astra-lib 0.0.4

A ML toolkit with code bits useful for our day to day research

"AI for Sustainability" Toolkit for Research and Analysis. ASTRA (अस्त्र) means a "tool" or "a weapon" in Sanskrit.

Design Principles

Since astra is developed for research purposes, we'd try to adhere to these principles:

What we will try to do:

1. Keep the API simple-to-use and standardized to enable quick prototyping via automated scripts.
2. Keep the API transparent to expose as many details as possilbe. Explicit should be preferred over implicit.
3. Keep the API flexible to allow users to stretch the limits of their experiments.

What we will try to avoid:

4. We will try not to reduce code repeatation at expence of transparency, flexibility and performance. Too much abstraction often makes the API complex to understand and thus becomes hard to adapt for custom use cases.

Examples

Points	Example
1 and 2	We have exactly same arguments for all strategies in astra.torch.al.strategies to ease the automation but we explicitely mention in the docstrings if an argument is used or ignored for a strategy.
2	predict functions in astra by default put the model on eval mode but also allow to set eval_mode to False. This can be useful for techniques like MC dropout.
3	train_fn from astra.torch.utils works for all types of models and losses which may or may not be from astra.
4	Though F1 score can be computed from precision and recall, we explicitely use F1 score formula to allow transparency and to avoid computing TP multiple times.

Install

Stable version:

pip install astra-lib

Latest version:

pip install git+https://github.com/sustainability-lab/ASTRA

Contributing

Please go through the contributing guidelines before making a contribution.

Useful Code Snippets

Data

Load Data

from astra.torch.data import load_mnist, load_cifar_10

data = load_cifar_10()
print(data)

Files already downloaded and verified
Files already downloaded and verified

CIFAR-10 Dataset
length of dataset: 60000
shape of images: torch.Size([3, 32, 32])
len of classes: 10
classes: ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
dtype of images: torch.float32
dtype of labels: torch.int64
range of image values: min=0.0, max=1.0
            

Models

MLPs

from astra.torch.models import MLPRegressor

mlp = MLPRegressor(input_dim=100, hidden_dims=[128, 64], output_dim=10, activation="relu", dropout=0.1)
print(mlp)

MLPRegressor(
  (featurizer): MLP(
    (dropout): Dropout(p=0.1, inplace=True)
    (input_layer): Linear(in_features=100, out_features=128, bias=True)
    (hidden_layer_1): Linear(in_features=128, out_features=64, bias=True)
  )
  (regressor): Linear(in_features=64, out_features=10, bias=True)
)

CNNs

from astra.torch.models import CNNClassifier

cnn = CNNClassifier(
    image_dims=(32, 32),
    kernel_size=5,
    input_channels=3,
    conv_hidden_dims=[32, 64],
    dense_hidden_dims=[128, 64],
    n_classes=10,
)
print(cnn)

CNNClassifier(
  (featurizer): CNN(
    (activation): ReLU()
    (max_pool): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (input_layer): Conv2d(3, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
    (hidden_layer_1): Conv2d(32, 64, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
    (aggregator): Identity()
    (flatten): Flatten(start_dim=1, end_dim=-1)
  )
  (classifier): MLPClassifier(
    (featurizer): MLP(
      (activation): ReLU()
      (dropout): Dropout(p=0.0, inplace=True)
      (input_layer): Linear(in_features=4096, out_features=128, bias=True)
      (hidden_layer_1): Linear(in_features=128, out_features=64, bias=True)
    )
    (classifier): Linear(in_features=64, out_features=10, bias=True)
  )
)

EfficientNets

import torch
from torchvision.models import efficientnet_b0, EfficientNet_B0_Weights
from astra.torch.models import EfficientNetClassifier

# Pretrained model
model = EfficientNetClassifier(model=efficientnet_b0, weights=EfficientNet_B0_Weights.DEFAULT, n_classes=10)
# OR without pretrained weights
# model = EfficientNetClassifier(model=efficientnet_b0, weights=None, n_classes=10)

x = torch.rand(10, 3, 224, 224)
out = model(x)
print(out.shape)

torch.Size([10, 10])

ViT

import torch
from torchvision.models import vit_b_16, ViT_B_16_Weights
from astra.torch.models import ViTClassifier

model = ViTClassifier(vit_b_16, ViT_B_16_Weights.DEFAULT, n_classes=10)
x = torch.rand(10, 3, 224, 224)  # (batch_size, channels, h, w)
out = model(x)
print(out.shape)

torch.Size([10, 10])

Training

Train Function Usage

import torch
import torch.nn as nn
import numpy as np
from astra.torch.utils import train_fn
from astra.torch.models import CNNClassifier

torch.autograd.set_detect_anomaly(True)

X = torch.rand(100, 3, 28, 28)
y = torch.randint(0, 2, size=(200,)).reshape(100, 2).float()

model = CNNClassifier(
    image_dims=(28, 28), kernel_size=5, input_channels=3, conv_hidden_dims=[4], dense_hidden_dims=[2], n_classes=2
)

# Let train_fn do the optimization for you
iter_losses, epoch_losses = train_fn(
    model, input=X, output=y, loss_fn=nn.CrossEntropyLoss(), lr=0.1, epochs=5, verbose=False
)
print(np.array(epoch_losses).round(2))

# OR

# Define your own optimizer

optimizer = torch.optim.Adam(model.parameters(), lr=0.1)
iter_losses, epoch_losses = train_fn(
    model,
    input=X,
    output=y,
    loss_fn=nn.MSELoss(),
    optimizer=optimizer,
    verbose=False,
    epochs=5,
)
print(np.array(epoch_losses).round(2))

# Get the state_dict of the model at each epoch

(iter_losses, epoch_losses), state_dict_history = train_fn(
    model,
    input=X,
    output=y,
    loss_fn=nn.MSELoss(),
    lr=0.1,
    epochs=5,
    verbose=False,
    return_state_dict=True,
)
print(np.array(epoch_losses).round(2))

[0.65 0.64 0.64 0.64 0.64]
[0.28 0.26 0.25 0.25 0.25]
[0.26 0.25 0.25 0.25 0.25]

Advanced Usage

import torch
import torch.nn as nn
import numpy as np
from astra.torch.utils import train_fn
from astra.torch.models import AstraModel


class CustomModel(AstraModel):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(2, 1)
        self.inp1_linear = nn.Linear(2, 1)

    def forward(self, x, inp1, fixed_bias):
        return self.linear(x) + self.inp1_linear(inp1) + fixed_bias


def custom_loss_fn(model_output, output, norm_factor):
    loss_fn = nn.MSELoss()
    loss_val = loss_fn(model_output, output)
    return loss_val / norm_factor


X = torch.randn(10, 2)
y = torch.randn(10, 1)
inp1 = torch.randn(10, 2)
bias = torch.randn(1)
norm_factor = torch.randn(1)

model = CustomModel()

optimizer = torch.optim.Adam(model.parameters(), lr=0.1)
(iter_losses, epoch_losses), state_dict_history = train_fn(
    model,
    input=X,  # Can be None if model.forward() does not require input
    model_kwargs={"inp1": inp1, "fixed_bias": bias},
    output=y,  # Can be None if loss_fn does not require output
    loss_fn=custom_loss_fn,
    loss_fn_kwargs={"norm_factor": norm_factor},
    optimizer=optimizer,
    epochs=5,
    shuffle=True,
    verbose=True,
    return_state_dict=True,
)

print("Epoch_losses", np.array(epoch_losses).round(2))

Epoch_losses [2.61 1.54 2.71 2.61 2.65]


  0%|          | 0/5 [00:00<?, ?it/s]
Loss: 2.60764122:   0%|          | 0/5 [00:00<?, ?it/s]
Loss: 2.60764122:  20%|██        | 1/5 [00:00<00:02,  1.49it/s]
Loss: 1.54055369:  20%|██        | 1/5 [00:00<00:02,  1.49it/s]
Loss: 2.71163487:  20%|██        | 1/5 [00:00<00:02,  1.49it/s]
Loss: 2.60861230:  20%|██        | 1/5 [00:00<00:02,  1.49it/s]
Loss: 2.64996576:  20%|██        | 1/5 [00:00<00:02,  1.49it/s]
Loss: 2.64996576: 100%|██████████| 5/5 [00:00<00:00,  7.39it/s]

Others

Count number of parameters in a model

from astra.torch.utils import count_params
from astra.torch.models import MLPRegressor

mlp = MLPRegressor(input_dim=2, hidden_dims=[5, 6], output_dim=1)

n_params = count_params(mlp)
print(n_params)

{'total_params': 58, 'trainable_params': 58, 'non_trainable_params': 0}