文档

How Do I Restore Model Training Parameters After Quantization Operators Are Inserted?

The list of quantization variable operators (quant_add_ops) to be added has been passed to the quantize_model API call. The variable values in the list cannot be found in the model training file. Therefore, an error indicating that the variables cannot be found is reported when the model training parameters are restored. In this case, you need to delete the variable values in the quant_add_ops list from the restoration list before restoring the model parameters.

  1. Restoration of shadow variables
     1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
    # 1. Obtain the dictionary variables_dict corresponding to {variable name:variable value} of all variables.
variables_ema = tf.train.ExponentialMovingAverage(moving_average_decay)
variables_dict = variables_ema.variables_to_restore()

# 2. Define the {variable name:variable value} of the variable to be restored in the dictionary params_need_load.
params_need_load = dict()

# 3. Search for the variables to be restored from variables_dict based on quant_add_ops.
for key, value in variables_dict.items():
    if value not in quant_add_ops:
        params_need_load[key] = value

# 4. Restore variables.
loader = tf.train.Saver(params_need_load)
loader.restore(sess, FLAGS.checkpoint)
  2. Recovery of non-shadow variables
     1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
    # 1. Obtain the dictionary variables_dict corresponding to {variable name:variable value} of all variables.
variables_global = tf.global_variables()
variables_dict = dict()
for var in variables_global:
    variables_dict[var.name[:-2]] = var

# 2. Define the {variable name:variable value} of the variable to be restored in the dictionary params_need_load.
params_need_load = dict()

# 3. Search for the variables to be restored from variables_dict based on quant_add_ops.
for key, value in variables_dict.items():
    if value not in quant_add_ops:
       params_need_load[key] = value

# 4. Restore variables.
loader = tf.train.Saver(params_need_load)
loader.restore(sess, FLAGS.checkpoint)
Parent topic: FAQ