Refine documentation format for ResNet-101 Accuracy Benchmark

kakaobrain · Sep 20, 2019 · 4004538 · 4004538
1 parent 1c0a887
commit 4004538
Show file tree

Hide file tree

Showing 2 changed files with 18 additions and 21 deletions.
diff --git a/docs/benchmarks.rst b/docs/benchmarks.rst
@@ -27,13 +27,13 @@ ResNet-101 Accuracy Benchmark
 -----------------------------
 
 ================  ===============
-Experiment        top-1 error (%)
+Experiment        Top-1 error (%)
 ================  ===============
-dataparallel-256  22.02&pm;0.11
-dataparallel-1k   22.04&pm;0.24
-pipeline-256      21.99&pm;0.13
-pipeline-1k       22.24&pm;0.19
-pipeline-4k       22.13&pm;0.09
+dataparallel-256       22.02±0.11
+dataparallel-1k        22.04±0.24
+pipeline-256           21.99±0.13
+pipeline-1k            22.24±0.19
+pipeline-4k            22.13±0.09
 ================  ===============
 
 The code is reproducible on Tesla P40 GPUs, and the experiment details

diff --git a/examples/resnet101_accuracy_benchmark/README.md b/examples/resnet101_accuracy_benchmark/README.md
@@ -1,24 +1,22 @@
 # ResNet-101 Accuracy Benchmark
 
 This example reproduces accuracy benchmark on ResNet-101, as stated by
-reported in Table 2(c) of the [Accurate, Large Minibatch SGD][] paper.
+reported in Table 2(c) of the [Accurate, Large Minibatch SGD][paper] paper.
 
 Every experiment setting is optimized for Tesla P40 GPUs.
 
 ## Result
 
-Experiment              | num gpus | kn  | learning rate | top-1 error (%) | throughput (samples/sec) | speed up
------------------------ | --------:| ---:|-------------: | ---------------:|-------------------------:|---------:
-reference-256 [paper][] |        8 | 256 |           0.1 |   22.08&pm;0.06 |                      N/A |      N/A
-reference-8k [paper][]  |      256 |  8k |           3.2 |   22.36&pm;0.09 |                      N/A |      N/A
-dataparallel-256        |        2 | 256 |           0.1 |   22.02&pm;0.11 |                  180.344 |   1.000x
-dataparallel-1k         |        8 |  1k |           0.4 |   22.04&pm;0.24 |                  606.916 |   3.365x
-dataparallel-4k         |        8 |  4k |           1.6 |             OOM |                      N/A |      N/A
-pipeline-256            |        2 | 256 |           0.1 |   21.99&pm;0.13 |                  117.432 |   0.651x
-pipeline-1k             |        8 |  1k |           0.4 |   22.24&pm;0.19 |                  294.739 |   1.634x
-pipeline-4k             |        8 |  4k |           1.6 |   22.13&pm;0.09 |                  378.746 |   2.100x
-
-
+Experiment                | GPUs | Batch size | Learning rate | Top-1 error (%) | Throughput          | Speedup
+------------------------- | ---: | ---------: | ------------: | --------------: |-------------------: | ------:
+reference-256 ([paper][]) |    8 |        256 |           0.1 |      22.08±0.06 |                 N/A |     N/A
+reference-8k ([paper][])  |  256 |         8k |           3.2 |      22.36±0.09 |                 N/A |     N/A
+dataparallel-256          |    2 |        256 |           0.1 |      22.02±0.11 | 180.344 samples/sec |  1.000x
+dataparallel-1k           |    8 |         1k |           0.4 |      22.04±0.24 | 606.916 samples/sec |  3.365x
+dataparallel-4k           |    8 |         4k |           1.6 |   Out of memory |                 N/A |     N/A
+pipeline-256              |    2 |        256 |           0.1 |      21.99±0.13 | 117.432 samples/sec |  0.651x
+pipeline-1k               |    8 |         1k |           0.4 |      22.24±0.19 | 294.739 samples/sec |  1.634x
+pipeline-4k               |    8 |         4k |           1.6 |      22.13±0.09 | 378.746 samples/sec |  2.100x
 
 ## Optimized Environment
 
@@ -38,6 +36,7 @@ $ pip install -r requirements.txt
 ```
 
 Prepare ImageNet dataset at `./imagenet`:
+
 ```sh
 $ python -c "import torchvision; torchvision.datasets.ImageNet('./imagenet', split='train', download=True)"
 $ python -c "import torchvision; torchvision.datasets.ImageNet('./imagenet', split='val', download=True)"
@@ -54,6 +53,4 @@ $ python main.py gpipe-8       # gpipie 1k
 $ python main.py gpipe-8-4k    # gpipie 4k
 ```
 
-
-[Accurate, Large Minibatch SGD]: https://arxiv.org/abs/1706.02677
 [paper]: https://arxiv.org/abs/1706.02677