Got it, let’s break down this issue step by step—this is a classic case of TensorFlow version-specific behavior changes that can completely throw off your model’s results, even with identical code. You’re absolutely right to trace this discrepancy back to tf.contrib.layers.fully_connected() between v1.3 and v1.4.

Key Behavior Changes That Cause the Gap #

The main culprit here is a default weight initializer update in TensorFlow 1.4:

• In TF 1.3, tf.contrib.layers.fully_connected() used xavier_initializer (Glorot initialization) by default, which scales weights based on the number of input and output units.
• In TF 1.4+, this default was swapped to variance_scaling_initializer (He initialization), which is optimized for ReLU activations and uses a different scaling factor.

This change directly impacts how your model converges: different initial weights lead to different gradient flows, resulting in vastly different final accuracy or loss values between the two versions.

Minimal Reproduction Code #

Here’s a stripped-down snippet that demonstrates the weight initialization difference clearly:

import tensorflow as tf

def build_simple_model():
    inputs = tf.placeholder(tf.float32, shape=[None, 10])
    # Use default parameters to highlight version differences
    output = tf.contrib.layers.fully_connected(inputs, 5)
    return inputs, output

with tf.Session() as sess:
    inputs, output = build_simple_model()
    sess.run(tf.global_variables_initializer())
    # Extract the layer's weights
    weights = [var for var in tf.trainable_variables() if 'weights' in var.name][0]
    print("Weight Initialization Stats:")
    print(f"Mean: {sess.run(tf.reduce_mean(weights)):.6f}")
    print(f"Std Deviation: {sess.run(tf.math.reduce_std(weights)):.6f}")

If you run this in TF 1.3 vs 1.4, you’ll see distinct mean and standard deviation values for the initialized weights—this is the root of your result mismatch.

Fixes to Align Results Across Versions #

To make your local code match the platform’s TF 1.3 results, explicitly override the initializer to use TF 1.3’s default:

# Modified layer to replicate TF 1.3's behavior in TF 1.4+
output = tf.contrib.layers.fully_connected(
    inputs,
    5,
    weights_initializer=tf.contrib.layers.xavier_initializer()
)

Alternatively, if you have control over the platform’s environment (unlikely for most courses), you could update the platform code to use TF 1.4’s default initializer to match your local setup:

# For TF 1.3, replicate TF 1.4's default behavior
output = tf.contrib.layers.fully_connected(
    inputs,
    5,
    weights_initializer=tf.contrib.layers.variance_scaling_initializer()
)

Pro Tips for Avoiding Version Issues #

• Pin your TensorFlow version: Use pip install tensorflow==1.3 (or the exact version the platform uses) to align your local environment perfectly.
• Avoid relying on defaults: Always explicitly define initializers, activations, and regularizers for critical layers. This makes your code reproducible across versions and environments.
• Check release notes: TensorFlow’s official release notes for v1.4 explicitly call out this initializer change in tf.contrib.layers—it’s always worth reviewing these when debugging version gaps.

内容的提问来源于stack exchange，提问作者LCasta