-
Notifications
You must be signed in to change notification settings - Fork 6.8k
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
minor wording change, given state of CI, feel free to ignore if build passes
NEWS.md
Outdated
@@ -1,5 +1,19 @@ | |||
MXNet Change Log | |||
================ | |||
## 1.2.1 | |||
### Deprecations | |||
- An incorrect [usage](https://github.com/apache/incubator-mxnet/issues/11091) of `save_params` was advertised in the gluon book which led to MXNet users depending on the incorrect usage and developing a hack around it. A change to internal structure of params file saved by `save_params` to resolve a bug, led to user scripts with the above mentioned hack to break. To fix this, we have reverted `save_params` and `load_params` to previous format and marked it as deprecated. We have added new APIs `save_parameters` and `load_parameters` for the new format. All scripts to save and load parameters for a Gluon model should now use the new API for `save_parameters` and `load_parameters`. If your model is hybridizable and you want to export a serialized structure of the model as well as parameters you need to use the `export` API and the newly added `imports` API instead of `save_params` and `load_params` API. For more details, Please see: [issue](https://github.com/apache/incubator-mxnet/issues/11091), [PR](https://github.com/apache/incubator-mxnet/pull/11127). |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
-> A change was made to the internal structure of the .params
file saved by save_params
to resolve a bug. This led to user scripts with the above mentioned hack to break
@ThomasDelteil Thanks! I have fixed it. |
ae48d71
to
2af551a
Compare
NEWS.md
Outdated
- Reduced memory consumption from inplace operation for ReLU activation (#10847). | ||
- Improved `slice` operator performance by 20x (#11124). | ||
- Improved performance of depthwise convolution by using cudnnv7 if available (#11076). | ||
- Improved performance and memory usage of Conv1D, by adding back cuDNN support for Conv1D (#11270). This adds a known issue: The cuDNN convolution operator may throw `CUDNN_STATUS_EXECUTION_FAILED` when `req == "add"` and `cudnn_tune != off` with large inputs(e.g. 64k channels). If you encounter this issue, please consider setting `cudnn_tune = "off"`. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Since cudnn_tune
requires changing users' python script, let's change to
If you encounter this issue, please consider setting environment variable MXNET_CUDNN_AUTOTUNE_DEFAULT
to 0.
2af551a
to
d9666c7
Compare
@piiswrong @ThomasDelteil cherry picked your changes. Please help review as i had to resolve some conflicts. This PR needs to be merged without squashing. |
I think 7632100 is too verbose. It would only confuse users. A simple pointer to export/import should be enough |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
please update the reference to the hybridize flags
docs/tutorials/gluon/hybrid.md
Outdated
```python | ||
net.hybridize(static_alloc=True) | ||
# or | ||
net.hybridize(static_alloc=True, static_shape=True) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This is slower in 1.2 #11293
let's not advertise it. I don't even think these flags exist in 1.2.0
let's just mention net.hybridize()
in 1.2
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Removed.
@piiswrong I think we can remove this sentence:
The rest I think is necessary to disambiguate the current API. |
7632100
to
4fa22ae
Compare
I agree that we need the rest of the content to inform users about the |
* enable cudnn for conv1d * add checks for backward * fix build * fix build * fix lint * Update convolution.cc
* Fixed armv7 wheel * Fixed bash default value substitution * Fixed whitespace errors * Added unzip installation to crosscompile docker images * Removed deprecated apt lists removal * Moved common parts to scripts for docker builds * Disabled bundled OpenMP for cross compilation * Added make based build for armv7 as option * Renamed install_openblas.sh -> arm_openblas.sh * Removed makefile build for armv7 * Trigger CI * Link OpenBLAS statically for cross compilations
* Fixes Scala memory leak (apache#10436) * Replaced the copy and disposed of sliced ndArray to resolve memory leak * Wrapped disposes in a finally to ensure they are called.
c00e673
to
b34d518
Compare
Closing this PR and will a new PR since I have messed up my git history. |
Description
Added NEWs and README.
@ThomasDelteil @piiswrong @zheng-da @eric-haibin-lin @nswamy @lebeg @lanking520 @andrewfayres
Checklist
Essentials
Please feel free to remove inapplicable items for your PR.
Changes
Comments