msiam / tfsegmentation Goto Github PK

Following the instruction in README.md, I download pretrained_weights.zip from the google drive.
However, it seems that shufflenet_weights.pkl , which is referenced by config/experiments_config/unet_shufflenet_train.yaml, is missing from the zip package. Is it correcly packed and uploaded?

I found it a duplicated issue, so close it.

Hi!

I tried to train by command run.sh.
($ python3 main.py --load_config=fcn8s_shufflenet_train.yaml train Train FCN8sShuffleNet)
I encount error as below.

$ ./run.sh
/home/sounansu/anaconda3/lib/python3.6/site-packages/h5py/init.py:36: FutureWarning: Conversion of the second argument of issubdtype from float to np.floating is deprecated. In future, it will be treated as np.float64 == np.dtype(float).type.
from ._conv import register_converters as _register_converters

Parsing Arguments..

Using this arguments check it

batch_size -- 8 --
batchnorm_enabled -- True --
bias -- 0.0 --
config_path -- fcn8s_shufflenet_train.yaml --
data_dir -- full_cityscapes_res --
data_mode -- experiment --
exp_dir -- fcn8s_shufflenet --
freeze_encoder -- False --
img_height -- 256 --
img_width -- 512 --
learning_decay -- 1e-07 --
learning_decay_every -- 100 --
learning_rate -- 0.0001 --
max_to_keep -- 1 --
mode -- train --
model -- FCN8sShuffleNet --
momentum -- 0.9 --
num_channels -- 3 --
num_classes -- 20 --
num_epochs -- 200 --
num_groups -- 3 --
operator -- Train --
out_dir -- fcn8s_shufflenet --
pretrained_path -- pretrained_weights/shufflenet_weights.pkl --
random_cropping -- False --
save_every -- 10 --
shuffle -- True --
test_every -- 10 --
verbose -- False --
weight_decay -- 0.0005 --
weighted_loss -- True --

Agent is running now...

2018-09-07 12:35:22.150963: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2018-09-07 12:35:22.150994: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2018-09-07 12:35:22.355806: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:893] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2018-09-07 12:35:22.356259: I tensorflow/core/common_runtime/gpu/gpu_device.cc:955] Found device 0 with properties:
name: GeForce GTX 1080 Ti
major: 6 minor: 1 memoryClockRate (GHz) 1.6705
pciBusID 0000:03:00.0
Total memory: 10.91GiB
Free memory: 10.19GiB
2018-09-07 12:35:22.356285: I tensorflow/core/common_runtime/gpu/gpu_device.cc:976] DMA: 0
2018-09-07 12:35:22.356295: I tensorflow/core/common_runtime/gpu/gpu_device.cc:986] 0: Y
2018-09-07 12:35:22.356308: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1045] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:03:00.0)

Building the MODEL...
Building the ShuffleNet..
Layer_name: network/shufflenet_encoder/conv1/Relu -Output_Shape: [8, 127, 255, 24]
Layer_name: network/shufflenet_encoder/max_pool -Output_Shape: [8, 63, 127, 24]
Layer_name: network/shufflenet_encoder/stage2_3/Relu -Output_Shape: [8, 32, 64, 240]
Layer_name: network/shufflenet_encoder/stage3_7/Relu -Output_Shape: [8, 16, 32, 480]
Layer_name: network/shufflenet_encoder/stage4_3/Relu -Output_Shape: [8, 8, 16, 960]

Encoder ShuffleNet is built successfully

The Model is built successfully

[-] build_model : 19.11881 sec, which is 0.31865 mins, which is 0.00531 hours

Training is initializing itself

Initializing the variables of the model
Initialization finished
[-] init_model : 3.71544 sec, which is 0.06192 mins, which is 0.00103 hours
Loading ImageNet pretrained weights...
ImageNet Pretrained Weights Loaded Initially

Pretrained weights of the encoder is loaded
Searching for a checkpoint

.. No ckpt, SO First time to train :D ..

[-] load_model : 4.51711 sec, which is 0.07529 mins, which is 0.00125 hours
Loading Training data..
Train-shape-x -- (2975, 256, 512, 3) 2975
Train-shape-y -- (2975, 256, 512)
Num of iterations on training data in one epoch -- 372
Training data is loaded
Loading Validation data..
Val-shape-x -- (500, 512, 1024, 3) 496
Val-shape-y -- (500, 512, 1024)
Num of iterations on validation data in one epoch -- 62
Validation data is loaded
[-] load_train_data : 22.91540 sec, which is 0.38192 mins, which is 0.00637 hours
Training mode will begin NOW ..
epoch-0-: 100%|????????????????????????????????????????????????????????????????????????????| 371/372 [01:31<00:00, 4.58it/s]
epoch-0-loss:14.767903- acc:0.6199
saving a checkpoint
Saved a checkpoint
Validation at step:372 at epoch:1 ..
Val-epoch-1-: 0%| | 0/62 [00:00<?, ?it/s]Traceback (most recent call last):
File "main.py", line 19, in
main()
File "main.py", line 15, in main
agent.run()
File "/home/sounansu/anaconda3/TFSegmentation/utils/misc.py", line 18, in timed
result = f(*args, **kwargs)
File "/home/sounansu/anaconda3/TFSegmentation/agent.py", line 96, in run
self.train()
File "/home/sounansu/anaconda3/TFSegmentation/agent.py", line 139, in train
self.operator.train()
File "/home/sounansu/anaconda3/TFSegmentation/train/train.py", line 583, in train
epoch=self.model.global_epoch_tensor.eval(self.sess))
File "/home/sounansu/anaconda3/TFSegmentation/train/train.py", line 636, in test_per_epoch
feed_dict=feed_dict)
File "/home/sounansu/anaconda3/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 895, in run
run_metadata_ptr)
File "/home/sounansu/anaconda3/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1100, in _run
% (np_val.shape, subfeed_t.name, str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (8, 512, 1024, 3) for Tensor 'network/input/Placeholder:0', which has shape '(8, 256, 512, 3)'

What happened?

Hello,
First of all, thank you for your work.
I am a beginer. I want to use my own .png format image to test or val. So i have some questions.

1.when i run it using the test command (python3 main.py --load_config=fcn8s_shufflenet_test.yaml test Train FCN8sShufflenet) have this question.

File "/home/mamingyu/TFSegmentation-master/train/train.py", line 741, in test
y_batch = self.test_data['Y'][idx:idx + 1]
KeyError: 'Y'

I found the fuction load_test_data in train.py line 389. The dict self.test_data did not load 'Y'.

2. Is X_train.npy corresponding to the train RGB image and Y_train.npy corresponding to the train label image? How to make a lot of png images become X_val.npy and Y_val.npy so that I simply replace my own data format from png to npy.

3. TFSegmentation/data/ has preprocess_cityscapes_tfrecords.py and preprocess_cityscapes.py . Is the first .py make png image become TFrecord format and the second .py make png image become npy format? So can i use TFrecord format as my input image data format and how to do that.

Please give me some suggestion. Thank you very much

I got this error code and I know why this error came up because my system environment has CUDA 9.0 CUDNN 7.0 version.

I would like to know how to run this code on my system (My situation is NOT possible to change CUDA version from 9.0 to 8.0).

I think I should change my tensorflow version to run on CUDA 9.0 environment but I do not know how to make it.

Please give me any advice. Thank you.

File "main.py", line 19, in
main()
File "main.py", line 15, in main
agent.run()
File "/home/testc/TFSegmentation/utils/misc.py", line 18, in timed
result = f(*args, **kwargs)
File "/home/testc/TFSegmentation/agent.py", line 85, in run
self.build_model()
File "/home/testc/TFSegmentation/utils/misc.py", line 18, in timed
result = f(*args, **kwargs)
File "/home/testc/TFSegmentation/agent.py", line 46, in build_model
self.model = self.model(self.args)
File "/home/testc/TFSegmentation/models/unet_vgg16.py", line 11, in init
super().init(args)
File "/home/testc/TFSegmentation/models/basic/basic_model.py", line 37, in init
self.params.class_weights= np.load(self.args.data_dir+'weights.npy')
File "/usr/local/lib/python3.5/dist-packages/numpy/lib/npyio.py", line 370, in load
fid = open(file, "rb")
FileNotFoundError: [Errno 2] No such file or directory: '/home/testc/TFSegmentation/data/full_cityscapes_res/weights.npy

FCN8sShuffleNet last layer is upscore with deconv stride 8 .this layer is very time comsuming .on jetson tx2 .this layer occupy 180 ms . any optimization ？

Hello,

Can you point me on how to install Tensorflow on Jetson TX2 board?

Every help will be appreciated.

Thank you.

Hi,
I trained unet-resnet18 on cityscapes using run.sh. But when I try to run the second command in optimize.sh:
python -m tensorflow.python.tools.optimize_for_inference --input graph_frozen.pb --output graph_optimized.pb --input_names=network/input/Placeholder --output_names=network/output/ArgMax

I got the a lot of warnings like
WARNING:tensorflow:Didn't find expected Conv2D input to 'network/output_block/batch_normalization_2/cond/FusedBatchNorm_1'

Later if I run infer_optimized, I will got
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/importer.py", line 493, in import_graph_def: raise ValueError(str(e))
ValueError: NodeDef expected inputs '' do not match 1 inputs specified

Has anyone seen this before?
Thanks

Hi,
I'm trying to train on my custom dataset with just two classes (person and none).
I've generated the X_train, Y_train, weights and mean files.
But while training I continuously see

loss: nan

My dataset is 512x640 images.

Hi. Firstly, thank you for this amazing project.
When I run this code:
python3 main.py --load_config=fcn8s_shufflenet_train.yaml train Train FCN8sShuffleNet
I got an error saying that validation data image size and training data image size is not equal. One of them is 512x1024 while other one is 256x512. I just simply resized whole data and save them. I fixed this issue but I just want to check if there is anyone facing this problem. If so, you should consider to edit the Y_val.npy and X_val.npy file.
If this issue is only the case for me then what could be the reason?
Thank you.

2018-08-06 08:21:31.998637: W tensorflow/core/common_runtime/bfc_allocator.cc:279] ****************************************************************************************************
2018-08-06 08:21:31.998655: W tensorflow/core/framework/op_kernel.cc:1318] OP_REQUIRES failed at gpu_swapping_kernels.cc:72 : Resource exhausted: OOM when allocating tensor with shape[5,32,512,1024] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
Traceback (most recent call last):
File "/root/anaconda3/envs/scc_py36/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1322, in _do_call
return fn(*args)
File "/root/anaconda3/envs/scc_py36/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1307, in _run_fn
options, feed_dict, fetch_list, target_list, run_metadata)
File "/root/anaconda3/envs/scc_py36/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1409, in _call_tf_sessionrun
run_metadata)
tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape[5,64,256,512] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
[[Node: network/train-operation/gradients/AddN_14-0-TransposeNHWCToNCHW-LayoutOptimizer = Transpose[T=DT_FLOAT, Tperm=DT_INT32, _device="/job:localhost/replica:0/task:0/device:GPU:0"](network/train-operation/gradients/network/upscale_4/upscale4/batch_normalization/cond/FusedBatchNorm_1/Switch_grad/cond_grad, PermConstNHWCToNCHW-LayoutOptimizer)]]
Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.

     [[Node: network/pixel_wise_accuracy/Mean/_1157 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_19236_network/pixel_wise_accuracy/Mean", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]

Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "main.py", line 19, in
main()
File "main.py", line 15, in main
agent.run()
File "/data/scc/TFSegmentation/utils/misc.py", line 18, in timed
result = f(*args, **kwargs)
File "/data/scc/TFSegmentation/agent.py", line 96, in run
self.train()
File "/data/scc/TFSegmentation/agent.py", line 139, in train
self.operator.train()
File "/data/scc/TFSegmentation/train/train.py", line 513, in train
feed_dict=feed_dict)
File "/root/anaconda3/envs/scc_py36/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 900, in run
run_metadata_ptr)
File "/root/anaconda3/envs/scc_py36/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1135, in _run
feed_dict_tensor, options, run_metadata)
File "/root/anaconda3/envs/scc_py36/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1316, in _do_run
run_metadata)
File "/root/anaconda3/envs/scc_py36/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1335, in _do_call
raise type(e)(node_def, op, message)
tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape[5,64,256,512] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
[[Node: network/train-operation/gradients/AddN_14-0-TransposeNHWCToNCHW-LayoutOptimizer = Transpose[T=DT_FLOAT, Tperm=DT_INT32, _device="/job:localhost/replica:0/task:0/device:GPU:0"](network/train-operation/gradients/network/upscale_4/upscale4/batch_normalization/cond/FusedBatchNorm_1/Switch_grad/cond_grad, PermConstNHWCToNCHW-LayoutOptimizer)]]
Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.

     [[Node: network/pixel_wise_accuracy/Mean/_1157 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_19236_network/pixel_wise_accuracy/Mean", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]

Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.

I change batch_size to 1, but it also occur OOM

Hello,

I tried to verify ShuffleSeg's real-time performance locally, but I didn't find the pretrained weights for ShuffleNet.

Did you release it?

Hello !!!
In your paper "SHUFFLESEG: REAL-TIME SEMANTIC SEGMENTATION NETWORK", what does the feed layer do? ? so the upsample methods can be very different according to decoder ?

I intend to train suffleseg model with my own dataset which has some unnecessary parts for training. (It does not mean 'background') So I would like to make ignore label for those parts to except training.
Do you have any idea for this? I do not exactly know which part I do have to modify or remove the code.
Thanks.

Thanks for your contribution in this work!
I have seen the metrics scripts in your project. However, are there some publicly evaluation tools for the fair comparison? If not, can I use your scripts to evaluate my model?
thanks ~

can you share shufflenet_weights.pkl ?

Hello
I want to try to use the mobilenet_v2 network. Do you have the pretrained weight (pkl format) for mobilenet_v2 network. And I can only find the ckpt format file on the internet. Could you tell me how to get pretrained weight (pkl format) or make it ( I found a mobilenet_v1_vanilla.pkl file in the mobilenet.py that it's converted to mobilenet_v1.pkl. How to get it?). Thank you for your help.

i run python3 main.py --load_config=unet_vgg16_test.yaml test Train UNetVGG16 is error

======================End of Report==========================
[-] build_model : 8.28218 sec, which is 0.13804 mins, which is 0.00230 hours

Training is initializing itself

Initializing the variables of the model
Initialization finished
[-] init_model : 3.16336 sec, which is 0.05272 mins, which is 0.00088 hours
Searching for a checkpoint

.. No ckpt, SO First time to train :D ..

[-] load_model : 0.00013 sec, which is 0.00000 mins, which is 0.00000 hours
ERROR Please select a proper data_mode BYE

good job,
what's the big about the file,I download it ,but failure. I have original cityscape dataset, how convert it to full_cityscapes_res?

Hi, I was really impressed by your works.

I wonder how to make the pretrained model using coarse annotation as described in your paper.

thanks in advance.

Hi @MSiam,

I'm trying to optimize the graph using the nchw_optimize_inference branch, but after I freeze the graph it is removing the input layer ['network/input/Placeholder'].

This is why optimize_for_inference throws an error
KeyError: "The following input nodes were not found: set(['network/input/Placeholder'])\n"
I also checked the frozen graph on TensorBoard and this node was absent there.

I'm using TensorFlow v1.4

How were you able to generate the frozen model for inference?

Thanks!

Hi awesome project! I have followed the instruction step by step and when I execute run.sh I encountering the following error:

======================End of Report==========================
   [-] build_model : 31.10391 sec, which is 0.51840 mins, which is 0.00864 hours

Training is initializing itself

Initializing the variables of the model
Initialization finished
   [-] init_model : 121.96605 sec, which is 2.03277 mins, which is 0.03388 hours
Loading ImageNet pretrained weights...
ImageNet Pretrained Weights Loaded Initially


Pretrained weights of the encoder is loaded
Searching for a checkpoint

.. No ckpt, SO First time to train :D ..

   [-] load_model : 6.38715 sec, which is 0.10645 mins, which is 0.00177 hours
Loading Validation data..
Validation-shape-x -- (500, 512, 1024, 3)
Validation-shape-y -- (500, 128, 256)
Validation data is loaded
   [-] load_val_data : 4.74056 sec, which is 0.07901 mins, which is 0.00132 hours
INFERENCE mode will begin NOW..
loading a checkpoint for BEST ONE
ERROR NO best checkpoint found

Observe that the structure of the folder pretrained weights:

$ ls TFSegmentation/pretrained_weights
ls -l                                                                                                                                                            (base) 
total 692156
-rw-rw-r-- 1 dgromov dgromov 130902443 sep 19 15:49 linknet_weights.pkl
-rw-rw-r-- 1 dgromov dgromov  17033751 sep 19 15:49 mobilenet_v1.pkl
-rw-rw-r-- 1 dgromov dgromov   7388146 sep 19 15:49 shufflenet_weights.pkl
-rw-rw-r-- 1 dgromov dgromov 553431761 sep 19 15:49 vgg16.npy

Is there something wrong that I have done ?

Thanks

Also What should I do with the Graph_optimized.pb file?

Hello, Thanks a ton for the code. However do you have pointers on how to run training on new dataset ? Any quick README would be of great help.

Hello everyone.
Your work looks really interesting so i downloaded everything to try i out. Unfortunetaly there are no pretrained weights for your shufflenet. Will you publish them at any time?
Regards!

edit: at the moment i only have a jetson tx2 avaible , so train it by my self would be a little tricky because of some lacking ressources, right?

Is this the ShuffleSeg model definition file?
https://github.com/MSiam/TFSegmentation/blob/master/models/unet_shufflenet.py

Hi.
I'm really interested in Real-time Semantic Segmentation. So, I'd like to check the latency and accuracy with webcam in real time. Can I check this with this program? and i want to know how to do.
Thanks.

Would you be able to provide a simple demo or interface for easy visualize your segmentation performance?

Thanks

I am trying to test the model with 7681024 images but it seems not that simple for me.
(I already converted my own images (7681024) to X_val.npy file.)
Please give me an advice which part of the code I need to modify.
Thanks.

I trained both fcn8s_mobilenet and fcn8s_shufflenet at the resolution 512x1024. After doing the graph optimization and deploying on Jetson (~7.3GB total, ~4GB available usually), I got the runtime error that Tensorflow ran out of the memory.

I checked that on my workstation both the models take around 7.8GB GPU memory. But for these two convolution neural networks, reducing input size should not affect the number of parameters too much, which means the memory needed shouldn't change too much. In this case, what other tricks shall I use to reduce the memory consumption?

Thanks!

where is the shuffleNet pretrained model , I cannot find it in google drive

When I run training codes, it seems not call best checkpoint file and just start from the initial value.

Which part of the codes do I should modify?

Hello
Thanks for your work. I have completed the first two steps and got graph.pb and graph_optimized.pb .

But when I run 'python3 infer_optimize.py --graph graph_optimized.pb'. I got a InvalidArgumentError : You must feed a value for placeholder tensor 'import/network/input/plaveholder_2' with dtype bool.

Please give me any advice, thank you

Hello,

First of all, your work is very impressive and helpful. Thank you!

I intend to visualize segmented output image (.png) using my own .png format image on fcn8s_mobilenet model.
And I already trained fcn8s_mobilenet model with cityscape dataset (X_train.npy, Y_train.npy is it right?)

The question is :

1. I would like to know how you make X_val.npy and Y_val.npy using png images so that I simply replace my own data format from png to npy. (Please explain about X_val.npy, Y_val.npy)
2. If I convert my own png image to X_val.npy and Y_val.npy successfully, then the test code will be worked without error? Or should I change some part of your code? (resolution or something else)

Thank you.

Here is my result below when Ive converted only one png(1024x512) image to npy file and replaced the file name to X_val.npy in data/full_cityscapes_res/ and run it using the test command (python3 main.py --load_config=fcn8s_mobilenet_test.yaml test Train FCN8sMobileNet)

testc@testc2018:~/TFSegmentation$ python3 main.py --load_config=unet_vgg16_train.yaml train Train VGG16UNET

Parsing Arguments..

Using this arguments check it

batch_size -- 3 --
batchnorm_enabled -- True --
bias -- 0.0 --
config_path -- unet_vgg16_train.yaml --
data_dir -- data/full_cityscapes_res --
data_mode -- experiment --
exp_dir -- unet_vgg16 --
freeze_encoder -- False --
img_height -- 512 --
img_width -- 1024 --
learning_decay -- 1e-07 --
learning_decay_every -- 100 --
learning_rate -- 0.0001 --
max_to_keep -- 2 --
mode -- train --
model -- VGG16UNET --
momentum -- 0.9 --
num_channels -- 3 --
num_classes -- 20 --
num_epochs -- 200 --
num_groups -- 3 --
operator -- Train --
out_dir -- unet_vgg16 --
pretrained_path -- pretrained_weights/vgg16.npy --
random_cropping -- False --
save_every -- 5 --
shuffle -- True --
test_every -- 5 --
verbose -- False --
weight_decay -- 4e-05 --
weighted_loss -- True --

Traceback (most recent call last):
File "main.py", line 19, in
main()
File "main.py", line 14, in main
agent = Agent(args)
File "/home/testc/TFSegmentation/agent.py", line 36, in init
self.model = globals()[args.model]
KeyError: 'VGG16UNET'

Sorry, I read the paper but I still don't know.

Hello everyone! I just wanted to remark that the folder full_cityscapes_res does not have a X_test.npy should I create one?

Thanks

Hello,

I am trying to test models which are skipnet-mobilenet and shuffleseg with my own dataset (1024x768 resolution)

I've seen in the paper that those models' can be used in real-time situation cause the inference speed is faster than 10fps on PC.

However, when I run your code for inference with my own data, the speed is about 1.5fps and both models speed are similar even though you mentioned shuffeseg is faster than skipnet-mobilenet here.

Please give me any advice for using the model in real-time.

Thank you.

Hello，

Can you point me on how to modify your code to accelerate training using multi-GPU?

Thanks.

What‘s the differences between RTSEG and SHUFFLRSEG? Those papers used the same methods and code.

I could not find pretrained weights for any of the models

Hi.
I ran infer_optimized.py with my webcam by reference to #20 and #42
But, i cant understand why out.shape = (1, 360, 640) (line38 : out = sess.run(y, feed_dict={x: img}).
i think this should be(360, 640, 3) because this is RGB image.
Can anyone explain this reason?

You provided the data in a npy format in your code which had a resolution of 5121024 . I want to train the model with cityscapes (10242048). Can you provide the data.npy with the resolution of 1024*2048. Thanks a lot!

Hi.
I want to try training model with my own coco format annotation data.
So, i have to convert my data into cityscapes format.
I think i can do the work with this directory(TFSegmentation-optimize_inference//data//coco).
Could you tell me how to do it ?
Thanks!

where is pretrained_weights/shufflenet_weights.pkl? I didn't find it in the pretrained_weights.zip.
Can you send it to my email ([email protected]) ?
Thanks.

Hello,
In one of the papers you mentioned that ShuffleSeg runs at 15.7 FPS on NVIDIA Jetson TX2. Was this result achieved by running images of size 2048x1024 (the size of an image from CityScapes data set)?

Hi @MSiam,

Where is X_test and related test like xnames_test and ynames_test npy dataset

Does your Cityscapes test dataset have labels?I have downloaded a label without a test data set from the official website.So I can't get the test mean IOU

I'm assuming your intention was to leave the folder empty as to not push weights to github. Do you have them hosted anywhere for the public to download?

Update: Should have said thank you for the excellent paper published on arXiv today!
Update 2: I should have been more specific, links to the pre-trained models used in your evaluation.

When I train the model with my own X_train.npy, Y_train.npy also I made Weights.npy and mean.npy using compute_weights.py file, there is no error but I see loss : nan-.
Ive checked the loss list values and every loss value shows nan.
(My dataset is 1024X768 and labeled 19 classes)

Do you plan to cover also this new model? Seems that it will arrive in the next PR. Check https://github.com/tensorflow/models/blob/master/research/deeplab/README.md