TensorFlow QuickDraw循环神经网络预测功能实现报错求助

Hey there, let's break down how to debug your prediction function issues with the QuickDraw recurrent tutorial. Since you mentioned evaluation works smoothly but prediction fails after modifying create_dataset.py and model_fn(), here are targeted steps to diagnose and fix the problem:

1. Validate Input Data Consistency #

First, double-check that your prediction input pipeline matches the evaluation pipeline exactly in terms of:

• Sequence handling: QuickDraw uses variable-length ink sequences—ensure your prediction code isn't truncating, padding, or reshaping sequences differently than during evaluation.
• Feature naming: The dictionary keys from your prediction input must match what model_fn() expects (e.g., if evaluation uses 'ink' as the feature key, prediction can't use a different label like 'drawing').
• Preprocessing: Confirm you're applying the same normalization (like scaling x/y coordinates to [-1, 1]) and pen-state encoding to prediction data as you did for training/evaluation.

2. Audit model_fn() Modifications for Prediction Mode #

When handling mode=tf.estimator.ModeKeys.PREDICT, your model_fn() needs to avoid evaluation-specific logic and return a properly structured EstimatorSpec:

• Skip loss calculation, metric tracking, and training-only ops (like gradient updates) in prediction mode.
• Ensure you're populating the predictions field correctly. A standard implementation looks like this:

  if mode == tf.estimator.ModeKeys.PREDICT:
      predictions = {
          'class_ids': tf.argmax(input=logits, axis=1),
          'probabilities': tf.nn.softmax(logits),
          'logits': logits,
      }
      return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)
  

• If your model uses batch normalization, make sure moving averages are properly loaded from the checkpoint in prediction mode (don't accidentally use training-mode batch stats).

3. Debug the Prediction Input Function #

Add logging to inspect the output of your custom prediction input function:

def predict_input_fn():
    # Your existing code to load/preprocess prediction data
    dataset = ... # Your dataset creation logic
    # Debug: Print first element's shape and dtype
    for elem in dataset.take(1):
        print("Prediction input shape:", elem['ink'].shape)
        print("Prediction input dtype:", elem['ink'].dtype)
    return dataset

Confirm the output matches the input signature your model was trained with—for QuickDraw, the ink feature is typically a 2D tensor of shape [None, 3] (sequence steps, x/y/pen state).

4. Check for Checkpoint-Graph Compatibility #

If you modified the model architecture after training, your saved checkpoint might not be compatible with the updated prediction graph. Try re-training the model with your revised model_fn() before running prediction to ensure consistency.

5. Capture and Analyze the Exact Error #

Even if you've tried multiple fixes, sharing the full error traceback will help pinpoint the issue. Common culprits include:

• Shape mismatch: Your input data doesn't match the model's expected input dimensions.
• NotFoundError: A variable or operation from the training graph is missing in the prediction graph (often from unplanned model changes).
• InvalidArgumentError: Incorrect data types or missing feature keys.

If you can share snippets of your modified predict_input_fn() and the PREDICT mode block in model_fn(), we can narrow this down even further!

内容的提问来源于stack exchange，提问作者Farooq Khan