From 0b7a03168ddcca3d99c35df50a2f286b5c5e6812 Mon Sep 17 00:00:00 2001 From: Thien Tran Date: Thu, 6 Jan 2022 20:25:01 +0800 Subject: [PATCH] update README - add info about webdataset - add VoVNet-57 results - add details about torchvision models --- README.md | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index 3786f18..b4bc461 100644 --- a/README.md +++ b/README.md @@ -27,6 +27,8 @@ python ./scripts/wds.py --data_dir ./ImageNet/ILSVRC/Data/CLS-LOC/train --save_d python ./scripts/wds.py --data_dir ./ImageNet/ILSVRC/Data/CLS-LOC/val --save_dir ./ImageNet/webdataset/val --shuffle False ``` +There should be 147 shards of training set, and 7 shards of validation set. Each shard is 1GB. + Reference training recipe: - https://github.com/pytorch/vision/blob/main/references/classification/train.py @@ -93,7 +95,7 @@ Backbone | Top-1 acc | #Params(M) | FLOPS(G)* ---------------|-----------|------------|---------- VoVNet-19-slim | 70.7 | 2.65 | 4.77 VoVNet-39 | 78.1 | 25.18 | 15.57 -VoVNet-57 | | 41.45 | 19.30 +VoVNet-57 | 79.2 | 41.45 | 19.30 VoVNet-99 | | 69.52 | 34.43 *FLOPS is measured with `(1,3,224,224)` input. @@ -102,6 +104,9 @@ VoVNet-99 | | 69.52 | 34.43 Some torchvision classification models are ported to use with the toolbox. They can output multiple feature map levels. +- For MobileNet and EfficientNet models, intermediate outputs are taken after the first 1x1 conv expansion layer of the strided MBConv block. See Section 6.2 of [MobileNetv2 paper](https://arxiv.org/abs/1801.04381) and Section 6.3 of [MobileNetv3 paper](https://arxiv.org/abs/1905.02244). +- To use weights from the new PyTorch training recipe, go to torchvision's [prototype](https://github.com/pytorch/vision/tree/main/torchvision/prototype/models) directory and copy weights URLs (labelled as `ImageNet1K_V2`) from their respective models' files. + ResNet: - ResNet-{18,34,50,101,152}