PeterJohnPyTorch
iPhone / developpeurs
Now You Can Use " PyTorch " on your iPhone even when you are in the train
or when your iPhone is OffLine.
****Why we Need " PyTorch " on the Edge Device such as iPhone??;
When "Torch" is a Lamp, "iPhone" becomes a Lamp Stand.
even When you are in the train or when your iPhone is OffLine.
So doNot put "Torch" on any Basket ( Hidden Place ) but put on your "iPhone".
So that the Light of the Lamp will Shine Before Others.
Scripture( Matthew 5:13;14-16 ) Says,
When the Light is Put on the Hill,
the City on the Hill canNot be Hidden.
****Matthew 5:13;14-16, ESV;
5:13 “You are the salt of the earth, but if salt has lost its taste,
how shall its saltiness be restored?
5:14 “You are the light of the world. A city set on a hill canNot be hidden.
5:15 Nor do people light a lamp and put it under a basket, but on a stand,
and it gives light to all in the house.
5:16 In the same way, let your light shine before others,
so that they may see your good works
and give glory to your Father who is in heaven.
****TutorialSeason007;
We Prepared some of Examples
to tell you what you Can do Using PeterJohnPyTorch.
s001QuestAnswerLibTorch.py;
This isNot PyTorch but just LibTorch.
QuestionAnswering Demo with LibTorch.
You canNot Customize almost Anything
'cause LibTorch is Called Via Swift
and 'cause Swift Needs to be Compiled
with XCode on MacOS.
This example Loads any Model from "qa360_quantized.ptl" file,
which includes "TorchScript".
s002ProvisionPyTorchTensor.py;
This Shows PeterJohnPyTorch Can Use Tensor.
s003ProvisionPyTorchAutoGradFoundation.py;
This Shows PeterJohnPyTorch Can Use AutoGrad.
s004ProvisionPyTorchAutoGrad.py;
This Shows PeterJohnPyTorch Can Use AutoGrad.
s005ProvisionPyTorchNN.py;
This Shows PeterJohnPyTorch Can Use torch.nn (NeuralNetwork).
s006ProvisionPyTorchNNoptimizer.py;
This Shows PeterJohnPyTorch Can Use torch.nn (NeuralNetwork)
and Optimizer.
This Uses SGD (Stochastic Gradient Descent ) as the Optimizer.
s007QuestAnswerPyTorch.py;
Now you Can See Not libTorch demo But PyTorch demo
about QuestionAnswering.
You can Customize, for example, Tokenizer
in "pjQuestionAnswering.py"
Using PyTorch on iPhone.
This example Loads any Model from "qa360_quantized.ptl" file,
which includes "TorchScript".
s008QuestAnswerTransformers.py;
QuestionAnswering Demo with PyTorch and Transformers.
After you touched "Run Script",
This example begins to Download "model.safetensors"(265.5MBytes),
We Recommend that you make a copy of "model.safetensors",
outside of this App "PeterJohnPyTorch",
using Apple's "Files.app".
So that, Even if you uninstalled this App,
After you installed this App "PeterJohnPyTorch" Again,
you can put back "model.safetensors" file to "/images" directory
of this App "PeterJohnPyTorch".
This example Loads any Model from "model.safetensors" file,
which is the Standard format of "HuggingFace".
****Some of Restrictions that We know Currently;
1) CanNot Trace any Model files using torch.jit.trace() Function;
Since Now "PeterJohnPyTorch" canNot Create any Traced Model,
Also it canNot Create either ".ptl"("PyTorch Lite" format) file
Nor ".pt"("PyTorch" format) file right now.
2) CanNot Script any Model files using torch.jit.script() Function;
Since Now "PeterJohnPyTorch" canNot Create any Scripted Model,
Also it canNot Create either ".ptl"("PyTorch Lite" format) file
Nor ".pt"("PyTorch" format) file right now.
Now you Need to
save any Model as "model.safetensors",
and use the Model via PyTorch or via Transformers,
Means Write the Logic via PyTorch or via Transformers.
3) Not Implement "MPS" (Metal Performance Shader) backends Yet;
So Now you Need to Specify "CPU" as Device.
The "CPU" backends is just the First Step Before the "GPU" backends.
****
Enjoy PeterJohnPyTorch even when you are in the train
or your iPhone is OffLIne.
--Yasushi Yassun Obata
Quoi de neuf dans la dernière version ?
'OpenCV-Python' ( OpenCV-Python Bindings ) without cv::dnn Available Version;
****Yes, You Can Use 'OpenCV-Python' without cv::dnn on 'PeterJohnPyTorch';
'without cv::dnn' Means that 'PeterJohnPyTorch' doesNot include OpenCV::DNN(DeepNeuralNetwork) module.
It is Due to 'Protobuf' Conflict Problem
'cause PyTorch uses its own Protobuf
and OpenCV::DNN(DeepNeuralNetwork) module has its own Protobuf,
and OpenCV::DNN(DeepNeuralNetwork) module is Tightly Combined with its own Protobuf.
Since there was No Way,
we built 'OpenCV-Python' without DNN module for 'PeterJohnPyTorch'.
but, Since 'PeterJohnPyTorch' can use 'PyTorch' for DNN(DeepNeuralNetwork),
So it will Not become any Big Problem.
Please Use 'PyTorch' for DNN(DeepNeuralNetwork)
and Just Use 'OpenCV-Python' for image processing.
****
if you have any Strong Reason Not to use PyTorch
But to Use OpenCV::DNN(DeepNeuralNetwork) module on iPhone,
please use 'PeterJohnNumPy' or 'PeterJohnNumPyMatplotlib'.
****Restriction in the 'PeterJohnThreeDim';
'OpenCV-Python' needs NumPy as the data representation,
so PeterJohnThreeDim which doesNot have NumPy,
canNot Support 'OpenCV-Python'.
if you need 'OpenCV-Python',
please Upgrade to 'PeterJohnNumPy', ''PeterJohnNumPyMatplotlib'
or 'PeterJohnPyTorch'.
****Two Examples;
We prepared two examples at this time;
TutorialSeason003/s006opencvPythonMachineLearning.py;
TutorialSeason003/s007opencvPythonVideoCaptureMachineLearning.py
****Two-Unimplemented Features;
OpenCV-Python on PeterJohnNumPy has Two-Unimplemented Features;
1) cap = cv2.VideoCapture()
OpenCV-Python on PeterJohnNumPy can't start
VideoCapture directly.
2) ret, img = cap.read()
OpenCV-Python on PeterJohnNumPy can't read frame
from VideoCamera directly.
But, instead of Abobe Two,
Use 'pjOpenCV as cvCPP';
So that you Can Edit the image captured from VideoCamera,
using 'OpenCV-Python'.
****How to Edit the image captured from VideoCamera;
1) setup VideoCapture;
import pjOpenCV as cvCPP
cvCPP.setVideoCaptureFPS(10)
##cvCPP.initVideoCaptureBackCamera()
cvCPP.initVideoCaptureFrontCamera()
##cvCPP.flipCamera() ## flipCamera()
2) start VideoCapture;
captured image is Stored as the UIImage 'srcImg';
cvCPP.startVideoCapture()
import cv2
3) To Get Captured image, use 'cvCPP.saveSrcImg()'.
cvCPP.saveSrcImg( srcAbsPath )
img = cv2.imread( srcAbsPath )
4) Edit the NumPy image, Using "OpenCV-Python".
5) To Reflect the edited NumPy image on the UIImage 'destImg',
use 'cvCPP.createDestImg';
cv2.imwrite(destAbsPath, img)
cvCPP.createDestImg(destAbsPath)
6) To Show the UIImage 'destImg', use 'cvCPP.showDestImg()'.
cvCPP.showDestImg()
Enjoy MachineLearning,
Enjoy DeepLearning,
--Yasushi Yassun Obata
