Thanks for sharing the great work.

I've some question regarding the self.crf(x, logits) usage from your tutorial code.

Does x is a normalized image tensors? Or should it be in a range of 0-255?

Also, is the pre-trained CRF available by any chance? Thanks a lot.

import torch

class CNNCRF(torch.nn.Module): """ Simple CNN-CRF model """ def init(self, cnn, crf): super().init() self.cnn = cnn self.crf = crf

def forward(self, x):
    """
    x is a batch of input images
    """
    logits = self.cnn(x)
    logits = self.crf(x, logits)
    return logits

Create a CNN-CRF model from given cnn and crf

This is a PyTorch module that can be used in a usual way

model = CNNCRF(cnn, crf)

First I train unET and save the model, then I load the trained UNET model and train UNET and CRFS. I found the loss stuck at 0.693147. Do you have any suggestions?

Hi, thanks for the great work.

I tried to install the pack via the python3 setup.py install, however as I import CRF I got no module name 'CRF'.

Do you have any insight regarding this? Thanks.

Here's what CRF-0.0.01-py3.6-linux-x86_64.egg directory has in my machine:

Hello, thank you for this amazing work!

I want to use this CRF package for distributed training. I wonder if it is OK to use it to train a segmentation model across many GPUs or processes. For example, does it contain any operations like batch normalization that needs to perform computations across batches?

import torch

class CNNCRF(torch.nn.Module): """ Simple CNN-CRF model """ def init(self, cnn, crf): super().init() self.cnn = cnn self.crf = crf

def forward(self, x):
    """
    x is a batch of input images
    """
    logits = self.cnn(x)
    logits = self.crf(x, logits)
    return logits

Create a CNN-CRF model from given cnn and crf

This is a PyTorch module that can be used in a usual way

model = CNNCRF(cnn, crf)

According to your usage, if my image input size is (1,3,512,512), the label shape is (1,2,512,512) where 2 is the category containing the background. The output logits shape after CNN processing is also (1,2,512,512). At this point, (1,3,512,512) and (1,2,512,512) will be input self.crf. I wonder if I can.

Hi, First of all, thanks for the impressive work.

We are experimenting with a novel post-processing method (aiming at CVPR) and would like to benchmark ours with Frank-Wolfe dense CRFs (FWCRF). To that end, we are not sure if we can use FWCRF directly as is (following chapter 5.2) for inference on Pascal VOC (as defined below, with adjusted alpha, beta, gamma according to section E.1) or whether we need to train it first on the training data of the respective dataset.

fw_params = frankwolfecrf.FrankWolfeParams(scheme='fixed',  # constant stepsize
                                             stepsize=1.0,
                                             regularizer='l2',
                                             lambda_=1.0,  # regularization weight
                                             lambda_learnable=False,
                                             x0_weight=0,  
                                             x0_weight_learnable=False)

  fw_crf = frankwolfecrf.DenseGaussianCRF(classes=21,
                                          alpha=80,
                                          beta=13,
                                          gamma=3,
                                          spatial_weight=1.0,
                                          bilateral_weight=1.0,
                                          compatibility=1.0,
                                          init='potts',
                                          solver='fw',
                                          iterations=5,
                                          params=fw_params)
  prediction_fwcrf = fw_crf(images, base_model_logits)

Thanks much in advance for getting back.

Best wishes, Lukas