<div align="center">
<img src="https://raw.githubusercontent.com/crybot/potatorch/main/docs/potatorch-banner.png" width="100%" role="img">
**PotaTorch is a lightweight PyTorch framework specifically designed to run on hardware with limited resources.**
______________________________________________________________________
<!-- [](https://pepy.tech/project/potatorch) -->
[](https://pypi.org/project/potatorch/)
[](https://badge.fury.io/py/potatorch)
[](https://github.com/crybot/potatorch/blob/main/LICENSE)
</div>
### Installation
PotaTorch is published on PyPI, you can install it through pip:
```bash
pip install potatorch
```
or you can install it from sources:
```bash
git clone --single-branch -b main https://github.com/crybot/potatorch
pip install -e potatorch
````
______________________________________________________________________
### Minimal Working Example
You can run the following example directly from `examples/mlp.py` if you already have pytorch installed, or you can run it with docker through the provided scripts:
```bash
./build.sh && ./run.sh
```
The example trains a feed forward network on a toy problem:
```python
import torch
from torch import nn
from potatorch.training import TrainingLoop, make_optimizer
from potatorch.callbacks import ProgressbarCallback
from torch.utils.data import TensorDataset
# Fix a seed for TrainingLoop to make non-deterministic operations such as
# shuffling reproducible
SEED = 42
device = 'cuda'
epochs = 100
lr = 1e-4
# Define your model as a pytorch Module
model = nn.Sequential(nn.Linear(1, 128), nn.ReLU(),
nn.Linear(128, 128), nn.ReLU(),
nn.Linear(128, 1))
# Create your dataset as a torch.data.Dataset
dataset = TensorDataset(torch.arange(1000).view(1000, 1), torch.sin(torch.arange(1000)))
# Provide a loss function and an optimizer
loss_fn = torch.nn.MSELoss()
optimizer = make_optimizer(torch.optim.Adam, lr=lr)
# Construct a TrainingLoop object.
# TrainingLoop handles the initialization of dataloaders, dataset splitting,
# shuffling, mixed precision training, etc.
# You can provide callback handles through the `callbacks` argument.
training_loop = TrainingLoop(
dataset,
loss_fn,
optimizer,
train_p=0.8,
val_p=0.1,
test_p=0.1,
random_split=False,
batch_size=None,
shuffle=False,
device=device,
num_workers=0,
seed=SEED,
val_metrics={'l1': nn.L1Loss(), 'mse': nn.MSELoss()},
callbacks=[
ProgressbarCallback(epochs=epochs, width=20),
]
)
# Run the training loop
model = training_loop.run(model, epochs=epochs)
```
______________________________________________________________________
### Automatic Hyperparameters Optimization
PotaTorch provides a basic set of utilities to perform hyperparameters optimization. You can choose among **grid search**, **random search** and **bayesian search**. All of them are provided by `potatorch.optimization.tuning.HyperOptimizer`. The following is a working example of a simple grid search on a toy problem. You can find the full script under `examples/grid_search.py`
```python
def train(dataset, device, config):
""" Your usual training function that runs a TrainingLoop instance """
SEED = 42
# `epochs` is a fixed hyperparameter; it won't change among runs
epochs = config['epochs']
# Define your model as a pytorch Module
model = nn.Sequential(nn.Linear(1, 128), nn.ReLU(),
nn.Linear(128, 128), nn.ReLU(),
nn.Linear(128, 1))
loss_fn = torch.nn.MSELoss()
# `lr` is a dynamic hyperparameter; it will change among runs
optimizer = make_optimizer(torch.optim.Adam, lr=config['lr'])
training_loop = TrainingLoop(
dataset,
loss_fn,
optimizer,
train_p=0.8,
val_p=0.1,
test_p=0.1,
random_split=False,
batch_size=None,
shuffle=False,
device=device,
num_workers=0,
seed=SEED,
val_metrics={'l1': nn.L1Loss(), 'mse': nn.MSELoss()},
callbacks=[
ProgressbarCallback(epochs=epochs, width=20),
]
)
model = training_loop.run(model, epochs=epochs, verbose=1)
# Return a dictionary containing the training and validation metrics
# calculated during the last epoch of the loop
return training_loop.get_last_metrics()
# Define your search configuration
search_config = {
'method': 'grid', # which search method to use: ['grid', 'bayes', 'random']
'metric': {
'name': 'val_loss', # the metric you're optimizing
'goal': 'minimize' # whether you want to minimize or maximize it
},
'parameters': { # the set of hyperparameters you want to optimize
'lr': {
'values': [1e-2, 1e-3, 1e-4] # a range of values for the grid search to try
}
},
'fixed': { # fixed hyperparameters that won't change among runs
'epochs': 200
}
}
def main():
device = 'cuda'
dataset = TensorDataset(torch.arange(1000).view(1000, 1), torch.sin(torch.arange(1000)))
# Apply additional parameters to the train function to have f(config) -> {}
score_function = partial(train, dataset, device)
# Construct the hyperparameters optimizer
hyperoptimizer = HyperOptimizer(search_config)
# Run the optimization over the hyperparameters space
config, error = hyperoptimizer.optimize(score_function, return_error=True)
print('Best configuration found: {}\n with error: {}'.format(config, error))
```
Raw data
{
"_id": null,
"home_page": null,
"name": "potatorch",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.8",
"maintainer_email": null,
"keywords": "deep learning, framework, limited resources, machine learning, python, pytorch",
"author": null,
"author_email": "Marco Pampaloni <marco.pampaloni1@gmail.com>",
"download_url": "https://files.pythonhosted.org/packages/49/19/f272c044c72cf8d95c019e3f202efcf0046462ba1ae21c0affb732ecd628/potatorch-0.0.4.tar.gz",
"platform": null,
"description": "\n<div align=\"center\">\n\n<img src=\"https://raw.githubusercontent.com/crybot/potatorch/main/docs/potatorch-banner.png\" width=\"100%\" role=\"img\">\n\n**PotaTorch is a lightweight PyTorch framework specifically designed to run on hardware with limited resources.**\n\n______________________________________________________________________\n\n<!-- [](https://pepy.tech/project/potatorch) -->\n[](https://pypi.org/project/potatorch/)\n[](https://badge.fury.io/py/potatorch)\n\n[](https://github.com/crybot/potatorch/blob/main/LICENSE)\n\n</div>\n\n### Installation\nPotaTorch is published on PyPI, you can install it through pip:\n```bash\npip install potatorch\n```\n\nor you can install it from sources:\n```bash\ngit clone --single-branch -b main https://github.com/crybot/potatorch\npip install -e potatorch\n````\n______________________________________________________________________\n\n### Minimal Working Example\nYou can run the following example directly from `examples/mlp.py` if you already have pytorch installed, or you can run it with docker through the provided scripts:\n```bash\n./build.sh && ./run.sh\n```\n\nThe example trains a feed forward network on a toy problem:\n```python\nimport torch\nfrom torch import nn\n\nfrom potatorch.training import TrainingLoop, make_optimizer\nfrom potatorch.callbacks import ProgressbarCallback\nfrom torch.utils.data import TensorDataset\n\n# Fix a seed for TrainingLoop to make non-deterministic operations such as\n# shuffling reproducible\nSEED = 42\ndevice = 'cuda'\n\nepochs = 100\nlr = 1e-4\n\n# Define your model as a pytorch Module\nmodel = nn.Sequential(nn.Linear(1, 128), nn.ReLU(), \n nn.Linear(128, 128), nn.ReLU(),\n nn.Linear(128, 1))\n\n# Create your dataset as a torch.data.Dataset\ndataset = TensorDataset(torch.arange(1000).view(1000, 1), torch.sin(torch.arange(1000)))\n\n# Provide a loss function and an optimizer\nloss_fn = torch.nn.MSELoss()\noptimizer = make_optimizer(torch.optim.Adam, lr=lr)\n\n# Construct a TrainingLoop object.\n# TrainingLoop handles the initialization of dataloaders, dataset splitting,\n# shuffling, mixed precision training, etc.\n# You can provide callback handles through the `callbacks` argument.\ntraining_loop = TrainingLoop(\n dataset,\n loss_fn,\n optimizer,\n train_p=0.8,\n val_p=0.1,\n test_p=0.1,\n random_split=False,\n batch_size=None,\n shuffle=False,\n device=device,\n num_workers=0,\n seed=SEED,\n val_metrics={'l1': nn.L1Loss(), 'mse': nn.MSELoss()},\n callbacks=[\n ProgressbarCallback(epochs=epochs, width=20),\n ]\n )\n# Run the training loop\nmodel = training_loop.run(model, epochs=epochs)\n```\n______________________________________________________________________\n\n### Automatic Hyperparameters Optimization\nPotaTorch provides a basic set of utilities to perform hyperparameters optimization. You can choose among **grid search**, **random search** and **bayesian search**. All of them are provided by `potatorch.optimization.tuning.HyperOptimizer`. The following is a working example of a simple grid search on a toy problem. You can find the full script under `examples/grid_search.py`\n\n```python\ndef train(dataset, device, config):\n \"\"\" Your usual training function that runs a TrainingLoop instance \"\"\"\n SEED = 42\n # `epochs` is a fixed hyperparameter; it won't change among runs\n epochs = config['epochs']\n\n # Define your model as a pytorch Module\n model = nn.Sequential(nn.Linear(1, 128), nn.ReLU(), \n nn.Linear(128, 128), nn.ReLU(),\n nn.Linear(128, 1))\n\n loss_fn = torch.nn.MSELoss()\n # `lr` is a dynamic hyperparameter; it will change among runs\n optimizer = make_optimizer(torch.optim.Adam, lr=config['lr'])\n\n training_loop = TrainingLoop(\n dataset,\n loss_fn,\n optimizer,\n train_p=0.8,\n val_p=0.1,\n test_p=0.1,\n random_split=False,\n batch_size=None,\n shuffle=False,\n device=device,\n num_workers=0,\n seed=SEED,\n val_metrics={'l1': nn.L1Loss(), 'mse': nn.MSELoss()},\n callbacks=[\n ProgressbarCallback(epochs=epochs, width=20),\n ]\n )\n model = training_loop.run(model, epochs=epochs, verbose=1)\n # Return a dictionary containing the training and validation metrics \n # calculated during the last epoch of the loop\n return training_loop.get_last_metrics()\n\n# Define your search configuration\nsearch_config = {\n 'method': 'grid', # which search method to use: ['grid', 'bayes', 'random']\n 'metric': {\n 'name': 'val_loss', # the metric you're optimizing\n 'goal': 'minimize' # whether you want to minimize or maximize it\n },\n 'parameters': { # the set of hyperparameters you want to optimize\n 'lr': {\n 'values': [1e-2, 1e-3, 1e-4] # a range of values for the grid search to try\n }\n },\n 'fixed': { # fixed hyperparameters that won't change among runs\n 'epochs': 200\n }\n }\n\ndef main():\n device = 'cuda'\n dataset = TensorDataset(torch.arange(1000).view(1000, 1), torch.sin(torch.arange(1000)))\n # Apply additional parameters to the train function to have f(config) -> {}\n score_function = partial(train, dataset, device)\n # Construct the hyperparameters optimizer\n hyperoptimizer = HyperOptimizer(search_config)\n # Run the optimization over the hyperparameters space\n config, error = hyperoptimizer.optimize(score_function, return_error=True)\n print('Best configuration found: {}\\n with error: {}'.format(config, error))\n```\n",
"bugtrack_url": null,
"license": null,
"summary": "Lightweight high-level PyTorch framework that runs on potato machines",
"version": "0.0.4",
"project_urls": {
"Bug Tracker": "https://github.com/crybot/potatorch/issues",
"Homepage": "https://github.com/crybot/potatorch"
},
"split_keywords": [
"deep learning",
" framework",
" limited resources",
" machine learning",
" python",
" pytorch"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "08cf475c9d568bfb37397d80ae373c8254beeb9491623eb2d00e9eed9ac105de",
"md5": "4976ea76550ae300a4e9238b28d4f5d4",
"sha256": "13509b2cdc3053b0740f699c145653f0c2e4b3342899f2e31541f7c43219829e"
},
"downloads": -1,
"filename": "potatorch-0.0.4-py3-none-any.whl",
"has_sig": false,
"md5_digest": "4976ea76550ae300a4e9238b28d4f5d4",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.8",
"size": 19628,
"upload_time": "2024-05-01T08:51:42",
"upload_time_iso_8601": "2024-05-01T08:51:42.763211Z",
"url": "https://files.pythonhosted.org/packages/08/cf/475c9d568bfb37397d80ae373c8254beeb9491623eb2d00e9eed9ac105de/potatorch-0.0.4-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "4919f272c044c72cf8d95c019e3f202efcf0046462ba1ae21c0affb732ecd628",
"md5": "4263e43d37389ba583265a6b06847b6d",
"sha256": "27ad825165f5ea77d43fef3da40f33c3397afd142ac00bf04b8b96f893d44d14"
},
"downloads": -1,
"filename": "potatorch-0.0.4.tar.gz",
"has_sig": false,
"md5_digest": "4263e43d37389ba583265a6b06847b6d",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.8",
"size": 1329485,
"upload_time": "2024-05-01T08:51:46",
"upload_time_iso_8601": "2024-05-01T08:51:46.482113Z",
"url": "https://files.pythonhosted.org/packages/49/19/f272c044c72cf8d95c019e3f202efcf0046462ba1ae21c0affb732ecd628/potatorch-0.0.4.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-05-01 08:51:46",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "crybot",
"github_project": "potatorch",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"lcname": "potatorch"
}
JSON
