light_head_rcnn
light_head_rcnn copied to clipboard
zengarden
the predict procedure don't support multi-batch
the predict procedure don't support multi-batch, and we test your code on Nvidia 1080Ti GPU which only achieve 2 images per seconds. Far lower than original paper 100 FPS.
Hi, About speed: We test Light-Head with resnet-101 on TITAN XP, it yields 40mAP at 10FPS. Following a common practice, testing time is only counting on network forward~(exclude post-processing). Maybe u need to modify test.py as follows to see a more reasonable time in testing. BTW, Tensorflow is slow at starting up, please check the detection speed after enough testing iteration. My testing log is shown in the attachment. result.txt
diff --git a/experiments/lizeming/light_head_rcnn.ori_res101.coco.ps_roialign/test.py b/experiments/lizeming/light_head_rcnn.ori_res101.coco.ps_roialign/test.py
index a48e336..7f67a4a 100644
--- a/experiments/lizeming/light_head_rcnn.ori_res101.coco.ps_roialign/test.py
+++ b/experiments/lizeming/light_head_rcnn.ori_res101.coco.ps_roialign/test.py
@@ -12,6 +12,7 @@ from config import cfg, config
import argparse
import dataset
+import time
import os.path as osp
import network_desp
import tensorflow as tf
@@ -64,10 +65,10 @@ def inference(val_func, inputs, data_dict):
feed_dict = {inputs[0]: resized_img[None, :, :, :], inputs[1]: im_info}
- #st = time.time()
+ st = time.time()
_, scores, pred_boxes, rois = val_func(feed_dict=feed_dict)
- #ed = time.time()
- #print(ed -st)
+ ed = time.time()
+ print(ed -st)
boxes = rois[:, 1:5] / scale
About multi-batch testing: I think it is not much hard to modify test.py to support multi-batch testing. BTW, We yield 100FPS with a small backbone network(xception like 145M FLOPs), not ResNet-101 7.6G FLOPs.
@zengarden I have changed the test.py for multi-batch,but I get a problem. How can I distinguish the output of the net to different image.the output of net don't show which image the box belong to? How can I konw it in multi-batch testing?
@zengarden For example, I change the test.py and batch_size is 2, but the output of net ,cls_score, is (1001,81).how can I know the order?each of this belong to which image?
@MaskVulcan For multi-batch testing, first you need to set attribute test_batch_per_gpu to number_of_batches(e.g. 2) in config.py. Then modify test.py to provide feed_dict with multi_image. If u finished them, scores and rois should have the shape of [2000, 81], [2000, 5]. The first dimension of rois is its batch_index.
@zengarden I get this. And I change the batch size to 2.and I find my output is not very good, The shape of rois is always changing. I print it, just like(2000,5),(577,5)(1844,5).......I am a little confused. I copy the code of get_data_for_singlegpu in your dataset.py to feed multi-batch images to test.
@MaskVulcan Since there are various number of boxes in each image, the shape of the rois may be changeable . The first column of rois is its batch_index, which is helpful for specifying rois of each image.
@zengarden I have changed the code. I copy the code in your dataset.py for train dataset and use it for test. But I find that when I set batch size to 2, the speed turns to 0.06s/img, map 26. But when I change it to 8, the speed is still 0.06s/img, map turns to 34. It is very strange. Is there any other code I need to change?
I ran test.py on gtx1070 and got a test speed of 0.18s/img. Is there any problem?
@MaskVulcan have you finished mutil-batch on prediction? when use mutil-batch, the speed is bigger than one-batch? how many?
