转移学习问题：使用我自己的训练的权重，或者是USIIGACI训练有素的MRCNN（Matterport实施）

发布于 2025-02-07 07:54:02 字数 979 浏览 3 评论 0原文

我的项目涉及细胞检测，所以我想我尝试使用 usiigaci预先训练的权重。但是，当我尝试时，我会收到以下错误： valueerror：第362页（命名为“锚”）期望1个权重，但是保存的权重为0个元素。例如，该培训适合预估计的可可丝。这是我用来加载权重的代码：

model = MaskRCNN(mode='training', model_dir='./', config=config) model.load_weights('Usiigaci_3.h5', by_name=True, exclude=["mrcnn_class_logits", "mrcnn_bbox_fc", "mrcnn_bbox", "mrcnn_mask"])

在尝试加载通过自己的照片训练我的模型生成的权重时，我也会遇到类似的问题，继续我上次停止的训练。收到的错误是：

ValueError: Layer #362 (named "anchors"), weight <tf.Variable 'Variable:0' shape=(4, 261888, 4) dtype=float32> has shape (4, 261888, 4), but the saved weight has shape (2, 261888, 4).

加载权重：

model.load_weights('new_weigths/40_epochs/mask_rcnn_cell_cfg_0040.h5', by_name=True, exclude=["mrcnn_class_logits", "mrcnn_bbox_fc", "mrcnn_bbox", "mrcnn_mask"])

如果您了解为什么会发生这种情况，请告诉我。谢谢！！

原文

my project involves cell detection, so I thought I'd try training my NN using Usiigaci pre-trained weights.
but when I try I get the following error:
ValueError: Layer #362 (named "anchors") expects 1 weight(s), but the saved weights have 0 element(s).
The training works fine for pretrained coco weigths for example.
this is the code i use to load the weights:

model = MaskRCNN(mode='training', model_dir='./', config=config) model.load_weights('Usiigaci_3.h5', by_name=True, exclude=["mrcnn_class_logits", "mrcnn_bbox_fc", "mrcnn_bbox", "mrcnn_mask"])

I also get a similar problem when trying to load the weights that were generated by training my model over my own photos, to continue the training where I stopped the last time.
the error received is:

ValueError: Layer #362 (named "anchors"), weight <tf.Variable 'Variable:0' shape=(4, 261888, 4) dtype=float32> has shape (4, 261888, 4), but the saved weight has shape (2, 261888, 4).

loading the weights:

model.load_weights('new_weigths/40_epochs/mask_rcnn_cell_cfg_0040.h5', by_name=True, exclude=["mrcnn_class_logits", "mrcnn_bbox_fc", "mrcnn_bbox", "mrcnn_mask"])

please let me know if you understand why is this happening.
thanks!!

分享到QQ

分享到微博