NiftyNet 'evaluation'动作输出异常及--dataset_to_infer选项失效求助

Let's break down the two problems you're facing and walk through practical fixes to diagnose what's going wrong:

1. Abnormal Evaluation Metrics (average_distance=0, Jaccard/Dice=1) #

These extreme, perfect scores are a clear sign the evaluation logic isn’t actually comparing your model’s predictions to the ground truth properly. Here are the most likely culprits to check:

• Mismatched prediction/ground truth paths: Double-check your evaluation configuration—if the tool is accidentally loading ground truth labels as both the "prediction" and "ground truth" inputs, it will naturally return perfect overlap scores (1 for Jaccard/Dice) and zero distance.
• Broken/unimplemented evaluation logic: Since you can’t find documentation for the evaluation action, it’s possible this feature is either half-baked or returns placeholder values. If this is an open-source tool, dive into the code handling evaluation: look for functions calculating these metrics and verify if they’re actually computing values vs. hardcoding 0/1.
• Format/dimension mismatches: Ensure your model’s output masks match the ground truth’s format exactly. For example, if labels are binary masks but predictions are raw probability maps, a misconfigured auto-threshold might accidentally turn all predictions into exact matches. Also confirm both have identical height/width/channel counts—mismatched dimensions could lead to incorrect alignment that somehow results in perfect scores.
• Debug mode artifacts: Some tools enable a debug mode that skips real evaluation and returns dummy metrics. Check if you’ve enabled any flags or environment variables that might trigger this behavior.

2. --dataset_to_infer=Validation Not Restricting to Validation Set #

If the tool is using the full dataset instead of just the validation split, try these targeted checks:

• Parameter name/format errors: Verify the exact parameter syntax by running your_command --help. Some tools use variations like --infer-dataset, --dataset, or require lowercase (validation) instead of capitalized. A tiny typo here could make the flag ignored.
• Configuration file override: If you’re using a config file alongside the command line, the config might be overriding your dataset_to_infer flag. Many tools prioritize config settings over command line arguments—open your config file, look for lines specifying the dataset split, and either modify it or use an --override-config flag (if available) to force the command line value to take precedence.
• Invalid validation split definition: Confirm your dataset’s validation split is properly set up:
  • Does your dataset directory have a Validation subfolder with the correct subset of data?
  • If using split files (like train.txt/val.txt), is the Validation split mapped to a subset of data instead of the full dataset?
• Incorrect parameter positioning: Some tools require flags to be placed in specific positions (e.g., before positional arguments). Make sure --dataset_to_infer=Validation is placed early in your command, before any paths or subcommands.

内容的提问来源于stack exchange，提问作者Ginesu_Kun