update log.

infer_benchmark.md

@@ -1,17 +1,43 @@
 # Xavier Nano Benchmark
-The latency is calculated between transfering input to gpu and fetch results back on cpu.
+The latency is calculated between transfering input to gpu and fetch results back on cpu with fp16.
 
-Network| Resolution  | Latency (ms) | GFlops
+Network| Resolution  | Latency (ms) 
 |----|----|----|
 DDRNet23-Slim_TRT(Backbone) | 768x384 | 33  |
 DDRNet23-Slim_TRT(Backbone)-Torch(Interpolate+Softmax+Max) | 768x384 | 54 
 DDRNet23-Slim_TRT(Backbone)-Torch(Interpolate+Softmax) | 768x384 | 95 
 DDRNet23-Slim_TRT(Backbone+Interpolate+Softmax) | 768x384 | 100  
 DDRNet23-Slim_TRT(Backbone+Interpolate+Softmax)-Torch(Max) | 768x384 | 60  
-
 OCR-DDRNet23-Slim_TRT(Backbone)| 768x384 | 78 ms
 OCR-DDRNet23-Slim_TRT(Backbone)-Torch(Interpolate+Softmax)| 768x384 | 136  ms
 OCR-DDRNet23-Slim_TRT(Backbone)-Torch(Interpolate+Softmax+Max)| 768x384 | 100  ms
 OCR-DDRNet23-Slim_TRT(Backbone+Interpolate+Softmax)-Torch(Max)| 768x384 | 126  ms
 
+# Xavier NX Benchmark
+Network| Resolution  | Latency (ms) | GFlops | Mode 
+|----|----|----|
+OCR-HRNet_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 151  ms  | fp16
+OCR-HRNet_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 109  ms  | int8
+OCR-DDRNet23-Slim_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 31  ms | fp16
+OCR-DDRNet23-Slim_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 29  ms | int8
+OCR-DDRNet23-Slim_TRT(Backbone+Interpolate+Softmax)-Torch(Max)| 768x384 | 23  ms | fp16
+OCR-DDRNet23-Slim_Mscale_0.5_1.0_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 42  ms | fp16
+OCR-HRNet_W18_SMALL_V2_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 72  ms | fp16
+OCR-HRNet_W18_SMALL_V2_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 61  ms | int8
+
 
+# 3090 Benchmark
+Network| Resolution  | Latency (ms) | mIOU
+|----|----|----|
+OCR-HRNet_Mscale_0.5_1.0_2.0_TRT(Backbone+Interpolate+Softmax+Max)| 1280x720 | 120  ms|
+OCR-HRNet_Mscale_0.5_1.0_TRT(Backbone+Interpolate+Softmax+Max)| 1280x720 | 30  ms | 0.606
+OCR-HRNet_Mscale_0.5_1.0_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 15  ms|
+OCR-HRNet_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 10  ms |
+OCR-HRNet_TRT(Backbone+Interpolate+Softmax)-Torch(Max)| 768x384 | 10  ms |
+OCR-DDRNet23-Slim_TRT(Backbone+Interpolate+Softmax)-Torch(Max)| 768x384 | 2  ms |
+OCR-DDRNet23-Slim_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 |  2  ms |
+HRNet_lite_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 6  ms |
+Naive_HRNet_lite_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 4  ms |
+OCR-HRNet_W18_SMALL_V2_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 5  ms |
+OCR-HRNet_lite_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 8  ms |
+OCR-Naive_HRNet_lite_TRT(Backbone+Interpolate+Softmax+Max)| 768x384 | 6  ms |

training_log.md

@@ -25,6 +25,9 @@ HRNet-OCR-dm-train | 2177 (flip,0.5,1.0,2.0) |0.69360286
 HRNet-OCR-dm-train  | 960 (flip,0.5,1.0,2.0) | 0.69965035
 HRNet-OCR-dm-train  | 768 (flip,0.5,1.0,2.0) | 0.6981542
 ddrnet23_slim_ocr                    | 1024 (0.5,1.0,2.0) | 0.6289
-ddrnet23_slim_ocr(use above weights w/o retrain) | 1024(1.0) | 59.44
+ddrnet23_slim_ocr(use above weights w/o retrain) | 1024(1.0) | 0.5944
 ddrnet23_slim | 1024(1.0) | 0.5391523
 ddrnet23_slim_augment |  1024(1.0) | 0.5040767
+naive_lite_hrnet_ocr               | 1024 (0.5,1.0,2.0) | 0.5454124
+lite_hrnet_ocr               | 1024 (0.5,1.0,2.0) | 0.5753439
+HRNet-OCR-dm-train-wo-rmi-loss | 2177 (flip,0.5,1.0,2.0) |0.6662466