The loss function in the original paper is composed of two parts. But you only use the output of feature fusion model to calculate the loss. And the loss they use is cross entropy, here you use is binary cross entropy. Is there any reason to these changes? Thanks!

When I first trained, the training data（MIOU） was normal. but second, miou has been fixed at around 0.12, I thought for a long time but did't solve this problem. So, I want to ask if you have encountered this problem.

load model from ./checkpoints/epoch_295.pth ... Traceback (most recent call last): File "c:\Users\Administrator\Desktop\BiSeNet-master\BiSeNet-master\demo.py", line 80, in main(params) File "c:\Users\Administrator\Desktop\BiSeNet-master\BiSeNet-master\demo.py", line 60, in main model.module.load_state_dict(torch.load(args.checkpoint_path)) File "C:\Program Files\Python\lib\site-packages\torch\nn\modules\module.py", line 518, in getattr type(self).name, name)) AttributeError: 'BiSeNet' object has no attribute 'module' PS C:\Users\Administrator\Desktop\BiSeNet-master\BiSeNet-master> Where do I put the dataset ?

when i train my data , it happend as followed:

os@os-l3:/disk3t-2/zym/BiSeNet-PyTorch$ python train.py epoch 0, lr 0.001000: 0%| | 0/4963 [00:00<?, ?it/s]Traceback (most recent call last): File "train.py", line 157, in main(params) File "train.py", line 141, in main train(args, model, optimizer, dataloader_train, dataloader_val, csv_path) File "train.py", line 56, in train output = model(data) File "/home/os/.local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 477, in call result = self.forward(*input, **kwargs) File "/home/os/.local/lib/python2.7/site-packages/torch/nn/parallel/data_parallel.py", line 121, in forward return self.module(*inputs[0], **kwargs[0]) File "/home/os/.local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 477, in call result = self.forward(*input, **kwargs) File "/disk3t-2/zym/BiSeNet-PyTorch/model/build_BiSeNet.py", line 97, in forward cx1 = self.attention_refinement_module1(cx1) File "/home/os/.local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 477, in call result = self.forward(*input, **kwargs) File "/disk3t-2/zym/BiSeNet-PyTorch/model/build_BiSeNet.py", line 40, in forward assert self.in_channels == x.size(1), 'in_channels and out_channels should all be {}'.format(x.size(1)) AssertionError: in_channels and out_channels should all be 256

Hi, I convert label from RGB mode to 'P' mode based on class_dict.csv, and found there are some points/coord with wrong label, aka its color is out of color defined in class_dict.csv. Here is how I convert RGB mode to 'P' mode and check label:

from PIL import Image
from utils import get_label_info

labelinfo = get_label_info(csvfile)
palette = []
for key in labelinfo:
    for item in labelinfo[key]:
        palette.append(item)

ref_image = Image.new(mode='P', size=(1, 1))
ref_image.putpalette(palette)

img_rgb = Image.open('rgb_mode.png')
img_p = img_rgb.quantize(palette=ref_image)

img_p_wrong = img_p[img > 31]
img_rgb_wrong = img_rgb[img > 31]

I'm not sure how to deal with this image, currently I simply remove it out of my training/val/testing. After convert RGB mode to 'P' mode, I don't need to use one hot encoding for label anymore. It seems that the one hot encoding & decoding slow down training speed.

And, the code you calculating accuracy seems weird, https://github.com/ooooverflow/BiSeNet/blob/master/utils.py#L103, pred & label are in shape of 1xCxHxW, where C is the channels, which is 3 in this case. pred[:, :, h, w] == label[:, :, h, w] means one pixel prediction right, not three.

Fisrt I met the same error with https://github.com/ooooverflow/BiSeNet/issues/1 using pytorch==0.4.0, and I met other error using pytorch==1.0.0 When I change the version into 0.4.1, it solved. But when I start training, it can work well with resnet101. When I change the build_contextpath into "resnet18", I met the same error with https://github.com/ooooverflow/BiSeNet/issues/3

i train with context_path: resnet101 and resnet 18. first question: both the validation accuracy hardly reach 0.9, Mostly stop as 0.88-0.89. second question: while the resnet101 training, the validation accuracy will fluctuate a lot after epoch 100, and will drop a lot after about epoch 180+. could you please share your training parameter like lr,batchsize,GPU num,crop_height,crop_width and some detailed trick?

RuntimeError: Error(s) in loading state_dict for BiSeNet: Missing key(s) in state_dict: "supervision1.weight", "supervision1.bias", "supervision2.weight", "supervision2.bias".

Hey, I git clone the repo, and download CamVid, change the path, and train for 300 epochs, but got poor performance, 0.182, not 94.1 or 93.2 reported in README.md. Any suggestion?

Here is the val curve.

我在训练过程中发现val过程很慢，比训练一个epoch都慢，请问有没有什么优化的方法？ 由于TypeError: can't convert CUDA tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.错误，我将train.py 中第32和38行改为了predict = predict.data.cpu().numpy()和label = label.data.cpu().numpy()，这是不是有影响？

1. support resnet18 and resnet101 both solve #3 #9
2. remove some errors on pytorch 1.0
3. support more optimizers, like sgd, adam(more smooth accuracy during training)

In 3.1 Spatial Path, there is a sentence "Based on this observation, we propose a Spatial Path to preserve the spatial size of the original input image and encode affluent spatial information. ". But in 3.1 Spatial Path, this path extracts the output feature maps that is 1/8 of the original image. What does it mean to preserve the spatial size of the original input image? Why does the paper say to preserve the spatial size of the original input image? I'm looking forward to your reply. I'm very confused about this question. Thank you very much！

class DiceLoss(nn.Module): def __init__(self): super().__init__() self.epsilon = 1e-5

def forward(self, output, target):
    # print(output.shape)
    # print(target.shape)

    assert output.size() == target.size(), "'input' and 'target' must have the same shape"
    # 在classes上做softmax
    output = F.softmax(output, dim=1)
    # 打平tensor
    output = flatten(output) # [num_classes,B*H*W]
    target = flatten(target) # [num_classes,B*H*W]
    # intersect = (output * target).sum(-1).sum() + self.epsilon
    # denominator = ((output + target).sum(-1)).sum() + self.epsilon

    intersect = (output * target).sum(-1)
    denominator = (output + target).sum(-1)
    # dice --(0-0.5)
    dice = intersect / denominator
    dice = torch.mean(dice)
    # 1-dice (0.5,1)???
    return 1 - dice
    # return 1 - 2. * intersect / denominator

double the intersection over union?

FileNotFoundError Traceback (most recent call last) in 5 #df 6 for site_name in df['SiteName'].unique(): ----> 7 df[df['SiteName'] == site_name].to_csv('{}.csv'.format(site_name))

~\AppData\Local\Continuum\anaconda3\lib\site-packages\pandas\core\generic.py in to_csv(self, path_or_buf, sep, na_rep, float_format, columns, header, index, index_label, mode, encoding, compression, quoting, quotechar, line_terminator, chunksize, tupleize_cols, date_format, doublequote, escapechar, decimal) 3018 doublequote=doublequote, 3019 escapechar=escapechar, decimal=decimal) -> 3020 formatter.save() 3021 3022 if path_or_buf is None:

~\AppData\Local\Continuum\anaconda3\lib\site-packages\pandas\io\formats\csvs.py in save(self) 155 f, handles = _get_handle(self.path_or_buf, self.mode, 156 encoding=self.encoding, --> 157 compression=self.compression) 158 close = True 159

~\AppData\Local\Continuum\anaconda3\lib\site-packages\pandas\io\common.py in _get_handle(path_or_buf, mode, encoding, compression, memory_map, is_text) 422 elif encoding: 423 # Python 3 and encoding --> 424 f = open(path_or_buf, mode, encoding=encoding, newline="") 425 elif is_text: 426 # Python 3 and no explicit encoding

FileNotFoundError: [Errno 2] No such file or directory: '/PI_Blackfriars_Sys_1_4/Room_34_SetTempHeat.csv'

can you please tell me why this is happening and how to solve this problem

Please see the function that come from utils.py: def one_hot_it_v11_dice(label, label_info): semantic_map = [] void = np.zeros(label.shape[:2]) for index, info in enumerate(label_info): color = label_info[info][:3] class_11 = label_info[info][3] if class_11 == 1: equality = np.equal(label, color) class_map = np.all(equality, axis=-1) semantic_map.append(class_map) else: equality = np.equal(label, color) class_map = np.all(equality, axis=-1) void[class_map] = 1 semantic_map.append(void) semantic_map = np.stack(semantic_map, axis=-1).astype(np.float) return semantic_map The variable "semantci_map" is a python list, but in the function ,the list only have twice append operations so that in compute loss pahse the error of output and target have different shape happend, because output shape is [batch, num class,w,h], but target shape is [batch ,2,w,h]. I can't guarantee I'm absolutely right.