torch_models

This is a repository where I personally save machine learning models, most of which are based on pytorch.

�Note

These models are not all fully tested. It is very likely the codes contain bugs. Be careful when you use the code.

Description

Typically, models have the following methods.
These methods are necessary to use the modules in my_utils such as Trainer and Evaluator.

def forward(self, inputs):
    '''
    Perform forward computation.
    '''
    return output

def fit(self, inputs, labels, optimizer):
    '''
    Caliculate loss and update parameters.
    '''
    return loss_item

def predict(self, inputs):
    '''
    Outputs predicted labels (or values) from inputs.
    '''
    return predicted

Usage Example

>>> train, test = get_dataset()

>>> from my_utils import get_device
>>> device = get_device()
===== Device =====
cpu

>>> from my_utils import get_device, DataLoader, torch_stack
>>> train_loader = DataLoader(train, batch_size=64, trans_func=torch_stack)
>>> test_loader = DataLoader(test, batch_size=64, trans_func=torch_stack)

>>> from torch_models import MLP
>>> model = MLP([784, 50, 10])
>>> print(model)
MLP(
  (fc_0): Linear(in_features=784, out_features=50, bias=True)
  (fc_out): Linear(in_features=50, out_features=10, bias=True)
  (criterion): CrossEntropyLoss()
  (activation): Tanh()
)

>>> from torch.optim import SGD
>>> from my_utils import Trainer, EvaluatorC

>>> optimizer = SGD(model.parameters(), lr=0.1)
>>> trainer = Trainer(model, train_loader)
>>> evaluator = EvaluatorC(model, test_loader)
>>> trainer.train_epoch(optimizer, max_epoch=10, evaluator=evaluator, show_log=True)
epoch 0  	loss: 0.1835	accuracy: 0.9491
epoch 1  	loss: 0.1595	accuracy: 0.9544
epoch 2  	loss: 0.1428	accuracy: 0.956
epoch 3  	loss: 0.1298	accuracy: 0.9598
epoch 4  	loss: 0.1188	accuracy: 0.9629
epoch 5  	loss: 0.1094	accuracy: 0.964
epoch 6  	loss: 0.1022	accuracy: 0.9663
epoch 7  	loss: 0.09551	accuracy: 0.9665
epoch 8  	loss: 0.08942	accuracy: 0.968
epoch 9  	loss: 0.08407	accuracy: 0.9706