keras4torch 0.9.6

An Easy-to-Use Wrapper for Training PyTorch Models

Keras4Torch

"An Easy-to-Use Wrapper for Training PyTorch Models❤"

Keras4Torch provides an easy way to train PyTorch models in Keras style. You can use keras4torch.Model to warp any torch.nn.Module to integrate core training features. With this framework, the training process can be considerably simplified.

• If you are a keras user, most of your training code can work well in Keras4Torch with little change.

• If you are a pytorch user, Keras4Torch can help you train pytorch models with far less code than basic pytorch.

Installation

pip install keras4torch

Keras4Torch supports PyTorch 1.6 or newer.

Quick Start

Let's start with a simple example of MNIST!

import torch
import torchvision
from torch import nn

import keras4torch as k4t

Step1: Preprocess data

mnist = torchvision.datasets.MNIST(root='./', download=True)
X, y = mnist.train_data, mnist.train_labels

X = X.float() / 255.0    # scale the pixels to [0, 1]

x_train, y_train = X[:40000], y[:40000]
x_test, y_test = X[40000:], y[40000:]

Step2: Define the model

model = torch.nn.Sequential(
    nn.Flatten(),
    nn.Linear(28*28, 512), nn.ReLU(),
    nn.Linear(512, 128), nn.ReLU(),
    nn.Linear(128, 10)
)

model = k4t.Model(model)    # attention this line

Alternatively, You can use KerasLayer for automatic shape inference, which can free you from calculating the input channels. Here is an equivalent to the model above.

model = torch.nn.Sequential(
    nn.Flatten(),
    k4t.layers.Linear(512), nn.ReLU(),
    k4t.layers.Linear(128), nn.ReLU(),
    k4t.layers.Linear(10)
)

A model containing KerasLayer needs an extra .build(input_shape) operation.

model = k4t.Model(model).build([28, 28])

Step3: Summary the model

model.summary()

==============================================================================
Layer (type:depth-idx)                   Output Shape              Param #
==============================================================================
├─Flatten: 1-1                           [-1, 784]                 --
├─Linear*: 1-2                           [-1, 512]                 401,920
├─ReLU: 1-3                              [-1, 512]                 --
├─Linear*: 1-4                           [-1, 128]                 65,664
├─ReLU: 1-5                              [-1, 128]                 --
├─Linear*: 1-6                           [-1, 10]                  1,290
==============================================================================
Total params: 468,874
Trainable params: 468,874
Non-trainable params: 0
Total mult-adds (M): 0.47
==============================================================================
Input size (MB): 0.00
Forward/backward pass size (MB): 0.00
Params size (MB): 1.79
Estimated Total Size (MB): 1.80
==============================================================================

Step4: Config optimizer, loss and metrics

model.compile(optimizer='adam', loss=nn.CrossEntropyLoss(), metrics=['acc'])

Step5: Training

history = model.fit(x_train, y_train,
                	epochs=30,
                	batch_size=512,
                	validation_split=0.2,
                	)

Train on 32000 samples, validate on 8000 samples:
Epoch 1/30 - 0.7s - loss: 0.7440 - acc: 0.8149 - val_loss: 0.3069 - val_acc: 0.9114 - lr: 1e-03
Epoch 2/30 - 0.5s - loss: 0.2650 - acc: 0.9241 - val_loss: 0.2378 - val_acc: 0.9331 - lr: 1e-03
Epoch 3/30 - 0.5s - loss: 0.1946 - acc: 0.9435 - val_loss: 0.1940 - val_acc: 0.9431 - lr: 1e-03
Epoch 4/30 - 0.5s - loss: 0.1513 - acc: 0.9555 - val_loss: 0.1663 - val_acc: 0.9524 - lr: 1e-03
... ...

Step6: Plot learning curve

history.plot(kind='line', y=['acc', 'val_acc'])

Step7: Evaluate on test set

model.evaluate(x_test, y_test)

{'loss': 0.121063925, 'acc': 0.9736}

Communication

If you have any problem when using Keras4Torch, please:

• open a Github Issue
• send email to [e-mail] or [e-mail].

Keras4Torch is still under development.

You can contribute new features by opening a Pull Request. (The details will be updated soon)

Any contribution to us would be more than welcome : )