hi @bfortuner , I' m interested in your pytorch_tiramisu.

I use the train.ipynb as the backbone to train the model. However, I cannot approach the performance as you do.

My FC-DenseNet67 result converge at 0.8357994871794872.

It will be appreciate to help me figure out my questions.
Here are my questions:

1. The origin code doesn't have weighted loss function. The weighted loss function do improve
   iou, but harm to pixel accuracy. How did you get those weight?

2. I see some past issue mention that the train.ipynb is slightly different from the original train
   code. Do you add a lot of things on it? Can you send me the original code for further
   comparison ?

Here is the modification from the code :

in train.ipynb

  1. N_EPOCHS -> 1000

  2. To use multi-gpu I add this line

  3. To implement "exclude the 'background' class" , I add this lines. Try to set 11 class, and 
      ignore void label loss

   4.  I train it directly. I didn't not finetune.

Please help me figure out these questions.🙏🙏🙏🙏🙇🙇🙇
My email is : [email protected]

Hello @bfortuner ,

Is it possible to share the weights you obtained with your training ? Everyone will be infinitely grateful :)

Thank you

hi @bfortuner ,
Is there any details for how to train a model?
I tried but there is following error ,

python train_densenet_pytorch.py
Traceback (most recent call last):
  File "train_densenet_pytorch.py", line 25, in <module>
    import utils.training as train_utils
  File "/pytorch_tiramisu-master/utils/training.py", line 47, in <module>
    def train(epoch, net, trainLoader, optimizer, trainF, sessionName=get_rand_str(5)):
  File "/pytorch_tiramisu-master/utils/training.py", line 21, in get_rand_str
    return ''.join(random.choices(string.ascii_uppercase + string.digits, k=n))
AttributeError: 'module' object has no attribute 'choices'

Hope for your help.
Thanks a ton!

thank you !

Hi, there comes a silly question to me.
I use this repo to train my dataset, which is a 2 classes dataset on Semantic Segmentation.
There is everything ok during Epoch1(I use the Epoch1.pth to predict pictures, and the result is ok. The Segmentation performs good at some place.)
And things go strange at Epoch2. The Loss decreases quickly to 0 at about in 100 steps.
I'm confused of it. Maybe because the num of classes of my dataset is too small? How can I modify it?
Thank you in advance.

Hi, @bfortuner ,

I found a strange term random.choices in this line. There is no random.choices (link) . Is it a typo or have special meaning?

First of all, congratulations on the good work. I got a size error in size of my data set (512,512.3). How do I fix this?
What changes do I have to make?

Hello!
I'm new for neural network and pytorch!
When I run this code on my own dataset , I have confusion about test() in train.py

1. I use model.load_state_dict( torch.load('weights-100-0.018-0.065.pth')) to load parameters before train_utils.test(model, test_loader, criterion, epoch=1)
   but i have error like this:
   KeyError: 'unexpected key "startEpoch" in state_dict'
2. How to use latest.th?

Thank you for your help!

Hello, first of all, thank you very much for your code, let me benefit a lot.
Second, I want to know how to modify the code to remove empty classes from the dataset. I post a slight change but always report error:
RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED.
But restore works again, and I suspect this isn't a cuda/cudnn issue, so I'm eager to find out.
Looking forward to your early reply!

why the model need the 'background' class? the paper author's code is 11 class

Hi, in the original paper, the parameters for FC-DenseNet56 is about 1.5 M
But your code requires about 4000MB gpu to do a forward computation.

I wonder why there is a huge difference between these.

Thank you so much @bfortuner for the implementation of this model in PyTorch. When I try to train the Network with the mentioned CamVid dataset, my training always starts with the Train Acc and Val Acc of 1.000. I have set the hyper-parameters as mentioned in the requirements section.

In other words training error and validation error is always 0. Wondering if am missing something ??

It has been long time since this code was updated so all libraries have upgraded. If you can provide requirements.txt file with dependencies, it will be very helpful.
Thanks

I tried batch_size=1, but no difference. I haven't attempted to calculate the actual memory requirement, but would you expect this to run on a 6GB GPU?

Although this repository claims to implement a paper about semantic segmentation, the code in this repository only does image classification based on the CIFAR dataset. Can this code actually be used for pixel-level semantic segmentation?

Hi;

After training a tiramisu, I would like to just use all the layers till the bottleneck layer for feature extraction. Is there a good and recommended way to do it. ?

Hi, can you share the weights of the network after training?
They would be useful for running simple inference or plugging the trained network into other systems.
I saw that in the original project there are the trained weights, but I think it's hard to convert them from Theano to PyTorch.

In update_viz_summary_plot fuction:

`
def update_viz_summary_plot(self):

    trn_loss = self.loss_history['train'][-1]

    val_loss = self.loss_history['val'][-1]

    trn_err = self.error_history['train'][-1]    

    val_err = self.error_history['val'][-1]

    txt = ("""Epoch: %d

        Train - Loss: %.3f Err: %.3f

        Test - Loss: %.3f Err: %.3f""" % (self.epoch, 

        trn_loss, trn_err, tst_loss, tst_err))

    window = self.visdom_plots['summary']

    return viz.text(

        txt,

        win=window,

        env=self.name

    )

`
there are no tst_loss and tst_err variables.

So I think they should be replaced by val_loss and val_err respectively, right?

did you training the whole network/weights in fine tuning or just some selected layers ?

I want to use Cityscape Dataset instead of CamVid ,How can it be done ?

What kind of results do you get? It would be really cool if you could post some learning curves, or even just the final mean IoU and accuracy on the camvid dataset for reference? :-)

The line 60 in your code, tiramisu.py, : cur_channels_count += prev_block_channels
actually does nothing, since the cur_channels_count is overwritten before and after. Please consider removing it, since it is confusing and misleading.

        #######################
       #   Upsampling path   #
       #######################

       self.transUpBlocks = nn.ModuleList([])
       self.denseBlocksUp = nn.ModuleList([])
       for i in range(len(up_blocks)-1):
           self.transUpBlocks.append(TransitionUp(prev_block_channels, prev_block_channels))
           cur_channels_count = prev_block_channels + skip_connection_channel_counts[i]

           self.denseBlocksUp.append(DenseBlock(
               cur_channels_count, growth_rate, up_blocks[i],
                   upsample=True))
           prev_block_channels = growth_rate*up_blocks[i]
           cur_channels_count += prev_block_channels

Hi @bfortuner ,

First of all thanks for the excellent implementation of the FCDenseNets.

I am trying to use your tiramisu implemetation for a different dataset and could really use your help. Particularly I need insight into how this is working

class LabelToLongTensor(object):
def __call__(self, pic):
    if isinstance(pic, np.ndarray):
        # handle numpy array
        label = torch.from_numpy(pic).long()
    else:
        label = torch.ByteTensor(torch.ByteStorage.from_buffer(pic.tobytes()))
        label = label.view(pic.size[1], pic.size[0], 1)
        label = label.transpose(0, 1).transpose(0, 2).squeeze().contiguous().long()
    return label

This is making a 1x224x224 label tensor for a label image of size 224x224x3. Now I am unable to adapt this for my dataset. I have 7 classes and each label image is 224x224x3. Should my label tensor be 1x224x224 with each value between 0-6 or 1-7 ? The nll_loss2d expects the output to be 7224224 if I am correct.

In your code you do DenseNet with the sum operation, which is the operation in paper ResNet. And DenseNet is the concatenation operation. This is probably why you have more code parameters than FC-DenseNet itself.