Hi, I'm currently using the eIQ Toolkit v1.13.1 to train an object detection model with the ssd_mobilenet_v3 plugin on BDD100K dataset. When I run the eiq-trainer command, I encounter the following error:
ValueError: Argument `name` must be a string and cannot contain character '/'. Received: name=Conv/BatchNorm (of type <class 'str'>)
From the traceback, it seems that the issue occurs during the initialization of a BatchNormalization layer, where the layer name includes a / character (e.g., Conv/BatchNorm). The specific line triggering this error appears to be related to layer naming inside the ssd_mobilenet_v3 plugin.
Command I used:
eiq-trainer --detection \
--input_shape=320,320,3 \
--batch_size=16 \
--epochs=50 \
--tune=ssd_mobilenet_v3 \
--dataset=/path/to/bdd100k.eiqp \
--optimizer=Adam \
--lr=0.001 \
--initialization=he \
--initialization_type=normal \
--initialization_seed=1 \
--loss=sigmoid_lossMy questions:
- Is there a way to prevent the layer names from including special characters like /?
- Should I modify the ssd_mobilenet_v3 plugin code to replace / with an underscore (_) in layer names?
- Is this a known issue with the current version of eIQ Toolkit, and would updating the toolkit solve this problem?
I would appreciate any guidance on how to resolve this issue. Thank you in advance!
