Deep Learning Experiment Toolkit

Provides infrastructure to help you set up deep learning experiments in a second.
Acknowledgement: This toolkit is mainly adapted and modified from https://github.com/tribhuvanesh/knockoffnets/tree/master/knockoff

Usage & Explanations

• config.py: Configuration file for setting:

  • Random seed
  • Default path for datasets, model zoo, etc.
  • Default parameters (i.g. mean and std of a dataset)

• model_utils.py: Training and evaluation functions that are capable to:

  • Create log file for training and testing
  • Create informative checkpoint.
  • Checkpoint struction:
    • epoch
    • arch
    • state_dict
    • best_test_acc
    • optimizer
    • created_on

• datasets: Manage all the datasets and corresponding default transformations durint training and testing. Current available datasets:

  • MNIST
  • CIFAR10
  • ImageNet

• modelzoo: Manage all the models trained on different datasets. Current available datasets:

  • MNIST
  • CIFAR10
  • ImageNet

Example

import modelzoo.zoo as zoo
import datasets
import config as cfg

params = {"model_name":"resnet50",
          "dataset_name":"ImageNet1k", # CIFAR10 or MNIST
          "num_classes":1000,
          "pretrained":"imagenet"}

# Load model
model_name = params["model_name"]
dataset_name = params["dataset_name"]
modelfamily = datasets.dataset_to_modelfamily[dataset_name]
num_classes = params["num_classes"]
pretrained = params["pretrained"]
model = zoo.get_net(model_name, modelfamily, num_classes=num_classes, pretrained=pretrained)

# Load dataset

modelfamily = datasets.dataset_to_modelfamily[dataset_name]
dataset = datasets.__dict__[dataset_name]

# Load model checkpoint
checkpoint = torch.load(model_path)
start_epoch = checkpoint["epoch"]
best_test_acc = checkpoint["best_acc"]
model.load_state_dict(checkpoint["state_dict"])
optimizer.load_state_dict(checkpoint["optimizer"])