Resnet npu

[ShawnXuan]

这个分支固定了数据集，去掉了随机。

需要准备一个可以加载的初始化模型到 .../models/Vision/classification/image/resnet50/examples/checkpoints/init

比如在910b上可以

cd .../models/Vision/classification/image/resnet50/examples
cp -r /data1/home/xiexuan/git-repos/models/Vision/classification/image/resnet50/examples/checkpoints .

然后就可以运行 ./npu_eager.sh 或 ./npu_graph.sh

目前npu eager和cuda eager/graph都对齐了，但npu graph还没有对齐，输出的pred都是 0.001，需要深入调查

loss
tensor(6.9073, placement=oneflow.placement(type="npu", ranks=[0]), sbp=(oneflow.sbp.partial_sum,),
dtype=oneflow.float32)
pred
tensor([[0.0010, 0.0010, 0.0010, ..., 0.0010, 0.0010, 0.0010],
[0.0010, 0.0010, 0.0010, ..., 0.0010, 0.0010, 0.0010],
[0.0010, 0.0010, 0.0010, ..., 0.0010, 0.0010, 0.0010],
...,
[0.0010, 0.0010, 0.0010, ..., 0.0010, 0.0010, 0.0010],
[0.0010, 0.0010, 0.0010, ..., 0.0010, 0.0010, 0.0010],
[0.0010, 0.0010, 0.0010, ..., 0.0010, 0.0010, 0.0010]],
placement=oneflow.placement(type="npu", ranks=[0]), sbp=(oneflow.sbp.split(dim=0),), dtype=oneflow.float32)
label
tensor([582, 209, 272, 331, 768, 626, 838, 202, 333, 754, 435, 955, 853, 943, 40, 723, 3, 104, 51, 60, 118,
762, 603, 353, 898, 69, 552, 824, 999, 217, 713, 334, 758, 818, 115, 1, 609, 238, 147, 446, 240, 455,
442, 257, 206, 200, 911, 355, 684, 419], placement=oneflow.placement(type="npu", ranks=[0]),
sbp=(oneflow.sbp.split(dim=0),), dtype=oneflow.int32)

[ShawnXuan]

暂时用cpu解码

[ShawnXuan]

待验证训练是否收敛。

[ShawnXuan]

这是flowvision里面的loss。
需要dim_gather

[Flowingsun007]

softmax_cross_entropy和dim_gather应该不难开发，我们可以列到npu开发计划里，后面等开发完成再试试

[Flowingsun007]

看了下npu的dim_gather已经支持了：oneflow_npu/kernels/dim_gather_kernel.cpp，应该再开发一个softmax_cross_entropy就行

[ShawnXuan]

softmax_cross_entropy 我开发了一个，可能反向还有问题。不过softmax_cross_entropy没有和torch对应，我倾向于开发和torch兼容的算子，所以选了flowvision的方案，不用softmax_cross_entropy。

我回头试试 dim_gather

[ShawnXuan]

目前npu不支持

[ShawnXuan]

进展：
查到了 Vision/classification/image/resnet50/models/resnet50.py 里面的 _make_layer 最后返回的是 nn.Sequential(*layers)，就没法深入对比了。

def_forward_impl(self, x: Tensor) ->Tensor:
ifself.pad_input:
ifself.channel_last:
# NHWCpaddings= (0, 1)
else:
# NCHWpaddings= (0, 0, 0, 0, 0, 1)
x=flow._C.pad(x, pad=paddings, mode="constant", value=0)
x=self.conv1(x)
ifself.fuse_bn_relu:
x=self.bn1(x, None)
else:
x=self.bn1(x)
x=self.relu(x)
x=self.maxpool(x)
# graph模式时前面都是对齐的，能够正常输出，说明算子没有问题x=self.layer1(x)
r=x# 到这里返回的值都是 0，深入 layer1发现，最后是nn.Sequential，不好对比了，可能需要构造一个nn.Sequential的graph 单测x=self.layer2(x)
x=self.layer3(x)
x=self.layer4(x)
x=self.avgpool(x)
x=flow.flatten(x, 1)
x=self.fc(x)

[ShawnXuan]

09.27 进展
nn.Sequential 测试通过，测试脚本如下：

importunittestimportnumpyasnpfromcollectionsimportOrderedDictimportoneflowasflowimportoneflow_npufromoneflow.test_utils.test_utilimportGenArgListdefcreate_model(shape, device, dtype):
model=flow.nn.Sequential(
flow.nn.Linear(shape[-1], shape[-1]),
flow.nn.ReLU(),
flow.nn.Linear(shape[-1], shape[-1]),
flow.nn.Softmax(dim=-1)
).to(device=device, dtype=dtype)
returnmodelclassEvalGraph(flow.nn.Graph):
def__init__(self, model):
super().__init__()
self.model=modeldefbuild(self, x):
logits=self.model(x)
returnlogitsclassTrainGraph(flow.nn.Graph):
def__init__(self, model):
super().__init__()
self.model=modelparam_group= {"params": [pforpinmodel.parameters() ifpisnotNone]}
optimizer=flow.optim.SGD([param_group], lr=0.1, momentum=0.9)
self.add_optimizer(optimizer)
defbuild(self, x):
logits=self.model(x)
loss=logits.sum()
loss.backward()
returnloss, logitsdef_test_sequential_forward(test_case, shape, device, dtype):
model_cpu=create_model(shape, "cpu", flow.float)
model_npu=create_model(shape, device, dtype)
model_npu.load_state_dict(model_cpu.state_dict())
arr=flow.rand(*shape)
arr_npu=arr.to(device=device, dtype=dtype)
out_cpu=model_cpu(arr.to(dtype=flow.float))
out_npu=model_npu(arr_npu)
tol=1e-3ifdtype==flow.float16else1e-5test_case.assertTrue(np.allclose(out_cpu.numpy(), out_npu.cpu().numpy(), tol, tol))
graph=EvalGraph(model_npu)
graph_out=graph(arr_npu)
test_case.assertTrue(np.allclose(graph_out.cpu().numpy(), out_cpu.numpy(), tol, tol))
def_test_sequential_grad(test_case, shape, device, dtype):
model_cpu=create_model(shape, "cpu", flow.float)
model_npu=create_model(shape, device, dtype)
model_npu.load_state_dict(model_cpu.state_dict())
graph=TrainGraph(model_npu)
np_arr=np.random.rand(*shape).astype(np.float32)
arr=flow.tensor(np_arr, dtype=flow.float, requires_grad=True)
arr_npu=flow.tensor(np_arr, dtype=dtype, device=device, requires_grad=True)
print("*"*100)
print("arr", arr)
print("arr_npu", arr_npu)
out_cpu=model_cpu(arr.to(dtype=flow.float))
out_cpu.sum().backward()
loss, out_npu=graph(arr_npu)
print("out_cpu", out_cpu)
print("out_npu", out_npu)
forparam_cpu, param_npuinzip(model_cpu.parameters(), model_npu.parameters()):
print("cpu", param_cpu)
print("npu", param_npu)
#tol = 1e-3 if dtype == flow.float16 else 1e-5#test_case.assertTrue(np.allclose(arr.grad.numpy(), arr_npu.grad.cpu().numpy(), tol, tol))@flow.unittest.skip_unless_1n1d()classTest_Sequential_Graph(flow.unittest.TestCase):
deftest_sequential_forward(test_case):
arg_dict=OrderedDict()
arg_dict["test_fun"] = [
_test_sequential_forward,
]
arg_dict["shape"] = [(2, 4), (2, 3, 4)]
arg_dict["device"] = ["npu"]
arg_dict["dtype"] = [flow.float32, flow.float16]
forarginGenArgList(arg_dict):
arg[0](test_case, *arg[1:])
deftest_sequential_grad(test_case):
arg_dict=OrderedDict()
arg_dict["test_fun"] = [
_test_sequential_grad,
]
arg_dict["shape"] = [(2, 4), (2, 3, 4)]
arg_dict["device"] = ["npu"]
arg_dict["dtype"] = [flow.float32]
forarginGenArgList(arg_dict):
arg[0](test_case, *arg[1:])
if__name__=="__main__":
unittest.main()