Add pr-curve support #8
Conversation
|
Just noticed, histogram support broke because of change in default_gernerate, I will try to fix it |
|
Do you knwo how to change this value I mean, Scale it down by 1000 like 0.003 to 0..007 something like that |
| @@ -140,9 +142,24 @@ class TensorBoardLogger { | |||
| const std::vector<std::string> &metadata = std::vector<std::string>(), | |||
| const std::string &metadata_filename = "", | |||
| int step = 1 /* no effect */); | |||
|
|
|||
| int prcurve(const std::string tag, | |||
| const std::vector<double>labels, | |||
| @@ -243,6 +248,121 @@ int TensorBoardLogger::add_embedding( | |||
| tensor_shape, step); | |||
| } | |||
|
|
|||
| std::vector<std::vector<double>> TensorBoardLogger::compute_curve( | |||
| const std::vector<double>labels, | |||
| { | ||
| weights.push_back(1.0); | ||
| } | ||
| generate_default_buckets({0, (double)num_thresholds - 1}, num_thresholds, true, true); |
There was a problem hiding this comment.
Can this num_thresholds be fixed to 127, so buckets only need to be generated once? I don't see the necessity to change it. (I will look deeper later)
| return data; | ||
| } | ||
| int TensorBoardLogger::prcurve( | ||
| const std::string tag, |
| auto *tensor = new TensorProto(); | ||
| tensor->set_dtype(tensorflow::DataType::DT_DOUBLE); | ||
| tensor->set_allocated_tensor_shape(tensorshape); | ||
| for(int i=0;i<data.size();i++) |
There was a problem hiding this comment.
Range-based for loop is better suited here.
| private: | ||
| int generate_default_buckets(); | ||
| std::vector<std::vector<double>> compute_curve( | ||
| const std::vector<double>labels, |
I have no idea, but I think maybe it's not necessary, since reasonable PR curve won't be such a flat line. |
|
I think bucket generation still have some issues, histogram looks kind of strange now. I also tested pr curve using data generated by pr curve demo code in TensorBoard repo, and the result looks different: vs.
I'm not sure whether I used it correctly, I'll try to find out. |
|
Ok, I will take a look again |
|
Hi can you quickly re run the script or test |
|
Sorry for the delay. The pr curve looks correct now, however the histogram still has some issue. I will try to find a fix. |
@RustingSword There you go
Not sure about the math, it probably looks correct but have a look at the math again :)
Changed
generate_bucket()a little bit to resemble ->
np.histogram()Note : https://github.com/reminisce/tensorboard-mxnet-logger/blob/ad7d6522e4010deaa3a2a95e93629d8b01751078/tensorboardX/summary.py#L219