Refactor export helpers #958
Conversation
daniil-lyakhov
requested review from
a team,
AlexanderDokuchaev,
alexsu52,
evgeniya-egupova,
mkaglins and
vshampor
|
@daniil-lyakhov Please, define what is closing prunable layer in PR description and add more details about why is it needed. |
Every prunable layer L should have closing prunable layer means next prunable layer after L which input channels are pruning. For example: Green group is group of prunable layers, which has same channel dimention. We are removing some output channels from them. And yellow prunable layer is closing prunable layer which input channels we are removing. Hope this helped |
daniil-lyakhov
force-pushed
the
dl/refactor_export_helpers
branch
from
ccb0597 to
7c40fc1
Compare
| if input_node.data.get('output_mask', None) is not None: | ||
| continue | ||
|
|
||
| source_nodes = get_sources_of_node(input_node, graph, cls.ConvolutionOp.get_all_op_aliases() + | ||
| cls.StopMaskForwardOp.get_all_op_aliases() + | ||
| cls.InputOp.get_all_op_aliases()) | ||
| sources_types = [node.node_type for node in source_nodes] + [input_node.node_type] |
There was a problem hiding this comment.
Please be aware of two changes here:
if input_node.data.get('output_mask', None) is not None:
instead of
if input_node.data.get('output_mask', None) is None:
and
sources_types = [node.node_type for node in source_nodes] + [input_node.node_type]
instead of
sources_types = [node.node_type for node in source_nodes]
There was a problem hiding this comment.
@daniil-lyakhov Please, describe how the second change (of sources_types) will impact how the algorithm works on known models with Concat.
And, please, test these changes carefully on these models.
There was a problem hiding this comment.
Tested locally, please check daniil-lyakhov#10
There was a problem hiding this comment.
Have run tests on CI #969
daniil-lyakhov
force-pushed
the
dl/refactor_export_helpers
branch
from
ba7140a to
e022a1f
Compare
|
Minor TODO:
|
|
Jenkins please retry a build |
| WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| See the License for the specific language governing permissions and | ||
| limitations under the License. | ||
| """ |
There was a problem hiding this comment.
Could we remove this module and use 'common module in the TensorFlow backend?
| if mask is None: | ||
| concat_axis = node.layer_attributes.axis | ||
| concat_dim = input_edges[i].tensor_shape[concat_axis] | ||
| device = cls._get_masks_device(input_masks) |
There was a problem hiding this comment.
Could you please explain what type of mask? Can we avoid functions as _get_masks_device?
daniil-lyakhov
force-pushed
the
dl/refactor_export_helpers
branch
from
23a4ff1 to
0712dff
Compare
daniil-lyakhov
force-pushed
the
dl/refactor_export_helpers
branch
from
0712dff to
9dd421a
Compare
daniil-lyakhov
force-pushed
the
dl/refactor_export_helpers
branch
4 times, most recently
from
648b4f2 to
4340499
Compare
…ated from non prunable node
daniil-lyakhov
force-pushed
the
dl/refactor_export_helpers
branch
from
0a972d2 to
5c46689
Compare
* Have separated common parts of pruning export_helpers * Implement base realization of `mask_propogation` methods to have opportunity to use mask propagation algo without any specific framework(tf, torch) functions. The main point of this refactoring is mask propagation algo with framework agnostic masks support. Thus we will be allowed to use mask propogation algo at model analyzing stage to check state: "Every prunable layer have correspondent closing prunable layer". Thus we will be able to propagate masks through reshape layers and will enable one of a realization of SE block in mobilenet_v3. For this, numpy arrays will be propagated by base realization of `mask_propogation` TODO: - [x] Implement tests to cover common and algo specific mask propogation
5 participants
mask_propogationmethods to have opportunity to use mask propagation algo without any specific framework(tf, torch) functions.The main point of this refactoring is mask propagation algo with framework agnostic masks support. Thus we will be allowed to use mask propogation algo at model analyzing stage to check state: "Every prunable layer have correspondent closing prunable layer". Thus we will be able to propagate masks through reshape layers and will enable one of a realization of SE block in mobilenet_v3. For this, numpy arrays will be propagated by base realization of
mask_propogationTODO: