Append ECG neural network tune example (#199)

* I supplemented the documentation with a paragraph about the work of the framework with the optimal selection of two real and one discrete parameters. Corrected the problem code for finding real and discrete parameters.

* correct target score

* Append new examples. Correct documentation

* Corrected documentation of examples

Author: Audice

.gitignore

@@ -134,3 +134,7 @@ dmypy.json
 
 # datasets
 benchmarks/data/datasets
+examples/Machine_learning/NeuralNetwork/Segmentation/data
+examples/Machine_learning/NeuralNetwork/Segmentation/models/*
+examples/Machine_learning/NeuralNetwork/Segmentation/lightning_logs
+examples/Machine_learning/NeuralNetwork/Segmentation/data.zip

examples/Machine_learning/NeuralNetwork/Segmentation/Problem/Cardio2D.py

@@ -0,0 +1,145 @@
+import random
+
+from examples.Machine_learning.NeuralNetwork.Segmentation.scripts.dataset import SegmentationDataset
+from examples.Machine_learning.NeuralNetwork.Segmentation.scripts.metric import AllMetricTracker
+from iOpt.trial import Point
+from iOpt.trial import FunctionValue
+from iOpt.problem import Problem
+from typing import Dict
+from datetime import datetime
+import os
+from sklearn.model_selection import train_test_split
+from torch.utils.data import DataLoader
+from lightning.pytorch import Trainer
+from lightning.pytorch.callbacks import ModelCheckpoint, EarlyStopping
+import torch
+import torch.nn as nn
+import numpy as np
+from lightning.pytorch import LightningModule
+from examples.Machine_learning.NeuralNetwork.Segmentation.scripts.metric import SegmentationMetric
+from examples.Machine_learning.NeuralNetwork.Segmentation.scripts.model import Encoder, Decoder, UNet
+
+
+class UnetModule(LightningModule):
+    def __init__(self, kernel_size=23, q=1.2, label_smoothing=0, p=0.75):
+        super().__init__()
+        self.save_hyperparameters()
+        encoder = Encoder(12, kernel_size=kernel_size, q=q, p=p)
+        decoder = Decoder(encoder, 4)
+
+        self.model = UNet(encoder, decoder)
+        self.loss = nn.CrossEntropyLoss(ignore_index=4, label_smoothing=label_smoothing)
+
+        self.p_metric = SegmentationMetric('p', 'all', return_type='f1', samples=150)
+        self.t_metric = SegmentationMetric('t', 'all', return_type='f1', samples=150)
+        self.qrs_metric = SegmentationMetric('qrs', 'all', return_type='f1', samples=150)
+
+    def predict(self, x):
+        if isinstance(x, np.ndarray):
+            x = torch.Tensor(x)
+            x = x.unsqueeze(0) if len(x.shape) == 2 else x
+        x = x.to(self.device)
+        logits = self.model(x)
+        y_pred = logits.argmax(axis=1)
+        return y_pred.cpu().detach().numpy()
+
+    def training_step(self, batch):
+        _, x, y = batch
+        logits = self.model(x)
+        loss = self.loss(logits, y)
+        dict_ = {'train_loss': loss}
+        self.log_dict(dict_, on_epoch=True, on_step=False)
+        return loss
+
+    def validation_step(self, batch):
+        _, x, y = batch
+        logits = self.model(x)
+        loss = self.loss(logits, y)
+        dict_ = {'val_loss': loss}
+
+        metrics = self.get_metric(x, y, 'val')
+        dict_.update(metrics)
+
+        self.log_dict(dict_, on_epoch=True, on_step=False)
+
+        return loss
+
+    def get_metric(self, x, y_true, prefix):
+        y_true = y_true.cpu().detach().numpy()
+        y_pred = self.predict(x)
+        p_f1_score = self.p_metric(y_pred, y_true)
+        qrs_f1_score = self.qrs_metric(y_pred, y_true)
+        t_f1_score = self.t_metric(y_pred, y_true)
+        dict = {f'{prefix}_p_wave': p_f1_score, f'{prefix}_qrs_wave': qrs_f1_score, f'{prefix}_t_wave': t_f1_score}
+        return dict
+
+    def configure_optimizers(self):
+        optimizer = torch.optim.AdamW(self.model.parameters())
+        scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=0.3, patience=50)
+        return [optimizer], [{"scheduler": scheduler,
+                              "interval": "epoch",
+                              "monitor": "train_loss"}]
+
+def get_dataset(paths):
+    return [np.load(f'data/signals/{x}') for x in paths], \
+           [np.load(f'data/masks/{x}') for x in paths]
+
+class Cardio2D(Problem):
+    def __init__(self, p_bound: Dict[str, float], q_bound: Dict[str, float]):
+        super(Cardio2D, self).__init__()
+        self.dimension = 2
+        self.number_of_float_variables = 2
+        self.number_of_discrete_variables = 0
+        self.number_of_objectives = 1
+        self.number_of_constraints = 0
+
+        ecg_list = sorted(os.listdir('data/signals/'))
+        ecg_list = [x for x in ecg_list if x.split('_')[-1] != 'unsupervised.npy']
+
+        train_list, test_list = train_test_split(ecg_list, test_size=0.2, shuffle=True, random_state=42)
+
+        for x in sorted(os.listdir('data/signals/')):
+            if x.split('_')[-1] == 'unsupervised.npy':
+                train_list.append(x)
+
+        x_train, y_train = get_dataset(train_list)
+        x_test, y_test = get_dataset(test_list)
+
+        train_dataset = SegmentationDataset('cpu', train_list, x_train, y_train, common_mask=True, for_train=True)
+        val_dataset = SegmentationDataset('cpu', test_list, x_test, y_test, common_mask=True)
+
+        self.train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)
+        self.val_loader = DataLoader(val_dataset, batch_size=32)
+
+        self.float_variable_names = np.array(["P parameter", "Q parameter"], dtype=str)
+        self.lower_bound_of_float_variables = np.array([p_bound['low'], q_bound['low']],
+                                                       dtype=np.double)
+        self.upper_bound_of_float_variables = np.array([p_bound['up'], q_bound['up']],
+                                                       dtype=np.double)
+
+    def calculate(self, point: Point, function_value: FunctionValue) -> FunctionValue:
+        p, q = point.float_variables[0], point.float_variables[1]
+
+        now = datetime.now().strftime('%d.%m.%Y_%H:%M:%S')
+
+        checkpoint = ModelCheckpoint(dirpath=f'models/',
+                                     filename=f"{random.uniform(1, 100):.9f}" + " " + f"{p:.9f}" + '_' + f"{q:.9f}" + '_' + '{epoch}_{val_p_wave:.6f}_{val_qrs_wave:.6f}_{val_t_wave:.6f}',
+                                     monitor='val_p_wave',
+                                     save_top_k=3,
+                                     mode='max')
+        early_stopping = EarlyStopping(monitor='val_loss',
+                                       patience=300)
+
+        cb = AllMetricTracker()
+        model = UnetModule(p=p, q=q)
+        trainer = Trainer(max_epochs=1_000_000, callbacks=[checkpoint, early_stopping, cb])
+        try:
+            trainer.fit(model, self.train_loader, self.val_loader)
+        except Exception as err:
+            print(f"Unexpected {err=}, {type(err)=}")
+
+        print('p ' + f"{p:.9f}")
+        print('q ' + f"{q:.9f}")
+        function_value.value = -cb.best_p_valscore
+        print(-cb.best_p_valscore)
+        return function_value

examples/Machine_learning/NeuralNetwork/Segmentation/UnetExample.py

@@ -0,0 +1,61 @@
+import shutil
+
+import numpy as np
+from examples.Machine_learning.NeuralNetwork.Segmentation.Problem.Cardio2D import Cardio2D
+from iOpt.output_system.listeners.console_outputers import ConsoleOutputListener
+from iOpt.solver import Solver
+from iOpt.solver_parametrs import SolverParameters
+import hashlib
+import os
+from pathlib import Path
+
+import requests
+from tqdm import tqdm
+
+
+def _get_hash(path: Path) -> str:
+    file_hash = hashlib.sha256()
+    with open(path, "rb") as f:
+        while chunk := f.read(8192):
+            file_hash.update(chunk)
+    return file_hash.hexdigest()
+
+
+def download(path: Path, public_key: str) -> None:
+    url = "https://cloud-api.yandex.net/v1/disk/public/resources"
+    params = {"public_key": f"https://disk.yandex.ru/d/{public_key}"}
+
+    response = requests.get(url, params=params).json()
+    download_url = response["file"]
+    file_size = response["size"]
+    sha256 = response["sha256"]
+
+    response = requests.get(download_url, stream=True)
+
+    if path.is_file() and os.path.getsize(path) == file_size:
+        print(f"File already downloaded: {path}")
+        if _get_hash(path) == sha256:
+            return
+
+    with tqdm(total=file_size, unit="B", unit_scale=True) as progress_bar:
+        with open(path, "wb") as f:
+            for data in response.iter_content(1024):
+                progress_bar.update(len(data))
+                f.write(data)
+
+
+if __name__ == "__main__":
+    if not os.path.exists('data'):
+        path = Path('data.zip')
+        download(path, 'Oqxcid6uX58kYQ')
+        shutil.unpack_archive('data.zip', 'data', format="zip")
+        os.remove('data.zip')
+
+    p_value_bound = {'low': 0.0, 'up': 1.0}
+    q_value_bound = {'low': 1.0, 'up': 1.6}
+    problem = Cardio2D(p_value_bound, q_value_bound)
+    method_params = SolverParameters(r=np.double(3.0), iters_limit=10)
+    solver = Solver(problem, parameters=method_params)
+    cfol = ConsoleOutputListener(mode='full')
+    solver.add_listener(cfol)
+    solver_info = solver.solve()

examples/Machine_learning/NeuralNetwork/Segmentation/__init__.py

@@ -0,0 +1,73 @@
+import torch
+import numpy as np
+
+class SegmentationDataset(torch.utils.data.Dataset):
+    def __init__(self, device, paths, signals, masks=None, common_mask=False, for_train=False):
+
+        self._device = device
+        self._paths = paths
+        self._signals = [torch.Tensor(x).to(device) for x in signals]
+        self._masks = [torch.LongTensor(x).to(device) for x in masks]
+
+        self.begin_noise, self.end_noise = 1e-3, 3e-3
+        self.begin_ampl, self.end_ampl = 0, 0.3
+
+        self.begin_freq, self.end_freq = 0, 0.009
+
+        self.prob_isoline = 0.7
+        self.prob_reverse = 0.5
+        self.sub_len = 4000
+
+        self.common_mask = common_mask
+        self.for_train = for_train
+
+    def reverse_ecg(self, signal):
+        result = torch.zeros_like(signal, device=self._device)
+        for i, x in enumerate(signal):
+            sign = 2 * (np.random.rand() < self.prob_reverse) - 1
+            result[i] = sign * x
+        return result
+
+    def __len__(self):
+        return len(self._signals)
+
+    def __getitem__(self, i):
+        if not self.for_train:
+            return self._paths[i], self._signals[i], self.skip_borders(self._masks[i][0])
+
+        shift = np.random.randint(0, 5000 - self.sub_len - 1)
+        noise = self.begin_noise + (self.end_noise - self.begin_noise) * np.random.rand()
+        signal = self._signals[i][:, shift:shift + self.sub_len] + torch.normal(0, noise,
+                                                                                size=(self.sub_len,),
+                                                                                device=self._device)
+
+        signal = self.reverse_ecg(signal)
+
+        if self._masks is None:
+            return self._paths[i], signal
+
+        mask = self._masks[i][:, shift: shift + self.sub_len]
+        indexes = torch.randperm(12, device=self._device)
+
+        if self.common_mask:
+            mask = mask[0]
+        else:
+            mask = mask[indexes]
+
+        return self._paths[i], signal[indexes], self.skip_borders(mask)
+
+    def skip_borders(self, mask):
+        wave_start = torch.logical_and(torch.roll(mask, 1) == 0, mask != 0).type(torch.uint8)
+        wave_finish = torch.logical_and(torch.roll(mask, -1) == 0, mask != 0).type(torch.uint8)
+
+        indexes_starts, = torch.where(wave_start == 1)
+        indexes_finish, = torch.where(wave_finish == 1)
+
+        left_skip = indexes_starts[indexes_starts > 500][0]
+        right_skip = indexes_finish[indexes_finish < len(mask) - 500][-1]
+
+        mask_copy = torch.clone(mask)
+        mask_copy[:left_skip] = 4
+        mask_copy[right_skip:] = 4
+
+        return mask_copy