如何在Python主程序中打印TensorFlow卷积函数的权重与卷积结果？

Hey there! Let's work through how to print both the weights inside your conv_fn and the convolution result properly. I'll start by pointing out a couple of issues in your current code, then show you the fixed version with clear explanations.

首先，修正代码里的核心问题 #

Your current main function has a mistake with feed_dict: p_input is a regular Python function parameter (a numpy array), not a TensorFlow placeholder, so you don't need to feed it. Also, to access weights1 from outside conv_fn, we need to make it accessible—either by returning it from the function or referencing it correctly.

1. 打印conv_fn中的weights1 #

There are two straightforward ways to get and print weights1:

方法一：让conv_fn返回权重和卷积结果（推荐） #

Modify conv_fn to return both the ReLU output and the weights tensor. This is the cleanest approach because it makes dependencies explicit and avoids relying on global variable names:

def conv_fn(p_input):
    # Convolutional Layer #1
    weights1 = tf.get_variable("weights1", [3, 3, 3, 32], initializer=tf.random_normal_initializer())
    conv1 = tf.nn.conv2d(
        input=p_input.astype(np.float32),
        filter=weights1,
        strides=[1, 1, 1, 1],
        padding="VALID"
    )
    conv1_relu = tf.nn.relu(conv1)
    # Return both the convolution result and the weights
    return conv1_relu, weights1

Then in your main function, capture both values and run them in the session:

def main(argv):
    sess = tf.InteractiveSession()
    sess.run(tf.global_variables_initializer())
    
    im = [1.0, 1.0, 2.0, 5.0, 1.0, 1.0, 2.0, 5.0, 5.0,
          1.0, 1.0, 2.0, 5.0, 1.0, 1.0, 2.0, 5.0, 5.0,
          1.0, 1.0, 2.0, 5.0, 1.0, 1.0, 2.0, 5.0, 5.0]
    image = np.reshape(im, [-1, 3, 3, 3])
    
    # Capture both the convolution result and weights tensor
    con, weights = conv_fn(image)
    
    # Fetch and print weights1
    weights_val = sess.run(weights)
    print("weights1 shape:", weights_val.shape)
    print("weights1 content:\n", weights_val)
    
    # Fetch and print convolution result
    con_val = sess.run(con)
    print("\nConvolution result shape:", con_val.shape)
    print("Convolution result content:\n", con_val)
    
    sess.close()

if __name__ == "__main__":
    tf.app.run()

方法二：通过变量名获取权重（不推荐，不够直观） #

Since weights1 is defined with tf.get_variable (which creates a global variable by default), you can fetch it from the TensorFlow graph using its name:

# In main function, after calling conv_fn(image)
weights_tensor = tf.get_default_graph().get_tensor_by_name("weights1:0")
weights_val = sess.run(weights_tensor)
print("weights1:\n", weights_val)

Note: This is less maintainable—if you ever rename the variable or add a scope, this will break. The first method is much safer for long-term use.

2. 打印卷积运算的结果 #

Once you remove the unnecessary feed_dict (your input is a numpy array that TensorFlow automatically converts to a tensor), you just need to run the con tensor in the session. The code above already includes this: we use sess.run(con) to get the actual numpy array of the convolution result, then print it.

额外提示：使用占位符提升灵活性（可选） #

If you want to reuse conv_fn with different inputs later, it's better to use a TensorFlow placeholder instead of passing numpy arrays directly. Here's how that would look:

def conv_fn(p_input):
    weights1 = tf.get_variable("weights1", [3, 3, 3, 32], initializer=tf.random_normal_initializer())
    conv1 = tf.nn.conv2d(
        input=p_input,  # p_input is now a placeholder tensor
        filter=weights1,
        strides=[1, 1, 1, 1],
        padding="VALID"
    )
    return tf.nn.relu(conv1), weights1

def main(argv):
    sess = tf.InteractiveSession()
    sess.run(tf.global_variables_initializer())
    
    # Create a placeholder for input images
    input_placeholder = tf.placeholder(tf.float32, shape=[None, 3, 3, 3])
    
    im = [1.0, 1.0, 2.0, 5.0, 1.0, 1.0, 2.0, 5.0, 5.0,
          1.0, 1.0, 2.0, 5.0, 1.0, 1.0, 2.0, 5.0, 5.0,
          1.0, 1.0, 2.0, 5.0, 1.0, 1.0, 2.0, 5.0, 5.0]
    image = np.reshape(im, [-1, 3, 3, 3])
    
    con, weights = conv_fn(input_placeholder)
    
    # Use feed_dict to pass the actual image data
    weights_val = sess.run(weights)
    con_val = sess.run(con, feed_dict={input_placeholder: image})
    
    print("weights1:\n", weights_val)
    print("Convolution result:\n", con_val)
    
    sess.close()

This setup lets you pass different images to conv_fn by updating the value in feed_dict, which aligns better with typical TensorFlow workflows.

内容的提问来源于stack exchange，提问作者V. Arbez