使用TF-slim评估微调后的Inception-V3模型时出现张量未找到报错

Hey there, let's work through this error you're hitting when evaluating your fine-tuned Inception-V3 model with TF-Slim's eval_image_classifier.py.

What's Causing This? #

The error boils down to one simple issue: your fine-tuned checkpoint doesn't contain the weights for the AuxLogits branch of Inception-V3. This usually happens for one of two reasons:

• You excluded the AuxLogits variables when loading the pre-trained model during fine-tuning, so they were never initialized or saved to your checkpoint.
• You disabled the AuxLogits branch during fine-tuning (e.g., set aux_logits=False), but the evaluation script is still trying to load these weights because it defaults to enabling AuxLogits.

Step-by-Step Fixes #

1. Skip Loading AuxLogits Weights During Evaluation

The quickest fix is to tell the evaluation script to ignore the missing AuxLogits variables. Add these flags when running eval_image_classifier.py:

python eval_image_classifier.py \
  --checkpoint_path=/path/to/your/fine-tuned/model.ckpt \
  --eval_dir=/path/to/eval/dir \
  --dataset_dir=/path/to/your/dataset \
  --dataset_name=your_dataset \
  --dataset_split_name=validation \
  --model_name=inception_v3 \
  --exclude_scopes=InceptionV3/AuxLogits \
  --checkpoint_exclude_scopes=InceptionV3/AuxLogits

The --exclude_scopes flag tells the model definition to skip creating those variables, and --checkpoint_exclude_scopes ensures the script doesn't try to load them from your checkpoint.

2. Verify Your Checkpoint's Variables

To confirm that AuxLogits weights are indeed missing from your checkpoint, use TensorFlow's checkpoint inspection tool:

python -m tensorflow.python.tools.inspect_checkpoint --file_name=/path/to/your/model.ckpt

This will print all variables stored in the checkpoint. If you don't see any entries starting with InceptionV3/AuxLogits, that confirms our initial diagnosis.

3. Fix the Issue at the Source (Re-Fine-Tune if Needed)

If you actually want the AuxLogits branch to be part of your model, you'll need to re-run fine-tuning with the correct settings:

• When using train_image_classifier.py, remove any --exclude_scopes=InceptionV3/AuxLogits flag you might have added. This ensures the pre-trained AuxLogits weights are loaded and updated during training.
• Make sure you're not setting aux_logits=False in your model configuration during training—keep it enabled (the default for Inception-V3) so the variables are created and saved to the checkpoint.

4. Match Model Configuration Between Training and Evaluation

Always ensure the model settings (like whether AuxLogits is enabled) are identical between training and evaluation. If you disabled AuxLogits during fine-tuning, add --aux_logits=False to your evaluation command to match:

python eval_image_classifier.py \
  # ... your other flags ...
  --aux_logits=False

内容的提问来源于stack exchange，提问作者Yunhao Li