Skip to content

Latest commit

 

History

History
179 lines (124 loc) · 6.65 KB

RELEASE.md

File metadata and controls

179 lines (124 loc) · 6.65 KB
Raw

TinyMS Release Notes

0.2.1

Released 2021-07-15.

Major Features and Improvements

  • Fix load_checkpoint interface bug in TinyMS 0.2.0 hub module. #96

0.2.0

Released 2021-06-07.

Major Features and Improvements

  • Add text module to provide the basic dataset loading and preprocessing in NLP scenarios. #53 #73
  • Upgrade the version of mindspore module dependencies from v1.1.1 to v1.2.0. #81 #84
  • Refactor the Client and Server communication interface in serving module. #76
  • Added server_path, start FlaskServer and add host and port parameters. #77
  • Implement TinyMS hub module to enable loading lots of pre-trained models, incluidng lenet5_v1, resnet50_v1, alexnet_v1, vgg16_v1, mobilenet_v2 and ssd300_v1. #86 #93
  • Publish the TinyMS Hub contributing guidelines in public to welcome pre-trained model from the comunity. #91
  • Refactor the model network entrypoint method to provide the unified interface. #85

Model Park

  • Add 5 models support: AlexNet, DenseNet100, VGG16, SentimentNet, Bert. #59 #89 #63 #67

API Change

  • Refactor the serving entrypoint function with Client and Server class interface.
v0.1.0 v0.2.0 
from tinyms.serving import start_server, server_started, list_servables, predict, shutdown

start_server()
if server_started():
    list_servables()
    predict('example.jpg', 'servable_name', dataset_name='mnist')
shutdown()
 
from tinyms.serving import Client, Server

server = Server()
server.start_server()
client = Client()
client.list_servables()
client.predict('example.jpg', 'servable_name', dataset_name='mnist')
server.shutdown()
  • Add a new interface load in model module to support load MindIR graph directly to perform model inference operation.
v0.2.0 
>>> import tinyms as ts
>>> import tinyms.layers as layers
>>> from tinyms.model import Model, load
>>>
>>> net = layers.Conv2d(1, 1, kernel_size=3)
>>> model = Model(net)
>>> input = ts.ones([1, 1, 3, 3])
>>> model.export(input, "net", file_format="MINDIR")
...
>>> net = load("net.mindir")
>>> print(net(input))
[[[[ 0.02548009  0.04010789  0.03120251]
    [ 0.00268656  0.02353744  0.03807815]
    [-0.00896441 -0.00303641  0.01502199]]]]
  • Add hub.load method to easily load pretrained model and apply model evaluation and inference operation.
v0.2.0 
from PIL import Image
from tinyms import hub
from tinyms.vision import mnist_transform
from tinyms.model import Model

img = Image.open(img_path)
img = mnist_transform(img)

# load LeNet5 pretrained model
net= hub.load('tinyms/0.2/lenet5_v1_mnist', class_num=10)
model = Model(net)

res = model.predict(ts.expand_dims(ts.array(img), 0)).asnumpy()
print("The label is:", mnist_transform.postprocess(res))

For the detailed API changes, please find TinyMS Python API in API Documentation.

Backwards Incompatible Change

None

Bug fixes

  • Fix some bugs when serving in Windows operating system. #74
  • Set batch_norm as True by default in VGG16 to fix the converge problem of accuracy. #90

Contributors

Great thanks go to these wonderful people:

@zjuter0126, @Mickls, @leonwanghui, @hannibalhuang, @hellowaywewe, @huxiaoman7

0.1.0

Released 2021-03-28.

Major Features and Improvements

  • Design the overall framework of TinyMS development toolkit. #3 #5 #12 #13
  • Support install TinyMS binary in Linux Ubuntu 18.04 and Window 10 environment, also provide TinyMS docker image to users. #2 #45
  • Enable document auto-generation using Sphinx. #35
  • Provide several end to end model development and deployment tutorials for machine learning beginners. #11 #24 #26 #34
  • Set up the initial CI pipeline (including cla-assistant, GitHub Actions, readthedocs) for TinyMS project. #1 #49 #50

Model Park

  • Add 5 models support: LeNet5, ResNet50, MobileNetV2, SSD300, CycleGAN. #5 #14 #17 #32

API Change

There is no API change for the first version of TinyMS. Please find TinyMS Python API in API Documentation.

Backwards Incompatible Change

None

Bug fixes

None

Contributors

Great thanks go to these wonderful people:

@leonwanghui, @lyd911, @hannibalhuang, @hellowaywewe, @Yikun, @huxiaoman7