Caffe设置propagate_down:0报错，提示ParamSpec无该字段求解决

Let’s work through your problem step by step—you’ve got two clear syntax/concept issues to fix, plus a better way to achieve your layer-freezing goal:

1. Fix the Non-ASCII Character Error #

First, look at your propagate_down：0 line: that colon is a Chinese full-width colon, not an English one. Protobuf is extremely strict about syntax, so replace it with a standard English colon:

param{
  propagate_down: 0
}

This is exactly what’s causing the Interpreting non ascii codepoint 239 error in your log.

2. Why propagate_down Doesn’t Work on Eltwise Layers #

The second error—Message type "caffe.ParamSpec" has no field named "propagate_down"—is totally valid. The propagate_down field only applies to layers that have trainable parameters (like Convolution, FullyConnected, BatchNorm, etc.). The Eltwise layer just performs element-wise operations (SUM in your case) and has no parameters of its own to train or freeze. Setting propagate_down here does nothing at all.

3. Correct Way to Freeze Preceding Layers & Fine-Tune Subsequent Ones #

Your goal is to freeze all layers before the Eltwise layer and fine-tune layers after it. Here are the most reliable methods to do this:

Option 1: Set Learning Rate Multipliers to 0 for Frozen Layers #

For every trainable layer before your Eltwise layer (e.g., conv1_1, conv1_2, ..., conv5_1, conv5_2_1x1_increase), add a param block to disable parameter updates:

layer {
  name: "conv5_1"
  type: "Convolution"
  # ... other layer settings ...
  param {
    lr_mult: 0       # No learning rate applied to this layer's weights
    decay_mult: 0    # No weight decay applied
  }
}

Repeat this for every trainable layer you want to keep fixed. This tells Caffe to skip updating these layers’ parameters during training.

Option 2: Use Solver-Level Layer Freezing (Caffe Version Dependent) #

Some custom Caffe builds support freezing layers directly in the solver config by listing layer names:

solver {
  # ... other solver settings ...
  freeze_layers: "conv1_1"
  freeze_layers: "conv1_2"
  # Add all layers you want to freeze here
}

Note that this isn’t supported in all vanilla Caffe versions, so option 1 is the most universally reliable.

Final Verification #

After fixing the colon and applying one of the layer-freezing methods above, your prototxt should parse without errors, and you’ll be able to fine-tune only the layers after your Eltwise layer as intended.

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