【PPSCI Doc No.12、13、14、15、16、17】ppsci.arch.Arch #752

megemini · 2024-01-15T05:49:35Z

PR types

Others

PR changes

Docs

Describe

补充 ppsci.arch.Arch 中的示例代码

关联 PR #686

@HydrogenSulfate

请评审～

paddle-bot · 2024-01-15T05:49:40Z

Thanks for your contribution!

HydrogenSulfate

辛苦大佬提交PR

HydrogenSulfate · 2024-01-15T06:03:11Z

ppsci/arch/base.py

+            >>> import paddle
+            >>> import ppsci
+            >>> model = ppsci.arch.Arch()
+            >>> # fetch one tensor
+            >>> out = model.concat_to_tensor({'x':paddle.to_tensor(123)}, ('x',))
+            >>> print(out)
+            Tensor(shape=[], dtype=int64, place=Place(gpu:0), stop_gradient=True,
+                   123)
+            >>> # fetch more tensors
+            >>> out = model.concat_to_tensor(
+            ...     {'x1':paddle.to_tensor([123]), 'x2':paddle.to_tensor([234])},
+            ...     ('x1', 'x2'),
+            ...     axis=0)
+            >>> print(out)
+            Tensor(shape=[2], dtype=int64, place=Place(gpu:0), stop_gradient=True,
+                   [123, 234])


感觉改成2D的Tensor能更好，因为2D更符合实际场景，用paddle.randn([8, 1])这种形状就行

输出改成这样：print(out.dtype, out.shape)，这样能避免cpu设备运行的问题，也重点展示了数据类型和数据形状。

这个 pydoctest 不太好用 ... ... 可以试试 xdoctest，或者用 paddle 前段时间的 doctester 来验证示例代码会比较好～

单纯的提个建议～～～ 😅

https://github.com/PaddlePaddle/Paddle/blob/develop/tools/sampcd_processor.py

这个 pydoctest 不太好用 ... ... 可以试试 xdoctest，或者用 paddle 前段时间的 doctester 来验证示例代码会比较好～

单纯的提个建议～～～ 😅

https://github.com/PaddlePaddle/Paddle/blob/develop/tools/sampcd_processor.py

谢谢大佬的建议，我研究一下，如果好用的话我就尽快换上

HydrogenSulfate · 2024-01-15T06:07:14Z

ppsci/arch/base.py

+        Examples:
+            >>> import paddle
+            >>> import ppsci
+            >>> model = ppsci.arch.Arch()
+            >>> # split one tensor
+            >>> out = model.split_to_dict(paddle.to_tensor(123), ('x',))
+            >>> print(out)
+            {'x': Tensor(shape=[], dtype=int64, place=Place(gpu:0), stop_gradient=True,
+                   123)}
+            >>> # split more tensors
+            >>> out = model.split_to_dict(paddle.to_tensor([123, 234]), ('x1', 'x2'), axis=0)
+            >>> print(out)
+            {'x1': Tensor(shape=[1], dtype=int64, place=Place(gpu:0), stop_gradient=True,
+                   [123]), 'x2': Tensor(shape=[1], dtype=int64, place=Place(gpu:0), stop_gradient=True,
+                   [234])}


同上，并且输出代码可以改成

for k, v in out.items(): print(f"{k} {v.dtype} {v.shape}")

HydrogenSulfate · 2024-01-15T06:11:25Z

ppsci/arch/base.py

+        Examples:
+            >>> import ppsci
+            >>> def transform_fn(in_):
+            ...     x = in_["x"]
+            ...     x = 2.0 * x
+            ...     input_trans = {"x": x}
+            ...     return input_trans
+            >>> model = ppsci.arch.Arch()
+            >>> model.register_input_transform(transform_fn)


这个example讲清楚了如何注册，但没有很好地展示transform_fn是如何作用的，所以：

可以使用MLP这个类作为model，并且接受"x"和"2x"这两个key作为输入关键字，输出y

然后register一个transform_in，它接受原始输入input = {"x": paddle.randn([4, 1])}，然后根据原始输入"x"，生成"2x"（=2*x），

运行model(input)，然后打印输出的y的类型和形状。

HydrogenSulfate · 2024-01-15T06:13:12Z

ppsci/arch/base.py

+            >>> import ppsci
+            >>> def transform_fn(in_, out):
+            ...     x = in_["x"]
+            ...     y = out["y"]
+            ...     u = 2.0 * x * y
+            ...     output_trans = {"u": u}
+            ...     return output_trans
+            >>> model = ppsci.arch.Arch()
+            >>> model.register_output_transform(transform_fn)


函数名改成transform_out

可以通过transform_out往输出dict中加入额外的内容，然后将结果打印出来，这样更容易让人理解

HydrogenSulfate · 2024-01-15T06:14:42Z

ppsci/arch/base.py

+        Examples:
+            >>> import ppsci
+            >>> model = ppsci.arch.Arch()
+            >>> # freeze all parameters and make model `eval`
+            >>> model.freeze()


末尾加上如下验证代码：

assert p.training is False for p in model.parameters(): assert p.stop_gradient is True

HydrogenSulfate · 2024-01-15T06:15:15Z

ppsci/arch/base.py

+        """Unfreeze all parameters.
+
+        Examples:
+            >>> import ppsci
+            >>> model = ppsci.arch.Arch()
+            >>> # unfreeze all parameters and make model `train`
+            >>> model.unfreeze()
+
+        """


与freeze类似加上验证代码

megemini · 2024-01-15T07:17:16Z

@HydrogenSulfate

已修改输出为 out.dtype, out.shape 样式
已修改 register_input_transform/register_output_transform，不过 register_input_transform 从输出貌似看不出啥不一样～ 😅

HydrogenSulfate · 2024-01-15T07:20:30Z

@HydrogenSulfate

已修改输出为 out.dtype, out.shape 样式

已修改 register_input_transform/register_output_transform，不过 register_input_transform 从输出貌似看不出啥不一样～ 😅

x这个字典会经过transform_in进行处理，所以通过以下方式体现transform_in具备“对x进行修改”这一能力

megemini · 2024-01-15T07:26:09Z

x这个字典会经过transform_in进行处理，所以通过以下方式体现transform_in具备“对x进行修改”这一能力 !

这个有点绕～～～ 🤣🤣🤣

帮忙看看现在对不对～

谢谢！

@HydrogenSulfate

HydrogenSulfate · 2024-01-15T07:28:32Z

ppsci/arch/base.py

+            >>> import ppsci
+            >>> model = ppsci.arch.Arch()
+            >>> # fetch one tensor
+            >>> out = model.concat_to_tensor({'x':paddle.to_tensor(123)}, ('x',))


这里的输入建议使用2D+float32 Tensor，比如[8,1]这种，0 DTensor不太有代表性

构造输入用paddle.randn(shape)更好

HydrogenSulfate · 2024-01-15T07:29:16Z

ppsci/arch/base.py

+            paddle.int64 []
+            >>> # fetch more tensors
+            >>> out = model.concat_to_tensor(
+            ...     {'x1':paddle.to_tensor([[1, 2], [2, 3]]), 'x2':paddle.to_tensor([[3, 4], [4, 5]])},


HydrogenSulfate · 2024-01-15T07:30:03Z

ppsci/arch/base.py

+            >>> out = model.split_to_dict(paddle.to_tensor(123), ('x',))
+            >>> for k, v in out.items():
+            ...     print(f"{k} {v.dtype} {v.shape}")
+            x paddle.int64 []
+            >>> # split more tensors
+            >>> out = model.split_to_dict(paddle.to_tensor([[1, 2], [2, 3]]), ('x1', 'x2'), axis=0)
+            >>> for k, v in out.items():
+            ...     print(f"{k} {v.dtype} {v.shape}")
+            x1 paddle.int64 [1, 2]
+            x2 paddle.int64 [1, 2]


同上，类型和形状可以改成二维+flaot32

megemini · 2024-01-15T07:58:18Z

已修改输入数据为 paddle.rand ~

HydrogenSulfate

LGTM

* [Add] arch examples * [Change] examples * [Change] examples * [Change] register_input_transform * [Change] data with rand --------- Co-authored-by: HydrogenSulfate <490868991@qq.com>

[Add] arch examples

1500192

paddle-bot bot added the contributor label Jan 15, 2024

HydrogenSulfate requested changes Jan 15, 2024

View reviewed changes

luotao1 mentioned this pull request Jan 15, 2024

【快乐开源】科学计算 API 文档补全 #686

Closed

[Change] examples

e30e098

[Change] examples

8084d34

megemini requested a review from HydrogenSulfate January 15, 2024 07:19

[Change] register_input_transform

aafc7ba

HydrogenSulfate requested changes Jan 15, 2024

View reviewed changes

[Change] data with rand

419cdb1

megemini requested a review from HydrogenSulfate January 15, 2024 07:58

HydrogenSulfate approved these changes Jan 15, 2024

View reviewed changes

Merge branch 'develop' into doc_arch

2a6468d

luotao1 added the HappyOpenSource 快乐开源活动issue与PR label Jan 16, 2024

HydrogenSulfate merged commit 4f191f7 into PaddlePaddle:develop Jan 16, 2024
3 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

【PPSCI Doc No.12、13、14、15、16、17】ppsci.arch.Arch #752

【PPSCI Doc No.12、13、14、15、16、17】ppsci.arch.Arch #752

megemini commented Jan 15, 2024

paddle-bot bot commented Jan 15, 2024

HydrogenSulfate left a comment

HydrogenSulfate Jan 15, 2024

megemini Jan 15, 2024

HydrogenSulfate Jan 15, 2024

HydrogenSulfate Jan 15, 2024

HydrogenSulfate Jan 15, 2024

HydrogenSulfate Jan 15, 2024

HydrogenSulfate Jan 15, 2024

HydrogenSulfate Jan 15, 2024

megemini commented Jan 15, 2024

HydrogenSulfate commented Jan 15, 2024

megemini commented Jan 15, 2024

HydrogenSulfate Jan 15, 2024

HydrogenSulfate Jan 15, 2024

HydrogenSulfate Jan 15, 2024

megemini commented Jan 15, 2024

HydrogenSulfate left a comment

【PPSCI Doc No.12、13、14、15、16、17】ppsci.arch.Arch #752

【PPSCI Doc No.12、13、14、15、16、17】ppsci.arch.Arch #752

Conversation

megemini commented Jan 15, 2024

PR types

PR changes

Describe

paddle-bot bot commented Jan 15, 2024

HydrogenSulfate left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

megemini commented Jan 15, 2024

HydrogenSulfate commented Jan 15, 2024

megemini commented Jan 15, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

megemini commented Jan 15, 2024

HydrogenSulfate left a comment

Choose a reason for hiding this comment