Hey, @sercant!

In the case of running the file evaluate.py or visualize.py with your weights (shufflenetv2_dpc_cityscapes_71_3), the following error occurs:

For visualize.py:
NotFoundError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a Variable name or other graph key that is missing from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Key aspp0/BatchNorm/beta not found in checkpoint
[[node save/RestoreV2 (defined at visualize.py:276) ]]
[[node save/RestoreV2 (defined at visualize.py:276) ]]

For evaluate.py:
NotFoundError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a Variable name or other graph key that is missing from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Key aspp0/BatchNorm/beta not found in checkpoint
[[node save/RestoreV2 (defined at evaluate.py:162) ]]

First of all I thank you for good job.

When the file is started visualize.py the following error occurs:

File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/client/session.py", line 1407, in _call_tf_sessionrun
run_metadata)
tensorflow.python.framework.errors_impl.InvalidArgumentError: Shape mismatch in tuple component 1. Expected [769,769,3], got [1024,2048,3]
[[{{node batch/padding_fifo_queue_enqueue}}]]

I tried to resize, but it did not help. When I change the size to [1024,2048] the following error appears.
However, the first image is successfully processed.

Traceback (most recent call last):
File "visualize.py", line 335, in
tf.app.run()
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/platform/app.py", line 125, in run
_sys.exit(main(argv))
File "visualize.py", line 320, in main
train_id_to_eval_id=train_id_to_eval_id)
File "visualize.py", line 183, in _process_batch
colormap_type=FLAGS.colormap_type)
File "/code/ShuffleNetV2/utils/save_annotation.py", line 33, in save_annotation
label, colormap_type)
File "/code/ShuffleNetV2/utils/get_dataset_colormap.py", line 389, in label_to_color_image
raise ValueError('label value too large.')
ValueError: label value too large.

Hello, author!
Thank you very much for your open source code.
I currently want to test my own atlas with the pre-training model you provided.
In the process of implementing this task, I first configured the version of tensorflow1.15 and converted the data of the citysacpes data set (png2tfrecord) according to your suggestion.
However, in the process of data conversion, I executed according to your code build_cityscapes_data.py, and the resulting tfrecord files were all empty, as follows:

I'd like you to help me with this problem. Thank you！

Hi,author.Your proposed network is very interesting and effective！
The first questions, the DPC encoder head have 5 franches.
Every franches have 256 channels, then we can get 1280 channels by concat, but in Table 1,
the DPC module's Output Channels are 512. Why?
The second, the rates in DPC is interesting, why are 1X6, 18X15, 6X3, 6X21?I see the deeplabv3+'s rate are 6, 12 ,18.
The last, the each franch in DPC troubles me.I can't understand the reasons.

I'm sorry for asking so many questions, I have an additional question. Why you preprocess the input images by standardizing each pixel to [-1, 1] rather than [0, 1] or not do this operation?

Hi Author!

I'm trying to train using my own data.
I made tfdata and tried to excute train.py.
But I can't define "tf_initial_checkpoint".
Could you please help me out?

Hi, @sercant !

Thank you for sharing this fantastic project!

I found that in core/shufflenet_v2.py line 69 "layer_stride = 1" will cause the parameter "rate" as 1 instead of 2 in shufflenet V2 stage 4.

Could you check this small issue?

Thanks

Hi! I‘m really interested in your paper. Do you plan to implement the code in pytorch? How to run the code on the mobile platform? Thanks!

Dear Sir,
Thank you for sharing your code, It do help me a lot.
The training and testing part of the model work very well.
But when I start to generate a pb and tflite file, problem comes.
Right now, I use tensorflow with a version of 2.0 beta. Just modify the directory of logs in order to execute the conversion script. As I look deeper into the code, the input node is "input_0" , the output node is "Cast" , which convert the data from int64 to int32. I also notices that, the outputs_to_scales_to_logits is generated by model.predict_labels, which include a operator "ArgMax", it also supported by the latest version of tflite. But the conversion progress did not report any errors, so I test the generated tflie file by the scripy below:

import numpy as np
import tensorflow as tf

Load TFLite model and allocate tensors.

interpreter = tf.lite.Interpreter(model_path="segmentation.tflite")
interpreter.allocate_tensors()

Get input and output tensors.

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

Test model on random input data.

input_shape = input_details[0]['shape']
print(input_shape)
input_data = np.array(np.random.random_sample(input_shape), dtype=np.float32)
print(input_data)
interpreter.set_tensor(input_details[0]['index'], input_data)

interpreter.invoke()
output_data = interpreter.get_tensor(output_details[0]['index'])
print(np.unique(output_data))

I just used a random picture to do the test, willing to see different labels [total 14 classes in my case] sorted by np.unique function. but unfortunately, just one number[4] or two numbers[3,4] is shown. Can you give me some ideas how to solve this problem? Thank you very much!

I also test the tflite file with a test picture, it also did not report expected results.

@sercant
the DPC module, the input is the ShuffleNet V2 stage4 (that means the input channel is 464) and here, the five branches conv's kernel out_channel is 256(as reffered in the paper "Searching for Efficient Multi-Scale Architectures for Dense Image Prediction") of DPC.
While, as for depthwise_conv2d, the out_channel of convs result is in_channel(inputs )* 256(kernel out_channel). So after 5 braches, we use tf.concat, the final result's out_channel is a very big number ?

Hi author，I see that you have used COCO 2017 data set for network pre-training processing.
I implemente a semantic segmentation network and download the COCO 2017 data set. I want to know how to get the label images(grey images) from the annotations'json.

Hi author, I want to know the forward time to test the FPS. How do I do ?

Hello, author!
At present, I want to test my own pictures through the model you trained. There is no label for my pictures. Can I test this? If so, how can I do it?

Hi ,author.
I really want to get the images like the "the folder of getfine in cityscapes data set".such as
'cityscapes/gtFine/train/aachen/aachen_000000_000019_gtFine_labelIds.png' and corresponding to the image such as 'cityscapes/leftImg8bit/train/aachen/aachen_000000_000019_leftImg8bit.png'.
I can make txt files to input data.

Folks awesome job with being able to do real time segmentation with shufflenet 2.

Do you mind sharing the pretrained weights