Differences with torch.optim.SparseAdam

torch.optim.SparseAdam

class torch.optim.SparseAdam(
    params,
    lr=0.001,
    betas=(0.9, 0.999),
    eps=1e-08
)

For more information, see torch.optim.SparseAdam.

mindspore.nn.LazyAdam

class mindspore.nn.LazyAdam(
    params,
    learning_rate=1e-3,
    beta1=0.9,
    beta2=0.999,
    eps=1e-8,
    use_locking=False,
    use_nesterov=False,
    weight_decay=0.0,
    loss_scale=1.0
)

For more information, see mindspore.nn.LazyAdam.

Differences

torch.optimize.SparseAdam is an Adam algorithm in PyTorch specifically for sparse scenarios.

mindspore.nn.LazyAdam can be used for both regular and sparse scenarios:

• mindspore.nn.LazyAdam is consistent with torch.optimize.SparseAdam with default parameters when the input gradient is a sparse Tensor, but mindspore.nn.LazyAdam currently only supports CPU backends;

• When the input gradient is a non-sparse Tensor, mindspore.nn.LazyAdam automatically executes the mindspore.nn.Adam algorithm, and supports CPU/GPU/Ascend backends.

Categories	Subcategories	PyTorch	MindSpore	Difference
Parameters	Parameter 1	params	params	Consistent function
	Parameter 2	lr	learning_rate	Same function, different parameter names
	Parameter 3	betas	beta1, beta2	Same function, different parameter names
	Parameter 4	eps	eps	Consistent function
	Parameter 5	-	use_locking	MindSpore use_locking indicates whether parameter updates are protected by locking, and PyTorch does not have this parameter
	Parameter 6	-	use_nesterov	Whether MindSpore use_nesterov uses the NAG algorithm to update the gradient, and PyTorch does not have this parameter
	Parameter 7	-	weight_decay	PyTorch does not have this parameter
	Parameter 8	-	loss_scale	MindSpore's loss_scale is the gradient scaling factor, and PyTorch does not have this parameter

Code Example

# MindSpore.
import mindspore
from mindspore import nn

net = nn.Dense(2, 3)
optimizer = nn.LazyAdam(net.trainable_params())
criterion = nn.MAELoss(reduction="mean")

def forward_fn(data, label):
    logits = net(data)
    loss = criterion(logits, label)
    return loss, logits

grad_fn = mindspore.value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True)

def train_step(data, label):
    (loss, _), grads = grad_fn(data, label)
    optimizer(grads)
    return loss

# PyTorch
import torch

model = torch.nn.Linear(2, 3)
criterion = torch.nn.L1Loss(reduction='mean')
optimizer = torch.optim.SparseAdam(model.parameters())
def train_step(data, label):
    optimizer.zero_grad()
    output = model(data)
    loss = criterion(output, label)
    loss.backward()
    optimizer.step()