Optimizers

Collection of Ivy optimizers.

class ivy.stateful.optimizers.Adam(lr=0.0001, beta1=0.9, beta2=0.999, epsilon=1e-07, inplace=True, stop_gradients=True, compile_on_next_step=False, device=None)[source]

Bases: Optimizer

__init__(lr=0.0001, beta1=0.9, beta2=0.999, epsilon=1e-07, inplace=True, stop_gradients=True, compile_on_next_step=False, device=None)[source]

Construct an ADAM optimizer.

Parameters
  • lr (float) – Learning rate, default is 1e-4. (default: 0.0001)

  • beta1 (float) – gradient forgetting factor, default is 0.9 (default: 0.9)

  • beta2 (float) – second moment of gradient forgetting factor, default is 0.999 (default: 0.999)

  • epsilon (float) – divisor during adam update, preventing division by zero, default is 1e-07 (default: 1e-07)

  • inplace (bool) – Whether to update the variables in-place, or to create new variable handles. (default: True) This is only relevant for frameworks with stateful variables such as PyTorch. Default is True, provided the backend framework supports it.

  • stop_gradients (bool) – Whether to stop the gradients of the variables after each gradient step. (default: True) Default is True.

  • compile_on_next_step (bool) – Whether to compile the optimizer on the next step. Default is False. (default: False)

  • device (Optional[Union[Device, NativeDevice]]) – Device on which to create the layer’s variables ‘cuda:0’, ‘cuda:1’, ‘cpu’ (default: None) etc. (Default value = None)

set_state(state)[source]

Set state of the optimizer.

Parameters

state (Container) – Nested state to update.

property state
class ivy.stateful.optimizers.LAMB(lr=0.0001, beta1=0.9, beta2=0.999, epsilon=1e-07, max_trust_ratio=10, decay_lambda=0, inplace=True, stop_gradients=True, compile_on_next_step=False, device=None)[source]

Bases: Optimizer

__init__(lr=0.0001, beta1=0.9, beta2=0.999, epsilon=1e-07, max_trust_ratio=10, decay_lambda=0, inplace=True, stop_gradients=True, compile_on_next_step=False, device=None)[source]

Construct an LAMB optimizer.

Parameters
  • lr (float) – Learning rate, default is 1e-4. (default: 0.0001)

  • beta1 (float) – gradient forgetting factor, default is 0.9 (default: 0.9)

  • beta2 (float) – second moment of gradient forgetting factor, default is 0.999 (default: 0.999)

  • epsilon (float) – divisor during adam update, preventing division by zero, default is 1e-07 (default: 1e-07)

  • max_trust_ratio (float) – The max value of the trust ratio; the ratio between the norm of the layer (default: 10) weights and norm of gradients update. Default is 10.

  • decay_lambda (float) – The factor used for weight decay. Default is zero. (default: 0)

  • inplace (bool) – Whether to update the variables in-place, or to create new variable handles. (default: True) This is only relevant for frameworks with stateful variables such as PyTorch. Default is True, provided the backend framework supports it.

  • stop_gradients (bool) – Whether to stop the gradients of the variables after each gradient step. (default: True) Default is True.

  • compile_on_next_step (bool) – Whether to compile the optimizer on the next step. Default is False. (default: False)

  • device (Optional[Union[Device, NativeDevice]]) – Device on which to create the layer’s variables ‘cuda:0’, ‘cuda:1’, ‘cpu’ (default: None) etc. (Default value = None)

set_state(state)[source]

Set state of the optimizer.

Parameters

state (Container) – Nested state to update.

property state
class ivy.stateful.optimizers.LARS(lr=<function LARS.<lambda>>, decay_lambda=0, inplace=True, stop_gradients=True, compile_on_next_step=False)[source]

Bases: Optimizer

__init__(lr=<function LARS.<lambda>>, decay_lambda=0, inplace=True, stop_gradients=True, compile_on_next_step=False)[source]

Construct a Layer-wise Adaptive Rate Scaling (LARS) optimizer.

Parameters
  • lr (float) – Learning rate, default is 1e-4. (default: <function LARS.<lambda> at 0x7fce0cf48af0>)

  • decay_lambda (float) – The factor used for weight decay. Default is zero. (default: 0)

  • inplace (bool) – Whether to update the variables in-place, or to create new variable handles. (default: True) This is only relevant for frameworks with stateful variables such as PyTorch. Default is True, provided the backend framework supports it.

  • stop_gradients (bool) – Whether to stop the gradients of the variables after each gradient step. (default: True) Default is True.

  • compile_on_next_step (bool) – Whether to compile the optimizer on the next step. Default is False. (default: False)

set_state(state)[source]

Set state of the optimizer.

Parameters

state (Container) – Nested state to update.

property state
class ivy.stateful.optimizers.Optimizer(lr, inplace=True, stop_gradients=True, init_on_first_step=False, compile_on_next_step=False, fallback_to_non_compiled=False, device=None)[source]

Bases: ABC

__init__(lr, inplace=True, stop_gradients=True, init_on_first_step=False, compile_on_next_step=False, fallback_to_non_compiled=False, device=None)[source]

Construct a general Optimizer. This is an abstract class, and must be derived.

Parameters
  • lr (float) – Learning rate.

  • inplace (bool) – Whether to update the variables in-place, or to create new variable handles. (default: True) This is only relevant for frameworks with stateful variables such as PyTorch. Default is True, provided the backend framework supports it.

  • stop_gradients (bool) – Whether to stop the gradients of the variables after each gradient step. (default: True) Default is True.

  • init_on_first_step (bool) – Whether the optimizer is initialized on the first step. Default is False. (default: False)

  • compile_on_next_step (bool) – Whether to compile the optimizer on the next step. Default is False. (default: False)

  • fallback_to_non_compiled (bool) – Whether to fall back to non-compiled forward call in the case that an error (default: False) is raised during the compiled forward pass. Default is True.

  • device (Optional[Union[Device, NativeDevice]]) – Device on which to create the layer’s variables ‘cuda:0’, ‘cuda:1’, ‘cpu’ (default: None) etc. (Default value = None)

abstract set_state(state)[source]

Set state of the optimizer.

Parameters

state (Container) – Nested state to update.

step(v, grads, ignore_missing=False)[source]

Update nested variables container v from overridden private self._step

Parameters
  • v (Container) – Nested variables to update.

  • grads (Container) – Nested gradients to update.

  • ignore_missing (bool) – Whether to ignore keys missing from the gradients which exist in (default: False) the variables. Default is False.

Returns

ret – The updated variables, following update step.

class ivy.stateful.optimizers.SGD(lr=<function SGD.<lambda>>, inplace=True, stop_gradients=True, compile_on_next_step=False)[source]

Bases: Optimizer

__init__(lr=<function SGD.<lambda>>, inplace=True, stop_gradients=True, compile_on_next_step=False)[source]

Construct a Stochastic-Gradient-Descent (SGD) optimizer.

Parameters
  • lr (float) – Learning rate, default is 1e-4. (default: <function SGD.<lambda> at 0x7fce0cf48820>)

  • inplace (bool) – Whether to update the variables in-place, or to create new variable handles. (default: True) This is only relevant for frameworks with stateful variables such as PyTorch. Default is True, provided the backend framework supports it.

  • stop_gradients (bool) – Whether to stop the gradients of the variables after each gradient step. (default: True) Default is True.

  • compile_on_next_step (bool) – Whether to compile the optimizer on the next step. Default is False. (default: False)

set_state(state)[source]

Set state of the optimizer.

Parameters

state (Container) – Nested state to update.

property state