Question

我的TF深度学习模型停留在50％的验证准确度上，而我的训练准确度是65％并在不断提高。以下是一些注意事项-

问题陈述-我使用（224 * 224）张面孔作为输入并尝试预测性别
数据大小-我每个人都有29,999张图像。培训数据为80％-47,998，验证数据为20％-11,999

下面是我的代码。可能有点让人不知所措，请随时问我一些具体问题-

def generator(positive_dataset, negative_dataset):
while True:
    for pos_rec, neg_rec in zip(positive_dataset, negative_dataset):
          pos_x, pos_y = pos_rec
          neg_x, neg_y = neg_rec
          x = tf.concat([pos_x, neg_x], axis=0)
          y = tf.concat([pos_y, neg_y], axis=0)
          yield x, y
train_generator = generator(train_dataset_female, train_dataset_male) 
test_generator = generator(test_dataset_female, test_dataset_male)
input_layer = layers.Input(shape=(IMG_SIZE, IMG_SIZE, 3))
mobilenet_layer = MobileNetV2(weights=None, include_top=False, alpha=0.35)(input_layer)
landmarks_layer = layers.GlobalAveragePooling2D()(mobilenet_layer)
landmarks_layer = layers.Flatten()(landmarks_layer)
landmarks_layer_flat = layers.Dense(68 * 2, activation='sigmoid', name='landmarks_flat')(landmarks_layer)
landmarks_layer = layers.Reshape(target_shape=(68, 2), name='landmarks')(landmarks_layer_flat)

landmarks_model = tf.keras.models.Model(inputs=input_layer, outputs=landmarks_layer)

prob_face_layer = layers.GlobalAveragePooling2D()(mobilenet_layer)
prob_face_layer = layers.Flatten()(prob_face_layer)
prob_face_layer = layers.concatenate([prob_face_layer, landmarks_layer_flat])
prob_face_layer = layers.Dense(256, activation='relu')(prob_face_layer)
prob_face_layer = layers.Dense(1, activation='sigmoid')(prob_face_layer)

prob_face_net = tf.keras.models.Model(inputs=input_layer, outputs=prob_face_layer)
print('Prob Face Net')
print(prob_face_net.summary())

print('Pose Model')
pose_model.summary()

print('Landmarks Model')
landmarks_model.summary()

combined_model = tf.keras.models.Model(inputs=input_layer,outputs[landmarks_layer, pose_layer])
print('Combined Model')
combined_model.summary()

full_model = tf.keras.models.Model(inputs=input_layer, outputs=[prob_face_layer, landmarks_layer, pose_layer])
print('Full Net')
print(full_model.summary())


combined_model.load_weights('/mnt/sachins/apprach2/models/bounding_box_based_model_faster_2.h5')
for layer in range(0, 8):
    prob_face_net.layers[layer].trainable = False 
prob_face_net.compile(loss='binary_crossentropy', optimizer='adam', metrics=[ 'accuracy'])


history = prob_face_net.fit_generator(
            train_generator,
            epochs=EPOCHS,
            steps_per_epoch=(num_train_imgs_male+ num_train_imgs_female)//(BATCH_SIZE//2),
            validation_data=test_generator,
            validation_steps=(29999*2 - (num_train_imgs_male+num_train_imgs_female))//(BATCH_SIZE//2),
            callbacks=callbacks,
            verbose = 1)

Answer 1

您应该考虑模型的道德问题。 https://venturebeat.com/2020/02/20/google-cloud-ai-removes-gender-labels-from-cloud-vision-api-to-avoid-bias/

深度学习模型的验证精度保持在0.5，而训练精度却在提高

1 个答案: