Does OpenVINO support Tensorflow, faster_rcnn_nas?

[Chard]

Hi,

Does OpenVINO support Tensorflow, faster_rcnn_nas?

The MO is done, but result is not correct.

[Severine_H_Intel]

Hi Chard,

yes, we support it. What is exactly your issue? Do you face accuracy issues in the Inference Engine phase? or the MO output is not good?

Best, 

Severine

[Chard]

Hi Severine,

It is accuracy issue.

1. I use below MO command.
python3 mo_tf.py \
--input_model "$tf_out/frozen_inference_graph.pb" \
--tensorflow_object_detection_api_pipeline_config "$tf_out/pipeline.config" \
--output_dir "fii_frcnn_nas" \
--output=detection_boxes,detection_scores,num_detections \
--tensorflow_use_custom_operations_config extensions/front/tf/faster_rcnn_support_api_v1.7.json

2. Running object_detection_sample_ssd, and getting below result
[ INFO ] InferenceEngine:
        API version ............ 1.2
        Build .................. 13911
Parsing input parameters
[ INFO ] Loading plugin

        API version ............ 1.2
        Build .................. lnx_20180510
        Description ....... MKLDNNPlugin
[ INFO ] Loading network files:
        fii_frcnn_nas/frozen_inference_graph.xml
        fii_frcnn_nas/frozen_inference_graph.bin
[ INFO ] Preparing input blobs
[ INFO ] Batch size is 1
[ INFO ] Preparing output blobs
[ INFO ] Loading model to the plugin
[ INFO ] Batch size is 1
[ INFO ] Start inference (1 iterations)
[ INFO ] Processing output blobs
[0,1] element, prob = 2.12854e-13    (123.812,165.823)-(359.565,480) batch id : 0
[1,1] element, prob = 2.61669e-15    (0,107.268)-(292.136,480) batch id : 0
[2,1] element, prob = 1.42357e-15    (120.562,154.796)-(156.424,288.469) batch id : 0
[3,1] element, prob = 6.25297e-18    (58.8543,238.316)-(338.631,480) batch id : 0
[4,1] element, prob = 4.76809e-18    (132.787,0)-(422.402,480) batch id : 0
[5,1] element, prob = 1.77602e-18    (116.802,206.061)-(156.071,251.131) batch id : 0
[6,2] element, prob = 1.04178e-11    (126.968,261.195)-(394.429,480) batch id : 0
[7,2] element, prob = 2.00192e-13    (0,230.624)-(354.162,480) batch id : 0
[8,2] element, prob = 7.30089e-14    (120.606,205.975)-(160.794,295.066) batch id : 0
[9,2] element, prob = 1.23133e-15    (122.605,158.209)-(477.77,480) batch id : 0
[10,2] element, prob = 9.34242e-16    (61.8106,348.39)-(405.385,480) batch id : 0
[11,2] element, prob = 3.56564e-16    (116.472,218.396)-(161.242,253.19) batch id : 0
[12,2] element, prob = 9.96831e-17    (27.8106,161.473)-(388.233,442.699) batch id : 0
[13,2] element, prob = 9.81133e-17    (0,45.3598)-(391.25,313.775) batch id : 0
[14,2] element, prob = 7.91892e-17    (185.71,339.464)-(446.858,480) batch id : 0
[15,2] element, prob = 4.81994e-17    (116.275,233.204)-(161.118,269.282) batch id : 0
[16,2] element, prob = 3.69921e-17    (171.694,69.3493)-(478.555,386.009) batch id : 0
[17,3] element, prob = 4.59317e-13    (162.877,238.344)-(425.884,480) batch id : 0
[18,3] element, prob = 4.77153e-15    (50.9165,200.037)-(385.176,480) batch id : 0
[19,3] element, prob = 2.36068e-15    (126.184,179.332)-(167.018,277.801) batch id : 0
[20,3] element, prob = 1.38449e-17    (195.662,133.574)-(521.794,480) batch id : 0
[21,3] element, prob = 1.05162e-17    (112.739,329.358)-(440.427,480) batch id : 0
[22,3] element, prob = 7.38237e-18    (11.7056,219.898)-(268.58,480) batch id : 0
[23,3] element, prob = 2.34051e-18    (122.938,212.778)-(169.502,247.384) batch id : 0
[24,3] element, prob = 1.29042e-18    (50.0903,37.843)-(428.754,297.894) batch id : 0
[25,4] element, prob = 1.61207e-11    (149.604,285.675)-(392.811,480) batch id : 0
[26,4] element, prob = 4.13695e-13    (124.327,205.218)-(160.826,282.79) batch id : 0
[27,4] element, prob = 3.85257e-13    (19.1442,254.324)-(343.899,480) batch id : 0
[28,4] element, prob = 2.01751e-15    (89.1127,384.477)-(405.261,480) batch id : 0
[29,4] element, prob = 1.54699e-15    (163.089,187.954)-(485.818,480) batch id : 0
[30,4] element, prob = 1.24294e-15    (0,283.709)-(235.351,480) batch id : 0
[31,4] element, prob = 6.87609e-16    (120.085,220.54)-(161.432,249.549) batch id : 0
[32,4] element, prob = 2.47142e-16    (24.1514,356.801)-(353.835,480) batch id : 0
[33,4] element, prob = 2.20097e-16    (9.24042,76.3342)-(388.134,312.018) batch id : 0
[34,4] element, prob = 1.87441e-16    (206.554,371.782)-(450.717,480) batch id : 0
[35,4] element, prob = 1.42175e-16    (54.8573,179.266)-(395.095,425.773) batch id : 0
[36,4] element, prob = 5.3766e-17    (119.469,235.226)-(161.887,266.29) batch id : 0
[37,5] element, prob = 4.39533e-13    (136.879,255.635)-(434.27,480) batch id : 0
[38,5] element, prob = 6.78363e-15    (7.06684,225.321)-(391.708,480) batch id : 0
[39,5] element, prob = 5.44006e-15    (124.618,188.223)-(166.957,274.034) batch id : 0
[40,5] element, prob = 1.55863e-17    (147.071,151.918)-(535.834,480) batch id : 0
[41,5] element, prob = 1.3509e-17    (63.2475,356.689)-(459.268,480) batch id : 0
[42,5] element, prob = 5.75718e-18    (120.636,216.432)-(170.542,244.558) batch id : 0
[43,6] element, prob = 2.63384e-13    (118.786,267.064)-(391.094,480) batch id : 0
[44,6] element, prob = 3.50807e-15    (119.228,198.197)-(161.134,275.565) batch id : 0
[45,6] element, prob = 3.40586e-15    (0,242.705)-(339.473,480) batch id : 0
[46,6] element, prob = 1.05949e-17    (38.5261,375.633)-(391.213,480) batch id : 0
[47,6] element, prob = 8.63083e-18    (107.123,186.956)-(464.447,480) batch id : 0
[48,6] element, prob = 6.98546e-18    (0,273.074)-(232.312,480) batch id : 0
[49,6] element, prob = 4.77018e-18    (113.739,220.207)-(163.353,245.788) batch id : 0
[ INFO ] Image out_0.bmp created!

total inference time: 1697.92
Average running time of one iteration: 1697.92 ms

Throughput: 0.588955 FPS

[ INFO ] Execution successful

[Severine_H_Intel]

Hi Chard,

For comparison, which result do you expect from your network on this image? What is the inference results through TF? 

Another question to get more context also, have you trained manually this network or is it from the TF zoo? 

Best, 

Severine

[Chard]

Hi Severine,

I trained this network by myself, and TF log is attached.

I expect there are 6 objects and all probability are 99%.

[Severine_H_Intel]

Hi Chard, 

we would like to examine your network in more details, would you be ok to send it via PM ? (Send Author A Message on top of my answer).

Best, 

Severine

[Chard]

Hi Severine,

I can not find out how to send you a PM. (Send Author A Message on top of my answer ==> I can not find it).

What files do you need to examine?

[Severine_H_Intel]

I sent you a PM, I think you should be able to read it ( look at the top of the page where your username is displayed, the inbox should be there.) 

[Severine_H_Intel]

Hi Chard, 

it just came to my mind that the error is maybe due to channel inversion. With the OpenVINO R3 release, we must also pass the parameter --reverse_input_channels to the MO. In general, check if you are passing the image with the channels order, usually, OpenCV uses with BGR order, while OpenVINO samples expect RGB.

Best, 

Severine

[Chard]

Hi Severine,

I tested passing the parameter --reverse_input_channels to the MO, the result is the same.
I also checked channels order, it is correct.

BTW, have you download the files I mentioned in PM?