processing/classification_model.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
Execution:
python3 -W ignore model.py
"""

import numpy as np
import torch


# Meta data
feedrate = ["477", "955", "1242"] # [mm/min]
depth_of_cut = ["0.1", "0.25", "0.5"] # [mm]
condition = ["new", "slightly used", "heavily used"]


# Trained model
model = None


def load_model(pathToModel: str):
    """
    Load torch model

    Args:
        pathToModel (str): (Full-)Path to pytorch model.
    """
    # Load torch model
    global model
    model = torch.jit.load(pathToModel)
    model.eval()
    model.cpu()


def classify(data: torch.Tensor):
    """
    Classifies data based on loaded model.

    Args:
        data (torch.Tensor): Data to classify. Structure [#tasks, 1, 500].

    Returns:
        An array with dim [9), separated by indices [0-3), [3,6), [6,9).
        Representing the classes "feedrate", "depth of cut" and "condition".
        "feerate": 477, 955, 1242 mm/min
        "depth of cut": 0.1, 0.25, 0.5 mm
        "condition": new, slightly used, heavily used
        Based on data separation, the "condition" was equally compressed among
        the data.
    """
    try:
        load_model("processing/model_d2_v0.6.pt")
    except:
        pass


    with torch.no_grad():
        # Data transformation and prediction
        data = torch.tensor(data, dtype=torch.float).to("cpu")
        prediction = model(data)
        prediction = torch.nn.functional.sigmoid(prediction)

        # Summed classification
        summed_pred = (prediction.sum(axis=0) / len(prediction)).numpy()[0]

        # Prediction class for label set
        res_pred = (np.argmax(summed_pred[0:3]).item(),
                    np.argmax(summed_pred[3:6]).item(),
                    np.argmax(summed_pred[6:len(summed_pred)]).item())

        res_con = f"Feedrate [mm/min]: {feedrate[res_pred[0]]}, "\
                  f"Depth of cut [mm]: {depth_of_cut[res_pred[1]]}, "\
                  f"Condition: {condition[res_pred[2]]}"

        # Label representation
        res_pred = (res_pred[0], res_pred[1] + 3, res_pred[2] + 6)
        res_label = torch.zeros(len(prediction[0][0]))
        for x in res_pred:
            res_label[x] = 1

        return summed_pred, res_label, res_con