FastMaskRCNN训练报错求助：TensorFlow弃用及张量转换警告

Hey there, let's break down and fix the two warnings you're hitting during FastMaskRCNN training. These aren't fatal errors—your training should still run—but resolving them will clean up your logs and align with modern TensorFlow practices.

1. Deprecation Warning for create_global_step #

What's happening #

The warning flags that create_global_step from tensorflow.contrib.framework is deprecated. TensorFlow moved this utility out of the contrib module into the core training tools a while back.

How to fix it #

• Open the train/train.py file in your FastMaskRCNN project
• Jump to line 224 where the deprecated function is called
• Replace the old usage with TensorFlow's official supported method:

  # Replace this existing line (or similar)
  # global_step = create_global_step()
  # With this updated code:
  global_step = tf.train.create_global_step()
  

• If you see an import for the deprecated function at the top of the file, remove it—you don't need it anymore:

  # Delete this line if present
  # from tensorflow.contrib.framework.python.ops.variables import create_global_step
  

2. UserWarning: Converting sparse IndexedSlices to a dense Tensor #

What's happening #

This warning pops up when TensorFlow automatically converts a sparse IndexedSlices object (common in gradient calculations for Mask RCNN's RoI-related layers) to a dense tensor. The conversion works, but it's flagged because it might hint at a chance to optimize tensor handling or could impact performance.

Fix options (pick one that fits your needs) #

• Option 1: Explicitly handle sparse tensors
  Find the gradient computation section in your training code (usually where the optimizer applies gradients). Add explicit conversion for sparse tensors before passing them to the optimizer:

  gradients = optimizer.compute_gradients(loss)
  # Convert sparse IndexedSlices to dense tensors manually
  gradients = [(tf.convert_to_tensor(grad) if isinstance(grad, tf.IndexedSlices) else grad, var) for grad, var in gradients]
  optimizer.apply_gradients(gradients, global_step=global_step)
  

• Option 2: Suppress the warning temporarily
  If training runs fine and you don't want to tweak core logic, add these lines at the very top of your training script to silence this specific warning:

  import warnings
  warnings.filterwarnings("ignore", category=UserWarning, message=r"Converting sparse IndexedSlices to a dense Tensor of unkn.*")
  

• Option 3: Match TensorFlow version to project requirements
  FastMaskRCNN was built for older TensorFlow 1.x versions (around 1.12-1.13). If you're using a newer TF 1.x release (like 1.14+), downgrading to a compatible version can eliminate these compatibility warnings. Do this via conda:

  conda install tensorflow-gpu==1.13.1  # Use tensorflow==1.13.1 for CPU-only setup
  

内容的提问来源于stack exchange，提问作者EmmaTang