Add switch order layer for FCN model #2788
Conversation
1. Add switch function for switching image dimensions order 2. Add CpuMatrix::backwardSoftmax function 3. Add pixel softmax layer, python wrapper and grad_test
… pixel_softmax_layer
paddle/function/SwitchOp.cpp
Outdated
| * how many zeros to add before and after the input in channel | ||
| * dimension. And the heightStart and heightEnd indicate padding | ||
| * in height dimension. The widthStart and widthEnd indicate the | ||
| * padding in width dimension. |
There was a problem hiding this comment.
The above comments are not correct for NCHW2NHWCFunc.
paddle/function/SwitchOp.cpp
Outdated
| size_t inW = inputs[0].shape()[3]; | ||
| typename Tensor<real, Device>::Vector vec(outputs[0].shape().getElements(), | ||
| outputs[0].data<real>()); | ||
| vec.zero(); |
There was a problem hiding this comment.
There is no need to set outputs to zero, since the assignment (=) is used in NCHW2NHWC.
paddle/function/SwitchOp.cpp
Outdated
| * Argument in this Function: | ||
| * \param pad_ The same meaning as it in PadFunc. | ||
| * \param inputs The gradient with respect to the output value of PadFunc. | ||
| * \param outputs The gradient with respect to the input value of PadFunc. |
There was a problem hiding this comment.
The above comments are not correct.
| Layer::forward(passType); | ||
| MatrixPtr input = inputLayers_[0]->getOutputValue(); | ||
| size_t batchSize = input->getHeight(); | ||
| // cout<<"useGpu:"<<useGpu(deviceId_)<<endl; |
| inW_ = img_conf.img_size(); | ||
| inC_ = img_conf.channels(); | ||
| createFunction(forward_, "NCHW2NHWC", FuncConfig()); | ||
| createFunction(backward_, "NHWC2NCHW", FuncConfig()); |
There was a problem hiding this comment.
Maybe, the name forward_ and backward_ is not quite good. nchw2nhwc_ and nhwc2nchw_?
There was a problem hiding this comment.
I can't rename forward_ and backward_ directly . Because they are defined in Layer.h. So i added std::vector<std::shared_ptr<FunctionBase>> nchw2nhwc_; and std::vector<std::shared_ptr<FunctionBase>> nhwc2nchw_; into PixelSoftmaxLayer.h to fix this problem. Thx for your suggestion.
| @wrap_name_default('pixel_softmax') | ||
| def pixel_softmax_layer(input, name=None, layer_attr=None): | ||
| """ | ||
| This layer calculate softmax in image channel dimension |
There was a problem hiding this comment.
Need more detailed comments.
| :param name: Name of this layer. | ||
| :type name: basestring | ||
| :param input: The input layer. | ||
| :type input: LayerOutput |
There was a problem hiding this comment.
It's better to put input in front of the name to keep the order same with input arguments.
| if isinstance(input, LayerOutput): | ||
| input = [input] | ||
| elif isinstance(input, Projection): | ||
| input = [input] |
There was a problem hiding this comment.
This layer can not support Projection. Only the mixed_layer can use the Projection as input.
| elif isinstance(input, Projection): | ||
| input = [input] | ||
| else: | ||
| assert isinstance(input, collections.Sequence) |
There was a problem hiding this comment.
This only support one input layer. Lines 5902 to 5907 should be:
assert isinstance(input, LayerOutput)
| assert isinstance(input, collections.Sequence) | ||
| l = Layer( | ||
| name=name, | ||
| inputs=[x.name for x in input], |
2. Make SwitchOrderLayer support for softmax activation 3. Fix bugs
… pixel_softmax_layer
wanghaoshuang
changed the title
wanghaoshuang
force-pushed
the
pixel_softmax_layer
branch
from
e35167a
to
ec236f4
Compare
… pixel_softmax_layer
… pixel_softmax_layer
|
what has_depth() means in SwitchOrderLayer(LayerBase) function, the function is in config_parser.py, i encounter an error: LayerConfig object has no attribute 'has_depth()', thx |
* add config for jde and deepsort, add engine metric tools * add mot_metrics * fix configs engine and others * fix configs others * move configs * fix ap_per_class, fix doc * fix test_mode, seqs * fix test_mode to metric * fix JDE arch metric * clean requirement * add to requirement
fix #2787