Pytorch implementation of Prototypical Networks Paper
To reproduce the working environment for this project create a new conda environment with the following command:
conda env create -f environment.yaml #environment with cuda dependencies
# to create an environment without cuda: conda env create -f environment-cpuonly.yaml
If you don't have the command conda install it from anaconda or miniconda
| Dataset | Train num classes | Train support samples | Train query samples | Test num classes | Test support samples | Test query samples | Paper Accuracy | Test accuracy | Test loss |
|---|---|---|---|---|---|---|---|---|---|
| cub | 50 | 0 | 10 | 50 | 0 | 10 | 54.60% | 53.87% | 13.5089 |
| miniimagenet | 20 | 5 | 15 | 5 | 1 | 1 | 49.42% | 45.35% | 28.6325 |
| miniimagenet | 20 | 5 | 15 | 5 | 5 | 5 | 68.20% | 66.60% | 13.6065 |
| miniimagenet | 30 | 1 | 15 | 5 | 1 | 1 | 49.42% | 49.75% | 20.1425 |
| miniimagenet | 30 | 1 | 15 | 5 | 5 | 5 | 68.20% | 63.91% | 15.5581 |
| omniglot | 60 | 1 | 5 | 5 | 1 | 1 | 98.80% | 98.73% | 0.809822 |
| omniglot | 60 | 1 | 5 | 5 | 5 | 5 | 99.70% | 99.67% | 0.219201 |
| omniglot | 60 | 1 | 5 | 20 | 1 | 1 | 96.00% | 95.74% | 2.65106 |
| omniglot | 60 | 1 | 5 | 20 | 5 | 5 | 98.90% | 99.01% | 0.61366 |
| omniglot | 60 | 5 | 5 | 5 | 1 | 1 | 98.80% | 97.79% | 1.76968 |
| omniglot | 60 | 5 | 5 | 5 | 5 | 5 | 99.70% | 99.65% | 0.282353 |
| omniglot | 60 | 5 | 5 | 20 | 1 | 1 | 96.00% | 93.55% | 5.15215 |
| omniglot | 60 | 5 | 5 | 20 | 5 | 5 | 98.90% | 98.84% | 0.794082 |
- Download link of table checkpoints
To reproduce the results of Zero Shot CUB dataset download the preprocessed dataset at this link.
The archive contains GoogleLeNet visual features of CUB images and class features.
To resume an interrupted training use the argument --run-path of train.py.
An example is: python train.py --run-path=run/0
Project structure is very straightforward and flatten to maintain simplicity.
├── data # where data will be downloaded
│ └── omniglot #contains omniglot data
├── environment-cpuonly.yaml #conda cpu environmnet
├── environment.yaml # conda environmnet to run this project
├── logger.py # logging function with tensorboard
├── Makefile
├── model.py
├── paper.pdf
├── paths.py
├── readme.md
├── run # where experiments will be stored
│ └── 0 # experiments are stored using index based folder name
├── test.py
├── trainer.py
├── train.py
└── utils.py
If you want to train by yourself the model you can train directly by running python3 train.py.
The script supports many arguments which are:
usage: train.py [-h] --dataset {omniglot,miniimagenet,cub} --classes NUM_CLASSES --support-samples SUPPORT_SAMPLES [--query-samples QUERY_SAMPLES] [--distance {euclidean,cosine}] [--epochs EPOCHS]
[--epoch-steps EPOCH_STEPS] [--seed SEED] [--device DEVICE] [--batch-size BATCH_SIZE] [--eval-steps EVAL_STEPS] [--run-path RUN_PATH]
optional arguments:
-h, --help show this help message and exit
--dataset {omniglot,miniimagenet,cub}, -d {omniglot,miniimagenet,cub}
Specify train dataset
--classes NUM_CLASSES, --num-classes NUM_CLASSES, -c NUM_CLASSES
Number of classes for each task in meta learning i.e. the N in N-way with K shots
--support-samples SUPPORT_SAMPLES, -s SUPPORT_SAMPLES
Number of training samples for each class in meta learning i.e. the K in N-way with K shots
--query-samples QUERY_SAMPLES, -q QUERY_SAMPLES
Number of test samples for each class in meta learning
--distance {euclidean,cosine}, --dst {euclidean,cosine}
Distance function to use inside PrototypicalNetwork
--epochs EPOCHS, -e EPOCHS
Number of training epochs. Set by default to a very high value because paper specify that train continues until validation loss continues to decrease.
--epoch-steps EPOCH_STEPS
--seed SEED
--device DEVICE
--batch-size BATCH_SIZE
--eval-steps EVAL_STEPS
Number of evaluation steps. By default is set to the number of steps to reach 600 episodes considering batch size as paper reports
--run-path RUN_PATH Set to resume a checkpoint
You can also test by yourself running python3 test.py. It supports also many arguments which are
usage: test.py [-h] --run-path RUN_PATH [--support-size SUPPORT_SAMPLES] [--query-size QUERY_SAMPLES] [--num-classes NUM_CLASSES] [--batch-size BATCH_SIZE] [--seed SEED] [--steps STEPS] [--device DEVICE]
optional arguments:
-h, --help show this help message and exit
--run-path RUN_PATH, -r RUN_PATH
--support-size SUPPORT_SAMPLES, -s SUPPORT_SAMPLES
--query-size QUERY_SAMPLES, -q QUERY_SAMPLES
--num-classes NUM_CLASSES, --nc NUM_CLASSES
--batch-size BATCH_SIZE, -b BATCH_SIZE
--seed SEED
--steps STEPS