tf_text_graph_faster_rcnn.py - Assertion error

[engineer1982]

System information (version)

• OpenCV => :3.4.3:
• Operating System / Platform => :Windows 10 64 bit:
• Compiler => :Visual Studio 2017:

Detailed description

I used Tensorflow Object Detection to train a "faster_rcnn_inception_resnet_v2_atrous_coco_2018_01_28" in my custom dataset. After trained, it detected well the new classes (just 3). I exported the new trained model to a folder, which contains the files "frozen_inference_graph.pb" and "pipeline.config"

Then, I used "tf_text_graph_mask_rcnn.py" in the following command:

python tf_text_graph_faster_rcnn.py --input faster_rcnn_inception_resnet_v2_atrous_coco_2018_01_28_step19391/frozen_inference_graph.pb --config faster_rcnn_inception_resnet_v2_atrous_coco_2018_01_28_step19391/pipeline.config --output faster_rcnn_inception_resnet_v2_atrous_coco_2018_01_28_step19391/graph.pbtxt

Here follows the output. Seem to run well, but got an error:

Number of classes: 3
Scales: [0.25, 0.5, 1.0, 2.0]
Aspect ratios: [0.5, 1.0, 2.0]
Width stride: 8.000000
Height stride: 8.000000
Features stride: 8.000000
2018-11-21 11:27:27.621794: I tensorflow/tools/graph_transforms/transform_graph.cc:318] Applying sort_by_execution_order
Traceback (most recent call last):
File "tf_text_graph_faster_rcnn.py", line 241, in
createFasterRCNNGraph(args.input, args.config, args.output)
File "tf_text_graph_faster_rcnn.py", line 59, in createFasterRCNNGraph
assert(graph_def.node[0].op == 'Placeholder')
AssertionError

Any ideas about what happened here?

Thank you guys.

[dkurt]

@engineer1982, please attach frozen_inference_graph.pb and pipeline.config if it's possible. You can also try the latest state of 3.4 or master branches because we did some fixes recently.

[engineer1982]

@engineer1982, please attach frozen_inference_graph.pb and pipeline.config if it's possible. You can also try the latest state of 3.4 or master branches because we did some fixes recently.

@dkurt , the .pb file has 216MB. Here goes the google drive link: https://drive.google.com/open?id=1MBeqFKd3jE9codu23cQJpHeahWHXVtiF

[dkurt]

@engineer1982, the problem is that OpenCV doesn't support any of _atrous network with dilated convolutions from TensorFlow Object Detection API for now.

[dkurt]

@engineer1982, Hi!
Please test the following PR: #13656.

For faster_rcnn_inception_resnet_v2_atrous_coco_2018_01_28 model we need to replace

height_stride: 8
width_stride: 8

to

height_stride: 16
width_stride: 16

However I don't actually know if this is a typo in pipeline.config or our script computes anchors in a wrong way. At least with 16x16 everything looks fine (1024x600 input image):

[engineer1982]

@engineer1982, Hi!
Please test the following PR: #13656.

For faster_rcnn_inception_resnet_v2_atrous_coco_2018_01_28 model we need to replace

height_stride: 8
width_stride: 8

to

height_stride: 16
width_stride: 16

However I don't actually know if this is a typo in pipeline.config or our script computes anchors in a wrong way. At least with 16x16 everything looks fine (1024x600 input image):

Thanks Dimitry! And where do I change it?

[dkurt]

@engineer1982, in the config file:

    first_stage_anchor_generator {
      grid_anchor_generator {
        height_stride: 16
        width_stride: 16
        scales: 0.25
        scales: 0.5
        scales: 1.0
        scales: 2.0
        aspect_ratios: 0.5
        aspect_ratios: 1.0
        aspect_ratios: 2.0
      }
    }

[engineer1982]

Thank you! gonna try it and update here.