Simclr

config.yaml

{
"batch_size": 256,
"dataset": "STL10",
"epochs": 100,
"fp16": false,
"fp16_opt_level": "O2",
"logistic_batch_size": 256,
"logistic_epochs": 100,
"model_num": 40,
"model_path": "logs/0",
"normalize": true,
"optimizer": "Adam",
"projection_dim": 64,
"resnet": "resnet50",
"seed": 42,
"start_epoch": 0,
"temperature": 0.5,
"workers": 16
}

config.yaml:

train options
seed: 42 sacred handles automatic seeding when passed in the config
batch_size: 256
workers: 16
start_epoch: 0
epochs: 100
dataset: "STL10" STL10

model options
resnet: "resnet18"
normalize: True
projection_dim: 64 "[...] to project the representation to a 128-dimensional latent space"

loss options
optimizer: "Adam" or LARS (experimental)
temperature: 0.5 see appendix B.7.: Optimal temperature under different batch sizes

reload options
model_path: "logs/0" set to the directory containing `checkpoint_.tar`
model_num: 40 set to checkpoint number

mixed-precision training
fp16: False
fp16_opt_level: "O2"


logistic regression options
logistic_batch_size: 256
logistic_epochs: 100

A pre-trained SimCLR model with the following parameters:
"batch_size": 256,
"epochs": 40,
"n_out": 64,
"normalize": true,
"resnet": "resnet18",
"seed": 634715003,
"start_epoch": 0,

"workers": 16

Accuracy with a logistic regression classifier trained on top of SimCLR on STL-10 test set: 0.72