CL_Thesis

Latent Replay for Continual Learning on Edge devices with Efficient Architectures.

Models

Available models:

• MobileNetV1:
  • mobilenetv1
• MobileNetV2:
  • mobilenetv2
  • 0.75_mobilenetv2
  • 0.5_mobilenetv2
• PhiNet:
  • phinet_2.3_0.75_5
  • phinet_1.2_0.5_6_downsampling
  • phinet_0.8_0.75_8_downsampling
  • phinet_1.3_0.5_7_downsampling
  • phinet_0.9_0.5_4_downsampling_deep
  • phinet_0.9_0.5_4_downsampling

Benchmarks

Available benchmarks:

• Split CIFAR 10:
  • split_cifar10
• CORe50:
  • core50
• Split MNIST:
  • split_mnist

Training

Expirience Replay

python ExpReplay.py --model_name phinet_0.8_0.75_8_downsampling --benchmark_name split_cifar10 --lr 0.0001 --latent_layer 1 --train_epochs 4 --rm_size 1500 --weight_decay 0 --split_ratio 0 --device 0

Latent Replay in Elements

python LatentReplay.py --model_name phinet_0.8_0.75_8_downsampling --benchmark_name split_cifar10 --lr 0.0001 --latent_layer 9 --train_epochs 4 --rm_size 1500 --weight_decay 0 --split_ratio 0 --device 0"

Latent Replay in MB

python LatentReplay.py --model_name phinet_0.8_0.75_8_downsampling --benchmark_name split_cifar10 --lr 0.0001 --latent_layer 9 --train_epochs 4 --rm_size_MB 0.5 --weight_decay 0 --split_ratio 0 --device 0"