如何正确使用tf.nn.embedding_lookup并查看Tensor对象的具体值？

Hey there, let's work through your TensorFlow 1.x embedding issue step by step—you're trying to inspect tensor values and use tf.nn.embedding_lookup correctly, right? Let's break down what's going wrong and fix it.

First, let's unpack your errors #

• The tf.constant([1][0]) syntax is invalid: [1][0] evaluates to the integer 1, so trying to do [0][0][1] is like trying to index an integer, which throws TypeError: 'int' object is not subscriptable. Similarly, [0][1][0] fails because the list [0] only has one element, so index 1 doesn't exist—hence the IndexError.
• You're misusing tf.nn.embedding_lookup: this function takes an embedding weight matrix and a tensor of indices to look up, not a pre-computed embedding tensor like X_embedded.
• You forgot to initialize variables: In TensorFlow 1.x, all trainable variables (like the weights in your Embedding layer) need to be initialized before you can evaluate their values in a session.

Fixed code example #

Let's rewrite your code to correctly build the embedding, use tf.nn.embedding_lookup, and inspect tensor values. I'll assume pos is your positional embedding tensor, and X is your input sequence of word indices:

import tensorflow as tf
from tensorflow import keras

# Your parameter dictionary (vocab size 30, embedding dim 128)
params = {'30': 30, '128': 128}

# 1. Define your input index tensor (example: 1 sample, 25 tokens)
# Replace this with your actual input indices
X = tf.constant([list(range(25))], dtype=tf.int32)

# 2. Build word embedding layer and generate embeddings
word_embedding_layer = keras.layers.Embedding(
    input_dim=params['30'],
    output_dim=params['128'],
    embeddings_initializer='glorot_normal'
)
word_embeddings = word_embedding_layer(X)

# 3. Add positional embedding (replace with your actual pos tensor)
# Example positional embedding matching shape (1,25,128)
pos = tf.random.normal(shape=(1, 25, 128))
final_embeddings = word_embeddings + pos

# 4. Correct usage of tf.nn.embedding_lookup
# Grab the embedding weight matrix from the layer
embedding_weights = word_embedding_layer.weights[0]
# Define indices you want to look up (e.g., indices 5 and 10)
lookup_indices = tf.constant([5, 10], dtype=tf.int32)
lookup_results = tf.nn.embedding_lookup(embedding_weights, lookup_indices)

# 5. Evaluate tensors in a session
with tf.Session() as sess:
    # Initialize all variables first!
    sess.run(tf.global_variables_initializer())
    
    # Inspect the final embedding tensor values
    final_embeds_val = sess.run(final_embeddings)
    print("Final Embeddings Shape:", final_embeds_val.shape)
    print("First token's embedding from the first sample:\n", final_embeds_val[0][0])
    
    # Inspect the embedding lookup results
    lookup_val = sess.run(lookup_results)
    print("\nEmbeddings for indices 5 and 10:\n", lookup_val)

Key takeaways to remember #

• Always initialize variables: Call sess.run(tf.global_variables_initializer()) before evaluating any tensors that depend on trainable variables (like Embedding weights).
• Use tf.nn.embedding_lookup properly: Pass the embedding weight matrix (from EmbeddingLayer.weights[0]) as the first argument, and your target indices as the second argument.
• Inspect tensors correctly: Use sess.run(tensor) or tensor.eval(session=sess) to get the actual numeric values of a tensor—just make sure all dependencies are initialized.
• Avoid invalid tensor construction: Don't chain list indices inside tf.constant()—if you need a constant tensor with values like [0,0,1], write tf.constant([[0,0,1]]) directly.

内容的提问来源于stack exchange，提问作者Carl Choi