slowtorch 31.5.2025

Yet another implementation of PyTorch from the ground up, but for real!

Logo generated using FLUX.1.

SlowTorch

SlowTorch is another personal pet-project of mine where I tried and implemented the basic and bare-bones functionality of PyTorch just using pure Python, similar to what I did with xsNumPy. This project is also a testament to the richness of PyTorch’s Tensor-oriented design. By reimplementing its core features in a self-contained and minimalistic fashion, this project aims to:

• Provide an educational tool for those seeking to understand tensor and automatic gradient (backpropagation) mechanics.

• Encourage developers to explore the intricacies of multidimensional array computation.

This project acknowledges the incredible contributions of the PyTorch team and community over decades of development. While this module reimagines PyTorch’s functionality, it owes its design, inspiration, and motivation to the pioneering work of the core PyTorch developers. If that’s obvious, this module is not a replacement for PyTorch by any stretch but an homage to its brilliance and an opportunity to explore its concepts from the ground up.

SlowTorch is a lightweight, pure-Python library inspired by PyTorch, designed to mimic essential tensor operations and auto-differentiation (backpropagation) features. This project is ideal for learning and experimentation with multidimensional tensor processing.

Installation

Install the latest version of SlowTorch using pip:

pip install -U git+https://github.com/xames3/slowtorch.git#egg=slowtorch

Features

As of now, SlowTorch offers the following features:

SlowTorch native Tensor object (slowtorch.Tensor)

• slowtorch.Tensor. The central data structure representing N-dimensional tensors with support for:

  • Arbitrary shapes and data types.

  • Broadcasting for compatible operations.

  • Arithmetic and comparison operations.

  >>> import slowtorch
  >>>
  >>> a = slowtorch.tensor([[1, 2, 3, 4, 5], [6, 7, 8, 9, 10]])
  >>> b = slowtorch.tensor([[4, 1, 5, 3, 2], [1, 3, 5, 7, 2]])
  >>>
  >>> a + b
  tensor([[ 5,  3,  8,  7,  7],
          [ 7, 10, 13, 16, 12]])

Tensor Creation Ops

• slowtorch.tensor. Create an N-dimensional tensor.

  >>> slowtorch.tensor([[0.1, 1.2], [2.2, 3.1], [4.9, 5.2]])
  tensor([[0.1, 1.2],
          [2.2, 3.1],
          [4.9, 5.2]])
  >>> slowtorch.tensor([[1, 3], [2, 3]])
  tensor([[1, 3],
          [2, 3]])

• slowtorch.empty. Create an uninitialised tensor of the given shape.

  >>> slowtorch.empty(2, 3)
  tensor([[ 0.,  0.,  0.],
          [ 0.,  0.,  0.]])

• slowtorch.zeros. Create a tensor filled with zeros.

  >>> slowtorch.zeros(2, 3)
  tensor([[ 0.,  0.,  0.],
          [ 0.,  0.,  0.]])

• slowtorch.ones. Create a tensor filled with ones.

  >>> slowtorch.ones(2, 3)
  tensor([[ 1.,  1.,  1.],
          [ 1.,  1.,  1.]])

• slowtorch.full. Create a tensor filled with fill_value.

  >>> slowtorch.full(2, 3, fill_value=3.141592)
  tensor([[3.1416, 3.1416, 3.1416],
          [3.1416, 3.1416, 3.1416]])

• slowtorch.arange. Generate evenly spaced values within a given range.

  >>> slowtorch.arange(5)
  tensor([0, 1, 2, 3, 4])
  >>> slowtorch.arange(1, 4)
  tensor([1, 2, 3])
  >>> slowtorch.arange(1, 2.5, 0.5)
  tensor([ 1., 1.5,  2.])

• slowtorch.linspace. Generate evenly spaced values from start to end, inclusive.

  >>> slowtorch.linspace(3, 10, steps=5)
  tensor([ 3.00000,  4.75000,  6.50000,  8.25000, 10.00000])
  >>> slowtorch.linspace(-10, 10, steps=5)
  tensor([-10.,  -5.,   0.,   5.,  10.])
  >>> slowtorch.linspace(start=-10, end=10, steps=5)
  tensor([-10.,  -5.,   0.,   5.,  10.])
  >>> slowtorch.linspace(start=-10, end=10, steps=1)
  tensor([-10.])

• slowtorch.cat. Concatenates the given sequence of tensors in tensors in the given dimension.

  >>> a = slowtorch.tensor([2.0, 4.5, -1.7])
  >>> slowtorch.cat((a, a))
  tensor([  2.,  4.5, -1.7,   2.,  4.5, -1.7])

Autograd Mechanics

• Automatic Differentiation. In lieu of mimicking PyTorch’s functionality, pivotal feature of this project is a simple Pythonic version of automatic differentiation, akin to PyTorch’s autograd. It allows for the computation of gradients automatically, which is essential for training neural networks.

  Note. To learn more about Autograd Mechanics, see this.

  >>> a = slowtorch.tensor(2.0, requires_grad=True)
  >>> b = slowtorch.tensor(3.0, requires_grad=True)
  >>> c = slowtorch.tensor(-7.0, requires_grad=True)
  >>> d = a + b * c
  >>> d
  tensor(-19.0, grad_fn=<AddBackward0>)
  >>> d.backward()
  >>> a.grad
  tensor(1.0)
  >>> b.grad
  tensor(-7.0, grad_fn=<AddBackward0>)
  >>> c.grad
  tensor(3.0, grad_fn=<AddBackward0>)

• Specialised Backward Functions. Like PyTorch, SlowTorch also implements some specialised backward functions for backpropagation. These functions are mainly for representing the derivative or gradient calculations of the said functions. SlowTorch supports a few backward functions when requires_grad is True:

  • AddBackward0. For addition operations.

  >>> a = slowtorch.tensor([2.0, 4.5, -1.7], requires_grad=True)
  >>> b = slowtorch.tensor([-0.24, 1.4, 7.2], requires_grad=True)
  >>> a + b
  tensor([1.76,  5.9,  5.5], grad_fn=<AddBackward0>)

  • SubBackward0. For subtraction operations.

  >>> a - b
  tensor([2.24,  3.1, -8.9], grad_fn=<SubBackward0>)

  • MulBackward0. For multiplication operations

  >>> a * b
  tensor([ -0.48,    6.3, -12.24], grad_fn=<MulBackward0>)

  • DivBackward0. For division operations.

  >>> a / b
  tensor([-8.3333,  3.2143, -0.2361], grad_fn=<DivBackward0>)
  >>> a // b
  tensor([-9.0, 3.00, -1.0], grad_fn=<DivBackward0>)

  • NegBackward0. For negation operations.

  >>> a = slowtorch.tensor([2.0, 4.5, -1.7], requires_grad=True)
  >>> -a
  tensor([ -2., -4.5,  1.7], grad_fn=<NegBackward0>)

  • DotBackward0. For matrix multiplication operations.

  >>> a @ b
  tensor(-6.42, grad_fn=<DotBackward0>)

  • PowBackward0. For exponentiation operations.

  >>> a ** 2
  tensor([4.000, 20.25,  2.89], grad_fn=<PowBackward0>)

  • LogBackward0. For logarithmic operations.

  >>> a.log()
  tensor([0.69315, 1.50408,    nan], grad_fn=<LogBackward0>)

  • CloneBackward0. For clone/copy operation.

  >>> a.clone()
  tensor([2.00,  4.5, -1.7], grad_fn=<CloneBackward0>)

  • SumBackward0. For calculating sum.

  >>> a.sum()
  tensor(4.8, grad_fn=<SumBackward0>)

  • MaxBackward0. For calculating maximum.

  >>> b.max()
  tensor(7.2, grad_fn=<MaxBackward0>)

  • MinBackward0. For calculating minimum.

  >>> a.min()
  tensor(-1.7, grad_fn=<MinBackward0>)

  • MeanBackward0. For calculating mean.

  >>> a.mean()
  tensor(1.6, grad_fn=<MeanBackward0>)

  • StdBackward0. For calculating standard deviation.

  >>> a.std()
  tensor(3.119294882, grad_fn=<StdBackward0>)

  • ExpBackward0. For exponentiation operation with respect to e.

  >>> a.exp()
  tensor([ 7.3891, 90.0171,  0.1827], grad_fn=<ExpBackward0>)

  • SqrtBackward0. For calculating square-roots.

  >>> a.sqrt()
  tensor([1.4142, 2.1213,    nan], grad_fn=<SqrtBackward0>)

  • ReluBackward0. When using ReLU non-linearity function.

  >>> a.relu()
  tensor([2.0, 4.5, 0.0], grad_fn=<ReluBackward0>)

  • EluBackward0. When using ELU non-linearity function.

  >>> a.elu(alpha=0.5)
  tensor([2.00000,     4.5, -0.4087], grad_fn=<EluBackward0>)

  • TanhBackward0. When using Tanh non-linearity function.

  >>> a.tanh()
  tensor([  0.964,  0.9998, -0.9354], grad_fn=<TanhBackward0>)

  • SigmoidBackward0. When using Sigmoid non-linearity function.

  >>> a.sigmoid()
  tensor([0.8808,  0.989, 0.1545], grad_fn=<SigmoidBackward0>)

  • AddmmBackward0. For calculating input @ weight.T + bias in Linear layer.

  >>> import slowtorch
  >>> import slowtorch.nn as nn
  >>>
  >>> xs = slowtorch.tensor(
  ...     [
  ...         [1.5, 6.2, 2.6, 3.1, 5.3, 5.3, 7.9, 2.8],
  ...         [3.9, 2.8, 9.3, 6.4, 8.5, 6.9, 3.8, 3.1],
  ...         [3.4, 6.0, 4.4, 8.7, 9.7, 7.7, 1.6, 7.5],
  ...         [6.7, 8.8, 7.5, 1.8, 3.3, 8.4, 4.7, 5.1],
  ...         [6.8, 0.6, 4.8, 2.9, 6.8, 3.6, 3.5, 5.6],
  ...         [4.3, 4.2, 3.7, 7.0, 3.5, 8.5, 2.4, 2.9],
  ...     ],
  ...     requires_grad=True
  ... )
  >>> ys = slowtorch.tensor(
  ...     [
  ...         [-1.0],
  ...         [+1.0],
  ...         [-1.0],
  ...         [+1.0],
  ...         [-1.0],
  ...         [-1.0],
  ...     ]
  ... )
  >>>
  >>> class NeuralNetwork(nn.Module):
  ...     def __init__(self, in_features, out_features):
  ...             super().__init__(in_features, out_features)
  ...             self.linear = nn.Linear(in_features, out_features)
  ...             self.out = nn.Linear(out_features, 1)
  ...     def forward(self, x):
  ...             return self.out(self.linear(x))
  ...
  >>> model = NeuralNetwork(8, 16)
  >>> ypred = model(xs)
  >>> ypred
  tensor([[1.5218],
          [1.5177],
          [1.8904],
          [3.6145],
          [1.7698],
          [2.0918]], grad_fn=<AddmmBackward0>)

  Note. The above demonstration is just for the forward pass through a linear layer without any activation. To get better results, you need to train the model with additional activation layer(s).

  • MSELossBackward0. When calculating Mean Squared Error loss.

  >>> criterion = nn.MSELoss()
  >>> loss = criterion(ypred, ys)
  >>> loss
  tensor(6.5081, grad_fn=<MSELossBackward0>)

Tensor class reference

• Tensor.device. Device where the tensor is.

  >>> a = slowtorch.tensor([[1, 2, 3, 4, 5], [6, 7, 8, 9, 10]])
  >>> a.device
  device(type='cpu', index=0)

• Tensor.grad. This attribute is None by default and becomes a Tensor the first time a call to backward() computes gradients for self.

• Tensor.ndim. Returns the number of dimensions of self tensor. Alias for Tensor.dim().

  >>> a = slowtorch.tensor([[1, 2, 3, 4, 5], [6, 7, 8, 9, 10]])
  >>> a.ndim
  2
  >>> b = slowtorch.zeros(2, 3, 4)
  >>> b.dim()
  3

• Tensor.nbytes. Total bytes consumed by the elements of the tensor.

  >>> a = slowtorch.zeros(3, 2, dtype=slowtorch.float64)
  >>> a
  tensor([[ 0.,  0.],
          [ 0.,  0.],
          [ 0.,  0.]])
  >>> a.nbytes
  48
  >>> b = slowtorch.zeros(1, 3, dtype=slowtorch.int64)
  >>> b
  tensor([[0, 0, 0]])
  >>> b.nbytes
  24

• Tensor.itemsize. Length of one tensor element in bytes. Alias for Tensor.element_size().

  >>> a = slowtorch.full(2, 3, fill_value=2.71253)
  >>> a
  tensor([[2.71253, 2.71253, 2.71253],
          [2.71253, 2.71253, 2.71253]])
  >>> a.itemsize
  8
  >>> b = slowtorch.tensor([1, 2, 3], dtype=slowtorch.int16)
  >>> b.element_size()
  2

• Tensor.shape. Size of the tensor as a tuple.

  >>> a = slowtorch.zeros(1, 3, dtype=slowtorch.int64)
  >>> a
  tensor([[0, 0, 0]])
  >>> a.shape
  (1, 3)
  >>> b = slowtorch.zeros(3, 5, 2, dtype=slowtorch.float64)
  >>> b.shape
  (3, 5, 2)
  >>> b.shape = (3, 10)
  >>> b
  tensor([[ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
          [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
          [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.]])

• Tensor.data. Python buffer object pointing to the start of the tensor’s data.

  >>> a = slowtorch.ones(2, 7)
  >>> a.data
  tensor([[ 1.,  1.,  1.,  1.,  1.,  1.,  1.],
          [ 1.,  1.,  1.,  1.,  1.,  1.,  1.]])

• Tensor.dtype. Data-type of the tensor’s elements.

  >>> a = slowtorch.ones(2, 7)
  >>> a.dtype
  slowtorch.float64
  >>> b = slowtorch.zeros(3, 5, 2, dtype=slowtorch.int16)
  >>> b.dtype
  slowtorch.int16
  >>> type(b.dtype)
  <class 'slowtorch.dtype'>

• Tensor.is_cuda. Is True if the Tensor is stored on the GPU, False otherwise.

  >>> a = slowtorch.tensor((1, 2, 3, 4, 5))
  >>> a.is_cuda
  False

• Tensor.is_quantized. Is True if the Tensor is quantized, False otherwise.

  >>> a = slowtorch.tensor((1, 2, 3))
  >>> a.is_quantized
  False

• Tensor.is_meta. Is True if the Tensor is a meta tensor, False otherwise.

  >>> a = slowtorch.zeros(1, 2, 3)
  >>> a.is_meta
  False

• Tensor.T. View of the transposed array.

  >>> a = slowtorch.tensor([[1, 2], [3, 4]])
  >>> a
  tensor([[1, 2],
          [3, 4]])
  >>> a.T
  tensor([[1, 3],
          [2, 4]])

Tensor class methods

• Tensor.to(). Copies a tensor to a specified data type. Alias for Tensor.type()

  >>> a = slowtorch.tensor((1, 2, 3, 4, 5))
  >>> a
  tensor([1, 2, 3, 4, 5])
  >>> a.to(slowtorch.float64)
  tensor([ 1.,  2.,  3.,  4.,  5.])
  >>> a.type(slowtorch.bool)
  tensor([True, True, True, True, True])

• Tensor.size(). Number of elements in the tensor.

  >>> a = slowtorch.tensor((1, 2, 3, 4, 5))
  >>> a.size()
  slowtorch.Size([5])
  >>> b = slowtorch.ones(2, 3)
  >>> b
  tensor([[ 1.,  1.,  1.],
          [ 1.,  1.,  1.]])
  >>> b.size()
  slowtorch.Size([2, 3])

• Tensor.stride(). Tuple of bytes to step in each dimension when traversing a tensor.

  >>> a = slowtorch.ones(2, 3)
  >>> a.stride()
  (3, 1)

• Tensor.nelement(). Return total number of elements in a tensor. Alias for Tensor.numel().

  >>> a = slowtorch.ones(2, 3)
  >>> a
  tensor([[ 1.,  1.,  1.],
          [ 1.,  1.,  1.]])
  >>> a.nelement()
  6
  >>> b = slowtorch.tensor((1, 2, 3, 4, 5))
  >>> b.numel()
  5

• Tensor.clone(). Return a deep copy of the tensor.

  >>> a = slowtorch.tensor((1, 2, 3, 4, 5))
  >>> b = a.clone()
  >>> b
  tensor([1, 2, 3, 4, 5])

• Tensor.fill_(). Fill the tensor with a scalar value.

  >>> a = slowtorch.tensor([1, 2])
  >>> a.fill_(0)
  >>> a
  tensor([0, 0])

• Tensor.flatten(). Return a copy of the tensor collapsed into one dimension.

  >>> a = slowtorch.tensor([[1, 2], [3, 4]])
  >>> a.flatten()
  tensor([1, 2, 3, 4])

• Tensor.item(). Copy an element of a tensor to a standard Python scalar and return it.

  >>> a = slowtorch.tensor((2,))
  >>> a
  tensor([2])
  >>> a.item()
  2

• Tensor.view(). Gives a new shape to a tensor without changing its data. Alias for Tensor.reshape().

  >>> a = slowtorch.arange(6).view(3, 2)
  >>> a
  tensor([[0, 1],
          [2, 3],
          [4, 5]])
  >>> a = slowtorch.tensor([[1, 2, 3], [4, 5, 6]])
  >>> a.reshape(6)
  tensor([1, 2, 3, 4, 5, 6])

• Tensor.transpose(). Returns a tensor with dimensions transposed. Alias for Tensor.swapaxes and Tensor.swapdims.

  >>> a = slowtorch.tensor([[1, 2], [3, 4]])
  >>> a
  tensor([[1, 2],
          [3, 4]])
  >>> a.transpose()
  tensor([[1, 3],
          [2, 4]])
  >>> a = slowtorch.tensor([1, 2, 3, 4])
  >>> a.swapaxes()
  tensor([1, 2, 3, 4])
  >>> a = slowtorch.ones((1, 2, 3))
  >>> a.swapdims((1, 0, 2)).shape
  (2, 1, 3)

Constants

• slowtorch.e. Euler’s constant.

  >>> slowtorch.e
  2.718281828459045

• slowtorch.inf. IEEE 754 floating point representation of (positive) infinity.

  >>> slowtorch.inf
  inf

• slowtorch.nan. IEEE 754 floating point representation of Not a Number (NaN).

  >>> slowtorch.nan
  nan

• slowtorch.newaxis. A convenient alias for None, useful for indexing tensors.

  >>> slowtorch.newaxis is None
  True

• slowtorch.pi. Pi…

  >>> slowtorch.pi
  3.141592653589793

SlowTorch In Action

Below is a small demonstration of what SlowTorch can do, albeit… slowly.

>>> import slowtorch
>>> import slowtorch.nn as snn
>>>
>>> xs = slowtorch.tensor(
...     [
...         [1.5, 6.2, 2.6, 3.1, 5.3, 5.3, 7.9, 2.8],
...         [3.9, 2.8, 9.3, 6.4, 8.5, 6.9, 3.8, 3.1],
...         [3.4, 6.0, 4.4, 8.7, 9.7, 7.7, 1.6, 7.5],
...         [6.7, 8.8, 7.5, 1.8, 3.3, 8.4, 4.7, 5.1],
...         [6.8, 0.6, 4.8, 2.9, 6.8, 3.6, 3.5, 5.6],
...         [4.3, 4.2, 3.7, 7.0, 3.5, 8.5, 2.4, 2.9],
...     ],
...     requires_grad=True
... )
>>> ys = slowtorch.tensor(
...     [
...         [0.558],
...         [0.175],
...         [0.152],
...         [0.485],
...         [0.232],
...         [0.0134],
...     ]
... )
>>>
>>>
>>> class NeuralNetwork(snn.Module):
...     def __init__(self):
...         super().__init__()
...         self.l1 = snn.Linear(8, 16)
...         self.l2 = snn.Linear(16, 32)
...         self.l3 = snn.Linear(32, 16)
...         self.l4 = snn.Linear(16, 8)
...         self.l5 = snn.Linear(8, 1)
...         self.tanh = snn.Tanh()
...     def forward(self, x):
...         x = self.tanh(self.l1(x))
...         x = self.tanh(self.l2(x))
...         x = self.tanh(self.l3(x))
...         x = self.tanh(self.l4(x))
...         x = self.tanh(self.l5(x))
...         return x
...
>>>
>>> model = NeuralNetwork()
>>> print(f"Parameters: {sum(p.nelement() for p in model.parameters())}")
Parameters: 1361
>>>
>>> epochs = 500
>>> criterion = snn.MSELoss()
>>> optimiser = slowtorch.optim.SGD(model.parameters(), 0.1, momentum=0.1)
>>>
>>> for epoch in range(epochs):
...     ypred = model(ys)
...     loss = criterion(ypred, ys)
...     optimiser.zero_grad()
...     loss.backward()
...     optimiser.step()
...     if epoch % 100 == 0:
...         print(f"New loss: {loss.item():.7f}")
...
New loss: 0.0403600
New loss: 0.0098700
New loss: 0.0002800
New loss: 0.0000100
New loss: 0.0000000
>>> ypred
tensor([[0.55807],
        [0.17516],
        [0.15148],
        [ 0.4849],
        [0.23193],
        [0.01396]], grad_fn=<TanhBackward0>)
>>>

Usage and Documentation

The codebase is structured to be intuitive and mirrors the design principles of PyTorch to a significant extent. Comprehensive docstrings are provided for each module and function, ensuring clarity and ease of understanding. Users are encouraged to delve into the code, experiment with it, and modify it to suit their learning curve.

Since, the implementation doesn’t rely on any external packages, it will work with any CPython build v3.10 and higher. Technically, it should work on 3.9 and below as well but due to some syntactical and type-aliasing changes, it will not support it directly. For instance, the typing module was significantly changed in 3.10, hence some features like types.GenericAlias and using native types like tuple, list, etc. will not work. If you choose to remove all the typing stuff, the code will work just fine, at least that’s what I hope.

Note. SlowTorch cannot and should not be used as an alternative to PyTorch.

Contributions and Feedback

Contributions to this project are warmly welcomed. Whether it’s refining the code, enhancing the documentation, or extending the current feature set, your input is highly valued. Feedback, whether constructive criticism or commendation, is equally appreciated and will be instrumental in the evolution of this educational tool.

Acknowledgments

This project is inspired by the remarkable work done by the PyTorch Development Team. It is a tribute to their contributions to the field of machine learning and the open-source community at large.

Note. This project also takes massive inspiration from excellent work done by Andrej Karpathy in his micrograd project.

License

SlowTorch is licensed under the MIT License. See the LICENSE file for more details.