Skip to content
Snippets Groups Projects

feat/release_pip

Closed feat/release_pip
feat/release_pip into dev
1 unresolved thread
Closed Grégoire Kubler requested to merge feat/release_pip into dev
1 unresolved thread
Compare
and
8 files
+ 622
− 70
Compare changes
  • Side-by-side
  • Inline
Files
8
aidge_interop_torch/utils.py
+ 2
− 2
@@ -118,13 +118,13 @@ class AidgeModule(torch.nn.Module):
# TODO: add a system to avoid creating a new node everytime
if self.input_nodes[0] == None:
if self.input_nodes[0] is None:
self.input_nodes[0] = aidge_core.Producer(
aidge_tensor, "Input_0")
# TODO: get datatype & backend from graph view
self.input_nodes[0].get_operator().set_datatype(
aidge_core.DataType.Float32)
aidge_core.dtype.float32)
self.input_nodes[0].get_operator().set_backend("cpu")
self.input_nodes[0].add_child(self._graph_view)
@@ -151,7+151,7 @@
@staticmethod
def backward(ctx, grad_output):
if not self.grad_compiled: aidge_core.compile_gradient(self._graph_view)
    • Grégoire Kubler
      Grégoire Kubler @gregkub ·
      Author Developer

      @cmoineau

      Running the tests in local to try to solve these issues returns an error on this line : aidge_core doesn't seem to have any compile_gradients function, I lloked for it in the whole codebase but cannot find it anywhere.

      • Cyril Moineau
        Cyril Moineau @cmoineau ·
        Maintainer

        Since aidge_core@ecc96977 grad is lazy init thus copile_gradient and init_grad have been removed.

      • Grégoire Kubler
        Grégoire Kubler @gregkub ·
        Author Developer

        I removed the call to compile gradients

      • Grégoire Kubler
        Grégoire Kubler @gregkub ·
        Author Developer

        Ok now the real issue : /

        https://gitlab.eclipse.org/eclipse/aidge/aidge_interop_torch/-/jobs/406724#L968

        It seems that the imported model contains 3 input nodes instead of 1 :

        $ p aidge_model.get_input_nodes()
{Node(name='input_24', optype='Conv', parents: [1, 1, 0], children: 1), Node(name='input_12', optype='Conv', parents: [1, 1, 0], children: 1), Node(name='input', optype='Conv', parents: [0, 1, 0], children: 1)}

        here is the model after being exported to onnx & re imported to aidge. We can clearly see that the export/import phase added 3 inputs & 2 inptuts which is not normal.

        %%{init: {'flowchart': { 'curve': 'monotoneY'}, 'fontFamily': 'Verdana' } }%%
flowchart TB

FC_1("data_37\n<sub><em>(FC#1)</em></sub>")
Producer_10("layer_15_bias\n<sub><em>(Producer#10)</em></sub>"):::producerCls
ReLU_0("Relu_1\n<sub><em>(ReLU#0)</em></sub>")
Producer_8("layer_13_bias\n<sub><em>(Producer#8)</em></sub>"):::producerCls
Producer_9("layer_15_weight\n<sub><em>(Producer#9)</em></sub>"):::producerCls
Producer_6("layer_10_weight\n<sub><em>(Producer#6)</em></sub>"):::producerCls
Conv_2("input_24\n<sub><em>(Conv#2)</em></sub>")
Producer_2("layer_2_bias\n<sub><em>(Producer#2)</em></sub>"):::producerCls
Producer_7("layer_13_weight\n<sub><em>(Producer#7)</em></sub>"):::producerCls
Producer_3("layer_5_weight\n<sub><em>(Producer#3)</em></sub>"):::producerCls
Producer_0("layer_0_weight\n<sub><em>(Producer#0)</em></sub>"):::producerCls_rootCls
MaxPooling_0("input_8\n<sub><em>(MaxPooling#0)</em></sub>")
Conv_1("input_12\n<sub><em>(Conv#1)</em></sub>")
ReLU_2("Relu_6\n<sub><em>(ReLU#2)</em></sub>")
BatchNormalization_1("BatchNormalization_7\n<sub><em>(BatchNormalization#1)</em></sub>"):::genericCls
ReLU_3("Relu_8\n<sub><em>(ReLU#3)</em></sub>")
MaxPooling_1("input_20\n<sub><em>(MaxPooling#1)</em></sub>")
ReLU_4("Relu_11\n<sub><em>(ReLU#4)</em></sub>")
FC_0("input_28\n<sub><em>(FC#0)</em></sub>")
ReLU_5("Relu_14\n<sub><em>(ReLU#5)</em></sub>")
Producer_5("layer_7_bias\n<sub><em>(Producer#5)</em></sub>"):::producerCls
Producer_4("layer_7_weight\n<sub><em>(Producer#4)</em></sub>"):::producerCls
Conv_0("input\n<sub><em>(Conv#0)</em></sub>")
BatchNormalization_0("BatchNormalization_2\n<sub><em>(BatchNormalization#0)</em></sub>"):::genericCls
ReLU_1("Relu_3\n<sub><em>(ReLU#1)</em></sub>")
Producer_1("layer_2_weight\n<sub><em>(Producer#1)</em></sub>"):::producerCls
Producer_10-->|"0 [10]&rarr;2"|FC_1
ReLU_0-->|"0&rarr;0"|BatchNormalization_0
Producer_8-->|"0 [84]&rarr;2"|FC_0
Producer_9-->|"0 [10, 84]&rarr;1"|FC_1
Producer_6-->|"0 [120, 16, 5, 5]&rarr;1"|Conv_2
Conv_2-->|"0&rarr;0"|ReLU_4
Producer_2-->|"0 [6]&rarr;2"|BatchNormalization_0
Producer_2-->|"0 [6]&rarr;2"|BatchNormalization_0
Producer_7-->|"0 [84, 120]&rarr;1"|FC_0
Producer_3-->|"0 [16, 6, 5, 5]&rarr;1"|Conv_1
Producer_0-->|"0 [6, 1, 5, 5]&rarr;1"|Conv_0
MaxPooling_0-->|"0&rarr;0"|Conv_1
Conv_1-->|"0&rarr;0"|ReLU_2
ReLU_2-->|"0&rarr;0"|BatchNormalization_1
BatchNormalization_1-->|"0&rarr;0"|ReLU_3
ReLU_3-->|"0&rarr;0"|MaxPooling_1
MaxPooling_1-->|"0&rarr;0"|Conv_2
ReLU_4-->|"0&rarr;0"|FC_0
FC_0-->|"0&rarr;0"|ReLU_5
ReLU_5-->|"0&rarr;0"|FC_1
Producer_5-->|"0 [16]&rarr;2"|BatchNormalization_1
Producer_5-->|"0 [16]&rarr;2"|BatchNormalization_1
Producer_4-->|"0 [16]&rarr;1"|BatchNormalization_1
Producer_4-->|"0 [16]&rarr;1"|BatchNormalization_1
Conv_0-->|"0&rarr;0"|ReLU_0
BatchNormalization_0-->|"0&rarr;0"|ReLU_1
ReLU_1-->|"0&rarr;0"|MaxPooling_0
Producer_1-->|"0 [6]&rarr;1"|BatchNormalization_0
Producer_1-->|"0 [6]&rarr;1"|BatchNormalization_0
input0((in#0)):::inputCls--->|"&rarr;0"|Conv_0
input1((in#1)):::inputCls--->|"&rarr;2"|Conv_0
input2((in#2)):::inputCls--->|"&rarr;2"|Conv_1
input3((in#3)):::inputCls--->|"&rarr;2"|Conv_2
FC_1--->|"0&rarr;"|output0((out#0)):::outputCls
BatchNormalization_1--->|"1&rarr;"|output1((out#1)):::outputCls
BatchNormalization_1--->|"2&rarr;"|output2((out#2)):::outputCls
BatchNormalization_0--->|"1&rarr;"|output3((out#3)):::outputCls
BatchNormalization_0--->|"2&rarr;"|output4((out#4)):::outputCls
classDef inputCls fill:#afa
classDef outputCls fill:#ffa
classDef externalCls fill:#ccc
classDef producerCls fill:#ccf
classDef genericCls fill:#f9f9ff,stroke-width:1px,stroke-dasharray: 5 5
classDef metaCls stroke-width:5px
classDef rootCls stroke:#f00
classDef producerCls_rootCls stroke:#f00,fill:#ccf
classDef genericCls_rootCls stroke:#f00,fill:#f9f9ff,stroke-width:1px,stroke-dasharray: 5 5
classDef metaCls_rootCls stroke:#f00,stroke-width:5px

        I dont know enough of pytorch & aidge_interop torch as of now to unerstand where is the issue, i'll sort this out tomorrow.

      • Cyril Moineau
        Cyril Moineau @cmoineau ·
        Maintainer

        Clearly this module is not a priority. I would be in favor of merging this in dev with the broken CI and mark this module as broken. The reasoning behind this being that this MR doesn't break the CI and just highlight issues. These issues are not on the roadmap :smile:

        @pierregaillard do you agree with this prioritization ?

      • Grégoire Kubler
        Grégoire Kubler @gregkub ·
        Author Developer

        ( On my side I agree with not having to fix a big I have no Idea how to fix )

      • Grégoire Kubler changed this line in version 8 of the diff

        changed this line in version 8 of the diff

      • Pierre Gaillard
        Pierre Gaillard @pierregaillard ·
        Reporter

        Totaly agree @cmoineau !

      • Please register or sign in to reply
Please register or sign in to reply
if self.multi_outputs_flag:
raise RuntimeError(

Copyright © Eclipse Foundation, Inc. All Rights Reserved.     Privacy Policy | Terms of Use | Copyright Agent