0bserver07 / one-hundred-layers-tiramisu Goto Github PK

I'm not sure if I'm reading your code right but is it perhaps missing the actual skip connections?

https://github.com/0bserver07/One-Hundred-Layers-Tiramisu/blob/master/model-tiramasu-103.py

I'd expect to see a line similar to the following:
https://github.com/farizrahman4u/keras-contrib/blob/master/keras_contrib/applications/densenet.py#L614

Hi,thanks for sharing your work, this helps me a lot.

When I scan your code, I have some confusion about your implement of denseblock.It is not implemented as a sensenet style.

The images are changed during this reshape.

hi @0bserver07 ,

When i use python camvid_data_loader.py, it showed the following error,

Traceback (most recent call last):
  File "camvid_data_loader.py", line 34, in <module>
    train_data, train_label = load_data("train")
  File "camvid_data_loader.py", line 27, in load_data
    data.append(np.rollaxis(normalized(cv2.imread(os.getcwd() + txt[i][0][7:])[136:,256:]),2))
TypeError: 'NoneType' object has no attribute '__getitem__'

I have installed all the libs in your requirement.txt .
Do you know what's the reason of that error and how can i deal with it ?

Thanks a ton!

layer_bottleneck = self.DenseBlock(15,408, layer_5a_down, enc_model_layers, 'layer_bottleneck') # m = 348 + 5*12 = 408

here DenseBlock(15,408... is 15 or 5 ?

when we count m = 348 + 5*12 = 408 use 5 , but , the parameter 15 should be used ?

Hi,
Thanks a lot for this easy to understand implementation. I wish to know if you'd predicted the output for test set and have any results for it?

current hacky fix:

import sys;
sys.setrecursionlimit(40000)

I want to train Tiramisu on my own custom dataset. Can I do it with different size images or I can only use images with size like the Camvid dataset?

I use the code to see the number of parameters for the model:
def parameter_num(model_structure):
with open(model_structure) as model_file:
model = models.model_from_json(model_file.read())
print(model.summary())
but the results are more larger than models in Simon Jegou et al 's paper:
4
Total params: 33,927,266
Trainable params: 33,902,114
Non-trainable params: 25,152

5
Total params: 59,332,562
Trainable params: 59,295,842
Non-trainable params: 36,720

6
Total params: 95,044,274
Trainable params: 94,993,778
Non-trainable params: 50,496

Hi,

How to fine-tune the model for a custom dataset with different number of classes?

Typically I should handle the pretrained model (I could not find the download link) and train the model with different number of classes on the custom dataset, but with the same format as original one.

Hi,
Maybe I'm not reading your code right, but I don't see a concatenation step after each convolution layer in the denseblock() function. I thought this was needed as shown in the paper.

THanks,

Thank you so much for you sharing.It really helps me a lot.Here is Another question. If I have two classes,how should I set the class weighting? Looking forward to your answer.
Thank you so much .

Hi,
You have done an awesome work. But i have a question to ask you. Have you reproduced the same result reported in the original paper, Mean_IoU=66.9%,acc=91.5 ?
If you have reproduce the same result, please tell me what kind of trick you used in your network?

Thank you in advance

run this:

sudo ln /dev/null /dev/raw1394

Hi,

I have installed exactly as described in the instructions but the training always starts in CPU mode. I'm using tensorflow backend. Why do we need to install both tensorflow-gpu as well as tensorflow?

Hello

Nice github repository.
I am trying your 103-layer tiramisu code. It seems to me that there is no skip-connections, or maybe I have misunderstood something.
Furthermore. When I train the model in keras the model summary displays 428,328,755 model parameters.
As far as I am informed the original tiramisu 103 layer uses 9.4 million paramers. Maybe I am misunderstanding something. I hope you can clarify it for me.

best regards

Might you be interested in a pull request of this code to the official keras-contrib repository which is the upstream source for Keras, and has a DensenetFCN implementation?

These keras-contrib issues are also relevant to this repository:

• keras-team/keras-contrib#47 incorporating coco + voc 2012
• keras-team/keras-contrib#59 dealing with gaps in FCN training performance
• keras-team/keras-contrib#63 training on other FCN networks

Keras-FCN, which I was planning to adapt for a merge into keras-contrib also has a SegDataGenerator implementation with some of the features you are looking for in your comments, plus additional models and experimental support for coco.

I figured it might be worth collaborating because it appears we are working on the same thing (training DenseNetFCN), and both running into the same accuracy limitations even with independent implementations.

Could you consider applying the MIT license to this repository, the same as keras, so people may make use of it?

about the MIT license and here is the text:

The MIT License (MIT)

Copyright (c) <year> <copyright holders>

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.