如何查看Torch预训练网络的代码结构与参数？

Absolutely! You can absolutely inspect the architecture and parameters (like filter sizes, channel counts, weight values) of your .t7 Torch pre-trained model. Let me walk you through simple, beginner-friendly steps to do this:

Step 1: Load the Model in Torch #

First, open up the Torch interactive shell by running this command in your terminal:

th

Once you're in the Torch shell, load your .t7 model file with this Lua code (replace your_model_file.t7 with your actual file path):

local model = torch.load('your_model_file.t7')

Step 2: Print the Full Network Architecture #

To get a high-level overview of the entire network structure, just print the model object:

print(model)

This will output a breakdown of every layer in your model—you'll see things like nn.Conv2d (convolutional layers) with details like kernel size, input/output channels, and nn.Linear (fully connected layers) with input/output neuron counts.

Step 3: Inspect Specific Layer Details #

If you want to dive deeper into a particular layer's parameters (like filter dimensions or weight values), you can access individual layers directly. For example, if your model is a nn.Sequential (the most common structure for simple networks), use get(index) to grab a layer:

-- Grab the first convolutional layer (adjust the index to target your desired layer)
local target_layer = model:get(1)

-- Print key details about the layer
print('Layer type:', target_layer.__typename)
print('Kernel size (width x height):', target_layer.kW, target_layer.kH)
print('Input channels:', target_layer.nInputPlane)
print('Output channels:', target_layer.nOutputPlane)

-- Check the shape of the layer's weight tensor (shows filter dimensions)
print('Weight tensor shape:', target_layer.weight:size())

For more complex models (like those with branched structures), you might need to traverse sub-modules—start by printing the full model first to understand how layers are nested.

Step 4: Visualize Parameters (Optional) #

If you want to see what a filter looks like, you can use Torch's built-in gnuplot tool to visualize weight tensors:

-- Load the gnuplot library
require 'gnuplot'

-- Visualize the first filter from your target convolutional layer
gnuplot.imagesc(target_layer.weight[1])

This will pop up a plot showing the pixel values of the filter.

Quick Tips for Beginners #

• If your .t7 file only contains weight values (not the full model structure), you'll need to first define the exact network architecture in Lua, then load the weights into that structure. Most pre-trained .t7 files, though, save the complete model object.
• If you're unsure about layer indices, the full model printout from Step 2 will list layers in order—just match the index to the layer you want to inspect.

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