TensorFlow变量已存在报错：tf.reset_default_graph()放置位置咨询

Hey there! Let's break down where to place tf.reset_default_graph() to resolve that frustrating variable duplication error in your Inverse Compositional CNN.

Why the Error Happens #

Your current code uses reuse=l > 0 in the variable scope to reuse variables after the first loop iteration, which works for a single run of the loop. But if you're running your code multiple times (like re-executing a notebook cell or rerunning the script), TensorFlow's default computation graph still holds the variables created from the previous run. When you try to build the model again, it tries to create duplicate variables with the same names—hence the ValueError.

Where to Place tf.reset_default_graph() #

The key is to run this function before any TensorFlow variables or model components are defined. Here are the best spots:

1. At the very start of your script/notebook
   Put it right after importing TensorFlow and other dependencies, before you define any options or start your loop:

   import tensorflow as tf
   import slim
   
   # Reset the default graph FIRST to clear any leftover variables from previous runs
   tf.reset_default_graph()
   
   # Now define your options and start building the model
   opt = ... # Your option setup code
   for l in range(opt.warpN):
       with tf.variable_scope("ICSTN", reuse=l > 0) as sc:
           end_points_collection = sc.original_name_scope + '_end_points'
           with slim.arg_scope([...])
           # Rest of your model construction code
   

2. Before re-running model code in interactive environments
   If you're using Jupyter Notebook or another interactive tool, execute tf.reset_default_graph() every time before you run the code block that builds your model. This ensures you're starting with a fresh graph each time.

Extra Check: Confirm Your Reuse Logic #

Your reuse=l > 0 setup is correct for reusing variables within the same loop run (first iteration creates variables, subsequent iterations reuse them). Combining this with tf.reset_default_graph() ensures that even across multiple runs of the entire code, you won't get duplicate variable conflicts.

内容的提问来源于stack exchange，提问作者Tuo Feng tfeng