search

chengjunwen / spp_net

spp net， tensorflow 实现
24 stars 13 forks source link

About the image size of Tensorflow #2

Open ideaRunner opened 6 years ago

ideaRunner commented 6 years ago

Hi, chengjunwen:

I think in tensorflow, the convolutional layer's input image size must be fixed. How you make it be capable to input different image size ?

Your reply will be really appriciated, thank you!

danielkaifeng commented 6 years ago

I agree. I think in tf.placeholder(), you must feed a fixed shape so that in later spp layer you can calculate the previous conv layer shape.

danielkaifeng commented 6 years ago

In spp_layer.py, if the size of not fixed previous conv layer will be None, so error occur when you try to devide it. self.batch_size = self.input.get_shape().as_list()[0]

WanliOuyang commented 6 years ago

@danielkaifeng How to solve this problem?

danielkaifeng commented 6 years ago

@WanliOuyang Unfortunately there is no way to solve this in tensorflow to my knowledge.

WanliOuyang commented 6 years ago

@danielkaifeng When I use a fixed batch size value(such as batch_size=4), the code can run, but when training, the precision of the training keeps fluctuating on a very small number(such as the mean accuracy is 0.16). Why is this?

tkone2018 commented 6 years ago

@danielkaifeng Do you sovle it and do you know how to use the multi_scale image inputs ?

tkone2018 commented 6 years ago

@ideaRunner Hi, do you solve the multi-scale image input size problem ?