Camvid数据集训练问题

[Li-Jiren]

您好，我在借用您的unet网络跑Camvid数据集的时候遇到了下面的问题，希望您可以指导一下：

`Total params: 31,051,200
Trainable params: 31,051,200
Non-trainable params: 0

---

Epoch 1/10
Found 100 images belonging to 1 classes.
Found 400 images belonging to 1 classes.
Found 100 images belonging to 1 classes.
Found 400 images belonging to 1 classes.
3/100 [========>.....................] - ETA: 3:58 - loss: 0.2298 - acc: 0.2382`

似乎所有图片都被判定属于同一个类别（belonging to 1 classes）？训练中acc确实会逐渐上升，最终达到1，但是使用test.py测试训练结果模型时得到的图片是全灰的（将data.py中COLOR_DICT的第一类颜色改为255，255，255的话会得到全白图像，感觉应该是把整个图像都分为第一类了）

看了之前几个issue，不知道是不是我的训练集图片路径设置错误，我是按如下路径设置的：
`
#path to images which are prepared to train a model

train_path = "E:/ljwgroupSRTP/unet/data/train"

image_folder = "image"

label_folder = "label"

valid_path =  "E:/ljwgroupSRTP/unet/data/valid"

valid_image_folder ="image"

valid_label_folder = "label"

log_filepath = 'E:/ljwgroupSRTP/unet/log'`

训练用的py文件都在E:/ljwgroupSRTP/unet文件夹下，而我的图片路径如下图：


不知这样的情况下我的路径设置是否正确？

如果不是图片路径的设置问题，我修改的参数是以下这些代码，希望您能指正一下错误：
train.py:
`
flag_multi_class = True

num_classes = 32

dp = data_preprocess(train_path=train_path,image_folder=image_folder,label_folder=label_folder,

                     valid_path=valid_path,valid_image_folder=valid_image_folder,valid_label_folder=valid_label_folder,

                     flag_multi_class=flag_multi_class,

                     num_classes=num_classes)

data.py中我在原本的12中颜色基础上，自己编了20种颜色并加入COLOR_DICT中（我猜测这个应该只是在test预测中用于染色的，并不一定要与label图像颜色相同？）
class data_preprocess:

def __init__(self, train_path=None, image_folder=None, label_folder=None,

             valid_path=None,valid_image_folder =None,valid_label_folder = None,

             test_path=None, save_path=None,

             img_rows=512, img_cols=512,

             flag_multi_class=True,

             num_classes = 32):

    self.img_rows = img_rows

    self.img_cols = img_cols

    self.train_path = train_path

    self.image_folder = image_folder

    self.label_folder = label_folder

    self.valid_path = valid_path

    self.valid_image_folder = valid_image_folder

    self.valid_label_folder = valid_label_folder

    self.test_path = test_path

    self.save_path = save_path

    self.data_gen_args = dict(rotation_range=0.2,

                              width_shift_range=0.05,

                              height_shift_range=0.05,

                              shear_range=0.05,

                              zoom_range=0.05,

                              vertical_flip=True,

                              horizontal_flip=True,

                              fill_mode='nearest')

    self.image_color_mode = "rgb"

    self.label_color_mode = "rgb"



    self.flag_multi_class = flag_multi_class

    self.num_class = num_classes

    self.target_size = (512, 512)

    self.img_type = 'png'

test.py中：
test_path = "E:/ljwgroupSRTP/unet/data/test/image"

# save the predict images

save_path = "E:/ljwgroupSRTP/unet/data/test/predict"



dp = data_preprocess(test_path=test_path,save_path=save_path,flag_multi_class=True,num_classes=32)`

请问我这几个文件中是不是有改错或者漏改的内容？或者是data_Pretreatment.py和label_visualization.py中也有需要修改的内容？

非常感谢您的指导！

[Li-Jiren]

已经解决了，不好意思

[wuyang0329]

已经解决了，不好意思

请问你是怎么解决，分享一下呗

[v-smwang]

mark

[lyc1995452-star]

您好，我在借用您的unet网络跑Camvid数据集的时候遇到了下面的问题，希望您可以指导一下：

`Total params: 31,051,200
Trainable params: 31,051,200
Non-trainable params: 0

Epoch 1/10
Found 100 images belonging to 1 classes.
Found 400 images belonging to 1 classes.
Found 100 images belonging to 1 classes.
Found 400 images belonging to 1 classes.
3/100 [========>.....................] - ETA: 3:58 - loss: 0.2298 - acc: 0.2382`

似乎所有图片都被判定属于同一个类别（belonging to 1 classes）？训练中acc确实会逐渐上升，最终达到1，但是使用test.py测试训练结果模型时得到的图片是全灰的（将data.py中COLOR_DICT的第一类颜色改为255，255，255的话会得到全白图像，感觉应该是把整个图像都分为第一类了）

看了之前几个issue，不知道是不是我的训练集图片路径设置错误，我是按如下路径设置的：
`
#path to images which are prepared to train a model

train_path = "E:/ljwgroupSRTP/unet/data/train"

image_folder = "image"

label_folder = "label"

valid_path =  "E:/ljwgroupSRTP/unet/data/valid"

valid_image_folder ="image"

valid_label_folder = "label"

log_filepath = 'E:/ljwgroupSRTP/unet/log'`

训练用的py文件都在E:/ljwgroupSRTP/unet文件夹下，而我的图片路径如下图：


不知这样的情况下我的路径设置是否正确？

如果不是图片路径的设置问题，我修改的参数是以下这些代码，希望您能指正一下错误：
train.py:
`
flag_multi_class = True

num_classes = 32

dp = data_preprocess(train_path=train_path,image_folder=image_folder,label_folder=label_folder,

                     valid_path=valid_path,valid_image_folder=valid_image_folder,valid_label_folder=valid_label_folder,

                     flag_multi_class=flag_multi_class,

                     num_classes=num_classes)

data.py中我在原本的12中颜色基础上，自己编了20种颜色并加入COLOR_DICT中（我猜测这个应该只是在test预测中用于染色的，并不一定要与label图像颜色相同？）
class data_preprocess:

def __init__(self, train_path=None, image_folder=None, label_folder=None,

             valid_path=None,valid_image_folder =None,valid_label_folder = None,

             test_path=None, save_path=None,

             img_rows=512, img_cols=512,

             flag_multi_class=True,

             num_classes = 32):

    self.img_rows = img_rows

    self.img_cols = img_cols

    self.train_path = train_path

    self.image_folder = image_folder

    self.label_folder = label_folder

    self.valid_path = valid_path

    self.valid_image_folder = valid_image_folder

    self.valid_label_folder = valid_label_folder

    self.test_path = test_path

    self.save_path = save_path

    self.data_gen_args = dict(rotation_range=0.2,

                              width_shift_range=0.05,

                              height_shift_range=0.05,

                              shear_range=0.05,

                              zoom_range=0.05,

                              vertical_flip=True,

                              horizontal_flip=True,

                              fill_mode='nearest')

    self.image_color_mode = "rgb"

    self.label_color_mode = "rgb"



    self.flag_multi_class = flag_multi_class

    self.num_class = num_classes

    self.target_size = (512, 512)

    self.img_type = 'png'

test.py中：
test_path = "E:/ljwgroupSRTP/unet/data/test/image"

# save the predict images

save_path = "E:/ljwgroupSRTP/unet/data/test/predict"



dp = data_preprocess(test_path=test_path,save_path=save_path,flag_multi_class=True,num_classes=32)`

请问我这几个文件中是不是有改错或者漏改的内容？或者是data_Pretreatment.py和label_visualization.py中也有需要修改的内容？

非常感谢您的指导！

请问你是怎么解决的，请指导一下，谢谢

[lyc1995452-star]

mark

请问你知道如何解决吗？

[Li-Jiren]

我是label图片用错了，应该用灰度图而不是这种彩色的可视化label，换成灰度图就解决了

[lyc1995452-star]

我是label图片用错了，应该用灰度图而不是这种彩色的可视化label，换成灰度图就解决了

Found 367 images belonging to 1 classes.
Found 101 images belonging to 1 classes.
Found 367 images belonging to 1 classes.
Found 101 images belonging to 1 classes.
我用的Camvid运行出现这个，你当时是怎么处理的呢？它的label是灰度图啊

[lyricgoal]

我是label图片用错了，应该用灰度图而不是这种彩色的可视化label，换成灰度图就解决了

请问label用灰度图的话data.py中label_color_mode要设置成grayscale吗？还是就保持rgb呢？

[wuyang0329]

需要改一下 在 2020-12-10 19:57:58，"lyricgoal" <[email protected]> 写道： 我是label图片用错了，应该用灰度图而不是这种彩色的可视化label，换成灰度图就解决了 请问label用灰度图的话data.py中label_color_mode要设置成grayscale吗？还是就保持rgb呢？ — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[RyanCCC]

我是label图片用错了，应该用灰度图而不是这种彩色的可视化label，换成灰度图就解决了

请问label用灰度图的话data.py中label_color_mode要设置成grayscale吗？还是就保持rgb呢？

您好！请问这个问题怎么解决，谢谢！

[RyanCCC]

我是label图片用错了，应该用灰度图而不是这种彩色的可视化label，换成灰度图就解决了

您好，能否指导一下怎么解决这个问题，非常感谢！