TorchFlare is a simple, beginner-friendly and an easy-to-use PyTorch Framework train your models without much effort. It provides an almost Keras-like experience for training your models with all the callbacks, metrics, etc
- A high-level module for Keras-like training.
- Off-the-shelf Pytorch style Datasets/Dataloaders for standard tasks such as Image classification, Image segmentation, Text Classification, etc
- Callbacks for model checkpoints, early stopping, and much more!
- Metrics and much more.
Currently, TorchFlare supports CPU and GPU training. DDP and TPU support will be coming soon!
pip install torchflare
The Documentation is available here
The library isn't mature or stable for production use yet.
The best of the library currently would be for non production use and rapid prototyping.
The core idea around TorchFlare is the Experiment class. It handles all the internal stuff like boiler plate code for training, calling callbacks,metrics,etc. The only thing you need to focus on is creating you PyTorch Model.
Also, there are off-the-shelf pytorch style datasets/dataloaders available for standard tasks, so that you don't have to worry about creating Pytorch Datasets/Dataloaders.
Here is an easy-to-understand example to show how Experiment class works.
import torch
import torch.nn as nn
from torchflare.experiments import Experiment
import torchflare.callbacks as cbs
import torchflare.metrics as metrics
#Some dummy dataloaders
train_dl = SomeTrainingDataloader()
valid_dl = SomeValidationDataloader()
test_dl = SomeTestingDataloader()Create a pytorch Model
model = nn.Sequential(
nn.Linear(num_features, hidden_state_size),
nn.ReLU(),
nn.Linear(hidden_state_size, num_classes)
)Define callbacks and metrics
metric_list = [metrics.Accuracy(num_classes=num_classes, multilabel=False),
metrics.F1Score(num_classes=num_classes, multilabel=False)]
callbacks = [cbs.EarlyStopping(monitor="accuracy", mode="max"), cbs.ModelCheckpoint(monitor="accuracy"),
cbs.ReduceLROnPlateau(mode="max" , patience = 2)]Define your experiment
# Set some constants for training
exp = Experiment(
num_epochs=5,
save_dir="./models",
model_name="model.bin",
fp16=False,
device="cuda",
compute_train_metrics=True,
seed=42,
)
# Compile your experiment with model, optimizer, schedulers, etc
exp.compile_experiment(
model=net,
optimizer="Adam",
optimizer_params=dict(lr=3e-4),
callbacks=callbacks,
criterion="cross_entropy",
metrics=metric_list,
main_metric="accuracy",
)
# Run your experiment with training dataloader and validation dataloader.
exp.run_experiment(train_dl=train_dl, valid_dl= valid_dl)For inference, you can use infer method, which yields output per batch. You can use it as follows
outputs = []
for op in exp.infer(test_loader=test_dl , path='./models/model.bin' , device = 'cuda'):
op = some_post_process_function(op)
outputs.extend(op)If you want to access your experiments history or plot it. You can do it as follows.
history = exp.history # This will return a dict
# If you want to plot progress of particular metric as epoch progress use this.
exp.plot_history(keys = ["loss" , "accuracy"] , save_fig = False , plot_fig = True)- Image Classification on CIFAR-10 using TorchFlare.
- Text Classification on IMDB data.
- Binary Classification of Tabular Data on previous kaggle competition
- Tutorial on using Hydra and TorchFlare for efficient workflow and parameter management.
Contributions are always welcomed, it would be great to have people use and contribute to this project to help users understand and benefit from the library.
- Create an issue: If you have a new feature in mind, feel free to open an issue and add some short description on what that feature could be.
- Create a PR: If you have a bug fix, enhancement or new feature addition, create a Pull Request and the maintainers of the repo, would review and merge them.
- Atharva Phatak - Atharva-Phatak