-
Summary
I’m trying to compile a custom ViTPose model (with CBAM modules) using hailomz parse. I have created a custom YAML file (custom_vit.yaml) and alls script (custom_vit.alls), but I’m encountering errors during the parsing stage. I suspect the issue might be related to the nodes configuration in the YAML file, but I’m also looking for general guidance on how to correctly configure the YAML file for a custom model. -
Questions
parser: nodes:
2-1. For a custom model, should the nodes section in the YAML file be left empty (), or should I specify specific nodes?
2-2. If I need to specify nodes, how do I determine the correct node paths for my custom model? (I have tried using Netron, but I’m not sure which nodes to select.)
2-3. I have encountered errors like “Expecting value” and “TypeError: ‘NoneType’ object is not subscriptable” when trying different nodes configurations.
2.4 Overall YAML configuration:
Are there any other settings in the YAML file that I should adjust
specifically for my custom ViTPose model?
Are the preprocessing, postprocessing, and info sections configured
correctly?
- Attempted Solution
3-1. Tried different nodes configurations (empty list, specific node paths, dummy nodes).
Verified Hailo Model Zoo installation and environment activation.
3-2. Compared my YAML file with the vit_pose_small.yaml example.
3-3. Reviewed the Hailo Model Zoo documentation.
-
Execution Command
hailomz parse --ckpt ./custom_vit_100.onnx --yaml custom_vit.yaml -
YAMl File
network:
network_name: “custom_vit”
paths:
alls_script: custom_vit.alls
network_path:
- custom_vit_100.onnx
parser:
nodes:
normalization_params:
normalize_in_net: true
std_list:
- 0.229
- 0.224
- 0.225
mean_list:
- 0.485
- 0.456
- 0.406
start_node_shapes: {}
preprocessing:
network_type: classification
meta_arch: null
padding_color: 0
quantization:
calib_set:
- custom_vit.tfrecord
calib_set_name: custom_vit
postprocessing:
network_type: null
meta_arch: null
postprocess_config_file: null
hpp: false
hn_editor:
yuv2rgb: false
flip: false
input_resize:
enabled: false
input_shape: [256, 192]
bgr2rgb: false