hi ! this is an amazing work ! I just wanna use this project to do some researches but my PC too slow. Can you support a trained model for this ? thanks.

When i try to run CUDA_VISIBLE_DEVICES=0 python decode_autodeeplab.py --dataset cityscapes --resume /AutoDeeplabpath/checkpoint.pth.tar There is an error: I can't find betas8, betas16.... Could you help me? Thanks a lot~ @NoamRosenberg @zhizhangxian

Before saying the problems, I show my running environment: pytorch version 1.0.1.post2 cuda version 8.0.61 cudnn version: 7102 python version: 3.7

I run the latest code, two problems bother me: 1.Traceback (most recent call last): File "train_autodeeplab.py", line 20, in import apex File "/home/zhangjw/anaconda3/lib/python3.7/site-packages/apex/init.py", line 18, in from apex.interfaces import (ApexImplementation, File "/home/zhangjw/anaconda3/lib/python3.7/site-packages/apex/interfaces.py", line 10, in class ApexImplementation(object): File "/home/zhangjw/anaconda3/lib/python3.7/site-packages/apex/interfaces.py", line 14, in ApexImplementation implements(IApex) File "/home/zhangjw/anaconda3/lib/python3.7/site-packages/zope/interface/declarations.py", line 483, in implements raise TypeError(_ADVICE_ERROR % 'implementer') TypeError: Class advice impossible in Python3. Use the @implementer class decorator instead.

Then I comment related code and set the APEX_AVAILABLE=False. The code continues to run， but the problems 2 occurs

2. Traceback (most recent call last): File "train_autodeeplab.py", line 421, in main() File "train_autodeeplab.py", line 414, in main trainer.training(epoch) File "train_autodeeplab.py", line 176, in training output = self.model(image) File "/home/zhangjw/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 489, in call result = self.forward(*input, **kwargs) File "/home/zhangjw/AutoML/auto_deeplab.py", line 282, in forward normalized_alphas) File "/home/zhangjw/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 489, in call result = self.forward(*input, **kwargs) File "/home/zhangjw/AutoML/cell_level_search.py", line 138, in forward s = sum(new_states) RuntimeError: The size of tensor a (13) must match the size of tensor b (14) at non-singleton dimension 3 It's so weird, so I print the new_states[0].shape, as following: torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 16, 14, 14]) torch.Size([2, 32, 7, 7]) torch.Size([2, 32, 7, 7]) torch.Size([2, 32, 7, 7]) torch.Size([2, 32, 7, 7]) torch.Size([2, 32, 7, 7]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 4, 56, 56]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 8, 28, 28]) torch.Size([2, 16, 13, 13])

when the last line torch.Size([2, 16, 13, 13]) comes, RuntimeError occurs.It's so weird. Have you encountered this situation? I would appreciate it if you tell me how to solve it.

Hi I'm running with the following command line: CUDA_VISIBLE_DEVICES=0,1 python train_autodeeplab.py --dataset coco --filter_multiplier 4 --resize 358 --crop_size 224 --batch-size 2 This gives me: RuntimeError: Given groups=1, weight of size 4 125 1 1, expected input[1, 128, 56, 56] to have 125 channels, but got 128 channels instead

I understand what the error about, but I can't really fix it, why I'm getting this error?

Thanks.

Hi, Thanks for your implementation! I have some confusion in the implementation of the decoded network. I knew that each cell has two inputs from prev_prev_cell and prev_cell. But what the case in the decoded network. I mean that: If layer_i has downsampling rate 16, but layer_{i-1} has downsampling rate 4. In this case, prev_prev_c is None and only has prev_c's output as input, right? (from the paper, I saw that prev_prev_c's output should have the same scale as the current layer). Could you give me some hints? Thanks in advance.

I'm tyring to run using this command line: CUDA_VISIBLE_DEVICES=0 python3 train_autodeeplab.py --dataset coco --filter_multiplier 4 --resize 358 --crop_size 224 --batch-size 2 This is after your post on how to solve memory issues, I'm still getting memory error, apparently, it's not enough for me.

After adding another GPU means running this command line: CUDA_VISIBLE_DEVICES=0,1 python3 train_autodeeplab.py --dataset coco --filter_multiplier 4 --resize 358 --crop_size 224 --batch-size 2 I'm getting this error: File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 547, in call result = self.forward(*input, **kwargs) File "/workdisk/AutoML/auto_deeplab.py", line 348, in forward aspp_result_4 = self.aspp_4 (self.level_4[-1]) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 547, in call result = self.forward(*input, **kwargs) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/container.py", line 92, in forward input = module(input) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 547, in call result = self.forward(*input, **kwargs) File "/workdisk/AutoML/operations.py", line 164, in forward conv_image_pool = self.conv_p(image_pool) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 547, in call result = self.forward(*input, **kwargs) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/container.py", line 92, in forward input = module(input) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 547, in call result = self.forward(*input, **kwargs) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/batchnorm.py", line 81, in forward exponential_average_factor, self.eps) File "/usr/local/lib/python3.6/dist-packages/torch/nn/functional.py", line 1652, in batch_norm raise ValueError('Expected more than 1 value per channel when training, got input size {}'.format(size)) ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 20, 1, 1])

Any Idea? Thanks!

viterbi 实现有问题吧

if (sample == 0 and rate == 2) or (sample == 3 and rate == 0):
      continue

第layer层，第0个细胞，不能进行的是0(上采样) 第layer层，第3个细胞，不能进行的是2(下采样) 应该是修改成这样的 if (sample == 0 and rate == 0) or (sample == 3 and rate == 2): continue

但是我修改成这样后，直接报错了。

https://github.com/NoamRosenberg/AutoML/issues/71 这里提到的sample == 3 rate == 0 denotes the information flow will come from the sample==4 in previous layer 第layer层，第3个细胞，不能由 第4层的细胞进行上采样 得到；

这里数据流向 和我不一样，到底哪个是对的呢

Hi,

Thanks for your works! But I have some question about Viterbi Algorithm in decoding_formulas.py

I think the shape of self.network_space is [layers, scales, choices], that is layers, samples, and rates in the for loops, right? If so, should be code in line 50 is

if (sample == 0 and  rate == 0) or (sample == 3 and rate == 2):
    continue

because the corresponding choices of rate 0, 1, 2 are ↗, →, and ↘. If I am wrong, could you give me some hints and further explanations? Thanks in advance.

Best regards, TTB

When trying to load data parallel model I get an error which I can't see because the command line gets flooded with model layers like such. I thought this might have something to do with data parallel because of the "module", but even removing the "module" from the state_dict names or trying to load with model.module.load_state_dict doesn't work.

. . . . 6.0.conv_p.1.running_var", "module.aspp_16.0.conv_p.1.num_batches_tracked", "module.aspp_16.0.concate_conv.0.weight", "module.aspp_16.0.concate_conv.1.weight", "module.aspp_16.0.concate_conv.1.bias", "module.aspp_16.0.concate_conv.1.running_mean", "module.aspp_16.0.concate_conv.1.running_var", "module.aspp_16.0.concate_conv.1.num_batches_tracked", "module.aspp_32.0.conv11.0.weight", "module.aspp_32.0.conv11.1.weight", "module.aspp_32.0.conv11.1.bias", "module.aspp_32.0.conv11.1.running_mean", "module.aspp_32.0.conv11.1.running_var", "module.aspp_32.0.conv11.1.num_batches_tracked", "module.aspp_32.0.conv33.0.weight", "module.aspp_32.0.conv33.1.weight", "module.aspp_32.0.conv33.1.bias", "module.aspp_32.0.conv33.1.running_mean", "module.aspp_32.0.conv33.1.running_var", "module.aspp_32.0.conv33.1.num_batches_tracked", "module.aspp_32.0.conv_p.0.weight", "module.aspp_32.0.conv_p.1.weight", "module.aspp_32.0.conv_p.1.bias", "module.aspp_32.0.conv_p.1.running_mean", "module.aspp_32.0.conv_p.1.running_var", "module.aspp_32.0.conv_p.1.num_batches_tracked", "module.aspp_32.0.concate_conv.0.weight", "module.aspp_32.0.concate_conv.1.weight", "module.aspp_32.0.concate_conv.1.bias", "module.aspp_32.0.concate_conv.1.running_mean", "module.aspp_32.0.concate_conv.1.running_var", "module.aspp_32.0.concate_conv.1.num_batches_tracked", "module.final_conv.weight", "module.final_conv.bias".

In the origin paper, the network level update is

But in your code:

           # normalized_betas[layer][ith node][0 : ➚, 1: ➙, 2 : ➘]
            for layer in range (len(self.betas)):
                if layer == 0:
                    normalized_betas[layer][0][1:] = F.softmax (self.betas[layer][0][1:].to(device=img_device), dim=-1)

                elif layer == 1:
                    normalized_betas[layer][0][1:] = F.softmax (self.betas[layer][0][1:].to(device=img_device), dim=-1)
                    normalized_betas[layer][1] = F.softmax (self.betas[layer][1].to(device=img_device), dim=-1)

                elif layer == 2:
                    normalized_betas[layer][0][1:] = F.softmax (self.betas[layer][0][1:].to(device=img_device), dim=-1)
                    normalized_betas[layer][1] = F.softmax (self.betas[layer][1].to(device=img_device), dim=-1)
                    normalized_betas[layer][2] = F.softmax (self.betas[layer][2].to(device=img_device), dim=-1)
                else :
                    `normalized_betas[layer][0][1:]` = F.softmax (self.betas[layer][0][1:].to(device=img_device), dim=-1)
                    normalized_betas[layer][1] = F.softmax (self.betas[layer][1].to(device=img_device), dim=-1)
                    normalized_betas[layer][2] = F.softmax (self.betas[layer][2].to(device=img_device), dim=-1)
                    normalized_betas[layer][3][:1] = F.softmax (self.betas[layer][3][:1].to(device=img_device), dim=-1)

in search You can use ABN and affine and by setting args

in re_train Now I get 0.74 miou by training 2000epochs(about0.4M iters) and 0.798 miou by training 5000epochs(about 1M iter) You can get the similar result by use 'python train.py' and set batch_size=2 on each gpu

I am wondering how can I visualise the best inner cell and outer network after search for my custom dataset. Something similar to the original paper in Figure 3.

I really appreciate any comment or help in advance.

Hey! Thank you for your implementation of Autodeeplab. It looks very promising.

I'm trying to use it out of the box (with default params) for Pascal dataset. But after 50 epochs I got Acc:0.6794524843482059, Acc_class:0.13206915050001747, mIoU:0.08714873731068112, fwIoU: 0.526273923425172

Is it normal?

I am wondering why training loss is increasing after alpha_epoch which in my case is the epoch 30, and accuracy is decreasing sharply? any suggestion would be appreciated. I have tried different arch-lr such as 1e-1, or 3e-3. I am doing segmentation on medical images.

Here are plots of loss and acc:

hello, i'm new to the autodeeplab, i want to ask, if the mask color definition in training file must be corresponding to the mask picture? because i don't exactly know the segmentation color in my segmentation image as ground truth, which i have automatically collected in simulation. Thank you in advance