This repository has been archived by the owner on Nov 17, 2023. It is now read-only.
Update documentation for deconvolution operation. #6184
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
|
see http://pytorch.org/docs/nn.html#convtranspose2d and https://keras.io/layers/convolutional/#conv2dtranspose and the corresponding tensorflow doc for how these ops should be documented. Note that our implementation is slightly different here and there. So you cannot simple copy their doc. In general we are trying to improve the doc to provide more information, instead of just normalizing punctuation. Please read the doc (and code if necessary) to understand what each operator is doing. |
|
This still falls short of the other docs. |
piiswrong
reviewed
May 24, 2017
| "If target_shape is set, " | ||
| "pad will be ignored and computed accordingly"); | ||
| .describe("The amount of implicit zero padding added during convolution for each " | ||
| "dimension of the input: " |
There was a problem hiding this comment.
This is pading output instead?
There was a problem hiding this comment.
The description says this is the padding that was used for the convolution input (not deconvolution input). Also, saying deconvolution pads the output could be misleading. Deconvolution doesn't actually set the borders to zero. Example:
>>> # The transpose of convolving a 4×4 kernel over a 5×5 input padded with a
... # 2 × 2 border of zeros using unit strides (i.e., i = 5, k = 4, s = 1 and p = 2).
... input_5x5 = mx.nd.normal(shape=[1, 16, 5, 5])
>>> kernel_4x4 = mx.nd.normal(shape=[16, 16, 4, 4])
>>> conv = mx.nd.Convolution(data=input_5x5, kernel=(4,4), weight=kernel_4x4, pad=(2,2), num_filter=16, no_bias=True)
>>> transpose = mx.nd.Deconvolution(data=conv, kernel=(4,4), weight=kernel_4x4, pad=(2,2), num_filter=16, no_bias=True)
>>> transpose[0][0].asnumpy()
array([[ 13.21707153, -174.19221497, 295.35494995, -482.07144165,
236.75886536],
[-166.75314331, 161.54397583, 344.41635132, -123.3405838 ,
-20.51438904],
[-991.8168335 , 14.92340851, -131.36260986, -170.46170044,
-283.42138672],
[ 165.29977417, -47.57701492, -143.96827698, -194.85569763,
-91.3994751 ],
[ 1.36075592, 244.04574585, 211.51608276, -40.39221573,
-149.74568176]], dtype=float32)
piiswrong
reviewed
May 25, 2017
python/mxnet/ndarray_doc.py
Outdated
| """ | ||
| Examples | ||
| -------- | ||
| >>> input_4x4 = mx.nd.normal(shape=[1, 16, 4, 4]) |
There was a problem hiding this comment.
Non-language specific examples should go to backend.
There was a problem hiding this comment.
I think making it language neutral will make it less readable and more difficult for users to copy paste code snippet.
cjolivier01
pushed a commit
to cjolivier01/mxnet
that referenced
this pull request
May 30, 2017
* Update documentation for deconvolution operation. * Add examples. * Remove the example.
piiswrong
pushed a commit
that referenced
this pull request
Jun 5, 2017
…hnorm (#6411) * Add channel_axis to batch norm, performance improvements * rearrange tests a bit * rearrange tests a bit * CR changes * cpp package link issue * Fix: MSVC wants all parallel omp to be int * CR comments, expand legal negative axes * lint * lint * Fix download link (#6431) * Fix Download Button * Small fix * Add release note (#6434) * Fixing tutorials. (#6436) Most of the fixes should be self evident. For tutorial on pre-trained models, one of the images doesn't exist anymore so selected a new one. Long-term, we should put such images on web-data repo but alas, some other day. For Handwritten digit tutorial, we are missing couple of imports in the test_utils.py that was recently created. Note that: for pre-trained model tutorial, we get a softmax_label warning and the probability scores are not really probabilities. Will deal with that issue in another PR. Testing: I've tried to test all the notebooks with this change and things look fine. * Formatting fixes (#6433) * Formatting fixes * lint fixed * fix * doc bash 2-5, for pack, unpack, pack_img and unpack_img (#6140) * doc bash for pack, unpack, pack_img and unpack_img * Add comments for labels could be 1d list * Update recordio.py * Update recordio.py * Update recordio.py fixing text * Update recordio.py fixing text * remove empty line * Improve style (#6445) * Correction (#6444) * CSVIter example correction * fix * Update documentation for MXNetDataIter in io.py (#6000) (#6113) * Update documentation for MXNetDataIter in io.py (#6000) * [DOC] Respond to feedback (#6113) * Fix minor issues with api pages. (#6410) 1. In the notes section for ndarray, references did not seem clear enough to be referring to mxnet.ndarray or numpy.ndarray. Added the package names as prefixes to make it more obvious. 2. "share the same C++ operator source codes" => "share the same code". Since we don't really need to throw in more details than required. 3. Other relatively minor language changes which will be obvious from the diff. Note that I'm relatively not sure about the need for 1/ since it makes things more verbose. Let me know if it unnecessary and I'll remove it. * fixing the early stop for maximize = T (#5915) close #4587 * Update documentation for mxnet.ndarray.GridGenerator. (#6430) * Update documentation for mxnet.ndarray.GridGenerator. Thanks @Lyken17 for #6147 * Fix lint error. * Minor fix. * Remove the example. * Update documentation for deconvolution operation. (#6184) * Update documentation for deconvolution operation. * Add examples. * Remove the example. * skip lines that have %matplotlib (#6451) * Fixing some more broken links before v0.10 release (#6449) * close #4838 (#6452) * Fix linear regression (#6432) * Fix Linear Regression Tutorial * Small fix * Pre-trained model tutorial fixes. (#6453) Before the change on running the tutorial for the first time: "UserWarning: Data provided by label_shapes don't match names specified by label_names ([] vs. ['softmax_label'])". It also showed probability of >>1 due to incorrect usage of np.argsort(). * Nightly test tutorial (#6447) * Add tutorial test * Fix pylint * Small fix * fix member variable name: make them end with underline (#6438) * [R] captcha example (#6443) * skip lines that have %matplotlib (#6459) * Fix cudnn_deconv not guarding no_bias (#6456) * Fixing up issues in install guide (#6463) * Fixing copy code functionality for bash command (#6465) * Residual unroll (#6397) * residual unroll * unroll for residual cell * merge_outputs fix * Linear regression Tutorial link (#6468) * Fixing a link in the linear regression tutorial. The link was initally going to mxnet-test.readthedocs.io. Changed it to mxnet.io/api. * More appropriate language. * bump up version number for release (#6462) * bump up version number for release * update version for scala/R/backend * [R][DOC] update R installation guide (#6457) * Use sphinx==1.3.5 in Dockerfile.doc (#6470) changed PR name * Add 0.10 release info to README.md and NEWS.md (#6471) @nswamy wants to merge it immediately, so i'm going to do it now. I also changed the PR title. * fix batchNorm cpp example (#6454) * Update im2rec.py (#6473) Updated Line 107 of 'im2rec.py'. Read an image as binary. * Change Interface of NDArray & TBlob for DLPack Compatible (#6345) * Change Interface of NDArray & TBlob for DLPack Compatible Fix for cudnn operator Fix cpp tests * Update nnvm * Fix for MKL mem * Fix for windows macro * Bump up version number to 0.10.1 * Update NDArray Save&Load * trigger update * Add test for legacy data load * Use LegacyTShapeLoad * trigger update * Update tensor_blob.h * change 'channel_axis' parameter to 'axis' * Change DEFAULT_CHANNEL_AXIS to DEFAULT_AXIS * wait for dlpack PR to go through * Trigger build
Guneet-Dhillon
pushed a commit
to Guneet-Dhillon/mxnet
that referenced
this pull request
Sep 13, 2017
* Update documentation for deconvolution operation. * Add examples. * Remove the example.
Guneet-Dhillon
pushed a commit
to Guneet-Dhillon/mxnet
that referenced
this pull request
Sep 13, 2017
…hnorm (apache#6411) * Add channel_axis to batch norm, performance improvements * rearrange tests a bit * rearrange tests a bit * CR changes * cpp package link issue * Fix: MSVC wants all parallel omp to be int * CR comments, expand legal negative axes * lint * lint * Fix download link (apache#6431) * Fix Download Button * Small fix * Add release note (apache#6434) * Fixing tutorials. (apache#6436) Most of the fixes should be self evident. For tutorial on pre-trained models, one of the images doesn't exist anymore so selected a new one. Long-term, we should put such images on web-data repo but alas, some other day. For Handwritten digit tutorial, we are missing couple of imports in the test_utils.py that was recently created. Note that: for pre-trained model tutorial, we get a softmax_label warning and the probability scores are not really probabilities. Will deal with that issue in another PR. Testing: I've tried to test all the notebooks with this change and things look fine. * Formatting fixes (apache#6433) * Formatting fixes * lint fixed * fix * doc bash 2-5, for pack, unpack, pack_img and unpack_img (apache#6140) * doc bash for pack, unpack, pack_img and unpack_img * Add comments for labels could be 1d list * Update recordio.py * Update recordio.py * Update recordio.py fixing text * Update recordio.py fixing text * remove empty line * Improve style (apache#6445) * Correction (apache#6444) * CSVIter example correction * fix * Update documentation for MXNetDataIter in io.py (apache#6000) (apache#6113) * Update documentation for MXNetDataIter in io.py (apache#6000) * [DOC] Respond to feedback (apache#6113) * Fix minor issues with api pages. (apache#6410) 1. In the notes section for ndarray, references did not seem clear enough to be referring to mxnet.ndarray or numpy.ndarray. Added the package names as prefixes to make it more obvious. 2. "share the same C++ operator source codes" => "share the same code". Since we don't really need to throw in more details than required. 3. Other relatively minor language changes which will be obvious from the diff. Note that I'm relatively not sure about the need for 1/ since it makes things more verbose. Let me know if it unnecessary and I'll remove it. * fixing the early stop for maximize = T (apache#5915) close apache#4587 * Update documentation for mxnet.ndarray.GridGenerator. (apache#6430) * Update documentation for mxnet.ndarray.GridGenerator. Thanks @Lyken17 for apache#6147 * Fix lint error. * Minor fix. * Remove the example. * Update documentation for deconvolution operation. (apache#6184) * Update documentation for deconvolution operation. * Add examples. * Remove the example. * skip lines that have %matplotlib (apache#6451) * Fixing some more broken links before v0.10 release (apache#6449) * close apache#4838 (apache#6452) * Fix linear regression (apache#6432) * Fix Linear Regression Tutorial * Small fix * Pre-trained model tutorial fixes. (apache#6453) Before the change on running the tutorial for the first time: "UserWarning: Data provided by label_shapes don't match names specified by label_names ([] vs. ['softmax_label'])". It also showed probability of >>1 due to incorrect usage of np.argsort(). * Nightly test tutorial (apache#6447) * Add tutorial test * Fix pylint * Small fix * fix member variable name: make them end with underline (apache#6438) * [R] captcha example (apache#6443) * skip lines that have %matplotlib (apache#6459) * Fix cudnn_deconv not guarding no_bias (apache#6456) * Fixing up issues in install guide (apache#6463) * Fixing copy code functionality for bash command (apache#6465) * Residual unroll (apache#6397) * residual unroll * unroll for residual cell * merge_outputs fix * Linear regression Tutorial link (apache#6468) * Fixing a link in the linear regression tutorial. The link was initally going to mxnet-test.readthedocs.io. Changed it to mxnet.io/api. * More appropriate language. * bump up version number for release (apache#6462) * bump up version number for release * update version for scala/R/backend * [R][DOC] update R installation guide (apache#6457) * Use sphinx==1.3.5 in Dockerfile.doc (apache#6470) changed PR name * Add 0.10 release info to README.md and NEWS.md (apache#6471) @nswamy wants to merge it immediately, so i'm going to do it now. I also changed the PR title. * fix batchNorm cpp example (apache#6454) * Update im2rec.py (apache#6473) Updated Line 107 of 'im2rec.py'. Read an image as binary. * Change Interface of NDArray & TBlob for DLPack Compatible (apache#6345) * Change Interface of NDArray & TBlob for DLPack Compatible Fix for cudnn operator Fix cpp tests * Update nnvm * Fix for MKL mem * Fix for windows macro * Bump up version number to 0.10.1 * Update NDArray Save&Load * trigger update * Add test for legacy data load * Use LegacyTShapeLoad * trigger update * Update tensor_blob.h * change 'channel_axis' parameter to 'axis' * Change DEFAULT_CHANNEL_AXIS to DEFAULT_AXIS * wait for dlpack PR to go through * Trigger build
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
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.
@mli, @zackchase, @madjam, @nswamy, @Roshrini, @jiajiechen, @piiswrong
Please review and merge.