U tweetove putem weba ili aplikacija drugih proizvođača možete dodati podatke o lokaciji, kao što su grad ili točna lokacija. Povijest lokacija tweetova uvijek možete izbrisati. Saznajte više
I ran a few #PyTorch ResNet experiments over the holiday season. Testing my (re)implementation of Google’s AugMix, I trained a vanilla ResNet50 to 78.994% top-1, just ImageNet, 224 resolution. With test time mean-max pooling at 288x288, the same weights manage 80% top-1.
Another result, modifying the ‘deep stem’ from Bag of Tricks. Tiering the stem to (24, 48, 64) or (24, 32, 64) shows benefit over the D stem of (32, 32, 64). 77.6% D vs 78.0% for tiered on an SE-ResNeXt26-32x4d base. Tiered stem inspired by @jeremyphoward
This is excellent! Are you gonna release code ?
Pushed all the relevant code and weights yesterday to https://github.com/rwightman/pytorch-image-models … ... the hparams (cmd line) for the ResNet50 run will be added in a day or two.
I will! Curious, the impl I see in your github, you replaced the stem MaxPool with MaxBlurPool, and the ResBlock downsample AvgPool with MaxBlurPool, and leave global AvgPool as is. The paper suggests replacing AvgPool with BlurPool (non max). What did you try so far?
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.
