I am trying to retrain the ResNet_18_threeClass_VGA_deploy_pruned model on the kitti-dataset to detect Car and Van. After getting some errors on Digits, I noticed the “*.prototxt” does not include any sort of loss function or an input for the label data which may be the root of the problem.
Could someone please provide advice on how to retrain this model? Is it intended to be trained on another software instead of Digits?
Below I include the setup parameters and errors I got on Digits:
Object Detection database:
Pad image (Width x Height): (None)
Resize image (Width x Height): 640 368
Channel conversion: RGB
Custom classes: DontCare, Car, Van
Object Detection model:
Subtract Mean: None
Batch size: 6
Batch Accumulation: 4
Solver type: Adam
Base Learning Rate: 0.0001
Custom Network > Caffe
Custom Network: Copied the contents of “ResNet_18_threeClass_VGA_deploy_pruned.prototxt” as provided on the DeepStream SDK
Pretrained model(s): Path to “ResNet_18_threeClass_VGA_pruned.caffemodel”
The model was trained without any errors, however no information was displayed on the Loss/Epoch graph.
When using the Test tool on an image of the dataset, the following error was displayed:
The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()