【PaddlePaddle Hackathon 4】add paddle linspace op #15835
Conversation
src/frontends/paddle/tests/test_models/gen_scripts/generate_linspace.py
Outdated
Show resolved
Hide resolved
src/frontends/paddle/tests/test_models/gen_scripts/generate_linspace.py
Outdated
Show resolved
Hide resolved
|
Hi, the source codes of paddle linspace API illustrate that the input scalar(int/float) will be transferred in to tensor.
I wonder if it is needed to provide the supporting of scalar input on linspace.cpp. Same question about unit-test samples also exist. |
| // generate a range of numbers [0, 1, ..., num) | ||
| auto const_zero = std::make_shared<default_opset::Constant>(start.get_element_type(), Shape{}, 0); | ||
| auto const_num = std::make_shared<default_opset::Squeeze>(num); | ||
| auto range0_n = std::make_shared<default_opset::Range>(const_zero, const_num, const_one, start.get_element_type()); |
There was a problem hiding this comment.
I just wonder can we use range for a direct output ? so we do not need Multiply and Add, just a guess
There was a problem hiding this comment.
great idea, it is important to simplify the operator code. but the behavior of Range is subtly different from Linspace even though their input parameters are similar with each other. Using Range for a direct output maybe require the overal re-design of linspace.
There was a problem hiding this comment.
Thanks, could you clarify what's gap between them? can we convert the num in Linspace into step in Range
There was a problem hiding this comment.
step can be calculated by num, which has already implemented in line 30 of linspace.cpp. The key point is the behavior of Range. For instance, same input (start: [0] stop: [1] num: [4]/step: [0.33]), and the result will be:
Range: [0, 0.33, 0.66]
Linspace: [0, 0.33, 0.66, 1]
And the input of Range must be scalar, which means must be taken before range operator.
I think it will become more complicate after adjusting the code.
There was a problem hiding this comment.
Great thanks for your illustration
There was a problem hiding this comment.
Can start and stop be negative here, and can you put this example ( (start: [0] stop: [1] num: [4]/step: [0.33])) into your test case, thanks?
There was a problem hiding this comment.
OK，test samples had been added
github-actions
bot
added
category: GPU
|
conflicts resoved |
ilya-lavrenov
added
ExternalPR
Hello could you also post a English version of Paddle Op description, thanks |
| namespace paddle { | ||
| namespace op { | ||
| NamedOutputs linspace(const NodeContext& node) { | ||
| auto start = node.get_input("Start"); |
There was a problem hiding this comment.
How about the inputs are variables, but not a tensor. e.g int or float
Hi, @OpenVINO-dev-contest
|
Sorry I missed it. but shall we consider the
Sorry, I missed it, but shall we consider the |
The python frontend of paddle could ensure that the input type and attr |
Can you update the uni-test with these case example ? for example input type is fp32 and dtype is int32. just ensure the Paddle front end will do conversion for us. |
sorry, I misunderstood the function of |
| auto stop = node.get_input("Stop"); | ||
| auto num = node.get_input("Num"); | ||
| std::unordered_map<int, element::Type> umap{{2, element::i32}, {3, element::i64}, {5, element::f32}}; | ||
| const auto dtype = node.get_attribute<int>("dtype", 5); |
There was a problem hiding this comment.
can we use auto dtype = node.get_attribute<ov::element::Type>("dtype"); to get the the type ?
There was a problem hiding this comment.
Pls add a screenshot of test results, thanks
|
Can one of the admins verify this patch? |
Details:
add linspace operation in Paddle front end
Reference:
Unit-Test