pytorch-retraining

Transfer Learning shootout for PyTorch's model zoo (torchvision).

• Load any pretrained model with custom final layer (num_classes) from PyTorch's model zoo in one line

model_pretrained, diff = load_model_merged('inception_v3', num_classes)

• Retrain minimal (as inferred on load) or a custom amount of layers on multiple GPUs. Optionally with _Cyclical Learning Rate_ (Smith 2017).

final_param_names = [d[0] for d in diff]stats = train_eval(model_pretrained, trainloader, testloader, final_params_names)

• Chart training_time, evaluation_time (fps), top-1 accuracy for varying levels of retraining depth (shallow, deep and from scratch)

|  | |:---:| | Transfer learning on example dataset Bee vs Ants with 2xK80 GPUs|

Results on more elaborate Dataset

num_classes = 23, slightly unbalanced, high variance in rotation and motion blur artifacts with 1xGTX1080Ti

|  | |:---:| | Constant LR with momentum |

|  | |:---:| | Cyclical Learning Rate |