Error propagation leads to precision issue

What commit version of aidge do you use

• aidge_core: 52596e70f80b83f047eea50f06a61da60423f97d
• aidge_backend_cpu: 4d100339
• aidge_onnx : ebd439bbbeaeac58ccd7cd6868e535c836ba953e

Problem description

Loading and inferring with the efficientNet model leads to precision issues.

I have analyzed with a script I have written the output errors that the precision errors come from intermediate layers of the network and propagate through it become non neglectable further in the network. Since the network is not a feed forward, the propagation is not constant. Here are the lines that test the absolute errors in the script (simplified)

    all_layers_identical = True
    layers_differ = []
    relative_precision=1e-3 
    absolute_presision=1e-4
    for idx, name in enumerate(onnx_node_names):
        aidge_node = find_node_with_name(aidge_nodes, aidge_node_to_search)
        onnx_node_output = onnx_outputs[idx]
        aidge_node_output = np.array(aidge_node.get_operator().get_output(0))

        error = np.abs(aidge_node_output - onnx_node_output)
        err_threshold=relative_precision * np.abs(onnx_node_output) + absolute_presision
        if np.any(error[error > err_threshold]):  # error tolerance :
            all_layers_identical = False
            layers_differ.append((idx, name, np.mean(error), np.mean(err_threshold)))

What I get is :

layer idx	layer name	mean error	mean error threshold
20	/blocks/blocks.1/blocks.1.0/se/ReduceMean_output_0	0.00017626438	0.0040660696
27	/blocks/blocks.1/blocks.1.0/conv_pwl/Conv_output_0	3.397078e-05	0.0073460503
34	/blocks/blocks.1/blocks.1.1/se/ReduceMean_output_0	0.00015502871	0.00081236113
41	/blocks/blocks.1/blocks.1.1/conv_pwl/Conv_output_0	0.0001423933	0.0053374995
42	/blocks/blocks.1/blocks.1.1/Add_output_0	0.00014138373	0.007889783
43	/blocks/blocks.2/blocks.2.0/conv_pw/Conv_output_0	2.2628392e-05	0.0016293061
46	/blocks/blocks.2/blocks.2.0/conv_dw/Conv_output_0	1.9270157e-05	0.0019656136
49	/blocks/blocks.2/blocks.2.0/se/ReduceMean_output_0	0.0006917381	0.0014792526
50	/blocks/blocks.2/blocks.2.0/se/conv_reduce/Conv_output_0	0.001630033	0.0031868948
52	/blocks/blocks.2/blocks.2.0/se/act1/Mul_output_0	0.0015957039	0.0030804675
53	/blocks/blocks.2/blocks.2.0/se/conv_expand/Conv_output_0	0.000783397	0.0026705994
55	/blocks/blocks.2/blocks.2.0/se/Mul_output_0	4.7445785e-05	0.0014161852
56	/blocks/blocks.2/blocks.2.0/conv_pwl/Conv_output_0	0.00032996858	0.004815449
57	/blocks/blocks.2/blocks.2.1/conv_pw/Conv_output_0	5.8389065e-05	0.0012063695
59	/blocks/blocks.2/blocks.2.1/bn1/act/Mul_output_0	2.4987983e-05	0.00059082295
60	/blocks/blocks.2/blocks.2.1/conv_dw/Conv_output_0	6.416062e-05	0.0014070298
61	/blocks/blocks.2/blocks.2.1/bn2/act/Sigmoid_output_0	9.393024e-06	0.0005374124
62	/blocks/blocks.2/blocks.2.1/bn2/act/Mul_output_0	2.2771315e-05	0.00051652733
63	/blocks/blocks.2/blocks.2.1/se/ReduceMean_output_0	0.0006387538	0.00045737767
64	/blocks/blocks.2/blocks.2.1/se/conv_reduce/Conv_output_0	0.0012412429	0.002316205
66	/blocks/blocks.2/blocks.2.1/se/act1/Mul_output_0	0.001327324	0.0021039292
67	/blocks/blocks.2/blocks.2.1/se/conv_expand/Conv_output_0	0.0007842614	0.0011331665
68	/blocks/blocks.2/blocks.2.1/se/gate/Sigmoid_output_0	0.00014541978	0.0006631331
69	/blocks/blocks.2/blocks.2.1/se/Mul_output_0	6.175164e-05	0.00035459618
70	/blocks/blocks.2/blocks.2.1/conv_pwl/Conv_output_0	0.0008929105	0.0022966256
71	/blocks/blocks.2/blocks.2.1/Add_output_0	0.00095355336	0.004889038
72	/blocks/blocks.3/blocks.3.0/conv_pw/Conv_output_0	0.00015333471	0.0014932235
73	/blocks/blocks.3/blocks.3.0/bn1/act/Sigmoid_output_0	2.2842714e-05	0.00042134782
74	/blocks/blocks.3/blocks.3.0/bn1/act/Mul_output_0	4.0460065e-05	0.00049278006
75	/blocks/blocks.3/blocks.3.0/conv_dw/Conv_output_0	9.4713614e-05	0.0017175112
77	/blocks/blocks.3/blocks.3.0/bn2/act/Mul_output_0	7.9814476e-05	0.0014430308
78	/blocks/blocks.3/blocks.3.0/se/ReduceMean_output_0	0.0025558283	0.0014156274
79	/blocks/blocks.3/blocks.3.0/se/conv_reduce/Conv_output_0	0.0031161667	0.0019700709
80	/blocks/blocks.3/blocks.3.0/se/act1/Sigmoid_output_0	0.00032758823	0.0002549808
81	/blocks/blocks.3/blocks.3.0/se/act1/Mul_output_0	0.0002366215	0.00032878076
84	/blocks/blocks.3/blocks.3.0/se/Mul_output_0	5.8025635e-05	0.0008889851
85	/blocks/blocks.3/blocks.3.0/conv_pwl/Conv_output_0	0.0009509306	0.0032336675
86	/blocks/blocks.3/blocks.3.1/conv_pw/Conv_output_0	0.00014966291	0.0011965767
87	/blocks/blocks.3/blocks.3.1/bn1/act/Sigmoid_output_0	2.3104172e-05	0.0006069027
88	/blocks/blocks.3/blocks.3.1/bn1/act/Mul_output_0	6.756534e-05	0.0005699682
89	/blocks/blocks.3/blocks.3.1/conv_dw/Conv_output_0	0.00019288302	0.0012409401
90	/blocks/blocks.3/blocks.3.1/bn2/act/Sigmoid_output_0	3.262623e-05	0.0004738825
91	/blocks/blocks.3/blocks.3.1/bn2/act/Mul_output_0	5.809163e-05	0.00040274422
92	/blocks/blocks.3/blocks.3.1/se/ReduceMean_output_0	0.0025280535	0.000365251
93	/blocks/blocks.3/blocks.3.1/se/conv_reduce/Conv_output_0	0.0057504224	0.0019475892
94	/blocks/blocks.3/blocks.3.1/se/act1/Sigmoid_output_0	0.0006946266	0.00092850963
95	/blocks/blocks.3/blocks.3.1/se/act1/Mul_output_0	0.005793375	0.0017197866
96	/blocks/blocks.3/blocks.3.1/se/conv_expand/Conv_output_0	0.0032542692	0.0013993264
97	/blocks/blocks.3/blocks.3.1/se/gate/Sigmoid_output_0	0.00048905966	0.0005417557
98	/blocks/blocks.3/blocks.3.1/se/Mul_output_0	0.00017608445	0.00023339101
99	/blocks/blocks.3/blocks.3.1/conv_pwl/Conv_output_0	0.0030686846	0.0014585877
100	/blocks/blocks.3/blocks.3.1/Add_output_0	0.0031812433	0.0035229283
101	/blocks/blocks.3/blocks.3.2/conv_pw/Conv_output_0	0.00046463092	0.0011035614
102	/blocks/blocks.3/blocks.3.2/bn1/act/Sigmoid_output_0	7.7532604e-05	0.0005926611
103	/blocks/blocks.3/blocks.3.2/bn1/act/Mul_output_0	0.0002076765	0.0005223205
104	/blocks/blocks.3/blocks.3.2/conv_dw/Conv_output_0	0.0005651282	0.0012037717
105	/blocks/blocks.3/blocks.3.2/bn2/act/Sigmoid_output_0	9.608331e-05	0.00048499275
106	/blocks/blocks.3/blocks.3.2/bn2/act/Mul_output_0	0.00017842928	0.00038385516
107	/blocks/blocks.3/blocks.3.2/se/ReduceMean_output_0	0.0025045415	0.00033949193
108	/blocks/blocks.3/blocks.3.2/se/conv_reduce/Conv_output_0	0.005862897	0.0018530544
109	/blocks/blocks.3/blocks.3.2/se/act1/Sigmoid_output_0	0.0007015884	0.0009064029
110	/blocks/blocks.3/blocks.3.2/se/act1/Mul_output_0	0.005586757	0.0016543369
111	/blocks/blocks.3/blocks.3.2/se/conv_expand/Conv_output_0	0.002790504	0.0012444104
112	/blocks/blocks.3/blocks.3.2/se/gate/Sigmoid_output_0	0.00041979848	0.0005083868
113	/blocks/blocks.3/blocks.3.2/se/Mul_output_0	0.00014819494	0.00021653684
114	/blocks/blocks.3/blocks.3.2/conv_pwl/Conv_output_0	0.0032936821	0.0015574635
115	/blocks/blocks.3/blocks.3.2/Add_output_0	0.00463647	0.00412507
116	/blocks/blocks.4/blocks.4.0/conv_pw/Conv_output_0	0.0007864241	0.0011596942
117	/blocks/blocks.4/blocks.4.0/bn1/act/Sigmoid_output_0	0.000116672076	0.0005367276
118	/blocks/blocks.4/blocks.4.0/bn1/act/Mul_output_0	0.00028260544	0.00044680314
119	/blocks/blocks.4/blocks.4.0/conv_dw/Conv_output_0	0.00056238647	0.0009743029
120	/blocks/blocks.4/blocks.4.0/bn2/act/Sigmoid_output_0	0.00010556815	0.00062841264
121	/blocks/blocks.4/blocks.4.0/bn2/act/Mul_output_0	0.00027722362	0.00054298557
122	/blocks/blocks.4/blocks.4.0/se/ReduceMean_output_0	0.0026416	0.0004904947
123	/blocks/blocks.4/blocks.4.0/se/conv_reduce/Conv_output_0	0.006663735	0.0019083644
124	/blocks/blocks.4/blocks.4.0/se/act1/Sigmoid_output_0	0.0006677717	0.0008596316
125	/blocks/blocks.4/blocks.4.0/se/act1/Mul_output_0	0.0062392363	0.0016416882
126	/blocks/blocks.4/blocks.4.0/se/conv_expand/Conv_output_0	0.0033047388	0.0014170706
127	/blocks/blocks.4/blocks.4.0/se/gate/Sigmoid_output_0	0.00054162584	0.00074971525
128	/blocks/blocks.4/blocks.4.0/se/Mul_output_0	0.0003330703	0.00038580506
129	/blocks/blocks.4/blocks.4.0/conv_pwl/Conv_output_0	0.005361787	0.002531155
130	/blocks/blocks.4/blocks.4.1/conv_pw/Conv_output_0	0.0011054997	0.00090794725
131	/blocks/blocks.4/blocks.4.1/bn1/act/Sigmoid_output_0	0.000205534	0.00054302946
132	/blocks/blocks.4/blocks.4.1/bn1/act/Mul_output_0	0.0004527371	0.00039086887
133	/blocks/blocks.4/blocks.4.1/conv_dw/Conv_output_0	0.0017248016	0.0013264744
134	/blocks/blocks.4/blocks.4.1/bn2/act/Sigmoid_output_0	0.00027664943	0.00045980973
135	/blocks/blocks.4/blocks.4.1/bn2/act/Mul_output_0	0.00052876485	0.00038345243
136	/blocks/blocks.4/blocks.4.1/se/ReduceMean_output_0	0.0027458246	0.0003202147
137	/blocks/blocks.4/blocks.4.1/se/conv_reduce/Conv_output_0	0.008337066	0.0018249564
138	/blocks/blocks.4/blocks.4.1/se/act1/Sigmoid_output_0	0.0009315759	0.0009176079
139	/blocks/blocks.4/blocks.4.1/se/act1/Mul_output_0	0.008538551	0.0016020554
140	/blocks/blocks.4/blocks.4.1/se/conv_expand/Conv_output_0	0.003670668	0.0010861835
141	/blocks/blocks.4/blocks.4.1/se/gate/Sigmoid_output_0	0.00061995053	0.0004947214
142	/blocks/blocks.4/blocks.4.1/se/Mul_output_0	0.0002776853	0.00021088583
143	/blocks/blocks.4/blocks.4.1/conv_pwl/Conv_output_0	0.0034398844	0.0010042868
144	/blocks/blocks.4/blocks.4.1/Add_output_0	0.0059230197	0.002727239
145	/blocks/blocks.4/blocks.4.2/conv_pw/Conv_output_0	0.0010920274	0.0010350886
146	/blocks/blocks.4/blocks.4.2/bn1/act/Sigmoid_output_0	0.00019354315	0.00051446556
147	/blocks/blocks.4/blocks.4.2/bn1/act/Mul_output_0	0.0004160102	0.00040520186
148	/blocks/blocks.4/blocks.4.2/conv_dw/Conv_output_0	0.0014268254	0.001469916
149	/blocks/blocks.4/blocks.4.2/bn2/act/Sigmoid_output_0	0.00021312047	0.00044073156
150	/blocks/blocks.4/blocks.4.2/bn2/act/Mul_output_0	0.0003990355	0.00038276153
151	/blocks/blocks.4/blocks.4.2/se/ReduceMean_output_0	0.002518116	0.0003317067
152	/blocks/blocks.4/blocks.4.2/se/conv_reduce/Conv_output_0	0.00727809	0.0018282518
153	/blocks/blocks.4/blocks.4.2/se/act1/Sigmoid_output_0	0.0008869405	0.0009299108
154	/blocks/blocks.4/blocks.4.2/se/act1/Mul_output_0	0.0075140693	0.001594181
155	/blocks/blocks.4/blocks.4.2/se/conv_expand/Conv_output_0	0.003148218	0.0010880264
156	/blocks/blocks.4/blocks.4.2/se/gate/Sigmoid_output_0	0.0005323714	0.0004367506
157	/blocks/blocks.4/blocks.4.2/se/Mul_output_0	0.00023109275	0.0002005821
158	/blocks/blocks.4/blocks.4.2/conv_pwl/Conv_output_0	0.0034760193	0.0009777565
159	/blocks/blocks.4/blocks.4.2/Add_output_0	0.0070682573	0.0028992943
160	/blocks/blocks.5/blocks.5.0/conv_pw/Conv_output_0	0.0011035419	0.001468014
161	/blocks/blocks.5/blocks.5.0/bn1/act/Sigmoid_output_0	0.00017246469	0.0003651205
162	/blocks/blocks.5/blocks.5.0/bn1/act/Mul_output_0	0.00023935655	0.00036948547
163	/blocks/blocks.5/blocks.5.0/conv_dw/Conv_output_0	0.0009335324	0.0016323271
164	/blocks/blocks.5/blocks.5.0/bn2/act/Sigmoid_output_0	0.0001503142	0.0008237422
165	/blocks/blocks.5/blocks.5.0/bn2/act/Mul_output_0	0.0007084055	0.0013622433
166	/blocks/blocks.5/blocks.5.0/se/ReduceMean_output_0	0.010306074	0.001346503
167	/blocks/blocks.5/blocks.5.0/se/conv_reduce/Conv_output_0	0.021179428	0.0036870483
168	/blocks/blocks.5/blocks.5.0/se/act1/Sigmoid_output_0	0.0007031831	0.00013924058
169	/blocks/blocks.5/blocks.5.0/se/act1/Mul_output_0	0.0013417543	0.00021202008
170	/blocks/blocks.5/blocks.5.0/se/conv_expand/Conv_output_0	0.00030736218	0.0018002571
171	/blocks/blocks.5/blocks.5.0/se/gate/Sigmoid_output_0	4.8304686e-05	0.00092346966
172	/blocks/blocks.5/blocks.5.0/se/Mul_output_0	0.0005672836	0.0011894618
173	/blocks/blocks.5/blocks.5.0/conv_pwl/Conv_output_0	0.006021776	0.0018752215
174	/blocks/blocks.5/blocks.5.1/conv_pw/Conv_output_0	0.001103883	0.000807528
175	/blocks/blocks.5/blocks.5.1/bn1/act/Sigmoid_output_0	0.00020608083	0.0005077854
176	/blocks/blocks.5/blocks.5.1/bn1/act/Mul_output_0	0.00035068433	0.00029885702
177	/blocks/blocks.5/blocks.5.1/conv_dw/Conv_output_0	0.0019906033	0.0014314817
178	/blocks/blocks.5/blocks.5.1/bn2/act/Sigmoid_output_0	0.00031967438	0.00043641028
179	/blocks/blocks.5/blocks.5.1/bn2/act/Mul_output_0	0.00057341723	0.00042094255
180	/blocks/blocks.5/blocks.5.1/se/ReduceMean_output_0	0.0099079115	0.00038134688
181	/blocks/blocks.5/blocks.5.1/se/conv_reduce/Conv_output_0	0.0344253	0.0014752947
182	/blocks/blocks.5/blocks.5.1/se/act1/Sigmoid_output_0	0.0056844726	0.0008371084
183	/blocks/blocks.5/blocks.5.1/se/act1/Mul_output_0	0.03026931	0.0012428119
184	/blocks/blocks.5/blocks.5.1/se/conv_expand/Conv_output_0	0.018655406	0.0013783332
185	/blocks/blocks.5/blocks.5.1/se/gate/Sigmoid_output_0	0.002813167	0.00038266054
186	/blocks/blocks.5/blocks.5.1/se/Mul_output_0	0.0010293577	0.00019056886
187	/blocks/blocks.5/blocks.5.1/conv_pwl/Conv_output_0	0.012285612	0.0007056559
188	/blocks/blocks.5/blocks.5.1/Add_output_0	0.013867586	0.0019946524
189	/blocks/blocks.5/blocks.5.2/conv_pw/Conv_output_0	0.0023675403	0.00091301044
190	/blocks/blocks.5/blocks.5.2/bn1/act/Sigmoid_output_0	0.00044476372	0.00048035663
191	/blocks/blocks.5/blocks.5.2/bn1/act/Mul_output_0	0.0006741842	0.00032202443
192	/blocks/blocks.5/blocks.5.2/conv_dw/Conv_output_0	0.0038887511	0.001523383
193	/blocks/blocks.5/blocks.5.2/bn2/act/Sigmoid_output_0	0.00062213914	0.0004099933
194	/blocks/blocks.5/blocks.5.2/bn2/act/Mul_output_0	0.0010981205	0.00039957504
195	/blocks/blocks.5/blocks.5.2/se/ReduceMean_output_0	0.010093547	0.0003625181
196	/blocks/blocks.5/blocks.5.2/se/conv_reduce/Conv_output_0	0.039611753	0.0015426408
197	/blocks/blocks.5/blocks.5.2/se/act1/Sigmoid_output_0	0.0061542415	0.0008513045
198	/blocks/blocks.5/blocks.5.2/se/act1/Mul_output_0	0.036021467	0.0013173624
199	/blocks/blocks.5/blocks.5.2/se/conv_expand/Conv_output_0	0.019549191	0.0015742334
200	/blocks/blocks.5/blocks.5.2/se/gate/Sigmoid_output_0	0.0027249856	0.0003322226
201	/blocks/blocks.5/blocks.5.2/se/Mul_output_0	0.00094588666	0.0001725451
202	/blocks/blocks.5/blocks.5.2/conv_pwl/Conv_output_0	0.0143884765	0.00067049795
203	/blocks/blocks.5/blocks.5.2/Add_output_0	0.021272337	0.002095741
204	/blocks/blocks.5/blocks.5.3/conv_pw/Conv_output_0	0.004865537	0.0010659747
205	/blocks/blocks.5/blocks.5.3/bn1/act/Sigmoid_output_0	0.00086632586	0.00044039768
206	/blocks/blocks.5/blocks.5.3/bn1/act/Mul_output_0	0.0012242291	0.00032635813
207	/blocks/blocks.5/blocks.5.3/conv_dw/Conv_output_0	0.0066965963	0.001564251
208	/blocks/blocks.5/blocks.5.3/bn2/act/Sigmoid_output_0	0.001018525	0.00039672412
209	/blocks/blocks.5/blocks.5.3/bn2/act/Mul_output_0	0.0017702777	0.0003774199
210	/blocks/blocks.5/blocks.5.3/se/ReduceMean_output_0	0.010352467	0.00034317243
211	/blocks/blocks.5/blocks.5.3/se/conv_reduce/Conv_output_0	0.045181204	0.0015816354
212	/blocks/blocks.5/blocks.5.3/se/act1/Sigmoid_output_0	0.0067019965	0.0008706029
213	/blocks/blocks.5/blocks.5.3/se/act1/Mul_output_0	0.043173436	0.0013603106
214	/blocks/blocks.5/blocks.5.3/se/conv_expand/Conv_output_0	0.022675313	0.0019136237
215	/blocks/blocks.5/blocks.5.3/se/gate/Sigmoid_output_0	0.0028843991	0.00026971917
216	/blocks/blocks.5/blocks.5.3/se/Mul_output_0	0.00097613776	0.00015066663
217	/blocks/blocks.5/blocks.5.3/conv_pwl/Conv_output_0	0.015474477	0.00053682236
218	/blocks/blocks.5/blocks.5.3/Add_output_0	0.028380187	0.002169295
219	/blocks/blocks.6/blocks.6.0/conv_pw/Conv_output_0	0.003017706	0.000740708
220	/blocks/blocks.6/blocks.6.0/bn1/act/Sigmoid_output_0	0.00061169197	0.0005338136
221	/blocks/blocks.6/blocks.6.0/bn1/act/Mul_output_0	0.0010064499	0.00037068935
222	/blocks/blocks.6/blocks.6.0/conv_dw/Conv_output_0	0.00964251	0.0014225851
223	/blocks/blocks.6/blocks.6.0/bn2/act/Sigmoid_output_0	0.0013903058	0.00044902234
224	/blocks/blocks.6/blocks.6.0/bn2/act/Mul_output_0	0.002510822	0.00040639512
225	/blocks/blocks.6/blocks.6.0/se/ReduceMean_output_0	0.010243341	0.00037751487
226	/blocks/blocks.6/blocks.6.0/se/conv_reduce/Conv_output_0	0.051814035	0.0009224626
227	/blocks/blocks.6/blocks.6.0/se/act1/Sigmoid_output_0	0.010513251	0.00067639345
228	/blocks/blocks.6/blocks.6.0/se/act1/Mul_output_0	0.035027497	0.0006653939
229	/blocks/blocks.6/blocks.6.0/se/conv_expand/Conv_output_0	0.040829808	0.00070964603
230	/blocks/blocks.6/blocks.6.0/se/gate/Sigmoid_output_0	0.009006625	0.0006644313
231	/blocks/blocks.6/blocks.6.0/se/Mul_output_0	0.0036383325	0.00027408556
232	/blocks/blocks.6/blocks.6.0/conv_pwl/Conv_output_0	0.028741943	0.0011057225
233	/conv_head/Conv_output_0	0.029740212	0.0022318913
234	/bn2/act/Sigmoid_output_0	0.0035772326	0.00031688085
235	/bn2/act/Mul_output_0	0.0067886924	0.0003421825
236	/global_pool/pool/GlobalAveragePool_output_0	0.0028113197	0.00029361926
238	648	0.029398264	0.00070788513

You can observe that the errors. To make it more digest I plotted the ratio of mean error of mean ratio :

layers_error.append(
    np.mean(error) / np.mean(err_threshold)
    if np.mean(err_threshold) != 0
    else 0
)

output :

You can observe spikes of error/threshold a very specific layers, since patterns is reccurring juste like the network node graph.

Reproducible example code

The script to run the tests is available on aidge_onnx but is not already merged : aidge_onnx!33 (merged)

changed the description

changed the description

I just pushed more numerically stable computation of mean for ReduceMean and GlobalAveragePooling operators in MR !45 (merged). Could you try again?

After merging with the dev branhc I cannot load the network anymore : It seems that reducemean2d is no longer supported.

Warning: ReduceMean2D is not supported.
- data__blocks_blocks_6_blocks_6_0_se_ReduceMean_output_0 (ReduceMean | GenericOperator)
        - axes : [2, 3]
        - keepdims : 1
- data__blocks_blocks_6_blocks_6_0_se_conv_reduce_Conv_output_0 (Conv)
- data__blocks_blocks_6_blocks_6_0_se_act1_Sigmoid_output_0 (Sigmoid)
- data__blocks_blocks_6_blocks_6_0_se_act1_Mul_output_0 (Mul)
- data__blocks_blocks_6_blocks_6_0_se_conv_expand_Conv_output_0 (Conv)
- data__blocks_blocks_6_blocks_6_0_se_gate_Sigmoid_output_0 (Sigmoid)
- data__blocks_blocks_6_blocks_6_0_se_Mul_output_0 (Mul)
- data__blocks_blocks_6_blocks_6_0_conv_pwl_Conv_output_0 (Conv)
- data__conv_head_Conv_output_0 (Conv)
- data__bn2_act_Sigmoid_output_0 (Sigmoid)
- data__bn2_act_Mul_output_0 (Mul)
- data__global_pool_pool_GlobalAveragePool_output_0 (GlobalAveragePool)
- data__global_pool_flatten_Flatten_output_0 (Flatten | GenericOperator)
        - axis : 1

Maybe you have an idea to solve the issue ?

mentioned in issue aidge_onnx#13 (closed)