Hey there, thanks for this great repo. Very easy to use and clean implementation.

This tiny PR cleans a couple things that might be helpful for others. It brings in the nice relative path imports in other notebooks that weren't in resnet_34_8s_train.ipynb. Also there was an import from adaptive_computation_time, but this module isn't needed.

If I get around to it later I might also help add some documentation on setting up VOC for use with this repo.

Cheers!

Hi @warmspringwinds I am getting - IOError: [Errno 2] No such file or directory: 'resnet_34_8s_66.pth' when running resnet_34_8s_demo.ipynb

Not able to figure out when and how the model will generate. Any clue on the next step .

Thanks Akash

Great script!

Btw, .apply() uses .children(), which may lead to skip convolutional layers in submodules. I replaced it for modules() for if you want to change it.

Hi Daniil, I'm trying to work on image segmentation of microscopy images using pytorch. I've been trying to work with your examples. But i'm having error on resnet_34_8s_train.

ImportError: No module named adaptive_computation_time

I wonder if it's something in the anaconda?

Hello Daniil, In the training process of your ResNet-8s, I notice that you crop all training images to 224x224 (RandomCropJoint(crop_size=(224, 224))). But you didn't adopt this approach when you train your FCN-32s model. Is it because the ResNet pretrained model is used as initial weights so we need to comply with its input image size (224x224) too? Do you think other input size can be used for training, without causing accuracy decline? Please advice. Thanks.

Hello, I am very impressed by your great work! However, I am a little confused when I look at your Resnet18_8s network. I assume Resnet18_8s follows your approach in your paper "Deep Residual Learning for Instrument Segmentation in Robotic Surgery", which employ dilated convolutions to keep resolution. But in resnet_dilated.py, I could not find any dilated convolutions in Class Resnet18_8s. Could you please give more detailed explanation on the structure of Resnet18_8s? Many thanks.

It says

TypeError Traceback (most recent call last) in () 34 img = Variable(img.cuda()) 35 ---> 36 fcn = resnet_dilated.Resnet34_8s(num_classes=19) 37 fcn.load_state_dict(torch.load('/home/sawyer/workspace/segmentation/resnet_34_8s_cityscapes_best.pth')) 38 fcn.cuda()

/home/sawyer/workspace/segmentation/pytorch-segmentation-detection/pytorch_segmentation_detection/models/resnet_dilated.pyc in init(self, num_classes) 293 pretrained=True, 294 output_stride=8, --> 295 remove_avg_pool_layer=True) 296 297 # Randomly initialize the 1x1 Conv scoring layer

/home/sawyer/workspace/segmentation/pytorch-segmentation-detection/vision/torchvision/models/resnet.pyc in resnet34(pretrained, **kwargs) 172 pretrained (bool): If True, returns a model pre-trained on ImageNet 173 """ --> 174 model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs) 175 if pretrained: 176 model.load_state_dict(model_zoo.load_url(model_urls['resnet34']))

TypeError: init() got an unexpected keyword argument 'fully_conv'

In both python2.7-3.5 torch version 1.0.1

Hey Daniil, thanks for your work on this repo.

I noticed that some places in the code use the output of the map function directly, which is a problem for Python3 since the function returns a map object instead of a list. I've just gone ahead and converted instances of map object to list in a bunch of places -- I noticed you did this for some files already, but it's currently breaking the cityscapes dataset, which I'm currently trying to use :-)

The Python2 docs guarantee the result of map is always a list so I assume this shouldn't break anything.

Very good repository.

How did you make this work ? it seems that vgg16 does not have the fully_conv keyword in torchvision

vgg16 = models.vgg16(pretrained=True,
                            fully_conv=True)

Hello, Can I know the optimizer and its specifications to use on unet model on Pascal Voc segmentation using FOCAL loss ? Should I have to use any learning rate schedulers? Also, is it better to take mean focal loss or sum focal loss ? I am training the model from scratch.

@warmspringwinds I am getting the following error for resnet_18_8s_59.pth

RuntimeError: Error(s) in loading state_dict for Resnet18_8s: size mismatch for resnet18_8s.fc.bias: copying a param with shape torch.Size([2]) from checkpoint, the shape in current model is torch.Size([21]). size mismatch for resnet18_8s.fc.weight: copying a param with shape torch.Size([2, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([21, 512, 1, 1]).

If I change num_classes=21 to num_classes=2, it generates the output without any segmentation (purple screen)

By using your fork of torchvision and default installation of pytorch for Linux-Python3.6-CUDA10:

1. init_weights argument in the class VGG was missing.

2. After fixing (1), the following error was generated:

In [1]: from torchvision import models                                                                               
In [2]: model = models.vgg16(pretrained=True, fully_conv=True)                                                       

RuntimeError                              Traceback (most recent call last)
<ipython-input-2-802ee77a237c> in <module>
----> 1 model = models.vgg16(pretrained=True, fully_conv=True)

~/repositories/github/pytorch-segmentation-detection/vision/torchvision/models/vgg.py in vgg16(pretrained, **kwargs)
    164     model = VGG(make_layers(cfg['D']), **kwargs)
    165     if pretrained:
--> 166         model.load_state_dict(model_zoo.load_url(model_urls['vgg16']))
    167     return model
    168 

~/.local/lib/python3.6/site-packages/torch/nn/modules/module.py in load_state_dict(self, state_dict, strict)
    767         if len(error_msgs) > 0:
    768             raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
--> 769                                self.__class__.__name__, "\n\t".join(error_msgs)))
    770 
    771     def _named_members(self, get_members_fn, prefix='', recurse=True):

RuntimeError: Error(s) in loading state_dict for VGG:
	size mismatch for classifier.0.weight: copying a param with shape torch.Size([4096, 25088]) from checkpoint, the shape in current model is torch.Size([4096, 512, 7, 7]).
	size mismatch for classifier.3.weight: copying a param with shape torch.Size([4096, 4096]) from checkpoint, the shape in current model is torch.Size([4096, 4096, 1, 1]).
	size mismatch for classifier.6.weight: copying a param with shape torch.Size([1000, 4096]) from checkpoint, the shape in current model is torch.Size([1000, 4096, 1, 1]).

Hi, I try to train model using python 3, but I got below issue:

/home/v2m/anaconda3/envs/my_env3/lib/python3.7/site-packages/torch/nn/_reduction.py:46: UserWarning: size│
_average and reduce args will be deprecated, please use reduction='sum' instead.                         │
  warnings.warn(warning.format(ret))                                                                     │
/home/v2m/anaconda3/envs/my_env3/lib/python3.7/site-packages/torch/nn/functional.py:2622: UserWarning: nn│
.functional.upsample_bilinear is deprecated. Use nn.functional.interpolate instead.                      │
  warnings.warn("nn.functional.upsample_bilinear is deprecated. Use nn.functional.interpolate instead.") │
0.4247354666311015                                                                                       │
0.5617590819998219                                                                                       │
0.5815541637890524                                                                                       │
0.6344758887881029                                                                                       │
Traceback (most recent call last):                                                                       │
  File "pytorch_segmentation_detection/recipes/pascal_voc/segmentation/psp_resnet_50_8s_train.py", line 3│
76, in <module>                                                                                          │
    optimizer.step()                                                                                     │
  File "/home/v2m/anaconda3/envs/my_env3/lib/python3.7/site-packages/torch/optim/adam.py", line 107, in s│
tep                                                                                                      │
    p.data.addcdiv_(-step_size, exp_avg, denom)                                                          │
RuntimeError: value cannot be converted to type float without overflow: (3.52033e-08,-1.14383e-08)

Can someone give me suggestion?

I'm new to implementing CNNs and I'm trying to understand how a model knows whether to perform semantic classification (pixelwise) or image classification (one class per image). As far as I can see, the only difference is in the models/resnet_dilated.py file in the lines resnet34_8s.fc = nn.Conv2d(resnet34_8s.inplanes, num_classes, 1)

whereas most other codes have it as resnet34_8s.fc = nn.Conv2d(resnet34_8s.fc.in_features, num_classes)

Is this the difference between returning a logits of shape [batch x num_classes x H x W] and [batch x num_classes]?