Use pytorch models

This notebook shows how to do inference with trained neural network from pytorch.

The first example is a 1D-CNN for wet-dry classification.

import xarray as xr
import matplotlib.pyplot as plt

import pycomlink as pycml
from pycomlink.processing.pytorch_utils import run_inference

Load and prepare data

data_path = pycml.io.examples.get_example_data_path()
cmls = xr.open_dataset(data_path + '/example_cml_data.nc')

# select 3 different CMLs to study
cmls = cmls.isel(cml_id = [0, 10, 370])

# Remove outliers, compute tl and interpolate missing values
cmls['tsl'] = cmls.tsl.where(cmls.tsl != 255.0)
cmls['rsl'] = cmls.rsl.where(cmls.rsl != -99.9)
cmls['tl'] = cmls.tsl - cmls.rsl # calculate total loss (previous TRSL)
cmls['tl'] = cmls.tl.interpolate_na(dim='time', method='linear', max_gap='5min')

# Dataarray shape is expected to be (time, channels, cml_id)
tl = cmls.tl.transpose('time', 'channel_id', 'cml_id')

# Normalisation
tl_normed = tl - tl.median(dim='time')

# Standardization should be part of the specific model, because it is trained with specific preprocessing

Load and run DL model

Option 1: Load DL model from local .pt file

# Set up your own local path:
model_path = 'your_local_path/model.pt'
result = run_inference.cnn_wd(model_path_or_url=model_path, data=tl_normed)

Option 2: Load the model from URL and cahe it

# Url adress:
model_URL = "https://github.com/jpolz/cml_wd_pytorch/raw/be2b15fa987838ea1f709dd0180917eebf66271a/data/dummy_model/best_model_jit.pt"
result = run_inference.cnn_wd(model_path_or_url=model_URL,data=tl_normed)

Results

for cml_id in cmls.cml_id.data:
    fig, axs = plt.subplots(3, 1, figsize=(14, 6))
    result.TL.sel(cml_id=cml_id).isel(channel_id=0).plot.line(x='time', ax=axs[0])
    result.predictions.sel(cml_id=cml_id).plot.line(x='time', ax=axs[1])
    (result.predictions.sel(cml_id=cml_id)>0.8).plot.line(x='time', ax=axs[2])