Benchmark task: Full-finetuning¶
Using a self-supervised model to initialize a network and finetune the full weights on the target task is a very common evaluation protocol.
This benchmark requires only initializing the model – no other settings in MODEL.FEATURE_EVAL_SETTINGS
are needed unlike other benchmark tasks.
Benchmark: ImageNet-1k¶
The configuration for full fine-tuning on Imagenet is available in benchmark/fulltune/imagenet1k and can be run as follows:
python tools/run_distributed_engines.py \
config=benchmark/fulltune/imagenet1k/eval_resnet_8gpu_transfer_in1k_fulltune \
config.MODEL.WEIGHTS_INIT.PARAMS_FILE=<my_weights.torch>
Configurations for fine-tuning on a sub-set of Imagenet are also available and can be run as follows:
# For fine-tuning on 1% of the dataset
python tools/run_distributed_engines.py \
config=benchmark/fulltune/imagenet1k/eval_resnet_8gpu_transfer_in1k_fulltune \
+config/benchmark/fulltune/imagenet1k/dataset=imagenet1k_1percent \
config.MODEL.WEIGHTS_INIT.PARAMS_FILE=<my_weights.torch>
# For fine-tuning on 10% of the dataset
python tools/run_distributed_engines.py \
config=benchmark/fulltune/imagenet1k/eval_resnet_8gpu_transfer_in1k_fulltune \
+config/benchmark/fulltune/imagenet1k/dataset=imagenet1k_10percent \
config.MODEL.WEIGHTS_INIT.PARAMS_FILE=<my_weights.torch>
Benchmark: Places205¶
The configuration for full fine-tuning on Places205 is available in benchmark/fulltune/places205 and can be run as follows:
python tools/run_distributed_engines.py \
config=benchmark/fulltune/places205/eval_resnet_8gpu_transfer_places205_fulltune \
config.MODEL.WEIGHTS_INIT.PARAMS_FILE=<my_weights.torch>
Note
Please see VISSL documentation on how to run a given training on 1-gpu, multi-gpu or multi-machine.
Note
Please see VISSL documentation on how to use the builtin datasets if you want to run this benchmark on a different target task..
Note
Please see VISSL documentation on how to use YAML comfiguration system in VISSL to override specific components like model of a config file. For example, in the above file, user can replace ResNet-50 model with a different architecture like RegNetY-256 etc. easily.