LIB: Add mindspore backend #169

chou-shun · 2021-08-24T02:35:27Z

Add mindspore backend
If both GPU and NPU exist in the environment, NPU is preferred

Signed-off-by: zhangjun [email protected]

JoeyHwong-gk · 2021-08-24T03:38:46Z

lib/sedna/backend/__init__.py

-    if 'CUDA_VISIBLE_DEVICES' in os.environ:
+
+    # NPU>GPU>CPU
+    if device_category == "ASCEND":


Maybe it's better to change "ASCEND" to "NPU".

This is because mindspore takes “ASCEND" as "device_target", is it necessary to modify it?
https://www.mindspore.cn/docs/api/zh-CN/r1.3/api_python/mindspore.context.html?highlight=Ascend#module-mindspore.context

In fact, this variable device_category does not affect the framework, only use to guide our contributors to distinguish between processing units, as usually juxtapose NPU with CPU/GPU, "ASCEND" is only provided by huawei.org.

In the current scenario, if MindSpore supports ASCEND only, I think it is best not to set the limit here, avoid special requirements for other products that may support NPU.

JoeyHwong-gk · 2021-08-24T03:44:03Z

lib/sedna/backend/mindspore/__init__.py

+        """todo: no support yet"""
+
+    def model_info(self, model, relpath=None, result=None):
+        _, _type = os.path.splitext(model)


It seems to be the same as the parent class, suggest to delete.

Got it, thanks.

1. Add mindspore backend 2. If both GPU and NPU environments exist in the environment, NPU environment is preferred 3. Modify parameter "use_ascend": "use_ascend"->"use_npu" 4. Delete function "model_info" in mindscore backend Signed-off-by: chou-shun <[email protected]>

chou-shun · 2021-08-24T06:43:05Z

@JoeyHwong-gk I have revised it according to your opinion. Please review it again. Thanks:)

JoeyHwong-gk · 2021-08-24T07:16:12Z

@JoeyHwong-gk I have revised it according to your opinion. Please review it again. Thanks:)

thx for your job, it looks good to me

ugvddm · 2021-08-24T14:10:15Z

lib/sedna/backend/mindspore/__init__.py

@@ -0,0 +1,74 @@
+# Copyright 2021 The KubeEdge Authors.


thanks for your great job! BTW, should we provide some example about mindspore to end user? joint inference or incremental learning?

OK, I'll upload a resnet example by tomorrow

@chou-shun please put your example in lib/examples/backend/mindspore/.

1. Develop a resnet example based on mindspore backend 2. scripts/: Script to start train, test and inference src/: ResNet network and model config train.py: Entrance to model training test.py: Entrance to model testing infrence.py Entrance to model inference interface.py: Implements class `Estimator`` Signed-off-by: chou-shun <[email protected]>

kubeedge-bot · 2021-08-27T03:58:34Z

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by:
To complete the pull request process, please assign jaypume after the PR has been reviewed.
You can assign the PR to them by writing /assign @jaypume in a comment when ready.

The full list of commands accepted by this bot can be found here.

Needs approval from an approver in each of these files:

examples/OWNERS
lib/OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

JoeyHwong-gk · 2021-08-27T04:03:52Z

examples/lib-samples/backend/mindspore/ResNet50/inference.py

+from mindspore import Tensor
+import mindspore.dataset.vision.c_transforms as C
+import numpy as np
+from lib.sedna.backend import set_backend


maybe change with from sedna.backend import set_backend better

JoeyHwong-gk · 2021-08-27T04:05:04Z

examples/lib-samples/backend/mindspore/ResNet50/inference.py

+import mindspore.dataset.vision.c_transforms as C
+import numpy as np
+from lib.sedna.backend import set_backend
+import cv2


import cv2 before sedna

JoeyHwong-gk · 2021-08-27T04:10:17Z

lib/sedna/backend/mindspore/__init__.py

+        if callable(self.estimator):
+            self.estimator = self.estimator()
+
+    def train(self, train_data, valid_data=None, **kwargs):


Based on existing specifications, the first parameter in train is lib.sedna.datasources.

That is because both formats (CSV, TXT) in datasource do not support cifar-10 datasets.
cifar-10: data_batch_1.bin, data_batch_2.bin, data_batch_3.bin,..., test_batch.bin

Sorry for the misunderstanding. datasource inherited of BaseDataSource，as that core feature of sedna require identifying the features and labels from data input, we specify that the first parameter for train/evaluate of the ML framework must be a specific object (inherited of BaseDataSource).

Ok, I get it. However, should I develop a specific datasource for the cifar-10 dataset? This will be a little complicated...

It would be a huge help to our community. And I would like you to give the professional advice on how to design DataSource.
In this example, it might be solved by:

from sedna.datasources import BaseDataSource mnist_ds = ds.MnistDataset(train_data_path) train_data = BaseDataSource(data_type="train") train_data.x = [] train_data.y = [] for item in mnist_ds.create_dict_iterator(): train_data.x.append(item["image"].asnumpy()) train_data.y.append(item["label"].asnumpy())

OK, I'll solve it as soon as possible.

I tried your method, but it didn't work very well. That is because, for method mindspore.Model.train, the data passed to it MUST BE a mindspore.Dataset, and can not be a BaseDataSource.
Here are some solutions I thought of：

Abandon class mindspore.Model and adopt a more flexible approach: for data in dataloader: loss = network(data), loss.update().

Repackage train_data into a mindspore.Dataset in interface.py.

For solution 1, it increases the difficulty of model development and is not a popular way.
For solution 2, actually it is counterintuitive.
Which do you prefer? Or do you have a more reasonable solution?

PTAL @jaypume

I suggest that BaseDataSorce is designed to compatible with existing Dataset such as mindspore.Dataset, torch.utils.data.Dataset, tf.data.Dataset. If the Dataset instance is passed to for example FederatedLearning.train(tfDatasetInstance), it should work.

I suggest that BaseDataSorce is designed to compatible with existing Dataset such as mindspore.Dataset, torch.utils.data.Dataset, tf.data.Dataset. If the Dataset instance is passed to for example FederatedLearning.train(tfDatasetInstance), it should work.

You can help us raise an issue so we can track it.

JoeyHwong-gk · 2021-08-27T07:34:59Z

lib/sedna/backend/mindspore/__init__.py

+            self.estimator.load_weights(model_path)
+
+    def get_weights(self):
+        """todo: no support yet"""


return self.estimator.parameters_dict()

JoeyHwong-gk · 2021-08-27T07:36:11Z

lib/sedna/backend/mindspore/__init__.py

+        """todo: no support yet"""
+
+    def set_weights(self, weights):
+        """todo: no support yet"""


for name, weight in weights.items():
weights[name] = mindspore.Parameter(weight, name=name)
mindspore.load_param_into_net(self.estimator, weights, strict_load=True)

self.estimator is not a net. get_weights and set_weights would to be developed later:)
mindspore.load_param_into_net(net, parameter_dict, strict_load=False)

1. Add directory description in README.md. 2. Modify "test" to "eval". 3. Add some comments. Signed-off-by: chou-shun <[email protected]>

ugvddm · 2021-08-30T11:02:36Z

examples/lib-samples/backend/mindspore/ResNet50/README.md

+    ├── run_standalone_train_cpu.sh        # launch cpu training
+  ├── src
+    ├── config.py                          # parameter configuration
+    ├── dataset.py                         # data preprocessing


should we move the way of data preprocessing for midnspore to sedna lib, instead of in examples? the other people may be want to use it when they develop a appliation based on mindspore.

Where do you think is the most appropriate ? sedna.datasource or sedna.backend.mindspore ?

Should we develop preprocessing methods for each model, or develop a general preprocessing method?
In my assumption, for a certain scene (such as image classification), we can predefine several fixed preprocessing methods (such as normalize, resize), and the user only needs to pass some parameters.

I suggest not to integrate data preprocessing functions into Sedna Lib, for the reason:

There are various preprocessing functions, and they cannot be covered by Sedna.

The data preprocessing functions are always already defined in developer's training code or AI framework.

JoeyHwong-gk · 2021-08-31T02:53:01Z

/assign @jaypume

Signed-off-by: chou-shun <[email protected]>

kubeedge-bot · 2022-11-01T06:49:16Z

@chou-shun: PR needs rebase.

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes/test-infra repository.

kubeedge-bot requested review from JimmyYang20 and TymonXie August 24, 2021 02:35

kubeedge-bot added the size/L Denotes a PR that changes 100-499 lines, ignoring generated files. label Aug 24, 2021

chou-shun force-pushed the main branch 2 times, most recently from 463b3b4 to 10a2ca7 Compare August 24, 2021 03:07

JoeyHwong-gk reviewed Aug 24, 2021

View reviewed changes

chou-shun force-pushed the main branch from 10a2ca7 to 99a4272 Compare August 24, 2021 06:33

kubeedge-bot added size/M Denotes a PR that changes 30-99 lines, ignoring generated files. and removed size/L Denotes a PR that changes 100-499 lines, ignoring generated files. labels Aug 24, 2021

chou-shun force-pushed the main branch from 99a4272 to d7ee67c Compare August 24, 2021 06:35

ugvddm reviewed Aug 24, 2021

View reviewed changes

kubeedge-bot added size/XXL Denotes a PR that changes 1000+ lines, ignoring generated files. and removed size/M Denotes a PR that changes 30-99 lines, ignoring generated files. labels Aug 27, 2021

JoeyHwong-gk reviewed Aug 27, 2021

View reviewed changes

EXAMPLES: Optimize some format problems

77aaded

1. Add directory description in README.md. 2. Modify "test" to "eval". 3. Add some comments. Signed-off-by: chou-shun <[email protected]>

ugvddm reviewed Aug 30, 2021

View reviewed changes

kubeedge-bot assigned jaypume Aug 31, 2021

kubeedge-bot added the needs-rebase Indicates a PR cannot be merged because it has merge conflicts with HEAD. label Jun 3, 2022

chou-shun added 2 commits July 6, 2022 21:11

bugfix

6dbc799

Signed-off-by: chou-shun <[email protected]>

Merge remote-tracking branch 'upstream/main'

82d614b

Signed-off-by: chou-shun <[email protected]>

kubeedge-bot removed the needs-rebase Indicates a PR cannot be merged because it has merge conflicts with HEAD. label Jul 6, 2022

kubeedge-bot added the needs-rebase Indicates a PR cannot be merged because it has merge conflicts with HEAD. label Nov 1, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

LIB: Add mindspore backend #169

LIB: Add mindspore backend #169

chou-shun commented Aug 24, 2021 •

edited

Loading

JoeyHwong-gk Aug 24, 2021

chou-shun Aug 24, 2021

JoeyHwong-gk Aug 24, 2021 •

edited

Loading

JoeyHwong-gk Aug 24, 2021

chou-shun Aug 24, 2021

chou-shun commented Aug 24, 2021

JoeyHwong-gk commented Aug 24, 2021

ugvddm Aug 24, 2021

chou-shun Aug 25, 2021

JimmyYang20 Aug 25, 2021

kubeedge-bot commented Aug 27, 2021

JoeyHwong-gk Aug 27, 2021

JoeyHwong-gk Aug 27, 2021

JoeyHwong-gk Aug 27, 2021

chou-shun Aug 27, 2021

JoeyHwong-gk Aug 27, 2021

chou-shun Aug 27, 2021

JoeyHwong-gk Aug 27, 2021

chou-shun Aug 27, 2021

chou-shun Aug 29, 2021

JoeyHwong-gk Aug 29, 2021

jaypume Sep 1, 2021

JoeyHwong-gk Sep 1, 2021

JoeyHwong-gk Aug 27, 2021

JoeyHwong-gk Aug 27, 2021

chou-shun Aug 27, 2021

ugvddm Aug 30, 2021

JoeyHwong-gk Aug 31, 2021

chou-shun Aug 31, 2021

jaypume Sep 1, 2021

JoeyHwong-gk commented Aug 31, 2021

kubeedge-bot commented Nov 1, 2022

LIB: Add mindspore backend #169

Are you sure you want to change the base?

LIB: Add mindspore backend #169

Conversation

chou-shun commented Aug 24, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JoeyHwong-gk Aug 24, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

chou-shun commented Aug 24, 2021

JoeyHwong-gk commented Aug 24, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kubeedge-bot commented Aug 27, 2021

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

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

Choose a reason for hiding this comment

JoeyHwong-gk commented Aug 31, 2021

kubeedge-bot commented Nov 1, 2022

chou-shun commented Aug 24, 2021 •

edited

Loading

JoeyHwong-gk Aug 24, 2021 •

edited

Loading