CNN模型预测测试数据报错：GPU资源耗尽/CPU输入类型错误求助

报错信息 #

GPU运行时错误 #

(0) Resource exhausted:  2 root error(s) found.
(0) Resource exhausted: SameWorkerRecvDone unable to allocate output tensor. 
Key: /job:localhost/replica:0/task:0/device:CPU:0;7a6cf13ce274a521;
/job:localhost/replica:0/task:0/device:GPU:0;ret_1;0:0

CPU运行时错误 #

File "/home//programs/Neural/model_testing.py", line 224, in <module>
prediction_prob_matrix = model(test_generator,len(df_test), verbose=0)
File "/home/.conda/envs/my_env/lib/python3.9/site-packages/keras/src/utils/traceback_utils.py", line 70, in error_handler
raise e.with_traceback(filtered_tb) from None
File "/home/.conda/envs/my_env/lib/python3.9/site-packages/keras/src/engine/input_spec.py", line 213, in assert_input_compatibility
raise TypeError(
TypeError: Inputs to a layer should be tensors. Got '<generator object generator at 0x7f7217fe3890>' (of type <class 'generator'>) as input for layer 'model'.

相关代码 #

train_generator = generator (df_train,tokenizer,onehot,label_encoder, n_classes,batch_size)
validation_generator = generator (df_valid,tokenizer,onehot,label_encoder, n_classes,batch_size)
test_generator = generator (df_test,tokenizer,onehot,label_encoder, n_classes,batch_size)
model.summary()

model.fit(train_generator, 
      validation_data=validation_generator,
      epochs=1,steps_per_epoch = len(df_train)//batch_size, validation_steps = len(df_valid)//batch_size, shuffle = True)

prediction = model.predict(test_generator, len(df_test), verbose=0)
# get the classes with the highest predicted probability, save them to our dataframe
df_test['lab'] = label_encoder.inverse_transform(prediction_prob_matrix)
# add the predicted probabilities
df_test['PREDICTED_PROB'] = prediction_prob_matrix.max(axis=1)
# take a look at what we've got
df_test.head()

解决方法 #

1. 修正CPU报错：规范model.predict用法 #

CPU报错核心是直接调用model(test_generator)导致的，必须使用model.predict()方法，同时代码存在变量名不一致问题（用prediction接收结果却调用prediction_prob_matrix），一并修正：

# 正确调用predict，设置steps参数并修正变量名
prediction_prob_matrix = model.predict(test_generator, steps=len(df_test)//batch_size, verbose=0)
# 若有剩余未整除的样本，可加1确保覆盖全部数据
# prediction_prob_matrix = model.predict(test_generator, steps=(len(df_test)//batch_size)+1, verbose=0)

注意：model.predict的第二个参数是steps（需执行的批次数），不是样本总数，需和生成器的batch_size匹配。

2. 解决GPU资源耗尽问题 #

GPU报错是显存不足导致，可通过以下方式缓解：

• 降低测试生成器的batch_size：生成测试数据时使用更小的批次尺寸，减少单批次显存占用。
• 开启TensorFlow显存增长模式：避免一次性占用全部GPU显存，按需分配：

import tensorflow as tf
gpus = tf.config.list_physical_devices('GPU')
if gpus:
    try:
        for gpu in gpus:
            tf.config.experimental.set_memory_growth(gpu, True)
    except RuntimeError as e:
        print(e)

• 清理冗余进程：关闭GPU上其他占用资源的程序，释放显存空间。

3. 验证生成器输出格式 #

确保自定义generator函数每次迭代返回**(输入张量, 标签张量)**的元组，即使测试阶段不需要标签，也需返回(inputs, None)，保证model.predict能正确解析输入数据。

内容的提问来源于stack exchange，提问作者borfo