no index, assign to void , I think there is a mistake that set idx_mat[h, w]=19.

Because trainID have 19, we can set it as 20 or others

In the torchvision,the vgg16.features has 43 layers,and the maxpool output should 6,13,23,33,43?I am not understand why do you use 5,10..?Thanks!

hello author,

i have some queries regarding your implementation.

1. why did you use the deconv operation 5 times in fcn32, slowly upsampling gives better resultsd than a single conv2d operation?
2. also, can you please give me some reference how to study about using pretrained weights for custom model. your usage of pretrained weights is good. I want to learn. Can you please give me some inputs. I built a simple fcn model and used pretrained models as follows:

model = fcn()
model.load_state_dict(model_zoo.load_url(model_urls['resnet50']), strict=False)

it gives better result, but i am not sure how it uses the weights. whether based on layer names or automatic concat weights based on kernel size.

thank you

Hello, I finished your tutorial, so I had a trained model. Thank you!

However, I do not know how to use this trained model.

Can you help me?

i can not find the file:label_colors.txt,can you share me?

Hi, I implement the code, but for the fcn.py on line 145. It says:
"The size of tensor a (44) must match the size of tensor b (45) at non-singleton dimension 2"

I found the score shape is [1, 512, 44, 60] and the x4 shape is [1, 512, 45, 60]

Curious how would that happen?

Thanks

i want to use this FCN to recognize my own picture,how can i do it?thanks a lot

I found the FCNs network structure(which has 2 more skip connections) is not included in the original FCN paper, but I do found some performance improvement by using it. I wonder is there any reference for this structure or it is just created by you?

Hello,
i tried to train the fcn on my own computer, and i set the batch to 2. Everything went well but i got this traceback error.

Traceback (most recent call last):
File "train.py", line 171, in
train()
File "train.py", line 101, in train
loss = criterion(outputs, labels)
File "C:\Users\83750\Anaconda3\lib\site-packages\torch\nn\modules\module.py", line 489, in call
result = self.forward(*input, **kwargs)
File "C:\Users\83750\Anaconda3\lib\site-packages\torch\nn\modules\loss.py", line 595, in forward
reduction=self.reduction)
File "C:\Users\83750\Anaconda3\lib\site-packages\torch\nn\functional.py", line 2075, in binary_cross_entropy_with_logits
raise ValueError("Target size ({}) must be the same as input size ({})".format(target.size(), input.size()))
ValueError: Target size (torch.Size([2, 32, 480, 640])) must be the same as input size (torch.Size([2, 20, 480, 640]))

how could this happen? how can i solve it? i'll be appreciated if you could help me.

Hello, I'm running Cityscapes_utils.py, I found that the running speed is a little slow. Is there any way to speed up the running speed? Thank you.

Have anybody tested on pytorch 0.4.1? I need some advice, thank you.

"nn.BCEWithLogitsLoss()" will this cause overlap between different classes?

Hello,thanks for your code.
But there may be some question in deconvolution weight initialize.In original paper,it initialize the deconvolution weight with bilinear interpoltion. And I can't find this in your code.

create a directory named "CamVid", and put data into it
the words above is a vague guidance for me , when i click in the camvid website, only to find all kinds of data provided, video ,image ...
can you be so kind to give some details

I think multi class can not using BCELoss here, why are u using that?

Hi
Do you have any pretrained models that you can share?
I am working on adversarial generation for semantic segmentation and a trained model will really speed things up for me.

Just confused in CamVid.py, it has the following code

h, w      = 720, 960
train_h   = int(h * 2 / 3)  # 480
train_w   = int(w * 2 / 3)  # 640
val_h     = int(h/32) * 32  # 704
val_w     = w               # 960

Why the training height and validation height has a different size and both of them are different from the original image size? What's the magic to make the network output image has a pixel to pixel corresponding to the 720*960 image?

Can you also mention about mIoU? how much did you acheive?

hello can you please tell me how we show the image with mask of the predicted data
i mean how you show these results ?

hello author,

can you please tell me your validation loss or iou. I implemented original fcn model and also tried yours. The training scores are good. But the validation iou score is less 0.1. I tried shuffling data and everything. Why this happens?
please please give me your advice.

Thank you
Soans

Hi,
why does CamVid_utils generate the following error on some files such as 0016E5_08460_L.png

Traceback (most recent call last):
File "python/CamVid_utils.py", line 141, in
parse_label()
File "python/CamVid_utils.py", line 122, in parse_label
assert color2label[tuple(color)] == test_ans[idx]
AssertionError

Hi. I'm demonstrating your code in docker my Mac. But I have this error. How can I solve this problem? I was running it sequentially with your README! Thanks!!

I am training the net and i wangt to know the final BCEloss value you guys got around,thanks!

i train about 100 epochs,and the miou is abot 25%,
i use CamVid and set the num_class to 32,is this the reason ?

can you tell me how to inference the mask with your network？i wrote one，but it‘s result is not good.

why you use the batch normalization after the Relu opration instead of the batch normalization before the Relu opration