如何将Keras各轮次训练权重合并保存为单个文件？

Hey there! Let's work through this problem together. I’ve run into similar scenarios when managing Keras model checkpoints, so here’s how you can merge those 50 weight files into a single one, depending on your needs:

Scenario 1: You Need a Single Final Weight File (Most Common Use Case) #

If your goal is to get a single weight file for converting to a TensorFlow checkpoint (which is typically all you need for deployment or further training), you don’t actually need to "merge" all files—you just need to pick the best/final epoch’s weights and save them as a single file. Here’s how:

• First, recreate your exact Keras model architecture (this has to match the model you used during training perfectly, down to layer types, shapes, and configurations):

  # Example model architecture (replace with your actual model code)
  from tensorflow.keras.models import Sequential
  from tensorflow.keras.layers import Conv2D, Flatten, Dense
  
  def build_your_model():
      model = Sequential([
          Conv2D(32, kernel_size=(3,3), activation='relu', input_shape=(28,28,1)),
          Flatten(),
          Dense(10, activation='softmax')
      ])
      return model
  
  # Initialize the model
  model = build_your_model()
  

• Next, load the weights from your desired epoch (usually the last one, i.e., checkpoint_epoch_49.hdf5 if you started counting from 0):

  model.load_weights('checkpoint_epoch_49.hdf5')
  

• Finally, save this single set of weights as your target checkpoint.h5 file:

  model.save_weights('checkpoint.h5')
  

That’s it! This checkpoint.h5 file is ready to use with TensorFlow just like any standard Keras weight file.

Scenario 2: You Want to Store All 50 Epochs’ Weights in One File #

If you need to keep all epoch weights for analysis (like tracking weight changes over training) instead of just one, you can bundle them into a single HDF5 file using the h5py library. Here’s how:

• First, install h5py if you haven’t already:

  pip install h5py
  

• Then run this code to load and merge all weights:

  import h5py
  # Use the same model architecture function from Scenario 1
  from your_model_script import build_your_model
  
  model = build_your_model()
  
  # Create a new HDF5 file to hold all epoch weights
  with h5py.File('all_epochs_merged_weights.h5', 'w') as merged_file:
      for epoch in range(50):
          # Load weights for the current epoch
          weight_path = f'checkpoint_epoch_{epoch}.hdf5'
          model.load_weights(weight_path)
          
          # Get the model's weight tensors
          layer_weights = model.get_weights()
          
          # Create a group in the HDF5 file for this epoch
          epoch_group = merged_file.create_group(f'epoch_{epoch}')
          
          # Save each layer's weights as a dataset in the group
          for layer_idx, weights in enumerate(layer_weights):
              epoch_group.create_dataset(f'layer_{layer_idx}', data=weights)
  

To retrieve weights from this merged file later (e.g., load epoch 10’s weights into your model):

import h5py
import numpy as np
from your_model_script import build_your_model

model = build_your_model()

with h5py.File('all_epochs_merged_weights.h5', 'r') as merged_file:
    # Fetch weights for epoch 10
    target_epoch_weights = []
    for layer_key in merged_file['epoch_10'].keys():
        target_epoch_weights.append(np.array(merged_file['epoch_10'][layer_key]))
    
    # Load weights into the model
    model.set_weights(target_epoch_weights)

内容的提问来源于stack exchange，提问作者Black Snow