Function Differences with torch.optim.Adagrad

View Source On Gitee

torch.optim.Adagrad

class torch.optim.Adagrad(
    params,
    lr=0.01,
    lr_decay=0,
    weight_decay=0,
    initial_accumulator_value=0,
    eps=1e-10
)

For more information, see torch.optim.Adagrad.

mindspore.nn.Adagrad

class mindspore.nn.Adagrad(
    params,
    accum=0.1,
    learning_rate=0.001,
    update_slots=True,
    loss_scale=1.0,
    weight_decay=0.0
)(grads)

For more information, see mindspore.nn.Adagrad.

Differences

PyTorch: Parameters to be optimized should be put into an iterable parameter then passed as a whole. The step method is also implemented to perform one single step optimization and return loss.

MindSpore: The ways of the same learning rate for all parameters and different values for different parameter groups are supported.

Code Example

# The following implements Adagrad with MindSpore.
import numpy as np
import torch
import mindspore.nn as nn
import mindspore as ms


net = Net()
#1) All parameters use the same learning rate and weight decay
optim = nn.Adagrad(params=net.trainable_params())

#2) Use parameter groups and set different values
conv_params = list(filter(lambda x: 'conv' in x.name, net.trainable_params()))
no_conv_params = list(filter(lambda x: 'conv' not in x.name, net.trainable_params()))
group_params = [{'params': conv_params, 'weight_decay': 0.01, 'grad_centralization':True},
                {'params': no_conv_params, 'lr': 0.01},
                {'order_params': net.trainable_params()}]
optim = nn.Adagrad(group_params, learning_rate=0.1, weight_decay=0.0)
# The conv_params's parameters will use default learning rate of 0.1 and weight decay of 0.01 and grad
# centralization of True.
# The no_conv_params's parameters will use learning rate of 0.01 and default weight decay of 0.0 and grad
# centralization of False.
# The final parameters order in which the optimizer will be followed is the value of 'order_params'.

loss = nn.SoftmaxCrossEntropyWithLogits()
model = ms.Model(net, loss_fn=loss, optimizer=optim)

# The following implements Adagrad with torch.
input_x = torch.tensor(np.random.rand(1, 20).astype(np.float32))
input_y = torch.tensor([1.])
net = torch.nn.Sequential(torch.nn.Linear(input_x.shape[-1], 1))
loss = torch.nn.MSELoss()
optimizer = torch.optim.Adagrad(net.parameters())
l = loss(net(input_x).view(-1), input_y) / 2
optimizer.zero_grad()
l.backward()
optimizer.step()
print(loss(net(input_x).view(-1), input_y).item() / 2)
# Out:
# 0.1830