syna package

Subpackages

• syna.algo package
  • syna.algo.dqn module
• syna.datasets package
  • syna.datasets.datasets module
• syna.functions package
  • syna.functions.activation module
  • syna.functions.function module
  • syna.functions.loss module
  • syna.functions.math module
• syna.layers package
  • syna.layers.layer module
  • syna.layers.rnn module
• syna.models package
  • syna.models.models module
• syna.optim package
  • syna.optim.adadelta module
  • syna.optim.adagrad module
  • syna.optim.adam module
  • syna.optim.optimizer module
  • syna.optim.sgd module
• syna.utils package
  • syna.utils.dataloaders module
  • syna.utils.rl module
  • syna.utils.transforms module
  • syna.utils.util module
  • syna.utils.viz module

Submodules

• syna.core module

Module contents

class syna.Config

Bases: object

Global config flags affecting backprop and training behavior.

enable_backprop = True

train = True

class syna.DataLoader(dataset: Iterable[Tuple[Any, Any]], batch_size: int, shuffle: bool = True)

Bases: object

Simple iterable data loader.

Args:

dataset: Sequence of (input, target) pairs. batch_size: Number of samples per batch. shuffle: If True, shuffle dataset order at the start of each epoch.

Behavior:

• Iterates over dataset in batches. The last batch may be smaller.

• Iteration raises StopIteration at epoch end and automatically resets for the next epoch.

next()

reset() → None

Reset iteration counters and prepare indices for the next epoch.

class syna.Dataset(train: bool = True, transform: Callable[[Any], Any] | None = None, target_transform: Callable[[Any], Any] | None = None)

Bases: object

Minimal base dataset.

Subclasses should override prepare() to populate self.data (and optionally self.label).

Args:

train (bool): Whether this is a training split. transform (callable, optional): Function applied to each data item. target_transform (callable, optional): Function applied to each label.

Attributes:

data: Sequence of data items (must support indexing and len()). label: Sequence of labels or None.

prepare()

Populate self.data (and optionally self.label). Must be implemented by subclasses.

class syna.Function

Bases: object

Base Function for forward/backward ops used in autograd.

Subclasses should implement forward (numpy arrays -> numpy arrays) and backward (Tensor grads -> Tensor grads).

backward(*args: Tensor) → Tensor | list[Tensor]

forward(*args: ndarray) → ndarray

class syna.Layer

Bases: object

Base layer class that tracks Parameters and sub-Layers.

Subclasses must implement forward(). Layers register any attribute that is a Parameter or Layer automatically.

cleargrads()

Clear gradients of all parameters.

forward(inputs)

Compute the forward pass. Must be implemented by subclasses.

load_weights(path: str)

Load parameters from a .npz file created by save_weights().

params()

Yield all Parameter objects in this layer (including nested layers).

save_weights(path: str)

Save layer parameters to a compressed .npz file.

class syna.Model

Bases: Layer

Base model class (also a Layer) with a convenience plot method.

The plot method runs forward on the provided inputs and writes a DOT graph of the computation to the given file.

plot(*inputs, to_file='model.png') → None

Run forward and save a DOT graph of the output computation.

class syna.Parameter(data, name: str | None = None)

Bases: Tensor

A thin wrapper for trainable parameters (keeps API separate).

class syna.SeqDataLoader(dataset: Iterable[Tuple[Any, Any]], batch_size: int)

Bases: DataLoader

Sequence-aware loader that yields batch_size sequences in parallel.

The loader divides the dataset into batch_size streams by computing a ‘jump’ = data_size // batch_size and, for each iteration step k, takes the elements at positions (i*jump + k) % data_size for i in range(batch_size).

Args:

dataset: Sequence of (input, target) pairs. batch_size: Number of parallel sequences (streams) per batch.

Behavior:

• shuffle is always disabled for sequence loader.

• Iteration yields exactly data_size // jump steps (i.e., max_iter inherited).

class syna.Tensor(data, name: str | None = None)

Bases: object

Simple Tensor container holding data, gradient and creator Function.

Most operator behavior delegates to syna.functions.* helpers so this class focuses on bookkeeping for autograd.

property T

backward(retain_grad=False, create_graph=False) → None

Run backpropagation to compute gradients of inputs.

Args:

retain_grad: if False, intermediate gradients are cleared to save memory. create_graph: if True, create graph for higher-order gradients.

cleargrad() → None

Clear stored gradient.

property dtype

max(**kwargs)

min(**kwargs)

property ndim

reshape(*shape: int)

set_creator(func: Function) → None

Mark this tensor as created by func (used for backprop ordering).

property shape

property size

sum(axis=None, keepdims=False)

transpose(*axes: int)

unchain() → None

Remove reference to creator to break the computational graph.

unchain_backward()

Remove creators for all upstream tensors (useful for freeing graph).

syna.arange(start, stop=None, step=1, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Return a 1-D tensor with values from start to stop (exclusive).

syna.as_array(x) → ndarray

Convert scalars to numpy arrays; leave arrays unchanged.

syna.as_tensor(obj) → Tensor

Ensure obj is a Tensor; convert scalars/arrays to Tensor if needed.

Delegates to tensor(…), which follows PyTorch-style semantics.

syna.eye(n, m=None, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Return a 2-D identity matrix (or rectangular identity if m is given).

syna.full(shape, fill_value, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Return a tensor of shape shape filled with fill_value.

shape may be an int or a tuple/list of ints (matches numpy/torch style).

syna.full_like(x, fill_value, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Return a tensor with the same shape as x filled with fill_value.

x may be a Tensor or array-like.

syna.no_grad()

Context manager to disable gradient tracking.

syna.ones(*shape, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Return a tensor filled with ones.

syna.ones_like(x, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Return a tensor of ones with the same shape as x (Tensor or array-like).

syna.randint(low, high=None, size=None, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Return random integers from low (inclusive) to high (exclusive).

syna.randn(*shape, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Return a tensor with samples from the standard normal distribution.

syna.range(start, stop=None, step=1, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Alias for arange to match torch.range-like convenience (behaves like arange).

syna.tensor(data, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Tensor factory.

Args:

data: array-like or scalar. dtype: numpy dtype (optional). requires_grad: if True, the returned Tensor is a leaf that will accumulate gradients. name: optional name for the Tensor.

syna.test_mode()

Context manager to set train flag to False.

syna.using_config(name: str, value: bool)

Temporarily set a Config attribute inside a context.

syna.zeros_like(x, dtype=None, requires_grad: bool = False, name: str | None = None) → Tensor

Return a tensor of zeros with the same shape as x (Tensor or array-like).