def part_ps_impl(self, dataset):
net = Menet(self.in_channels, self.out_channels, self.kernel_size, self.vocab_size,
self.embedding_size, self.output_channels, self.target, self.sparse)
net.embedding_lookup.set_param_ps()
net.conv.conv2d.add_prim_attr('primitive_target', 'CPU')
net.conv.bias_add.add_prim_attr('primitive_target', 'CPU')
net.set_train()
loss = SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
opt = Adam(params=filter(lambda x: x.requires_grad, net.get_parameters()))
opt.target = 'CPU'
model = Model(net, loss, opt)
model.train(self.epoch_size, dataset, dataset_sink_mode=False)
input_me = Tensor(self.input_np)
out_me = model.predict(input_me)
return out_me.asnumpy()
def compile_net(net):
optimizer = Adam(net.trainable_params(),
learning_rate=0.1,
loss_scale=1024.0,
weight_decay=0.9)
train_net = TrainOneStepCell(net, optimizer)
_executor.compile(train_net, _x, _b)
def compile_net(net, shape):
x = Tensor(np.ones(shape), dtype=ms.int32)
y = Tensor(np.ones(shape), dtype=ms.float32)
z = Tensor(np.ones(shape), dtype=ms.int32)
optimizer = Adam(net.trainable_params(), learning_rate=0.1)
train_net = TrainOneStepCell(net, optimizer)
train_net.set_auto_parallel()
train_net.set_train()
_executor.compile(train_net, x, y, z)
context.reset_auto_parallel_context()
def _model_train_and_save_ckpt(self, net, dataset, epoch):
self.opt = Adam(params=net.get_parameters())
if self.target == 'CPU':
self.opt.target = self.target
if self.sparse:
context.set_context(enable_sparse=True)
self.model = Model(network=net,
loss_fn=self.loss_fn,
optimizer=self.opt)
ckpt_config = CheckpointConfig(keep_checkpoint_max=1)
ckpt_path = './rank_{}_ckpt'.format(self.global_rank_id)
ckpt_callback = ModelCheckpoint(prefix='parallel',
directory=ckpt_path,
config=ckpt_config)
clean_all_ckpt_files(ckpt_path)
self.model.train(epoch=epoch,
train_dataset=dataset,
callbacks=[ckpt_callback],
dataset_sink_mode=False)
newest_ckpt_file = find_newest_ckpt_file(ckpt_path)
return load_checkpoint(newest_ckpt_file)
base_params = filter(lambda p: id(p) not in ignored_params,
net.get_parameters())
optimizer_P = Adam(
[
{
'params': base_params,
'lr': 0.1 * args.lr
},
{
'params': net.bottleneck.get_parameters(),
'lr': args.lr
},
{
'params': net.classifier.get_parameters(),
'lr': args.lr
},
{
'params': net.wpa.get_parameters(),
'lr': args.lr
},
# {'params': net.attention_0.parameters(), 'lr': args.lr},
# {'params': net.attention_1.parameters(), 'lr': args.lr},
# {'params': net.attention_2.parameters(), 'lr': args.lr},
# {'params': net.attention_3.parameters(), 'lr': args.lr},
# {'params': net.out_att.parameters(), 'lr': args.lr} ,
],
learning_rate=args.lr,
weight_decay=5e-4)
########################################################################
# Start Training