Refactor tensorflow import backend #499
Conversation
| if activation_op.type == 'Relu': | ||
| layer_type = 'ReLU' | ||
|
|
||
| if activation_op.type == 'LeakyRelu': |
| layer_params['alpha'] = 1 | ||
| layer_type = 'ELU' | ||
|
|
||
| elif activation_op.type == 'Softplus': |
There was a problem hiding this comment.
For all layers below this can't we directly assign activation_op.type?
| conv2d_op = next((x for x in layer_ops if x.type == 'Conv2D'), None) | ||
|
|
||
| layer_params = {} | ||
|
|
| def import_conv2d(layer_ops): | ||
|
|
||
| conv2d_op = next((x for x in layer_ops if x.type == 'Conv2D'), None) | ||
|
|
|
|
||
|
|
||
| def import_conv3d(layer_ops): | ||
|
|
There was a problem hiding this comment.
It would be good to have comments in every method
|
|
||
| def import_deconvolution(layer_ops): | ||
| deconv_op = next((x for x in layer_ops if x.type == 'Conv2DBackpropInput'), None) | ||
|
|
| layer_params = {} | ||
|
|
||
| if '3D' in depthwise_conv_op.type: | ||
| return |
There was a problem hiding this comment.
We should throw some sort of error in such case so that user knows what was the issue
|
|
||
| def import_batchnorm(layer_ops): | ||
| layer_params = {} | ||
|
|
There was a problem hiding this comment.
Please follow same styles for newlines, I mean add newlines by following some style guidelines
abhigyan7
force-pushed
the
refactor_import_graphdef
branch
from
d9a39bd
to
c3c7386
Compare
|
@haabhi while merging we will squash the commits anyway so if all your changes are present in the PR then it shouldn't be an issue. |
abhigyan7
force-pushed
the
refactor_import_graphdef
branch
from
99abc64
to
8a23946
Compare
|
@Ram81 Made the changes. I have added docstrings and normal comments where things were not very clear. As for new lines, I have deleted a lot of them and grouped related lines tightly and separated them with single blank lines. |
|
@Ram81 updated the layer addition tutorial to match these changes. |
|
|
||
|
|
||
| def get_input_layers(layer_ops): | ||
| ''' return the name of the layers directly preceeding the layer of layer_ops. |
There was a problem hiding this comment.
Can you move the comment to new line
| ''' return the name of the layers directly preceeding the layer of layer_ops. | ||
| layer_ops is a list of all ops of the layer we want the inputs of. | ||
| ''' | ||
| ret = [] |
| placeholder_op = layer_ops[0] | ||
| layer_params = {} | ||
| layer_dim = [int(dim.size) for dim in placeholder_op.get_attr('shape').dim] | ||
| # make batch size 1 if it is -1 |
| - Create a function name import_<layer_name> that takes one parameter, layer_ops, that is a list of all ops in the layer being imported. | ||
| - Get layer parameters from the operations in layer_ops and build a dictionary mapping parameter names to values. | ||
| - Get a list of input layers to the layer being processed using `get_input_layers(layer_ops)`. | ||
| - Create and return a json layer for your layer, calling `jsonLayer` with your layer_type, parameters, input layers and output_layers(optional). |
There was a problem hiding this comment.
change this to Create and return a json layer for new layer replace your with new in the statement
* Refactor tensorflow import code * Remove unnecessary newlines and make code concise * Update documentation for adding new layers to tensorflow import * Minor changes
This PR includes a clean, modular code base for tensorflow import. I have split the import logic into two files, much like with the keras and caffe import ones.
All the examples inside example/tensorflow work exactly like they used to.