TensorFlow中GraphKeys.INIT_OP的作用？如何添加初始化时执行的Assign OP？

Hey there! Let's clear up your confusion around GraphKeys.INIT_OP and the right way to add one-time initialization Assign ops in TensorFlow.

What's tf.GraphKeys.INIT_OP Used For? #

Even though it's missing from the official docs, INIT_OP is a core internal collection in TensorFlow that gathers all operations related to initializing variables and other graph state. Think of it as a registry for every op that needs to run once at startup to set up your graph.

When you call tf.global_variables_initializer(), TensorFlow doesn't just magically know which ops to run—it actually creates a grouped operation that executes every op in the INIT_OP collection. This includes the auto-generated assign ops from variable initializers, plus any custom ops you add to the collection yourself.

It's worth noting that this collection is mostly intended for TensorFlow's internal use, but understanding it helps you hook into the standard initialization flow cleanly.

Best Way to Add an Assign Op That Runs Only Once #

Your initial thought to add it to GLOBAL_VARIABLES is a common misstep—GLOBAL_VARIABLES is for storing Variable objects, not operations. Adding an Assign op there will cause errors, since the collection expects variable instances, not ops.

Here are the two cleanest approaches:

1. Add the Assign Op to INIT_OP (Integrate with Global Initialization) #

If your Assign op is part of your core initialization logic (you want it to run alongside variable initializers), add it directly to the INIT_OP collection:

import tensorflow as tf

# Define your variable
my_custom_var = tf.Variable(0.0)

# Create your Assign op
assign_initial_value = tf.assign(my_custom_var, 42.0)

# Add it to the INIT_OP collection
tf.add_to_collection(tf.GraphKeys.INIT_OP, assign_initial_value)

# Later, when initializing
with tf.Session() as sess:
    # This will run ALL ops in INIT_OP, including your custom Assign op
    sess.run(tf.global_variables_initializer())
    print(sess.run(my_custom_var))  # Outputs 42.0

This ensures your op runs exactly once when you call the global initializer—no extra steps needed.

2. Create a Custom Initialization Group (Independent Control) #

If you want to run your Assign op separately from the global variable initialization (e.g., for conditional setup), create a grouped operation:

import tensorflow as tf

my_var = tf.Variable(0.0)
assign_op = tf.assign(my_var, 100.0)

# Group your custom init ops
custom_init = tf.group(assign_op)

with tf.Session() as sess:
    # Run global variable init first if needed
    sess.run(tf.global_variables_initializer())
    # Run your custom init only when needed
    sess.run(custom_init)
    print(sess.run(my_var))  # Outputs 100.0

This gives you explicit control over when the op runs, while still ensuring it only executes once (unless you call custom_init again).

Key Takeaway #

• INIT_OP is the internal collection that powers TensorFlow's global initialization flow—use it to hook custom init ops into the standard startup process.
• Never add ops to GLOBAL_VARIABLES (that's for variables, not operations).
• Choose between adding to INIT_OP (for integrated setup) or creating a custom group (for independent control) based on your use case.

内容的提问来源于stack exchange，提问作者David Parks