TensorFlow 1.14中使用tf.compat.v1.keras.utils.Progbar更新训练Loss和Accuracy时出现AttributeError问题的求助

Let's break down the issue and fix it step by step:

What's Causing the Error? #

Two key mistakes are triggering that AttributeError:

1. Incorrect average calculation for accuracy: You're calling np.mean(batch_train_acc) instead of using the accumulated train_meta_acc list. This means you're trying to compute the mean of a single Tensor object instead of a list of values.
2. Trying to process Tensor objects directly with NumPy: TensorFlow 1.x's Tensor objects can't be directly passed to NumPy functions like np.mean()—you need to convert them to plain numeric values first. The error about DType comes from NumPy trying to handle the Tensor's internal type system incorrectly.

Solution Code #

Here's the corrected version of your code, with explanations:

import numpy as np
import tensorflow as tf

train_progbar = tf.compat.v1.keras.utils.Progbar(train_data.steps)
train_meta_loss = []
train_meta_acc = []

for i in range(train_data.steps):
    # Get batch metrics (Tensor objects)
    batch_train_loss, batch_train_acc = maml.train_on_batch(
        train_data.get_one_batch(), 
        inner_optimizer, 
        inner_step=1, 
        outer_optimizer=outer_optimizer
    )
    
    # Convert Tensors to NumPy-compatible values (TF 1.x compatible)
    batch_loss_val = tf.keras.backend.eval(batch_train_loss)
    batch_acc_val = tf.keras.backend.eval(batch_train_acc)
    
    # Add to accumulated lists
    train_meta_loss.append(batch_loss_val)
    train_meta_acc.append(batch_acc_val)
    
    # Calculate averages from the accumulated lists
    mean_train_meta_loss = np.mean(train_meta_loss)
    mean_train_meta_acc = np.mean(train_meta_acc)  # Fixed: use the accumulated list
    
    # Update Progbar with numeric values
    train_progbar.update(i+1, [('loss', mean_train_meta_loss), ('accuracy', mean_train_meta_acc)])

Even Better: Let Progbar Handle Averages #

Progbar is designed to automatically compute running averages for you, so you don't need to manually track and average the accumulated lists. This is more efficient and cleaner:

train_progbar = tf.compat.v1.keras.utils.Progbar(train_data.steps)
train_meta_loss = []
train_meta_acc = []

for i in range(train_data.steps):
    batch_train_loss, batch_train_acc = maml.train_on_batch(
        train_data.get_one_batch(), 
        inner_optimizer, 
        inner_step=1, 
        outer_optimizer=outer_optimizer
    )
    
    batch_loss_val = tf.keras.backend.eval(batch_train_loss)
    batch_acc_val = tf.keras.backend.eval(batch_train_acc)
    
    # Save history if needed
    train_meta_loss.append(batch_loss_val)
    train_meta_acc.append(batch_acc_val)
    
    # Let Progbar calculate the running average automatically
    train_progbar.update(i+1, [('loss', batch_loss_val), ('accuracy', batch_acc_val)])

Key Notes for TF 1.14 #

• If you've enabled eager execution (via tf.enable_eager_execution()), you can replace tf.keras.backend.eval(tensor) with tensor.numpy() for simpler syntax.
• Always ensure you're working with numeric values (not Tensors) when passing data to NumPy functions or Progbar updates.

内容的提问来源于stack exchange，提问作者Ken