能否修改Keras训练后的Embedding层并扩展词向量维度？

Absolutely, this is totally doable! Let me walk you through exactly how to pull this off:

Step 1: Extract the trained embedding matrix #

First, you need to get hold of the word embedding weights your model learned. The Embedding layer's weights are stored as a numpy array, which you can access like this:

import numpy as np

# Get the original 8-dimensional embedding matrix (shape: (100, 8))
original_embedding = model.layers[0].get_weights()[0]

Note: Since your Embedding layer is the first layer in the model, we use layers[0] here. Adjust the index if your layer order is different.

Step 2: Prepare your extra 2-dimensional data #

Assuming you've already calculated the 2 additional values for each of your 100 words, store them in a numpy array with shape (100, 2). For example:

# Replace this with your actual computed values
extra_dims = np.random.rand(100, 2)  # Simulating 2 extra dimensions for 100 words

Make sure the first dimension (100) matches the number of words in your original embedding—this is critical for successful concatenation.

Step 3: Combine the original embedding with extra dimensions #

Use numpy's hstack to append the extra dimensions to each word vector:

# New embedding matrix with shape (100, 10) (8 original + 2 extra)
new_embedding_matrix = np.hstack((original_embedding, extra_dims))

Step 4: Update your model with the new embedding #

You have two options here, depending on whether you want to keep your original model structure or build a new one:

Option 1: Build a new model with the updated embedding layer

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Embedding, Flatten, Dense

new_model = Sequential()
# Define the new Embedding layer with output_dim=10 (instead of 8)
new_model.add(Embedding(
    input_dim=100,
    output_dim=10,
    input_length=max_length,
    weights=[new_embedding_matrix]  # Load the combined embedding matrix
))
new_model.add(Flatten())
new_model.add(Dense(1, activation='sigmoid'))

# Recompile the model
new_model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['acc'])
print(new_model.summary())

Option 2: Modify the original model in-place

If you want to tweak your existing model instead of creating a new one:

# Remove the original Embedding layer from the model
model.pop(0)

# Insert the new Embedding layer at the front
model.insert(
    0,
    Embedding(
        input_dim=100,
        output_dim=10,
        input_length=max_length,
        weights=[new_embedding_matrix]
    )
)

# Recompile to apply changes
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['acc'])
print(model.summary())

Key Notes #

• Double-check that the shape of your extra_dims array matches the number of words in your embedding (100 in your case). Mismatched shapes will cause errors during concatenation.
• When updating the model, ensure the output_dim of the new Embedding layer matches the size of your combined embedding matrix (10, in this case).
• The Flatten layer will automatically adapt to the new input size (from max_length*8 to max_length*10), so you don't need to adjust subsequent Dense layers unless they had hard-coded input dimensions.

内容的提问来源于stack exchange，提问作者Ziqi