TensorNeko

Tensor Neural Engine Kompanion. An util library based on PyTorch and PyTorch Lightning.

Install

pip install tensorneko

Neko Layers and Modules

Build an MLP with linear layers. The activation and normalization will be placed in the hidden layers.

784 -> 1024 -> 512 -> 10

import tensorneko as neko
import torch.nn

mlp = neko.module.MLP(
    neurons=[784, 1024, 512, 10],
    build_activation=torch.nn.ReLU,
    build_normalization=[
        lambda: torch.nn.BatchNorm1d(1024),
        lambda: torch.nn.BatchNorm1d(512)
    ],
    dropout_rate=0.5
)

Build a Conv2d with activation and normalization.

import tensorneko as neko
import torch.nn

conv2d = neko.layer.Conv2d(
    in_channels=256,
    out_channels=1024,
    kernel_size=(3, 3),
    padding=(1, 1),
    build_activation=torch.nn.ReLU,
    build_normalization=lambda: torch.nn.BatchNorm2d(256),
    normalization_after_activation=False
)

All modules and layers

layers:

Concatenate
Conv2d
Linear
Log
PatchEmbedding2d
PositionalEmbedding
Reshape

modules:

DenseBlock
InceptionModule
MLP
ResidualBlock and ResidualModule
AttentionModule, TransformerEncoderBlock and TransformerEncoder

Neko modules

All tensorneko.layer and tensorneko.module are NekoModule. They can be used in fn.py pipe operation.

from tensorneko.layer import Linear
from torch.nn import ReLU
import torch

linear0 = Linear(16, 128, build_activation=ReLU)
linear1 = Linear(128, 1)

f = linear0 >> linear1
print(f(torch.rand(16)).shape)
# torch.Size([1])

Neko reader

Easily load different modal data.

import tensorneko as neko
from tensorneko.io.text.text_reader import json_data
from typing import List

# read video (Temporal, Channel, Height, Width)
video_tensor = neko.io.read.video.of("path/to/video.mp4")
# read audio (Channel, Temporal)
audio_tensor = neko.io.read.audio.of("path/to/audio.wav")
# read image (Channel, Height, Width)
image_tensor = neko.io.read.image.of("path/to/image.png")
# read plain text
text_string = neko.io.read.text.of("path/to/text.txt")
# read json as DataFrame
json_df = neko.io.read.text.of_json("path/to/json.json", to_df=True)
# read json as a object
@json_data
class JsonData:
    x: int
    y: int

json_obj: List[JsonData] = neko.io.read.text.of_json("path/to/json.json", cls=List[JsonData])

Neko preprocessing

import tensorneko as neko

# A video tensor with (120, 3, 720, 1280)
video = neko.io.read.video.of("example/video.mp4")
# Get a resized tensor with (120, 3, 256, 256)
neko.preprocess.resize_video(video, (256, 256))

All preprocessing utils

resize_video
resize_image

Neko Model

Build and train a simple model for classifying MNIST with MLP.

from typing import Optional, Union, Sequence, Dict, List

import torch.nn
from torch import Tensor
from torch.optim import Adam
from torchmetrics import Accuracy
from pytorch_lightning.callbacks import ModelCheckpoint

import tensorneko as neko
from tensorneko.util import get_activation, get_loss


class MnistClassifier(neko.NekoModel):

    def __init__(self, name: str, mlp_neurons: List[int], activation: str, dropout_rate: float, loss: str,
        learning_rate: float, weight_decay: float
    ):
        super().__init__(name)
        self.weight_decay = weight_decay
        self.learning_rate = learning_rate

        self.flatten = torch.nn.Flatten()
        self.mlp = neko.module.MLP(
            neurons=mlp_neurons,
            build_activation=get_activation(activation),
            dropout_rate=dropout_rate
        )
        self.loss_func = get_loss(loss)()
        self.acc_func = Accuracy()

    def forward(self, x):
        # (batch, 28, 28)
        x = self.flatten(x)
        # (batch, 768)
        x = self.mlp(x)
        # (batch, 10)
        return x

    def training_step(self, batch: Optional[Union[Tensor, Sequence[Tensor]]] = None, batch_idx: Optional[int] = None,
        optimizer_idx: Optional[int] = None, hiddens: Optional[Tensor] = None
    ) -> Dict[str, Tensor]:
        x, y = batch
        logit = self(x)
        prob = logit.sigmoid()
        loss = self.loss_func(logit, y)
        acc = self.acc_func(prob.max(dim=1)[1], y)
        return {"loss": loss, "acc": acc}

    def validation_step(self, batch: Optional[Union[Tensor, Sequence[Tensor]]] = None, batch_idx: Optional[int] = None,
        dataloader_idx: Optional[int] = None
    ) -> Dict[str, Tensor]:
        x, y = batch
        logit = self(x)
        prob = logit.sigmoid()
        loss = self.loss_func(logit, y)
        acc = self.acc_func(prob.max(dim=1)[1], y)
        return {"loss": loss, "acc": acc}

    def predict_step(self, batch: Tensor, batch_idx: int, dataloader_idx: Optional[int] = None) -> Tensor:
        x, y = batch
        logit = self(x)
        return logit

    def configure_optimizers(self):
        optimizer = Adam(self.parameters(), lr=self.learning_rate, betas=(0.5, 0.9), weight_decay=self.weight_decay)
        return {
            "optimizer": optimizer
        }


model = MnistClassifier("mnist_mlp_classifier", [784, 1024, 512, 10], "ReLU", 0.5, "CrossEntropyLoss", 1e-4, 1e-4)

dm = ...  # The MNIST datamodule from PyTorch Lightning

trainer = neko.NekoTrainer(log_every_n_steps=0, gpus=1, logger=model.name, precision=32,
    checkpoint_callback=ModelCheckpoint(dirpath="./ckpt",
        save_last=True, filename=model.name + "-{epoch}-{val_acc:.3f}", monitor="val_acc", mode="max"
    ))

trainer.fit(model, dm)

Neko utilities

StringGetter: Get PyTorch class from string.

import tensorneko as neko
activation = neko.util.get_activation("leakyRelu")()

__: The arguments to pipe operator

from tensorneko.util import __, _
result = __(20) >> (_ + 1) >> (_ * 2) >> __.get
print(result)
# 42

Utilities list:

reduce_dict_by
summarize_dict_by
generate_inf_seq
compose
listdir
with_printed
with_printed_shape
is_bad_num
ifelse
dict_add
count_parameters
as_list
Configuration
get_activation
get_loss
__

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

tensorneko-0.1.17.tar.gz (925.4 kB view details)

Uploaded Oct 14, 2021 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

tensorneko-0.1.17-py3-none-any.whl (949.8 kB view details)

Uploaded Oct 14, 2021 Python 3

File details

Details for the file tensorneko-0.1.17.tar.gz.

File metadata

Download URL: tensorneko-0.1.17.tar.gz
Upload date: Oct 14, 2021
Size: 925.4 kB
Tags: Source
Uploaded using Trusted Publishing? No
Uploaded via: twine/3.4.2 importlib_metadata/4.8.1 pkginfo/1.7.1 requests/2.26.0 requests-toolbelt/0.9.1 tqdm/4.62.3 CPython/3.9.7

File hashes

Hashes for tensorneko-0.1.17.tar.gz
Algorithm	Hash digest
SHA256	`c76799848f3e29766507c30b0e0b6488e11afd450f7c3b198e94ae413d4edf5f`
MD5	`b87e80bf05a2094b4dda5c81ea3da2ee`
BLAKE2b-256	`074a267b3118bb07093c499ad98f5520937e23864d5372c69e5674afa7582db0`

See more details on using hashes here.

File details

Details for the file tensorneko-0.1.17-py3-none-any.whl.

File metadata

Download URL: tensorneko-0.1.17-py3-none-any.whl
Upload date: Oct 14, 2021
Size: 949.8 kB
Tags: Python 3
Uploaded using Trusted Publishing? No
Uploaded via: twine/3.4.2 importlib_metadata/4.8.1 pkginfo/1.7.1 requests/2.26.0 requests-toolbelt/0.9.1 tqdm/4.62.3 CPython/3.9.7

File hashes

Hashes for tensorneko-0.1.17-py3-none-any.whl
Algorithm	Hash digest
SHA256	`c5449cef1a340f131ea048b780d9d9a8e62673782864678fb0b41cb42327960d`
MD5	`fb071e07536098137a236ac39516daac`
BLAKE2b-256	`45c351beb26417d1a850ab0f787e50425c45ba8b63342f9a918f4a4071e84f45`