TensorFlow LinearClassifier预测时出现KeyError: 'latitude'问题求助

Let's break down why this error is happening and how to fix it quickly:

The Root Cause #

Your LinearClassifier was trained expecting features named latitude and longitude (check your training input_fn—it creates a dictionary with these exact keys). But when making predictions, you're passing in a dictionary with a single key "x" instead. The model has no idea what "x" refers to, so it throws a KeyError when looking for the latitude feature it was trained on.

Fix Steps #

You need to align the input keys in your prediction function with the feature column names the model expects. Here are two clean ways to do this:

Option 1: Split your new samples into matching feature arrays

new_samples = np.array(([39.8070525,-5.698599],[43.8800776,4.654769]),dtype=np.float64)

# Create an input dictionary with keys matching your feature columns
my_input_fn = tf.estimator.inputs.numpy_input_fn(
    x={
        "latitude": new_samples[:, 0],  # Extract first column as latitude values
        "longitude": new_samples[:, 1]  # Extract second column as longitude values
    },
    y=None,
    num_epochs=1,  # No need for multiple epochs when predicting once
    shuffle=False
)

# Now this will execute without errors
pred = model.predict_classes(input_fn=my_input_fn)

Option 2: Pass separate feature lists directly

If you already have latitude and longitude values as separate lists, you can pass them directly to the input function:

new_lats = [39.8070525, 43.8800776]
new_lons = [-5.698599, 4.654769]

my_input_fn = tf.estimator.inputs.numpy_input_fn(
    x={
        "latitude": np.array(new_lats, dtype=np.float64),
        "longitude": np.array(new_lons, dtype=np.float64)
    },
    y=None,
    num_epochs=1,
    shuffle=False
)

Quick Optimization Note #

I adjusted num_epochs=2 to num_epochs=1 in the prediction input function—since you're predicting on these samples once, there's no need to repeat the process multiple times.

内容的提问来源于stack exchange，提问作者user2159982