Keras-Flask文档分类API无输出问题排查求助

Hey there! I’ve seen this exact set of issues pop up a lot when folks first put a Keras model behind a Flask endpoint—let’s work through them one by one to get your /predict endpoint running smoothly.

1. Fixing the NameError: name 'texts' is not defined #

This error means your code is trying to use a variable texts that hasn’t been created or assigned a value before it’s referenced. The most common culprit here is missing a step where you convert the incoming request text into a format Keras can process.

Here’s a corrected snippet for your endpoint that addresses this:

from flask import Flask, request, jsonify
import tensorflow as tf
from tensorflow.keras.preprocessing.text import Tokenizer
from tensorflow.keras.preprocessing.sequence import pad_sequences
import json

app = Flask(__name__)

# Load your model and tokenizer ONCE at app startup (critical for performance!)
model = tf.keras.models.load_model('your_cnn_model.h5')
tokenizer = Tokenizer(num_words=5000)
# Load the word index you saved during training (don't skip this!)
tokenizer.word_index = json.load(open('word_index.json', 'r'))

@app.route('/predict', methods=['POST'])
def predict():
    try:
        # Extract text from the incoming JSON request
        request_data = request.get_json()
        input_text = request_data.get('text', '')
        
        # Create the 'texts' variable by wrapping the input in a list (Keras expects batches!)
        texts = [input_text]
        
        # Preprocess the text exactly as you did during training
        sequences = tokenizer.texts_to_sequences(texts)
        padded_sequences = pad_sequences(sequences, maxlen=100) # Use YOUR training maxlen here
        
        # Run prediction
        prediction = model.predict(padded_sequences)
        predicted_class = tf.argmax(prediction, axis=1).numpy()[0]
        
        return jsonify({
            'predicted_class': int(predicted_class),
            'confidence': float(prediction[0][predicted_class])
        })
    except Exception as e:
        return jsonify({'error': str(e)}), 400

Common mistakes to avoid here:

• Forgetting to wrap the single input text in a list (Keras models are built to handle batches of data, even if it’s just one sample)
• Using the wrong method to extract request data (use request.get_json() for JSON payloads, request.form for form data)
• Not reloading the tokenizer’s word index from training—this ensures consistent text encoding

2. Fixing the TensorFlow ValueError #

Once you resolve the NameError, this error almost always stems from an input shape mismatch between what your model expects and what you’re feeding it.

CNN text models typically expect input in the shape (batch_size, max_sequence_length). To fix this:

• Double-check that the maxlen value in pad_sequences matches exactly what you used during training. If you trained with maxlen=150, don’t use maxlen=100 in your Flask app.
• Verify the input shape of your model’s first layer. For example, if your first layer is Embedding(5000, 128, input_length=100), your padded sequences must be of length 100.
• Ensure you’re passing a 2D array to model.predict(). If your padded sequence is a 1D array (shape (100,) instead of (1,100)), add a batch dimension with:

  import numpy as np
  padded_sequences = np.expand_dims(padded_sequences, axis=0)
  

Quick Windows-Specific Tips #

• Stick to tensorflow.keras instead of standalone Keras to avoid version compatibility issues.
• Never load your model inside the /predict endpoint—this will reload it on every request, causing slowdowns and errors. Load it once when the app starts.
• Test your preprocessing pipeline independently first: take a sample text, run it through tokenization and padding, then check if the output shape matches what your model expects before integrating with Flask.

Content of the question originates from Stack Exchange, question author StatguyUser