Verify the correctness of detection_output API in SSD.

[qingqing01]

• Target: this verification is to make the detection consistent with Caffe.

• Method: similar to Verify the correctness of SSD loss. #8626 (comment)

• The main code of fluid is as follows, which will check the output label, scores, boxes and output LoD.

  import paddle.v2 as paddle
  import paddle.fluid as fluid
  import numpy as np
  
  def load(file_name, shape):
      with open(file_name, 'rb') as f:
          return np.fromfile(f, dtype=np.float32).reshape(shape)
  
  def load_data(loc_shape, conf_shape, prior_shape, out_shape):
      loc = load('detection_data/detect_loc_data', loc_shape)
      conf = load('detection_data/detect_conf_data', conf_shape)
      prior = load('detection_data/detect_prior_data', prior_shape)
      pb = prior[0, :]
      pbv = prior[1, :]
  
      decode_box = load('detection_data/detect_decode_bboxes_data', loc_shape)
      out = load('detection_data/detection_output_data', out_shape)
  
      return loc, conf, pb, pbv, decode_box, out
  
  def detection_output_check():
      p_num = 8732
      num_class = 21
  
      conf = fluid.layers.data(name='conf', shape=[p_num, num_class], dtype='float32')
      loc = fluid.layers.data(name='loc', shape=[p_num, 4], dtype='float32')
      prior = fluid.layers.data(name='prior', shape=[p_num, 4], dtype='float32', append_batch_size=False)
      prior_var = fluid.layers.data(name='prior_var', shape=[p_num, 4], dtype='float32', append_batch_size=False)
  
      out = fluid.layers.detection_output(loc, conf, prior, prior_var, nms_threshold=0.45)
      out.persistable = True
  
      place = fluid.CPUPlace()
      exe = fluid.Executor(place)
  
      loc_shape = [-1, p_num, 4]
      conf_shape = [-1, p_num, num_class]
      proir_shape = [2, p_num, 4]
      out_shape = [-1, 7]
  
      loc_t, conf_t, pb_t, pbv_t, decode_box, expect_out = load_data(loc_shape, conf_shape, proir_shape, out_shape)
  
      feeding = {'loc': loc_t, 'conf': conf_t, 'prior': pb_t, 'prior_var': pbv_t}
      #print fluid.default_main_program()
      out_v, = exe.run(fluid.default_main_program(), feed=feeding, fetch_list=[out], return_numpy=False)
  
      out_arr = np.array(out_v)
      out_var = fluid.global_scope().find_var('detection_output_0.tmp_0').get_tensor()
      fluid_lod = out_var.lod()
  
      fluid_label = out_arr[:, 0]
      fluid_score = out_arr[:, 1]
      fluid_box = out_arr[:, 2:6]
  
      caffe_iterm = expect_out[:, 0].astype(int).tolist()
      caffe_label = expect_out[:, 1]
      caffe_score = expect_out[:, 2]
      caffe_box = expect_out[:, 3:7]
  
      batch_size = max(caffe_iterm) + 1 # iterm index is from 0
      lod = [0 for i in range(batch_size + 1)]
      for i in range(batch_size):
          lod[i + 1] = lod[i] + caffe_iterm.count(i)
  
      score_diff = caffe_score - fluid_score
      box_diff = caffe_box - fluid_box
  
      caffe_lod = np.array(lod).astype(int)
      fluid_lod = np.array(fluid_lod).astype(int)
  
      assert caffe_label.any() == fluid_label.any()
      assert caffe_lod.any() == fluid_lod[0].any()
      print 'the shape of fluid output', out_arr.shape
      print 'the shape of caffe output', expect_out.shape
      print 'score diff', score_diff[score_diff > 0.]
      print 'box diff', box_diff[box_diff > 0.]
  
  if __name__ == '__main__':
      detection_output_check()

• Results:

  • batch_size=2
  • one mini-batch testing data
  • same inputs, the outputs of above code are:

  the shape of fluid output (400, 6)
  the shape of caffe output (400, 7)
  score diff []
  box diff [  2.98023224e-08   7.45058060e-09   5.96046448e-08   7.45058060e-09
  7.45058060e-09   7.45058060e-09   7.45058060e-09   7.45058060e-09
  3.72529030e-09   2.98023224e-08   3.72529030e-09   2.98023224e-08
  7.45058060e-09   2.98023224e-08   7.45058060e-09   7.45058060e-09
  7.45058060e-09   3.72529030e-09   5.96046448e-08   7.45058060e-09
  1.49011612e-08   5.96046448e-08   3.72529030e-09   2.98023224e-08
  2.98023224e-08   5.96046448e-08   7.45058060e-09   2.98023224e-08
  7.45058060e-09   1.49011612e-08   3.72529030e-09   1.49011612e-08
  2.98023224e-08   2.98023224e-08   5.96046448e-08   2.98023224e-08
  2.98023224e-08   2.98023224e-08   2.98023224e-08   2.98023224e-08
  2.98023224e-08   5.96046448e-08   2.98023224e-08   2.98023224e-08
  3.72529030e-09   2.98023224e-08   1.49011612e-08   5.96046448e-08
  7.45058060e-09   2.98023224e-08   7.45058060e-09   3.72529030e-09
  3.72529030e-09   7.45058060e-09   5.96046448e-08   1.49011612e-08
  5.96046448e-08   1.49011612e-08   1.49011612e-08   1.49011612e-08
  3.72529030e-09   7.45058060e-09]
  

  From this results, we can see that there is no diff for output label index and score and LoD. Some boxes have diff, but the absolute error is about 0.000006 % ~ 0.000007 %, this is very small.

• Conlusion

  There is no problem of detection_output API, which is consistent with Caffe.