Closed mohammadx0098 closed 11 months ago
After learning with these hyperparameters :
hyperparameters["epochs"] = 5 hyperparameters["input_shape"] = (100, 1) # based on the window size hyperparameters["optimizer"] = "keras.optimizers.Adam" hyperparameters["learning_rate"] = 0.0005 hyperparameters["latent_dim"] = 20 hyperparameters["batch_size"] = 64
the model learned the time serie but why th amplitudes are not the same?
Now just plot the reconstructed signal :
Hi @mohammadx0098, thanks for using Orion!
To do that, you need to add a visualization
option to the pipeline json. please add the following block to the tadgan_without_gpu_dropout
pipeline after the output_names
block.
"outputs": {
"default": [
{
"name": "events",
"variable": "orion.primitives.timeseries_anomalies.find_anomalies#1.y"
}
],
"visualization": [
{
"name": "generated_timeseries",
"variable": "orion.primitives.tadgan.score_anomalies#1.predictions"
}
]
}
Here is matplotlib version that I use
matplotlib==3.5.3
matplotlib-inline==0.1.6
you can also let pip resolve this compatibility by installing both packages together using pip install orion-ml matplotlib
There are extreme values in the original time series that are probably affecting the learning process and shifting your reconstructed time series.
Hope this helps, let me know if you have any further questions!
@sarahmish Thanks.
3.I train the model witjout extreme values. The shift is solved but I have still have the problem with amplitute of the signal with is not the same as original one. I will send the results
Hi @mohammadx0098!
The generated_timeseries
contains 5 dimensions representing the minimum, 25th, 50th, 75th, and maximum values respectively. To plot the results, I made a quick snippet here to visualize the generated time series
import matplotlib.pyplot as plt
signal = viz['generated_timeseries']
signal = signal.squeeze(1)
x = len(signal)
median = signal[:, 2]
y1 = signal[:, 1]
y2 = signal[:, 3]
plt.plot(median)
plt.fill_between(range(x), y1, y2, alpha=0.3)
plt.show()
Hi
1. I want to use tadgan_without_gpu_dropout and have the visualized option.
How can I do that?
I am training my model and I think my problem is related to ability of my model to reconstruct the time serie.
2. When I'm usiing orion-ml in my windows local machine I have problem with mathplotlib and I think it's because of the conflict between your package and matplotlib in numpy or pandas version.
Which version of matplotlib is compatible with your package?