Torchest

Pytorch tools and utilities (Trainers, data generators, functions, and more...)

Trainers

One of the most common tasks you will do in pytorch is define training loops, which is a hassle because they are almost always the same. Torchest has trainers you can use for your projects

from torchest.trainer import SimpleTrainer

# previous definition of model and data preparation
# train_dataloader = DataLoader(train_data)
# dev_dataloader = DataLoader(dev_data)
# test_dataloader = DataLoader(test_data)
# model = nn.Sequential(...)
...

"""
Prepare Trainer
"""
loss_fn = nn.CrossEntropyLoss()
learning_rate= 1e-2
optimizer = optim.Adam(model.parameters(), lr=learning_rate)

trainer = SimpleTrainer(model, loss_fn, optimizer)

"""
Train
"""
epochs = 500
trainer.train(data_train=train_dataloader, data_dev=dev_dataloader, data_test=test_dataloader, epochs=epochs)

Training loop progress

31%|██████████████████▍      | 156/500 [00:02<00:05, 58.01epoch/s, train_accuracy=76.2, train_cost=0.238290]

Trainer also saves the cost for the train, dev and test passes. You can simply call trainer.plot_costs() to display the graph with the costs

Wandb visualization

Trainers support wandb to visualize loss and acurracy as well as to save model versions. if you want to enable you have to set 2 parameters in your trainer initialization, name and wandb

trainer = SimpleTrainer(model, loss_fn, optimizer, wandb_project_name="MyProject", wandb=True)

On the first run you will have to enter your API key, which you can obtain at https://wandb.ai/settings

Data generators

Spiral data

This will create a 2d matrix with points in a dataplot. Ideal for testing non-linearity in your network

from torchest.datagen import spiral_datagen

class_num =  3
X, Y = spiral_datagen(450, class_num) # 450 elements per class