tramac / lightweight-segmentation Goto Github PK

thanks for your excellent work!

why must the data be crop into squares(768*768), and the original aspect ratio cannot be maintained?.

When I observe the result you get based on your model, I find the best one for MobileNetV3 small is still about 7% worse. Could you let me know what the reason might be?

Hi:
Do you have plans to implement mobileNet_0.25?
Thanks

Great work!

to_onnx->onnx_to_trt
Unsupported: ONNX export of operator adaptive_avg_pool2d, since output size is not factor of input size

@Tramac
hope that the dataset part of this link will help you load some new datasets.
https://github.com/xiaoyufenfei/EDDSNet/tree/master

Thanks for your excellent work !

I print the shapes of c2 and c4 from here, [1, 112, 64, 128] and [1, 960, 64, 128] respectively.

As shown above, c2 should be 1/8 of the input, shape should be [1, 40, 128, 256]. c4 should be 1/16 of the input picture, and shape should be [1, 160, 64, 128]

In addition, layer3_setting and layer4_setting are inconsistent with those in the paper:

Is my understanding wrong? hope for help

by the way, mode == 'small' might be wrong too

I have added mscoco.py from the project that you did suggest.

I guess the problem below arises because of the incompatibility between python's version and pycocotools. Is there any suggestion to modify the code? thanks in advance.

train set
loading annotations into memory...
Done (t=0.89s)
creating index...
index created!
Traceback (most recent call last):
File "train.py", line 291, in
trainer = Trainer(args)
File "train.py", line 106, in init
args.iters_per_epoch = len(trainset) // (args.num_gpus * args.batch_size)
TypeError: object of type 'COCOSegmentation' has no len()

Hi,
I am using PyTorch version '1.4.0.dev20191118' and I am trying to train mobilenet v3 large.
I am getting this error related to the lr_scheduler:

Traceback (most recent call last):
File "C:/CP/DL_REPOSITORIES/Lightweight-Segmentation/scripts/train.py", line 291, in
trainer = Trainer(args)
File "C:/CP/DL_REPOSITORIES/Lightweight-Segmentation/scripts/train.py", line 157, in init
warmup_method=args.warmup_method)
File "C:\CP\DL_REPOSITORIES\Lightweight-Segmentation\light\utils\lr_scheduler.py", line 154, in init
super(WarmupPolyLR, self).init(optimizer, last_epoch)
File "C:\ProgramData\Anaconda3\lib\site-packages\torch\optim\lr_scheduler.py", line 77, in init
self.step()
File "C:\ProgramData\Anaconda3\lib\site-packages\torch\optim\lr_scheduler.py", line 150, in step
for param_group, lr in zip(self.optimizer.param_groups, values):
UnboundLocalError: local variable 'values' referenced before assignment

Should I downgrade my PyTorch version if I want the training to work or will you support latter PyTorch versions?

Thanks.

Could you give me some tips, thanks a lot!

I got the error when i use the model of mobilenetv3_small to train,and i used my own dataset like as VOC. Would you give me some advice?
And,would you tell me the torch version and torchvision version?

Acoording to Searching for MobileNetV3, there is a connection from feature 1/8 to the end of the segmentation head. Is it implemented? I did not find a shortcut.

I am trying to train my own dataset with two classes ( background and foreground ). I get confused of the output from seg net:

class MobileNetV3Seg(BaseModel):
    def __init__(self, nclass, aux=False, backbone='mobilenetv3_small', pretrained_base=False, **kwargs):
        super(MobileNetV3Seg, self).__init__(nclass, aux, backbone, pretrained_base, **kwargs)
        mode = backbone.split('_')[-1]
        self.head = _Head(nclass, mode, **kwargs)
        if aux:
            inter_channels = 40 if mode == 'large' else 24
            self.auxlayer = nn.Conv2d(inter_channels, nclass, 1)

    def forward(self, x):
        size = x.size()[2:]
        _, c2, _, c4 = self.base_forward(x)
        outputs = list()
        x = self.head(c4)
        x = F.interpolate(x, size, mode='bilinear', align_corners=True)
        outputs.append(x)  # Why output is a list which only append one elemenet?

        if self.aux:  # What's the aux for?
            auxout = self.auxlayer(c2)
            auxout = F.interpolate(auxout, size, mode='bilinear', align_corners=True)
            outputs.append(auxout)
        return tuple(outputs)

Furthermore I wonder the form of segmentation ground truth. My dataset comes with ground truth of dimension (Height, Width, 1) and last aixs sets 0 for background and 1 for foreground. But as my observation of your code the ground truth of cityscapes seems split into NUMCLASS ( 19 for cityscapes ) channels, something like (Height, Width, NUMCLASS ). So I wonder how to adapt the segmentation ground truth of my dataset to such kind of form.

rt, I want to eval at a CPU machine, how to do it?

there is aux=False in network,but it still have problem with log_softmax
AttributeError: 'tuple' object has no attribute 'log_softmax'.
can you help me ?thank you!

@Tramac thanks for the wonderful code base , i have few following queries

1. Can we use the architectures like efficient net , shuffle net for instance segmentation
2. How much work around will it be there if we have to use the following for instance segmentation

Hi, Tramac!

Your repo is really helpful to do some interesting experiments : )
I'm new in semantic segmentation, and I have some questions w.r.t. training of MobileNetV3+LR-ASPP.
Could you please tell how to use 768x768 resolution to train LR-ASPP framework ? (There is AvgPool (49, 49) with stride [16, 20], 768x768 input will get error).

BTW, my WeChat Id is: UsedToBe97. Maybe we can talk through it more smoothly.

    for i, (image, target) in enumerate(self.val_loader):
        image = image.to(self.device)
        target = target.to(self.device)

        with torch.no_grad():
            outputs = model(image)
        self.metric.update(outputs[0], target)
        pixAcc, mIoU = self.metric.get()
        logger.info("Sample: {:d}, validation pixAcc: {:.3f}, mIoU: {:.3f}".format(
            i + 1, pixAcc * 100, mIoU * 100))

        if self.args.save_pred:
            pred = torch.argmax(outputs[0], 1)
            pred = pred.cpu().data.numpy()

            predict = pred.squeeze(0)
            mask = get_color_pallete(predict, self.args.dataset)
            mask.save(os.path.join(outdir,  str(i) + '.png'))
    synchronize()

就这一段代码里面，原来for循环里面不是有一个filename的吗？然后他总是报错。后来我就迫不得已把他删掉， mask.save(os.path.join(outdir, str(i) + '.png'))改成用i循环命名的方法？有什么办法可以解决吗？求助了！

Can you please share it again?