CNN训练触发TypeError：float()参数类型异常求解决方案

Hey there, let's dig into your problem step by step and fix those issues one by one!

1. The Direct Cause of Your TypeError: Uncalled flatten() Method #

Looking at your generator code, this line is the root of the float() conversion error:

labels.append((np.array(batch_sample[1].flatten)).transpose())

flatten is a method of numpy arrays, not a property. You need to call it with parentheses flatten()—right now you're passing the method object itself instead of the flattened numerical values. Fix it to:

labels.append(np.array(batch_sample[1].flatten()).transpose())

2. Wrong Loss Function for Your Task #

You mentioned this is an image detection (regression) task (output corresponds to 1D image size parameters), but you're using sparse_categorical_crossentropy—this loss is designed for classification tasks, not regression. Swap it for a regression-friendly loss function:

model.compile(loss='mean_squared_error', optimizer='adam', metrics=['mae'], options=run_opts)

mean_squared_error is the standard choice for regression tasks, and mae (Mean Absolute Error) is a more interpretable metric for evaluating regression performance.

3. Redundant/Incorrect Layer in Model Structure #

• The Activation('softmax') layer added after your last MaxPooling2D is completely unnecessary here. Applying softmax to feature maps doesn't make sense for a regression task, and it will distort the feature distribution before flattening. Delete this layer entirely.
• Your output Dense(19316) layer is correct for regression (linear activation is the default, which is what we want for continuous value prediction).

4. Wrong Step Counts in fit_generator #

You set steps_per_epoch = len(training_data) and validation_steps = len(val_data)—this will make the training run way more steps than needed. The correct calculation is total samples divided by batch size:

batch_size = 20
steps_per_epoch = len(training_data) // batch_size
validation_steps = len(val_data) // batch_size

model.fit_generator(
    train_generator, 
    steps_per_epoch=steps_per_epoch, 
    epochs=nb_epoch, 
    validation_data=validation_generator, 
    validation_steps=validation_steps
)

Note: If you're using TensorFlow 2.1 or later, fit_generator() is deprecated—you can directly use model.fit() with your generator input.

5. Small Optimization for Your Generator #

You can simplify the generator logic and add data shuffling to improve training stability:

def generator(data_arr, batch_size=10):
    num_samples = len(data_arr)
    while True:
        # Shuffle data each epoch to avoid order bias
        np.random.shuffle(data_arr)
        for offset in range(0, num_samples, batch_size):
            batch_samples = data_arr[offset:offset+batch_size]
            samples = []
            labels = []
            for batch_sample in batch_samples:
                samples.append(batch_sample[0])
                labels.append(batch_sample[1].flatten().transpose())
            # Simplify dimension expansion
            X_ = np.array(samples)[..., np.newaxis]
            Y_ = np.array(labels)
            yield (X_, Y_)

内容的提问来源于stack exchange，提问作者lr99