Replies: 3 comments 9 replies
-
|
Should we set Thanks. |
Beta Was this translation helpful? Give feedback.
-
|
Even setting use_conv as False yields the same result |
Beta Was this translation helpful? Give feedback.
-
|
May be @wyli can suggest something. He worked on the PR for torchvisionfc |
Beta Was this translation helpful? Give feedback.
-
|
yes, it's not possible to make a single config that suits all popular torchvision models, as there are some inconsistencies in the model variable names and arrangements. for the inceptionv3, here are the test cases for your references: {
"model_name": "inception_v3",
"num_classes": 5,
"use_conv": True,
"pool": "",
"in_channels": 2048,
"node_name": "Mixed_7c.cat_2",
},
(2, 3, 299, 299),
(2, 5, 8, 8),there are some other use cases in the test file https://github.com/Project-MONAI/MONAI/blob/dev/tests/test_torchvision_fc_model.py#L29 |
Beta Was this translation helpful? Give feedback.
-
|
It is kind of weird that if I try to input wrong shape intentionally as
I will keep digging and post |
Beta Was this translation helpful? Give feedback.
-
|
I think I found the problem and for this method to work a small tweak needs to be done on the But if I use an and it fails to instantiate the network. Here is the network block that causes the problem So, yes this is the issue. Sorry @Nic-Ma @wyli for bringing out so many issues :( |
Beta Was this translation helpful? Give feedback.
-
|
Its fixed now after I made the changes I am starting a pull request |
Beta Was this translation helpful? Give feedback.
-
|
I see, the class can already pass the extra params to the torchvision model: so you can write the json section like this: "network_def": {
"_target_": "TorchVisionFCModel",
"model_name": "inception_v3",
"num_classes": 4,
"in_channels": 2048,
"use_conv": true
"pool": "",
"node_name": "Mixed_7c.cat_2",
"aux_logits": false,
}, |
Beta Was this translation helpful? Give feedback.
-
|
Thanks @wyli, that would have been my guess too that we can pass extra parameters and I was surprised when it was not doing so. that is why I started the pull request. But maybe you can ignore it or merge it (upto you), it is a non-breaking change. It works now. |
Beta Was this translation helpful? Give feedback.
-
Hello,
I think there is some issue with how the error checking happens in the model zoo parser. Please see attached the json script I created.after my discussions with @Nic-Ma and @yiheng-wang-nv. I tested the script and it works great when used with Densenet without any issues. However when used with
TorchvisionFCModelafter @wyli integrated inception_v3 in the latest monai TorchvisionCode. I am using it as followsI get the following error:
I checked the shaped of the input data and I can assure its [4, 3, 299, 299] where 4 is the batch size. I do not understand why it is using an input size of [4, 1000]. I also tried the following
and I get the same error
Any thoughts ?
train.txt
Beta Was this translation helpful? Give feedback.
All reactions