【WIP】fleet support dygraph in mnist/resnet/transformer

[danleifeng]

main change:

from paddle.distributed import fleet

fleet.init(is_collective=True)
adam = fleet.distributed_optimizer(adam)
# call after distributed_optimizer so as to apply dist_strategy
mnist = fleet.build_distributed_model(mnist)

Sample code:

import paddle
import paddle.nn as nn
import paddle.optimizer as opt
from paddle.distributed import fleet

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear1 = nn.Linear(10, 10)
        self._linear2 = nn.Linear(10, 1)
        
    def forward(self, x):
        return self._linear2(self._linear1(x))

def train():
    # 1. enable dynamic mode
    paddle.disable_static()
    
    # 2. create layer & optimizer
    layer = LinearNet()
    loss_fn = nn.MSELoss()
    adam = opt.Adam(
        learning_rate=0.001, parameters=layer.parameters())

    # 3. get data_parallel model using fleet
    fleet.init(is_collective=True)
    adam = fleet.distributed_optimizer(adam)
    # call after distributed_optimizer so as to apply dist_strategy
    dp_layer = fleet.build_distributed_model(layer)
    
    # 4. run layer
    inputs = paddle.randn([10, 10], 'float32')
    outputs = dp_layer(inputs)
    labels = paddle.randn([10, 1], 'float32')
    loss = loss_fn(outputs, labels)
    
    loss = dp_layer.scale_loss(loss)
    loss.backward()
    dp_layer.apply_collective_grads()

    adam.step()
    adam.clear_grad()

if __name__ == '__main__':
   train()