search

ApolloAuto / apollo

An open autonomous driving platform
Apache License 2.0
25.19k stars 9.71k forks source link

CAMERATEST:Could not find tensor res2a_branch2b_scale2a_branch2b in tensorScales. #8243

Closed gengqx closed 5 years ago

gengqx commented 5 years ago

base on current cuda8 docker ,update to cuda10,or ubuntu18.04 docker, modified 

modules/perception/inference/tensorrt/rt_net.h

add RTX 2080 ,RTX 2080 Ti device, the camera test could not pass

I0509 15:49:34.664520 30000 omt_obstacle_tracker.cc:41] [] load omt parameters from /apollo/modules/perception/production/data/perception/camera/models/omt_obstacle_tracker/config.pt 
img_capability: 14 
lost_age: 3 
reserve_age: 5 
border: 20 
target_thresh: 0.6 
correct_type: 0
I0509 15:49:34.669407 30000 multicue_obstacle_transformer.cc:35] [] Load transformer parameters from /apollo/modules/perception/production/data/perception/camera/models/multicue_obstacle_transformer/confi
g.pt 
min dimension: 0 
do template search: 1
I0509 15:49:34.669517 30000 location_refiner_obstacle_postprocessor.cc:40] [] Load postprocessor parameters from /apollo/modules/perception/production/data/perception/camera/models/location_refiner_obstac
le_postprocessor/config.pt 
min_dist_to_camera: 30 
roi_h2bottom_scale: 0.5
I0509 15:49:34.669528 30000 obstacle_camera_perception.cc:133] [] No feature config found
I0509 15:49:34.669536 30000 obstacle_camera_perception.cc:201] [] lane_detector_name: DenselineLaneDetector
I0509 15:49:34.669765 30000 denseline_lane_detector.cc:40] [] denseline param: model_param {
  model_name: "denseline"
  proto_file: "deploy.prototxt"
  weight_file: "deploy.caffemodel"
  resize_scale: 0.8
  input_offset_y: 440
  input_offset_x: 0
  crop_height: 640
  crop_width: 1920
  mean_b: 95
  mean_g: 99
  mean_r: 96
  is_bgr: true
  model_type: "RTNetInt8"
}
net_param {
  in_blob: "data"
  out_blob: "conv_out"
  internal_blob_int8: "reduce1_add"
  internal_blob_int8: "reduce2_add"
  internal_blob_int8: "reduce3_add"
  internal_blob_int8: "reduce4_add"
  internal_blob_int8: "res6b_input_64"
  internal_blob_int8: "res5b_input_32"
  internal_blob_int8: "res4d_input_16"
  internal_blob_int8: "res3c_input_8"
}
I0509 15:49:34.669775 30000 denseline_lane_detector.cc:61] [] input_height: 1080
I0509 15:49:34.669780 30000 denseline_lane_detector.cc:62] [] input_width: 1920
I0509 15:49:34.670006 30000 denseline_lane_detector.cc:93] [] GPU: GeForce RTX 2080
I0509 15:49:34.670019 30000 denseline_lane_detector.cc:103] [] net input blobs: data
I0509 15:49:34.670022 30000 denseline_lane_detector.cc:106] [] net output blobs: conv_out
I0509 15:49:34.670025 30000 denseline_lane_detector.cc:106] [] net output blobs: reduce1_add
I0509 15:49:34.670027 30000 denseline_lane_detector.cc:106] [] net output blobs: reduce2_add
I0509 15:49:34.670029 30000 denseline_lane_detector.cc:106] [] net output blobs: reduce3_add
I0509 15:49:34.670032 30000 denseline_lane_detector.cc:106] [] net output blobs: reduce4_add
I0509 15:49:34.670034 30000 denseline_lane_detector.cc:106] [] net output blobs: res6b_input_64
I0509 15:49:34.670037 30000 denseline_lane_detector.cc:106] [] net output blobs: res5b_input_32
I0509 15:49:34.670039 30000 denseline_lane_detector.cc:106] [] net output blobs: res4d_input_16
I0509 15:49:34.670042 30000 denseline_lane_detector.cc:106] [] net output blobs: res3c_input_8
I0509 15:49:34.670043 30000 denseline_lane_detector.cc:110] [] model_type: RTNetInt8
I0509 15:49:34.869321 30000 denseline_lane_detector.cc:126] [] input_reshape: 1, 3, 512, 1536
I0509 15:49:34.904882 30000 rt_net.cc:648] [] Device Works on Int8 Mode.
I0509 15:49:34.949389 30000 rt_net.cc:30] [] ../builder/cudnnBuilder2.cpp (1508) - Misc Error in buildEngine: -1 (Could not find tensor res2a_branch2b_scale2a_branch2b in tensorScales.)
I0509 15:49:34.954591 30000 rt_net.cc:30] [] ../builder/cudnnBuilder2.cpp (1508) - Misc Error in buildEngine: -1 (Could not find tensor res2a_branch2b_scale2a_branch2b in tensorScales.)
external/bazel_tools/tools/test/test-setup.sh: line 169: 30000 Segmentation fault      (core dumped) "${TEST_PATH}" "$@"
KaWaiTsoiBaidu commented 5 years ago

Hi @gengqx. Thanks for the feedback. But the current version Apollo does not support RTX 2080 yet.

KaWaiTsoiBaidu commented 5 years ago

@gengqx Or could you try to set this line to be model_type: "CaffeNet" and see if that fix the problem?

gengqx commented 5 years ago

@KaWaiTsoiBaidu Yes,CaffeNet can pass the test ,maybe when the master code update to 18.04,it would need to release new tensorrt model to support Nvidia new arc: Volta

KaWaiTsoiBaidu commented 5 years ago

@gengqx We are working on that. Thanks for your feedback. Closing this issue.