Sign up on our console for pilot access!
Startai is an open-source machine learning framework that enables you to:
- π Convert code into any framework: Use and build on top of any model, library, or device by converting any code from one framework to another using
startai.transpile
. - βοΈ Write framework-agnostic code: Write your code once in
startai
and then choose the most appropriate ML framework as the backend to leverage all the benefits and tools.
Join our growing community π to connect with people using Startai. Let's khulnasoft.com together π¦Ύ
Startai's transpiler helps you convert code between different ML frameworks. To get pilot access to the transpiler, sign up and generate an API key. The Get Started notebook should help you set up your API key and the Quickstart notebook should give you a brief idea of the features!
The most important notebooks are:
Beyond that, based on the frameworks you want to convert code between, there are a few more examples further down this page π which contain a number of models and libraries transpiled between PyTorch, JAX, TensorFlow and NumPy.
The easiest way to set up Startai is to install it using pip:
pip install startai
Docker Images
Given the challenges of maintaining installations of various frameworks in a single environment,
users who would want to test startai
with multiple frameworks at once can use our Docker images for a seamless experience.
You can pull the images from:
docker pull khulnasoft/startai:latest # CPU
docker pull khulnasoft/startai:latest-gpu # GPU
From Source
You can also install Startai from source if you want to take advantage of the latest changes, but we can't ensure everything will work as expected π
git clone https://github.com/khulnasoft/startai.git
cd startai
pip install --user -e .
If you want to set up testing and various frameworks it's probably best to check out the Setting Up page, where OS-specific and IDE-specific instructions and video tutorials to do so are available!
After installing Startai, you can start using it straight away, for example:
Transpiling any code from one framework to another
import startai
import torch
import jax
def jax_fn(x):
a = jax.numpy.dot(x, x)
b = jax.numpy.mean(x)
return x * a + b
jax_x = jax.numpy.array([1., 2., 3.])
torch_x = torch.tensor([1., 2., 3.])
torch_fn = startai.transpile(jax_fn, source="jax", to="torch", args=(jax_x,))
ret = torch_fn(torch_x)
Running your code with any backend
import startai
import torch
import jax
startai.set_backend("jax")
x = jax.numpy.array([1, 2, 3])
y = jax.numpy.array([3, 2, 1])
z = startai.add(x, y)
startai.set_backend('torch')
x = torch.tensor([1, 2, 3])
y = torch.tensor([3, 2, 1])
z = startai.add(x, y)
The Examples page features a wide range of
demos and tutorials showcasing the functionalities of Startai along with
multiple use cases, but feel free to check out some shorter
framework-specific examples here β¬οΈ
I'm using PyTorchβ
You can use Startai to get PyTorch code from:Any model
From TensorFlow
import startai import torch import tensorflow as tf # Get a pretrained keras model eff_encoder = tf.keras.applications.efficientnet_v2.EfficientNetV2B0( include_top=False, weights="imagenet", input_shape=(224, 224, 3) ) # Transpile it into a torch.nn.Module with the corresponding parameters noise = tf.random.normal(shape=(1, 224, 224, 3)) torch_eff_encoder = startai.transpile(eff_encoder, source="tensorflow", to="torch", args=(noise,)) # Build a classifier using the transpiled encoder class Classifier(torch.nn.Module): def __init__(self, num_classes=20): super().__init__() self.encoder = torch_eff_encoder self.fc = torch.nn.Linear(1280, num_classes) def forward(self, x): x = self.encoder(x) return self.fc(x) # Initialize a trainable, customizable, torch.nn.Module classifier = Classifier() ret = classifier(torch.rand((1, 244, 244, 3)))From JAX
import startai import jax import torch # Get a pretrained haiku model # https://github.com/khulnasoft/demos/blob/15c235f/scripts/deepmind_perceiver_io.py from deepmind_perceiver_io import key, perceiver_backbone # Transpile it into a torch.nn.Module with the corresponding parameters dummy_input = jax.random.uniform(key, shape=(1, 3, 224, 224)) params = perceiver_backbone.init(rng=key, images=dummy_input) startai.set_backend("jax") backbone = startai.transpile( perceiver_backbone, source="jax", to="torch", params_v=params, kwargs={"images": dummy_input} ) # Build a classifier using the transpiled backbone class PerceiverIOClassifier(torch.nn.Module): def __init__(self, num_classes=20): super().__init__() self.backbone = backbone self.max_pool = torch.nn.MaxPool2d((512, 1)) self.flatten = torch.nn.Flatten() self.fc = torch.nn.Linear(1024, num_classes) def forward(self, x): x = self.backbone(images=x) x = self.flatten(self.max_pool(x)) return self.fc(x) # Initialize a trainable, customizable, torch.nn.Module classifier = PerceiverIOClassifier() ret = classifier(torch.rand((1, 3, 224, 224)))Any library
From Tensorflow
import startai import torch import os os.environ["SM_FRAMEWORK"] = "tf.keras" import segmentation_models as sm # transpile sm from tensorflow to torch torch_sm = startai.transpile(sm, source="tensorflow", to="torch") # get some image-like arrays output = torch.rand((1, 3, 512, 512)) target = torch.rand((1, 3, 512, 512)) # and use the transpiled version of any function from the library! out = torch_sm.metrics.iou_score(output, target)From JAX
import startai import rax import torch # transpile rax from jax to torch torch_rax = startai.transpile(rax, source="jax", to="torch") # get some arrays scores = torch.tensor([2.2, 1.3, 5.4]) labels = torch.tensor([1.0, 0.0, 0.0]) # and use the transpiled version of any function from the library! out = torch_rax.poly1_softmax_loss(scores, labels)From NumPy
import startai import torch import madmom # transpile madmon from numpy to torch torch_madmom = startai.transpile(madmom, source="numpy", to="torch") # get some arrays freqs = torch.arange(20) * 10 # and use the transpiled version of any function from the library! out = torch_madmom.audio.filters.hz2midi(freqs)Any function
From Tensorflow
import startai import tensorflow as tf import torch def loss(predictions, targets): return tf.sqrt(tf.reduce_mean(tf.square(predictions - targets))) # transpile any function from tf to torch torch_loss = startai.transpile(loss, source="tensorflow", to="torch") # get some arrays p = torch.tensor([3.0, 2.0, 1.0]) t = torch.tensor([0.0, 0.0, 0.0]) # and use the transpiled version! out = torch_loss(p, t)From JAX
import startai import jax.numpy as jnp import torch def loss(predictions, targets): return jnp.sqrt(jnp.mean((predictions - targets) ** 2)) # transpile any function from jax to torch torch_loss = startai.transpile(loss, source="jax", to="torch") # get some arrays p = torch.tensor([3.0, 2.0, 1.0]) t = torch.tensor([0.0, 0.0, 0.0]) # and use the transpiled version! out = torch_loss(p, t)From NumPy
import startai import numpy as np import torch def loss(predictions, targets): return np.sqrt(np.mean((predictions - targets) ** 2)) # transpile any function from numpy to torch torch_loss = startai.transpile(loss, source="numpy", to="torch") # get some arrays p = torch.tensor([3.0, 2.0, 1.0]) t = torch.tensor([0.0, 0.0, 0.0]) # and use the transpiled version! out = torch_loss(p, t)
I'm using TensorFlowβ
You can use Startai to get TensorFlow code from:Any model
From PyTorch
import startai import torch import timm import tensorflow as tf # Get a pretrained pytorch model mlp_encoder = timm.create_model("mixer_b16_224", pretrained=True, num_classes=0) # Transpile it into a keras.Model with the corresponding parameters noise = torch.randn(1, 3, 224, 224) mlp_encoder = startai.transpile(mlp_encoder, to="tensorflow", args=(noise,)) # Build a classifier using the transpiled encoder class Classifier(tf.keras.Model): def __init__(self): super().__init__() self.encoder = mlp_encoder self.output_dense = tf.keras.layers.Dense(units=1000, activation="softmax") def call(self, x): x = self.encoder(x) return self.output_dense(x) # Transform the classifier and use it as a standard keras.Model x = tf.random.normal(shape=(1, 3, 224, 224)) model = Classifier() ret = model(x)From JAX
import startai import jax import tensorflow as tf # Get a pretrained haiku model # https://khulnasoft.com/demos/scripts/deepmind_perceiver_io.py from deepmind_perceiver_io import key, perceiver_backbone # Transpile it into a tf.keras.Model with the corresponding parameters dummy_input = jax.random.uniform(key, shape=(1, 3, 224, 224)) params = perceiver_backbone.init(rng=key, images=dummy_input) backbone = startai.transpile( perceiver_backbone, to="tensorflow", params_v=params, args=(dummy_input,) ) # Build a classifier using the transpiled backbone class PerceiverIOClassifier(tf.keras.Model): def __init__(self, num_classes=20): super().__init__() self.backbone = backbone self.max_pool = tf.keras.layers.MaxPooling1D(pool_size=512) self.flatten = tf.keras.layers.Flatten() self.fc = tf.keras.layers.Dense(num_classes) def call(self, x): x = self.backbone(x) x = self.flatten(self.max_pool(x)) return self.fc(x) # Initialize a trainable, customizable, tf.keras.Model x = tf.random.normal(shape=(1, 3, 224, 224)) classifier = PerceiverIOClassifier() ret = classifier(x)Any library
From PyTorch
import startai import kornia import requests import numpy as np import tensorflow as tf from PIL import Image # transpile kornia from torch to tensorflow tf_kornia = startai.transpile(kornia, source="torch", to="tensorflow") # get an image url = "http://images.cocodataset.org/train2017/000000000034.jpg" raw_img = Image.open(requests.get(url, stream=True).raw) # convert it to the format expected by kornia img = np.array(raw_img) img = tf.transpose(tf.constant(img), (2, 0, 1)) img = tf.expand_dims(img, 0) / 255 # and use the transpiled version of any function from the library! out = tf_kornia.enhance.sharpness(img, 5)From JAX
import startai import rax import tensorflow as tf # transpile rax from jax to tensorflow tf_rax = startai.transpile(rax, source="jax", to="tensorflow") # get some arrays scores = tf.constant([2.2, 1.3, 5.4]) labels = tf.constant([1.0, 0.0, 0.0]) # and use the transpiled version of any function from the library! out = tf_rax.poly1_softmax_loss(scores, labels)From NumPy
import startai import madmom import tensorflow as tf # transpile madmom from numpy to tensorflow tf_madmom = startai.transpile(madmom, source="numpy", to="tensorflow") # get some arrays freqs = tf.range(20) * 10 # and use the transpiled version of any function from the library! out = tf_madmom.audio.filters.hz2midi(freqs)Any function
From PyTorch
import startai import torch import tensorflow as tf def loss(predictions, targets): return torch.sqrt(torch.mean((predictions - targets) ** 2)) # transpile any function from torch to tensorflow tf_loss = startai.transpile(loss, source="torch", to="tensorflow") # get some arrays p = tf.constant([3.0, 2.0, 1.0]) t = tf.constant([0.0, 0.0, 0.0]) # and use the transpiled version! out = tf_loss(p, t)From JAX
import startai import jax.numpy as jnp import tensorflow as tf def loss(predictions, targets): return jnp.sqrt(jnp.mean((predictions - targets) ** 2)) # transpile any function from jax to tensorflow tf_loss = startai.transpile(loss, source="jax", to="tensorflow") # get some arrays p = tf.constant([3.0, 2.0, 1.0]) t = tf.constant([0.0, 0.0, 0.0]) # and use the transpiled version! out = tf_loss(p, t)From NumPy
import startai import numpy as np import tensorflow as tf def loss(predictions, targets): return np.sqrt(np.mean((predictions - targets) ** 2)) # transpile any function from numpy to tensorflow tf_loss = startai.transpile(loss, source="numpy", to="tensorflow") # get some arrays p = tf.constant([3.0, 2.0, 1.0]) t = tf.constant([0.0, 0.0, 0.0]) # and use the transpiled version! out = tf_loss(p, t)
I'm using Jaxβ
You can use Startai to get JAX code from:Any model
From PyTorch
import startai import timm import torch import jax import haiku as hk # Get a pretrained pytorch model mlp_encoder = timm.create_model("mixer_b16_224", pretrained=True, num_classes=0) # Transpile it into a hk.Module with the corresponding parameters noise = torch.randn(1, 3, 224, 224) mlp_encoder = startai.transpile(mlp_encoder, source="torch", to="haiku", args=(noise,)) # Build a classifier using the transpiled encoder class Classifier(hk.Module): def __init__(self, num_classes=1000): super().__init__() self.encoder = mlp_encoder() self.fc = hk.Linear(output_size=num_classes, with_bias=True) def __call__(self, x): x = self.encoder(x) x = self.fc(x) return x def _forward_classifier(x): module = Classifier() return module(x) # Transform the classifier and use it as a standard hk.Module rng_key = jax.random.PRNGKey(42) x = jax.random.uniform(key=rng_key, shape=(1, 3, 224, 224), dtype=jax.numpy.float32) forward_classifier = hk.transform(_forward_classifier) params = forward_classifier.init(rng=rng_key, x=x) ret = forward_classifier.apply(params, None, x)From TensorFlow
import startai import jax import haiku as hk import tensorflow as tf jax.config.update("jax_enable_x64", True) # Get a pretrained keras model eff_encoder = tf.keras.applications.efficientnet_v2.EfficientNetV2B0( include_top=False, weights="imagenet", input_shape=(224, 224, 3) ) # Transpile it into a hk.Module with the corresponding parameters noise = tf.random.normal(shape=(1, 224, 224, 3)) hk_eff_encoder = startai.transpile(eff_encoder, source="tensorflow", to="haiku", args=(noise,)) # Build a classifier using the transpiled encoder class Classifier(hk.Module): def __init__(self, num_classes=1000): super().__init__() self.encoder = hk_eff_encoder() self.fc = hk.Linear(output_size=num_classes, with_bias=True) def __call__(self, x): x = self.encoder(x) x = self.fc(x) return x def _forward_classifier(x): module = Classifier() return module(x) # Transform the classifier and use it as a standard hk.Module rng_key = jax.random.PRNGKey(42) dummy_x = jax.random.uniform(key=rng_key, shape=(1, 224, 224, 3)) forward_classifier = hk.transform(_forward_classifier) params = forward_classifier.init(rng=rng_key, x=dummy_x) ret = forward_classifier.apply(params, None, dummy_x)Any library
From PyTorch
import startai import kornia import requests import jax.numpy as jnp from PIL import Image jax.config.update("jax_enable_x64", True) # transpile kornia from torch to jax jax_kornia = startai.transpile(kornia, source="torch", to="jax") # get an image url = "http://images.cocodataset.org/train2017/000000000034.jpg" raw_img = Image.open(requests.get(url, stream=True).raw) # convert it to the format expected by kornia img = jnp.transpose(jnp.array(raw_img), (2, 0, 1)) img = jnp.expand_dims(img, 0) / 255 # and use the transpiled version of any function from the library! out = jax_kornia.enhance.sharpness(img, 5)From TensorFlow
import startai import jax import os os.environ["SM_FRAMEWORK"] = "tf.keras" import segmentation_models as sm # transpile sm from tensorflow to jax jax_sm = startai.transpile(sm, source="tensorflow", to="jax") # get some image-like arrays key = jax.random.PRNGKey(23) key1, key2 = jax.random.split(key) output = jax.random.uniform(key1, (1, 3, 512, 512)) target = jax.random.uniform(key2, (1, 3, 512, 512)) # and use the transpiled version of any function from the library! out = jax_sm.metrics.iou_score(output, target)From NumPy
import startai import madmom import jax.numpy as jnp # transpile madmon from numpy to jax jax_madmom = startai.transpile(madmom, source="numpy", to="jax") # get some arrays freqs = jnp.arange(20) * 10 # and use the transpiled version of any function from the library! out = jax_madmom.audio.filters.hz2midi(freqs)Any function
From PyTorch
import startai import torch import jax.numpy as jnp def loss(predictions, targets): return torch.sqrt(torch.mean((predictions - targets) ** 2)) # transpile any function from torch to jax jax_loss = startai.transpile(loss, source="torch", to="jax") # get some arrays p = jnp.array([3.0, 2.0, 1.0]) t = jnp.array([0.0, 0.0, 0.0]) # and use the transpiled version! out = jax_loss(p, t)From TensorFlow
import startai import tensorflow as tf import jax.numpy as jnp def loss(predictions, targets): return tf.sqrt(tf.reduce_mean(tf.square(predictions - targets))) # transpile any function from tf to jax jax_loss = startai.transpile(loss, source="tensorflow", to="jax") # get some arrays p = jnp.array([3.0, 2.0, 1.0]) t = jnp.array([0.0, 0.0, 0.0]) # and use the transpiled version! out = jax_loss(p, t)From NumPy
import startai import numpy as np import jax import jax.numpy as jnp jax.config.update('jax_enable_x64', True) def loss(predictions, targets): return np.sqrt(np.mean((predictions - targets) ** 2)) # transpile any function from numpy to jax jax_loss = startai.transpile(loss, source="numpy", to="jax") # get some arrays p = jnp.array([3.0, 2.0, 1.0]) t = jnp.array([0.0, 0.0, 0.0]) # and use the transpiled version! out = jax_loss(p, t)
I'm using NumPyβ
You can use Startai to get NumPy code from:Any library
From PyTorch
import startai import kornia import requests import numpy as np from PIL import Image # transpile kornia from torch to np np_kornia = startai.transpile(kornia, source="torch", to="numpy") # get an image url = "http://images.cocodataset.org/train2017/000000000034.jpg" raw_img = Image.open(requests.get(url, stream=True).raw) # convert it to the format expected by kornia img = np.transpose(np.array(raw_img), (2, 0, 1)) img = np.expand_dims(img, 0) / 255 # and use the transpiled version of any function from the library! out = np_kornia.enhance.sharpness(img, 5)From TensorFlow
import startai import numpy as np import os os.environ["SM_FRAMEWORK"] = "tf.keras" import segmentation_models as sm # transpile sm from tensorflow to numpy np_sm = startai.transpile(sm, source="tensorflow", to="numpy") # get some image-like arrays output = np.random.rand(1, 3, 512, 512).astype(dtype=np.float32) target = np.random.rand(1, 3, 512, 512).astype(dtype=np.float32) # and use the transpiled version of any function from the library! out = np_sm.metrics.iou_score(output, target)From Jax
import startai import rax import numpy as np # transpile rax from jax to numpy np_rax = startai.transpile(rax, source="jax", to="numpy") # get some arrays scores = np.array([2.2, 1.3, 5.4]) labels = np.array([1.0, 0.0, 0.0]) # and use the transpiled version of any function from the library! out = np_rax.poly1_softmax_loss(scores, labels)Any function
From PyTorch
import startai import torch import numpy as np def loss(predictions, targets): return torch.sqrt(torch.mean((predictions - targets) ** 2)) # transpile any function from torch to numpy np_loss = startai.transpile(loss, source="torch", to="numpy") # get some arrays p = np.array([3.0, 2.0, 1.0]) t = np.array([0.0, 0.0, 0.0]) # and use the transpiled version! out = np_loss(p, t)From TensorFlow
import startai import tensorflow as tf import numpy as np def loss(predictions, targets): return tf.sqrt(tf.reduce_mean(tf.square(predictions - targets))) # transpile any function from tf to numpy np_loss = startai.transpile(loss, source="tensorflow", to="numpy") # get some arrays p = np.array([3.0, 2.0, 1.0]) t = np.array([0.0, 0.0, 0.0]) # and use the transpiled version! out = np_loss(p, t)From JAX
import startai import jax.numpy as jnp import numpy as np def loss(predictions, targets): return jnp.sqrt(jnp.mean((predictions - targets) ** 2)) # transpile any function from jax to numpy np_loss = startai.transpile(loss, source="jax", to="numpy") # get some arrays p = np.array([3.0, 2.0, 1.0]) t = np.array([0.0, 0.0, 0.0]) # and use the transpiled version! out = np_loss(p, t)
I'm using Startaiβ
Or you can use Startai as a framework, breaking yourself (and your code) free from deciding which community to support, allowing anyone to run your code in their framework of choice!import startai
# A simple image classification model
class StartaiNet(startai.Module):
def __init__(
self,
h_w=(32, 32),
input_channels=3,
output_channels=512,
num_classes=2,
data_format="NCHW",
device="cpu",
):
self.h_w = h_w
self.input_channels = input_channels
self.output_channels = output_channels
self.num_classes = num_classes
self.data_format = data_format
super().__init__(device=device)
def _build(self, *args, **kwargs):
self.extractor = startai.Sequential(
startai.Conv2D(self.input_channels, 6, [5, 5], 1, "SAME", data_format=self.data_format),
startai.GELU(),
startai.Conv2D(6, 16, [5, 5], 1, "SAME", data_format=self.data_format),
startai.GELU(),
startai.Conv2D(16, self.output_channels, [5, 5], 1, "SAME", data_format=self.data_format),
startai.GELU(),
)
self.classifier = startai.Sequential(
# Since the padding is "SAME", this would be image_height x image_width x output_channels
startai.Linear(self.h_w[0] * self.h_w[1] * self.output_channels, 512),
startai.GELU(),
startai.Linear(512, self.num_classes),
)
def _forward(self, x):
x = self.extractor(x)
# flatten all dims except batch dim
x = startai.flatten(x, start_dim=1, end_dim=-1)
logits = self.classifier(x)
probs = startai.softmax(logits)
return logits, probs
After building your model in Startai, you can set your favourite framework as the backend to use its operations under the hood!
startai.set_backend("torch")
model = StartaiNet()
x = torch.randn(1, 3, 32, 32)
logits, probs = model(x)
startai.set_backend("tensorflow")
model = StartaiNet()
x = tf.random.uniform(shape=(1, 3, 32, 32))
logits, probs = model(x)
startai.set_backend("jax")
model = StartaiNet()
x = jax.random.uniform(key, shape=(1, 3, 32, 32))
logits, probs = model(x)
startai.set_backend("numpy")
model = StartaiNet()
x = np.random.uniform(size=(1, 3, 32, 32))
logits, probs = model(x)
Last but not least, we can also build the training pipeline in pure startai β¬οΈ
Let's define some helper functions first
# helper function for loading the dataset in batches
def generate_batches(images, classes, dataset_size, batch_size=32):
if batch_size > dataset_size:
raise startai.utils.exceptions.StartaiError("Use a smaller batch size")
for idx in range(0, dataset_size, batch_size):
yield images[idx : min(idx + batch_size, dataset_size)], classes[
idx : min(idx + batch_size, dataset_size)
]
# helper function to get the number of current predictions
def num_correct(preds, labels):
return (preds.argmax() == labels).sum().to_numpy().item()
# define a loss function
def loss_fn(params):
v, model, x, y = params
_, probs = model(x, v=v)
return startai.cross_entropy(y, probs), probs
And train this model!
# train the model on gpu if it's available
device = "gpu:0" if startai.gpu_is_available() else "cpu"
# training hyperparams
optimizer = startai.Adam(1e-4)
batch_size = 4
num_epochs = 20
num_classes = 10
model = StartaiNet(
h_w=(28, 28),
input_channels=1,
output_channels=120,
num_classes=num_classes,
device=device,
)
images = startai.random_uniform(shape=(16, 1, 28, 28))
classes = startai.randint(0, num_classes - 1, shape=(16,))
# training loop
def train(images, classes, epochs, model, device, num_classes=10, batch_size=32):
# training metrics
epoch_loss = 0.0
metrics = []
dataset_size = len(images)
for epoch in range(epochs):
train_correct = 0
train_loop = tqdm(
generate_batches(images, classes, len(images), batch_size=batch_size),
total=dataset_size // batch_size,
position=0,
leave=True,
)
for xbatch, ybatch in train_loop:
xbatch, ybatch = xbatch.to_device(device), ybatch.to_device(device)
# Since the cross entropy function expects the target classes to be in one-hot encoded format
ybatch_encoded = startai.one_hot(ybatch, num_classes)
# update model params
loss_probs, grads = startai.execute_with_gradients(
loss_fn,
(model.v, model, xbatch, ybatch_encoded),
)
model.v = optimizer.step(model.v, grads["0"])
batch_loss = startai.to_numpy(loss_probs[0]).mean().item() # batch mean loss
epoch_loss += batch_loss * xbatch.shape[0]
train_correct += num_correct(loss_probs[1], ybatch)
train_loop.set_description(f"Epoch [{epoch + 1:2d}/{epochs}]")
train_loop.set_postfix(
running_loss=batch_loss,
accuracy_percentage=(train_correct / dataset_size) * 100,
)
epoch_loss = epoch_loss / dataset_size
training_accuracy = train_correct / dataset_size
metrics.append([epoch, epoch_loss, training_accuracy])
train_loop.write(
f"\nAverage training loss: {epoch_loss:.6f}, Train Correct: {train_correct}",
end="\n",
)
# assuming the dataset(images and classes) are already prepared in a folder
train(
images,
classes,
num_epochs,
model,
device,
num_classes=num_classes,
batch_size=batch_size,
)
For a more comprehensive overview, head over to the Demos section with more on the basics, a few guides and a wide-ranging set of examples that demonstrate the transpilation of various popular models. We continue to expand on that list, let us know what demos you'd like us to add next π―
Let's take a look at how Startai works both as a transpiler and a framework in a bit more detail to get an idea of why and where to use it.
Startai as a transpiler
When should I use Startai as a transpiler?
If you want to use building blocks published in other frameworks (neural networks, layers, array computing libraries, training pipelines...), you want to integrate code developed in various frameworks, or maybe straight up move code from one framework to another, the transpiler is definitely the tool π§ for the job! As the output of transpilation is native code in the target framework, you can use the converted code just as if it was code originally developed in that framework, applying framework-specific optimizations or tools, instantly exposing your project to all of the unique perks of a different framework.
Startai's transpiler allows you to use code from any other framework (or from any other version of the same framework!) in your own code, by just adding one line of code. Under the hood, Startai traces a computational graph and leverages the frontends and backends to link one framework to another.
This way, Startai makes all ML-related projects available for you, independently of the framework you want to use to research, develop, or deploy systems. Feel free to head over to the docs for the full API reference, but the functions you'd most likely want to use are:
# Traces an efficient fully-functional graph from a function, removing all wrapping and redundant code
startai.trace_graph()
# Converts framework-specific code to a different framework
startai.transpile()
# Converts framework-specific code to Startai
startai.unify()
These functions can be used eagerly or lazily. If you pass the necessary arguments for function tracing, the graph tracing/transpilation step will happen instantly (eagerly). Otherwise, the graph tracing/transpilation will happen only when the returned function is first invoked.
import startai
import jax
startai.set_backend("jax")
# Simple JAX function to transpile
def test_fn(x):
return jax.numpy.sum(x)
x1 = startai.array([1., 2.])
# Arguments are available -> transpilation happens eagerly
eager_graph = startai.transpile(test_fn, source="jax", to="torch", args=(x1,))
# eager_graph is now torch code and runs efficiently
ret = eager_graph(x1)
# Arguments are not available -> transpilation happens lazily
lazy_graph = startai.transpile(test_fn, source="jax", to="torch")
# The transpiled graph is initialized, transpilation will happen here
ret = lazy_graph(x1)
# lazy_graph is now torch code and runs efficiently
ret = lazy_graph(x1)
If you want to learn more, you can find more information in the Startai as a transpiler section of the docs!
Startai as a framework
When should I use Startai as a framework?
As Startai supports multiple backends, writing code in Startai breaks you free from framework limitations. If you want to publish highly flexible code for everyone to use, independently of the framework they are using, or you plan to develop ML-related tools and want them to be interoperable with not only the already existing frameworks, but also with future frameworks, then Startai is for you!
The Startai framework is built on top of various essential components,
mainly the Backend
Handler,
which manages what framework is being used behind the scenes and the
Backend Functional
APIs,
which provide framework-specific implementations of the Startai functions.
Likewise, classes such as startai.Container
or startai.Array
are also
available, facilitating the use of structured data and array-like
objects (learn more about them
here!).
All of the functionalities in Startai are exposed through the
Startai functional API
and the Startai stateful API
. All functions in the
Functional
API
are Framework Agnostic Functions, which means that we can use them
like this:
import startai
import jax.numpy as jnp
import tensorflow as tf
import numpy as np
import torch
def mse_loss(y, target):
return startai.mean((y - target)**2)
jax_mse = mse_loss(jnp.ones((5,)), jnp.ones((5,)))
tf_mse = mse_loss(tf.ones((5,)), tf.ones((5,)))
np_mse = mse_loss(np.ones((5,)), np.ones((5,)))
torch_mse = mse_loss(torch.ones((5,)), torch.ones((5,)))
In the example above we show how Startai's functions are compatible with tensors from different frameworks. This is the same for ALL Startai functions. They can accept tensors from any framework and return the correct result.
The Startai Stateful API, on the other hand, allows you to define trainable modules and layers, which you can use alone or as a part of any other framework code!
import startai
class Regressor(startai.Module):
def __init__(self, input_dim, output_dim):
self.input_dim = input_dim
self.output_dim = output_dim
super().__init__()
def _build(self, *args, **kwargs):
self.linear0 = startai.Linear(self.input_dim, 128)
self.linear1 = startai.Linear(128, self.output_dim)
def _forward(self, x):
x = self.linear0(x)
x = startai.functional.relu(x)
x = self.linear1(x)
return x
If we put it all together, we'll have something like this. This example uses PyTorch as the backend, but this can easily be changed to your favorite frameworks, such as TensorFlow, or JAX.
import startai
class Regressor(startai.Module):
def __init__(self, input_dim, output_dim):
self.input_dim = input_dim
self.output_dim = output_dim
super().__init__()
def _build(self, *args, **kwargs):
self.linear0 = startai.Linear(self.input_dim, 128)
self.linear1 = startai.Linear(128, self.output_dim)
def _forward(self, x):
x = self.linear0(x)
x = startai.functional.relu(x)
x = self.linear1(x)
return x
startai.set_backend('torch') # set backend to PyTorch (or any other backend!)
model = Regressor(input_dim=1, output_dim=1)
optimizer = startai.Adam(0.3)
n_training_examples = 2000
noise = startai.random.random_normal(shape=(n_training_examples, 1), mean=0, std=0.1)
x = startai.linspace(-6, 3, n_training_examples).reshape((n_training_examples, 1))
y = 0.2 * x ** 2 + 0.5 * x + 0.1 + noise
def loss_fn(v, x, target):
pred = model(x, v=v)
return startai.mean((pred - target) ** 2)
for epoch in range(40):
# forward pass
pred = model(x)
# compute loss and gradients
loss, grads = startai.execute_with_gradients(lambda params: loss_fn(*params), (model.v, x, y))
# update parameters
model.v = optimizer.step(model.v, grads)
# print current loss
print(f'Epoch: {epoch + 1:2d} --- Loss: {startai.to_numpy(loss).item():.5f}')
print('Finished training!')
The model's output can be visualized as follows:
As always, you can find more information about Startai as a framework in the docs!
You can find Startai's documentation on the Docs page, which includes:
-
Motivation: This contextualizes the problem Startai is trying to solve by going over
- The current ML Explosion.
- Explaining why it is important to solve this problem.
- Explaining how we adhere to existing standards to make this happen.
- Related Work: Which paints a picture of the role Startai plays in the ML stack, comparing it to other existing solutions in terms of functionalities and abstraction level.
- Design: A user-focused guide about the design decision behind the architecture and the main building blocks of Startai.
- Deep Dive: Which delves deeper into the implementation details of Startai and is oriented towards potential contributors to the code base.
We believe that everyone can contribute and make a difference. Whether it's writing code π», fixing bugs π, or simply sharing feedback π¬, your contributions are definitely welcome and appreciated π
Check out all of our Open Tasks, and find out more info in our Contributing guide in the docs!
Join our amazing community as a contributor, and help accelerate our journey to unify all ML frameworks!
In order to achieve the ambitious goal of unifying AI, we definitely need as many hands as possible on it! Whether you are a seasoned developer or just starting out, you'll find a place here! Join the Startai community on our Discord πΎ server, which is the perfect place to ask questions, share ideas, and get help from both fellow developers and the Startai Team directly!
Also! Feel free to follow us on Twitter π¦ as well, we use it to share updates, sneak peeks, and all sorts of relevant news, certainly a great way to stay in the loop π
Can't wait to see you there!
If you use Startai for your work, please don't forget to give proper credit by including the accompanying paper π in your references. It's a small way to show appreciation and help to continue to support this and other open source projects π
@article{lenton2021startai,
title={Startai: Templated deep learning for inter-framework portability},
author={Lenton, Daniel and Pardo, Fabio and Falck, Fabian and James, Stephen and Clark, Ronald},
journal={arXiv preprint arXiv:2102.02886},
year={2021}
}