【Hackathon 5th No.57】Neural networks for topology optimization #597
Conversation
|
Thanks for your contribution! |
|
辛苦大佬提交代码～ |
| nn.Sigmoid(), | ||
| ) | ||
|
|
||
| def forward(self, x): |
There was a problem hiding this comment.
参照UNetEx，是否可以尽量使用for循环，或调用UNetEx中的encode，decode来写forward？
| x1 = x[self.input_keys[0]][:, k, :, :] | ||
| x2 = x[self.input_keys[0]][:, k - 1, :, :] | ||
| x = paddle.stack((x1, x1 - x2), axis=1) | ||
| # Layer 1 (bs, 2, 40, 40) -> (bs, 16, 40, 40) |
|
|
||
| def augmentation(input_dict, label_dict, weight_dict=None): | ||
| """Apply random transformation from D4 symmetry group | ||
| # Arguments |
There was a problem hiding this comment.
type hint的格式不对，可参考ppsci下文件修改
| @@ -0,0 +1,409 @@ | |||
| { | |||
There was a problem hiding this comment.
当前github代码中不放ipynb，需要改成代码。最终案例文档会附加AI Studio链接，可以把这个ipynb放到AI Studio项目中
|
|
||
| if __name__ == "__main__": | ||
| ### CASE 1: poisson(5) | ||
| SIMP_stop_point_sampler = poisson_sampler(5) |
There was a problem hiding this comment.
这几个例子是不是除了这一行都一样？如果是的话，只写一个文件就行，其他的备注写一下，或者写一个函数，如
| VOL_COEFF = 1 | ||
| LEARNING_RATE = 0.001 / (1 + NUM_EPOCHS // 15) | ||
| ITERS_PER_EPOCH = int(N_SAMPLE * TRAIN_TEST_RATIO / BATCH_SIZE) | ||
| NUM_PARAMS = 192113 # the number given in paper |
There was a problem hiding this comment.
参考其他的案例的开头，如example/bracket.py，增加一些必要部分，调整一下参数顺序，以及加一点点说明参数是干什么的备注，不用加太多
| @@ -0,0 +1,31 @@ | |||
| import os | |||
There was a problem hiding this comment.
这个文件在case中似乎没有被import，看Readme是生成数据用的，代码中没有OriginalData，所以这个其实也跑不了。数据集可以先上传到AI Studio，这个文件就删掉不用提交了
| # optimizer | ||
| optimizer = ppsci.optimizer.Adam(learning_rate=LEARNING_RATE, epsilon=1e-07)( | ||
| model | ||
| ) # epsilon = 1e-07 is the default in tf |
| iters_per_epoch=ITERS_PER_EPOCH, | ||
| ) | ||
|
|
||
| solver.train() |
There was a problem hiding this comment.
参照其他案例如bracket，在solver这边加几句注释，不用加多
jointContribution/TopOpt/README.md
Outdated
| @@ -0,0 +1,38 @@ | |||
| # TopOpt | |||
There was a problem hiding this comment.
正式提交的时候需要写md文档，但不是README
examples/topopt/eval.py
Outdated
| ) | ||
| ), | ||
| dtype="float32", | ||
| ) |
There was a problem hiding this comment.
w00 = paddle.sum(paddle.multiply(paddle.equal(paddle.round(output), 0.0),paddle.equal(paddle.round(y), 0.0),),dtype=paddle.get_default_dtype())应该就行
examples/topopt/eval.py
Outdated
| paddle.equal(paddle.round(y), 1.0), | ||
| ) | ||
| ), | ||
| dtype="float32", |
examples/topopt/eval.py
Outdated
| paddle.equal(paddle.round(y), 0.0), | ||
| ) | ||
| ), | ||
| dtype="float32", |
examples/topopt/eval.py
Outdated
| paddle.equal(paddle.round(y), 1.0), | ||
| ) | ||
| ), | ||
| dtype="float32", |
examples/topopt/train.py
Outdated
| "_vol_coeff", | ||
| str(VOL_COEFF), | ||
| ] | ||
| ), |
There was a problem hiding this comment.
可以写成 OUTPUT_DIR = (f"{OUTPUT_DIR}/{sampler_key}{num}_vol_coeff{VOL_COEFF}" if num is not Noneelse f"{OUTPUT_DIR}/{sampler_key}_vol_coeff{VOL_COEFF}")
examples/topopt/eval.py
Outdated
| model_path, | ||
| "".join([model_name, "_vol_coeff", str(VOL_COEFF)]), | ||
| "checkpoints", | ||
| "latest", |
There was a problem hiding this comment.
model_path = f"{model_path}/{model_name}_vol_coeff{VOL_COEFF}/checkpoints/latest"
examples/topopt/eval.py
Outdated
| from ppsci.utils import logger | ||
|
|
||
|
|
||
| def evaluation_and_plot(dataloader): |
There was a problem hiding this comment.
这个函数里eval的部分是否可以用validator写？用ppsci.loss.FunctionalLoss这个API自定义一下loss，想存4个实验的结果也可以写成类似于这样？
examples/topopt/eval.py
Outdated
| current_iou_results = [] | ||
|
|
||
| # only evaluate for NUM_VAL_STEP times of iteration | ||
| for x, y, _ in iter(dataloader): |
There was a problem hiding this comment.
这边只是取了n个batch的数据？n相当于iteration？
There was a problem hiding this comment.
是的，这里evaluate的逻辑是，对于每一个训练的模型，用16组不同的数据做16次评估(不同的数据是指分别取原始数据的第5，10，15，20，...，80通道作为输入，对应代码中的 stop_iter ， stop_iter 这里指SIMP的初始迭代次数——原始数据有100个通道对应的是SIMP算法100次的迭代结果，这个模型想做的就是用SIMP中间某一步的迭代结果直接预测最后一步的迭代结果，所以他是这样来评估的)，每一次评估的时候只取 NUM_VAL_STEP 个batch的数据。
examples/topopt/eval.py
Outdated
| iou_results = [] | ||
|
|
||
| # evaluation for different fixed iteration stop times | ||
| for stop_iter in iterations_stop_times: |
examples/topopt/TopOptModel.py
Outdated
|
|
||
|
|
||
| # NCHW data format | ||
| class TopOptNN(ppsci.arch.UNetEx): |
HydrogenSulfate
assigned HydrogenSulfate, luotao1 and lijialin03 and unassigned HydrogenSulfate
|
辛苦～
|
|
There was a problem hiding this comment.
辛苦～
- 代码有一些需要修改的地方。
- 需要确定一下指标，以便进行ci检查，一般是eval的指标，如 l2 error，acc等。同时为了确认复现效果（参考官网），将指标和原代码的对比贴到comment中（不是md文档）吧，格式类似这样：
| total acc | possion5 | xxx | |
|---|---|---|---|
| paper | xx | xx | xx |
| ppsci | xx | xx | xx |
| diff | xx% | xx% | xx% |
- 文档内容大体没什么问题，主要需要检查一下用英文符号、公式是否正确显示、改代码行数（等代码完全确认后）
- 写完文档后提交之前，可以通过下图方法（可能需要pip安装一些包），用mkdocs提前预览显示效果，以确保正确显示
6. 另外需要在PaddleScience/mkdocs.yml中添加相应部分，以确保该案例文档可以通过官网主页进入
谢谢
examples/topopt/conf/topopt.yaml
Outdated
| hydra: | ||
| run: | ||
| # dynamic output directory | ||
| dir: outputs_topopt/ |
There was a problem hiding this comment.
改成：outputs_topopt/${now:%Y-%m-%d}/${now:%H-%M-%S}/${hydra.job.override_dirname}
这样子会生成类似下面这样的目录结构，就不用担心训练结果被覆盖了
| dir: outputs_topopt/ | ||
| job: | ||
| name: ${mode} # name of logfile | ||
| chdir: false # keep current working direcotry unchaned |
There was a problem hiding this comment.
参考bracket，增加exclude_keys，这样才能允许用户指定参数，如用 "EVAL.pretrained_model_path"指定预训练模型
| layers: 2 | ||
|
|
||
| # training settings | ||
| TRAIN: |
There was a problem hiding this comment.
参照bracket，这边主要放跟训练有直接关系的值，比如epoch等，然后改一下名字。其他的值放到TRAIN这一层外面吧。
| shuffle: true | ||
|
|
||
| # evaluation settings | ||
| EVAL: |
examples/topopt/conf/topopt.yaml
Outdated
| CASE_PARAM: [[Poisson, 5], [Poisson, 10], [Poisson, 30], [Uniform, null]] | ||
|
|
||
| # set data path | ||
| DATA_PATH: ./Dataset/PreparedData/top_dataset.h5 |
There was a problem hiding this comment.
改成./datasets/top_dataset.h5吧
docs/zh/examples/topopt.md
Outdated
| --8<-- | ||
| ``` | ||
|
|
||
| 优化器选用 Adam，训练代码如下： |
docs/zh/examples/topopt.md
Outdated
| --8<-- | ||
| ``` | ||
|
|
||
| ### 3.9 metric构建 |
There was a problem hiding this comment.
loss和metric构建，loss既然也是自定义的，也介绍一下
docs/zh/examples/topopt.md
Outdated
| 本案例选择 binary accuracy 和 IoU 进行评估: | ||
| $\text{Bin. Acc.} = \frac{w_{00}+w_{11}}{n_{0}+n_{1}}$ | ||
| $\text{IoU} = \frac{1}{2}\left[\frac{w_{00}}{n_{0}+w_{10}} + \frac{w_{11}}{n_{1}+w_{01}}\right]$ | ||
| $n_{0} = w_{00} + w_{01}, \quad n_{1} = w_{10} + w_{11}$ |
docs/zh/examples/topopt.md
Outdated
| --8<-- | ||
| ``` | ||
|
|
||
| ### 3.10 评估模型 |
There was a problem hiding this comment.
主要介绍一下评估是怎么做的，评估的是哪些指标。然后可以在介绍之后加个 #### 3.10.1 评估器构建，把validator的代码单独写一下，避免展示重复代码
| --8<-- | ||
| ``` | ||
|
|
||
| ## 5. 结果展示 |
There was a problem hiding this comment.
参照bracket，增加一点对图片内容的介绍，主要是让其他人能看懂。
参照tempoGAN，把指标（eval结果）用表格放上
examples/topopt/conf/topopt.yaml
Outdated
| hydra: | ||
| run: | ||
| # dynamic output directory | ||
| dir: outputs_topopt/ |
| dir: outputs_topopt/ | ||
| job: | ||
| name: ${mode} # name of logfile | ||
| chdir: false # keep current working direcotry unchaned |
examples/topopt/conf/topopt.yaml
Outdated
| CASE_PARAM: [[Poisson, 5], [Poisson, 10], [Poisson, 30], [Uniform, null]] | ||
|
|
||
| # set data path | ||
| DATA_PATH: ./Dataset/PreparedData/top_dataset.h5 |
| layers: 2 | ||
|
|
||
| # training settings | ||
| TRAIN: |
| shuffle: true | ||
|
|
||
| # evaluation settings | ||
| EVAL: |
examples/topopt/topopt.py
Outdated
| # define metric | ||
| def val_metric(output_dict, label_dict, weight_dict=None): | ||
| output = output_dict["output"] | ||
| y = label_dict["output"] |
examples/topopt/topopt.py
Outdated
| # evaluation for different fixed iteration stop times | ||
| for stop_iter in iterations_stop_times: | ||
| # only evaluate for NUM_VAL_STEP times of iteration | ||
| X_data, Y_data = generate_train_test( |
examples/topopt/topopt.py
Outdated
| "drop_last": cfg.EVAL.drop_last, | ||
| "shuffle": cfg.EVAL.shuffle, | ||
| }, | ||
| "transforms": ( |
examples/topopt/topopt.py
Outdated
| if cfg.mode == "train": | ||
| train(cfg) | ||
| elif cfg.mode == "eval": | ||
| evaluate_and_plot(cfg) |
examples/topopt/topopt.py
Outdated
| iou_results_summary["thresholding"] = th_iou_results | ||
|
|
||
| # plot and save figures | ||
| plt.figure(figsize=(12, 6)) |
There was a problem hiding this comment.
已修改用ppsci.utils.misc.plot_curve()
examples/topopt/functions.py
Outdated
| label_dict (Dict[str, np.adarray]): label dict of np.adarray size `(batch_size, 1, height, width)` | ||
| weight_dict (Dict[str, np.adarray]): weight dict if any | ||
|
|
||
| Returns: |
examples/topopt/functions.py
Outdated
| label_dict: Dict[str, np.ndarray], | ||
| weight_dict: Dict[str, np.ndarray] = None, | ||
| ) -> Tuple[ | ||
| Dict[str, paddle.Tensor], Dict[str, paddle.Tensor], Dict[str, paddle.Tensor] |
There was a problem hiding this comment.
这个函数的返回dict里的值应该是np.darray形式，其他案例写错了
examples/topopt/functions.py
Outdated
| Returns: | ||
| Tuple[Dict[str, paddle.Tensor], Dict[str, paddle.Tensor], Dict[str, paddle.Tensor]]: (transformed_input_dict, transformed_label_dict, transformed_weight_dict) | ||
| """ | ||
| inputs = paddle.to_tensor(input_dict["input"], dtype=paddle.get_default_dtype()) |
examples/topopt/functions.py
Outdated
| new_perm = list(range(len(inputs.shape))) | ||
| new_perm[1], new_perm[2] = new_perm[2], new_perm[1] | ||
| inputs = paddle.transpose(inputs, perm=new_perm) | ||
| labels = paddle.transpose(labels, perm=new_perm) |
examples/topopt/topopt.py
Outdated
| cfg, model, model_path, data_iters, data_targets, iterations_stop_times | ||
| ) | ||
|
|
||
| model_name = model_path.split("\\")[-3].split("_")[0] |
There was a problem hiding this comment.
这个"\\"只是windows的，linux就会报错。
还写成类似sampler_key, num = cfg.CASE_PARAM[i]这样的吧
There was a problem hiding this comment.
嗯嗯，那这样的话就让用户传一个 EVAL.pretrained_model_path 然后这样写？
There was a problem hiding this comment.
这样确实很合理，但是参考其他单一模型的案例，其实预训练模型是从网上下载的，把train和evaluate分开，也是为了让用户可以不必一定先训练，而是可以直接加载训练好的模型
所以如果要一次性加载多个，虽然很难受，但是似乎必须让用户在运行时输入命令后面加EVAL.pretrained_model_path=[path1,path2,...]或者EVAL.pretrained_model_path_possion5=path1 EVAL.pretrained_model_path_xx=path2，而不能直接指定xx/checkpoints/latest这样的路径
There was a problem hiding this comment.
明白了，那改成这样可以吗？让用户在运行时输入命令后面加EVAL.pretrained_model_path=[path1,path2,...]，同时加上对应的名字EVAL.case_name=[case_name1, case_name2,...]这样
There was a problem hiding this comment.
明白了，那改成这样可以吗？让用户在运行时输入命令后面加EVAL.pretrained_model_path=[path1,path2,...]，同时加上对应的名字EVAL.case_name=[case_name1, case_name2,...]这样
这里现在改成了下面这样
可以这样输入参数：
python topopt.py 'mode=eval' 'EVAL.pretrained_model_path_dict={"Uniform": "path1", "Poisson5": "path2", "Poisson10": "path3", "Poisson30": "path4"}'
examples/topopt/topopt.py
Outdated
| current_iou_results = [] | ||
|
|
||
| # only calculate for NUM_VAL_STEP times of iteration | ||
| for _ in range(10): |
There was a problem hiding this comment.
10是哪里来的数字呀，是cfg.EVAL.num_val_step吗？
There was a problem hiding this comment.
哦哦这里应该是 cfg.EVAL.num_val_step，已修改
examples/topopt/topopt.py
Outdated
| acc_results_summary["thresholding"] = th_acc_results | ||
| iou_results_summary["thresholding"] = th_iou_results | ||
|
|
||
| # # plot and save figures |
| data_iters, data_targets, 1.0, cfg.EVAL.batch_size * cfg.EVAL.num_val_step | ||
| ) | ||
|
|
||
| sup_validator = ppsci.validate.SupervisedValidator( |
There was a problem hiding this comment.
validator一定要放这里吗？放这里运行的时候会报错，错误比较像是程序运行结束，但是data_loader还在等着读取数据
如果实在不行，evaluate里不用validator也行，可以参考deephpms最新的evaluate代码
There was a problem hiding this comment.
是因为 validator 放在循环里面了吗，如果要把 validator 放在循环外面就需要它包含所有的10000个数据，但是我不知道怎么设置 solver 让 solver.eval() 运行的时候只跑 cfg.EVAL.num_val_step 个batch (比如10个)。我试了如果把 validator 放在循环外面，然后 solver.epochs 设为1，solver.iters_per_epoch 设为 cfg.EVAL.num_val_step (比如10)，validator 里 drop_last 设为 True，solver.eval() 还是会跑 10000/batch_size 个 iteration； solver.epochs 设为 cfg.EVAL.num_val_step，solver.iters_per_epoch 设为1，solver.eval() 还是会跑 10000/batch_size 个 iteration
There was a problem hiding this comment.
先说一下结论，我们又查了一下发现，只要设置一下这个就不会报错了，估计是底层数据读取的时候自动开了多线程
然后说一下，eval的逻辑是运行n(1)个epoch，每个epoch 跑 iterations(10)次，每次会从所有数据(160)中读取batch_size(16)个
我理解现在是每个 sampler 调用 solver.eval() 80/5=16次，每次的iteration是10(cfg.EVAL.num_val_step)，batch_size是16(cfg.EVAL.batch_size)，数据量是10x16=160?其实validator放外面之后，还按照现在这样设置，只是在里面调solver.eval应该就行。但是这样有个问题是按照原先的写法，validator输入的数据是每次循环新生成的，这部分数据应该是所有数据随机后，其中的前160*1.0个，也就是每次循环都会改变，使用不同的数据，这个是个问题
所以就按照第一行，加一下"num_workers": 0吧
There was a problem hiding this comment.
另外上次漏了，现在代码里有几个dtype="float32"，都改成dtype=paddle.get_default_dtype()吧
examples/topopt/conf/topopt.yaml
Outdated
| CASE_PARAM: [[Poisson, 5], [Poisson, 10], [Poisson, 30], [Uniform, null]] | ||
|
|
||
| # set data path | ||
| DATA_PATH: ./Dataset/top_dataset.h5 |
There was a problem hiding this comment.
跟其他案例统一，改成./datasets/top_dataset.h5吧
examples/topopt/conf/topopt.yaml
Outdated
| learning_rate: 0.001 | ||
| epsilon: | ||
| log_loss: 0.0000001 | ||
| optimizer: 0.0000001 |
感谢！ 又发现了一个问题：在AI Studio上拉取最新的 ppsci 之后构建 SupervisedConstraint 这一步就有问题了，报错是这样子 看起来是这个更新导致的 把报错里面的 data 打印出来是这种形式：( ( {'input': input_dataset}, {'output': output_dataset}, {} ), ) |
之前有这个错误，是代码中多加了一个括号导致的，这个pr已经去掉了括号pr619，改成了： |
已解决，感谢！ |
|
指标和原代码的对比：
|
|
代码冲突了，看mkdocs.yml文件会有这么多修改，应该是拉的不是最新的，要拉最新分支修改。看你仓库里这个分支的commit，最后一个非你进行的修改在10.23，需要更新一下。 |
NKNaN
force-pushed
the
ayase-develop2
branch
from
e0df189
to
386b929
Compare
docs/zh/examples/topopt.md
Outdated
| ``` | ||
|
|
||
|
|
||
| ### 3.4 transform构建 |
There was a problem hiding this comment.
现在文档里的transform都是指网络输入输出transform，所以这边改成data transform吧
docs/zh/examples/topopt.md
Outdated
|
|
||
| ### 3.6 采样器构建 | ||
|
|
||
| 原始数据有100个通道对应的是 SIMP 算法 100 次的迭代结果，本案例模型目标是用 SIMP 中间某一步的迭代结果直接预测 SIMP 最后一步的迭代结果，而论文原始代码中的模型输入是原始数据对通道进行采样后的数据，为应用 PaddleScience API，本案例将采样步骤放入模型的 forward 方法中，所以需要传入不同的采样器。 |
There was a problem hiding this comment.
文档主要是把内容介绍清楚就行，不用太强调跟原论文的比较，修改一下这边，只将现在的代码是怎么做的，做了什么，为什么这么做，讲清楚就行
docs/zh/examples/topopt.md
Outdated
| <figcaption>IoU结果</figcaption> | ||
| </figure> | ||
|
|
||
| 结果与[原始代码结果](https://github.com/ISosnovik/nn4topopt/blob/master/results.ipynb)基本一致 |
docs/zh/examples/topopt.md
Outdated
|
|
||
| 结果与[原始代码结果](https://github.com/ISosnovik/nn4topopt/blob/master/results.ipynb)基本一致 | ||
|
|
||
| 此外将这些计算的指标与论文中展示的对应指标 (`table 1` 与 `table2` 中的) 对比，指标相对误差均小于 10%，以下表格是所有的指标计算结果： |
There was a problem hiding this comment.
这边也是，改成“用表格表示上图指标为：”这类的吧
examples/topopt/topopt.py
Outdated
| dtype=paddle.get_default_dtype(), | ||
| ) | ||
| n0 = paddle.add(w01, w00) | ||
| n1 = paddle.add(w11, w10) |
There was a problem hiding this comment.
这边wxx/nx等变量名改一下吧，让名字有点具体含义，我看这边是求iou计算里的TP、FP、TN、FN？对照改一下吧
examples/topopt/TopOptModel.py
Outdated
| @@ -0,0 +1,157 @@ | |||
| # Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved. | |||
There was a problem hiding this comment.
这个文件名字改成小写，topopt里的import也一起改一下
examples/topopt/conf/topopt.yaml
Outdated
| learning_rate: 0.001 | ||
| epsilon: | ||
| log_loss: 1.0e-7 | ||
| optimizer: 1.0e-7 |
There was a problem hiding this comment.
这两个epsilon是不是调试的时候基本不会改，那就直接在程序里写死吧，不用特别拿出来写到yaml文件里了
examples/topopt/topopt.py
Outdated
| import paddle | ||
| from functions import augmentation | ||
| from functions import generate_sampler | ||
| from functions import generate_train_test |
There was a problem hiding this comment.
参考tempogan，改成import functions as func_module, 用的时候func_module.augmentation这样吧
| === "模型训练命令" | ||
|
|
||
| ``` sh | ||
| python topopt.py |
There was a problem hiding this comment.
数据集我已上传到 https://paddle-org.bj.bcebos.com/paddlescience/datasets/topopt/top_dataset.h5，参考 别的案例比如hpinns 添加一下数据下载命令
docs/zh/examples/topopt.md
Outdated
| === "模型评估命令" | ||
|
|
||
| ``` sh | ||
| python topopt.py 'mode=eval' 'EVAL.pretrained_model_path_dict={"Uniform": "path1", "Poisson5": "path2", "Poisson10": "path3", "Poisson30": "path4"}' |
There was a problem hiding this comment.
path1/2/3/4可以先写上，比如 https://paddle-org.bj.bcebos.com/paddlescience/models/topopt/uniform_pretrained.pdparams之后我再上传就行
如果你现在有现成的，方便的话可以传到aistudio上放数据集那里吗，我下下来上传，没有的话就我之后跑一下
There was a problem hiding this comment.
已修改，模型放到aistudio数据集那里了
…ePaddle#597) * update hackthon no. 57 * delete TopOpt * add example/topopt * update * modify hydra * add md * fix code * fix yaml * fix yaml * update code * recommit * fix code * fix docs * remove TopOptModel.py * add topoptmodel.py
PR types
Others
PR changes
Others
Describe
按照意见修改了之前的pr: #559