基于InceptionV3多标签分类预测值过小及阈值设置咨询

Hey there! Let's break down your two key issues with fine-tuning InceptionV3 for multi-class multi-label classification:

1. Why are my predictions all tiny values? #

Your current setup has a few critical mismatches for multi-label tasks that are likely causing those near-zero predictions. Let's fix them:

a. Wrong class_mode in data generators #

By default, flow_from_directory uses class_mode='categorical', which is for single-label tasks (each image belongs to exactly one class). For multi-label classification (images can have multiple classes), you need to set class_mode='binary' so the generator outputs binary labels (0/1) for each class—this matches your binary_crossentropy loss function.

Update your generator code:

train_generator = train_datagen.flow_from_directory(
    args.train_dir,
    target_size=(IM_WIDTH, IM_HEIGHT),
    batch_size=batch_size,
    class_mode='binary'  # Add this line!
)

validation_generator = test_datagen.flow_from_directory(
    args.val_dir,
    target_size=(IM_WIDTH, IM_HEIGHT),
    batch_size=batch_size,
    class_mode='binary'  # Add this line!
)

b. Dataset structure doesn't fit multi-label #

flow_from_directory assumes each subfolder is a single class, and images in that folder only belong to that class. If you're doing multi-label (one image has multiple classes), this structure won't work. Instead, use flow_from_dataframe with a CSV file that maps each image to its multiple labels.

Example CSV format (train_labels.csv):

Then adjust your data loading:

import pandas as pd

train_df = pd.read_csv('train_labels.csv')
train_generator = train_datagen.flow_from_dataframe(
    dataframe=train_df,
    x_col='image_path',
    y_col=['cat', 'dog', 'bird'],  # List all your class columns
    target_size=(IM_WIDTH, IM_HEIGHT),
    batch_size=batch_size,
    class_mode='raw'  # For multi-label, use 'raw' mode
)

c. Too few training epochs #

NB_EPOCHS = 3 is way too low for transfer learning + fine-tuning. Models need time to adapt to your custom data. Try starting with 10-20 epochs for transfer learning, then 20-30 for fine-tuning, and adjust based on validation loss/accuracy.

d. Validation set checks #

While the validation set isn't the direct cause of tiny predictions, double-check:

• It has the same number of classes as the training set
• Its images use the same preprocessing as training
• Labels are correctly annotated (no mismatches)

2. How to set a threshold for detecting classes? #

In multi-label classification, each output value is a probability (0-1) representing how likely the class is present. Here's how to use thresholds effectively:

a. Basic thresholding (start with 0.5) #

A default threshold of 0.5 works for many cases. Use it to filter which classes are considered "present":

import numpy as np

# Assume you've preprocessed your image to match InceptionV3's requirements
preprocessed_img = ...  # Shape: (1, 299, 299, 3)
predictions = model.predict(preprocessed_img)[0]  # Get the prediction array for your image

threshold = 0.5
# Get indices of classes where probability > threshold
predicted_class_indices = np.where(predictions > threshold)[0]

# Map indices to class names (use your generator's class_indices)
class_names = list(train_generator.class_indices.keys())
detected_classes = [class_names[idx] for idx in predicted_class_indices]

print("Detected classes in the image:", detected_classes)

b. Optimize the threshold for your task #

0.5 isn't always optimal. Adjust based on what matters more:

• Reduce false positives (don't label a class that's not there): Raise the threshold (e.g., 0.6-0.7)
• Reduce false negatives (don't miss a class that is there): Lower the threshold (e.g., 0.3-0.4)

To find the best threshold, calculate the F1-score (balances precision and recall) on your validation set:

from sklearn.metrics import f1_score

# Get validation set true labels and predictions
val_true = validation_generator.labels  # Or load from CSV if using flow_from_dataframe
val_pred = model.predict(validation_generator)

best_threshold = 0.5
best_f1 = 0

# Test thresholds from 0.1 to 0.9 in 0.1 steps
for threshold in np.arange(0.1, 1.0, 0.1):
    # Convert probabilities to binary labels
    pred_labels = (val_pred > threshold).astype(int)
    # Calculate macro F1-score (treats all classes equally)
    current_f1 = f1_score(val_true, pred_labels, average='macro')
    if current_f1 > best_f1:
        best_f1 = current_f1
        best_threshold = threshold

print(f"Best threshold: {best_threshold} | Best F1-score: {best_f1}")

内容的提问来源于stack exchange，提问作者ou2105