6 times, most recently from 7fa470e to ba39b58 Compare February 6, 2015 03:07

shelhamer added ready for review and removed in progress labels Feb 6, 2015

longjon reviewed Feb 11, 2015
View reviewed changes

philkr added a commit to philkr/caffe that referenced this pull request Feb 15, 2015

Merge pull request BVLC#1313 from shelhamer/reshape-data-layer

9462bc1

Reshape single input batches for inputs of varying dimension

shelhamer force-pushed the reshape-data-layer branch from ba39b58 to 7cbd4ff Compare February 17, 2015 03:45

reshape DATA + IMAGE_DATA for inputs of varying dimension

82551ad

To feed inputs of varying dimension, the `DATA` and `IMAGE_DATA` layer reshapes its prefetch and top blobs when the batch size is 1. The `BasePrefetchingDataLayer` always reshapes on forward.

shelhamer force-pushed the reshape-data-layer branch from 7cbd4ff to 46ae0f9 Compare February 17, 2015 04:10

shelhamer mentioned this pull request Feb 17, 2015

Next: release candidater #1849

Merged

test reshaping DATA and IMAGE_DATA

d3b2010

shelhamer force-pushed the reshape-data-layer branch from 46ae0f9 to d3b2010 Compare February 17, 2015 04:31

shelhamer added a commit that referenced this pull request Feb 17, 2015

Merge pull request #1313 from shelhamer/reshape-data-layer

0b30aab

Reshape single input batches for inputs of varying dimension

shelhamer merged commit 0b30aab into BVLC:dev Feb 17, 2015

shelhamer deleted the reshape-data-layer branch February 17, 2015 04:31

philkr mentioned this pull request Feb 25, 2015

Decoding the datum before feeding it into the reshaping data layer #1955

Merged

shelhamer mentioned this pull request Mar 24, 2015

Spatial Pyramid Pooling Layer #2177

Merged

shelhamer mentioned this pull request Apr 5, 2015

text classification and dynamic input / output sizes #2244

Closed

melgor mentioned this pull request Apr 10, 2015

Sliding Window, Varying input/output size and Dense, multiscale extraction #189

Closed

longjon mentioned this pull request Sep 7, 2015

On-the-fly net resizing, without reallocation (where possible) #594

Merged

Reshape single input batches for inputs of varying dimension #1313

Reshape single input batches for inputs of varying dimension #1313

Uh oh!

Conversation

shelhamer commented Oct 17, 2014

Uh oh!

sguada Oct 18, 2014

Choose a reason for hiding this comment

Uh oh!

sguada Oct 18, 2014

Choose a reason for hiding this comment

Uh oh!

shelhamer Oct 18, 2014

Choose a reason for hiding this comment

Uh oh!

bhack commented Dec 22, 2014

Uh oh!

longjon Dec 24, 2014

Choose a reason for hiding this comment

Uh oh!

longjon commented Dec 24, 2014

Uh oh!

shelhamer commented Jan 7, 2015

Uh oh!

bhack commented Jan 8, 2015

Uh oh!

shelhamer commented Feb 5, 2015

Uh oh!

shelhamer commented Feb 5, 2015

Uh oh!

shelhamer commented Feb 6, 2015

Uh oh!

longjon Feb 11, 2015

Choose a reason for hiding this comment

Uh oh!

dxj19831029 commented Apr 10, 2015

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants