我有一个受过训练的应用程序MobileNetV2 using FasterRcnn in Tensorflow.
相同的训练数据和相同的程序,我用 训练MobileNetv1
,结果很好。
我使用此页面中的预训练模型both MobileNetV1 and MobileNetV2.
唯一的问题是我无法在从预训练模型的检查点训练 MobileNetV2 时初始化所有微调变量。使用 FasterRCNN 训练的 MobileNetV2 模型需要这些变量。但它们在预训练模型中不可用(检查点的所有变量都在这里)
WARNING:root:Variable [MobilenetV2/expanded_conv_15/expand/biases] is not available in checkpoint
WARNING:root:Variable [MobilenetV2/expanded_conv_15/expand/weights] is not available in checkpoint
WARNING:root:Variable [MobilenetV2/expanded_conv_15/project/biases] is not available in checkpoint
WARNING:root:Variable [MobilenetV2/expanded_conv_16/depthwise/biases] is not available in checkpoint
WARNING:root:Variable [MobilenetV2/expanded_conv_16/expand/biases] is not available in checkpoint
我不能只初始化五个变量,但为什么结果如此糟糕。会不会是其他问题?