Train ECG localization network#
[1]:
# Author: Ben Dai
# Licensed under the Apache License, Version 2.0 (the "License");
# Train ECG localization network
[1]:
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras.layers import Dense, Activation, Flatten, Convolution1D, Dropout,MaxPooling1D,GlobalAveragePooling1D
from tensorflow.keras import Model, layers,Sequential,regularizers
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.callbacks import EarlyStopping,ReduceLROnPlateau
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.callbacks import LearningRateScheduler
2022-10-14 21:52:36.322750: I tensorflow/core/util/util.cc:169] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
[2]:
## Load data and pretrained model
discriminator=tf.keras.models.load_model('./tests/ECG_model/pretrained_model.h5')
# discriminator.summary()
mit_train_path="./dataset/mitbih_train.csv"
mit_test_path="./dataset/mitbih_test.csv"
def create_pd(train_path,test_path):
train=pd.read_csv(train_path)
test=pd.read_csv(test_path)
train.columns=[x for x in range(188)]
test.columns=[x for x in range(188)]
return pd.concat([train,test], axis=0, join='inner').sort_index()
mit= create_pd(mit_train_path,mit_test_path)
X = np.asarray(mit.iloc[:,:187].values)
y = mit.iloc[:,187].values
y = to_categorical(y)
X = X.reshape(-1, 187, 1)
input_shape = X.shape[1:]
from sklearn.model_selection import train_test_split
X, X_test, y, y_test = train_test_split(X, y, test_size=0.33, random_state=42)
2022-10-14 21:52:39.025442: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:975] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-14 21:52:39.030579: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:975] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-14 21:52:39.030931: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:975] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-14 21:52:39.031751: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: AVX2 AVX512F AVX512_VNNI FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2022-10-14 21:52:39.032337: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:975] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-14 21:52:39.032622: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:975] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-14 21:52:39.032886: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:975] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-14 21:52:39.390903: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:975] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-14 21:52:39.391197: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:975] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-14 21:52:39.391438: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:975] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-14 21:52:39.391670: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1532] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 3647 MB memory: -> device: 0, name: NVIDIA GeForce RTX 2060, pci bus id: 0000:01:00.0, compute capability: 7.5
[3]:
## Model
from dnn_locate import loc_model
## define the backend localizer before TRELU activation
localizer_backend = tf.keras.Sequential(
[
layers.Input(shape=(input_shape[0], input_shape[1])),
layers.Conv1D(
filters=32, kernel_size=5, padding="same", strides=1, activation="relu"
),
layers.Dropout(rate=0.2),
layers.Conv1D(
filters=16, kernel_size=5, padding="same", strides=1, activation="relu"
),
layers.Conv1DTranspose(
filters=16, kernel_size=5, padding="same", strides=1, activation="relu"
),
layers.Dropout(rate=0.2),
layers.Conv1DTranspose(
filters=32, kernel_size=5, padding="same", strides=1, activation="relu"
),
layers.Conv1DTranspose(filters=1, kernel_size=5, padding="same"),
]
)
es_detect1 = ReduceLROnPlateau(monitor="loss", factor=0.382, min_lr=1e-6,
verbose=1, patience=3, mode="min")
es_detect2 = EarlyStopping(monitor='loss', mode='min', min_delta=.0001,
verbose=1, patience=10, restore_best_weights=True)
fit_params={'callbacks': [es_detect1, es_detect2],
'epochs': 100, 'batch_size': 64}
tau_range = [10., 20, 30]
## define framework
cue = loc_model(input_shape=input_shape,
localizer_backend=localizer_backend,
discriminator=discriminator,
target_r_square='auto',
r_metric='acc',
# r_metric='loss',
tau_range=tau_range)
/home/ben/tf/lib/python3.10/site-packages/keras/optimizers/optimizer_v2/gradient_descent.py:108: UserWarning: The `lr` argument is deprecated, use `learning_rate` instead.
super(SGD, self).__init__(name, **kwargs)
[4]:
cue.fit(X_train=X, y_train=y,
fit_params=fit_params,
optimizer=Adam(learning_rate=.01)
# optimizer=SGDW(learning_rate=.1, weight_decay=.0001, momentum=.9)
)
2022-10-14 21:52:44.663196: I tensorflow/stream_executor/cuda/cuda_dnn.cc:384] Loaded cuDNN version 8101
2292/2292 [==============================] - 4s 1ms/step - loss: 0.0483 - accuracy: 0.9861 - auc: 0.9990
2292/2292 [==============================] - 3s 1ms/step - loss: 5.1039 - accuracy: 0.7042 - auc: 0.8160
Epoch 1/100
1146/1146 [==============================] - 7s 5ms/step - loss: -5.3217 - lr: 0.0100
Epoch 2/100
1146/1146 [==============================] - 4s 4ms/step - loss: -7.3224 - lr: 0.0100
Epoch 3/100
1146/1146 [==============================] - 4s 4ms/step - loss: -7.9352 - lr: 0.0100
Epoch 4/100
1146/1146 [==============================] - 4s 4ms/step - loss: -8.1900 - lr: 0.0100
Epoch 5/100
1146/1146 [==============================] - 4s 4ms/step - loss: -8.1250 - lr: 0.0100
Epoch 6/100
1146/1146 [==============================] - 4s 4ms/step - loss: -8.2364 - lr: 0.0100
Epoch 7/100
1146/1146 [==============================] - 4s 4ms/step - loss: -8.4304 - lr: 0.0100
Epoch 8/100
1146/1146 [==============================] - 4s 4ms/step - loss: -8.7386 - lr: 0.0100
Epoch 9/100
1146/1146 [==============================] - 5s 4ms/step - loss: -8.6743 - lr: 0.0100
Epoch 10/100
1146/1146 [==============================] - 4s 4ms/step - loss: -8.7296 - lr: 0.0100
Epoch 11/100
1146/1146 [==============================] - 4s 4ms/step - loss: -8.8596 - lr: 0.0100
Epoch 12/100
1146/1146 [==============================] - 4s 4ms/step - loss: -8.9157 - lr: 0.0100
Epoch 13/100
1146/1146 [==============================] - 4s 4ms/step - loss: -8.7552 - lr: 0.0100
Epoch 14/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.0108 - lr: 0.0100
Epoch 15/100
1146/1146 [==============================] - 4s 4ms/step - loss: -8.7194 - lr: 0.0100
Epoch 16/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.1780 - lr: 0.0100
Epoch 17/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.0624 - lr: 0.0100
Epoch 18/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.2963 - lr: 0.0100
Epoch 19/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.2474 - lr: 0.0100
Epoch 20/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.2477 - lr: 0.0100
Epoch 21/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.3848 - lr: 0.0100
Epoch 22/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.4511 - lr: 0.0100
Epoch 23/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.5072 - lr: 0.0100
Epoch 24/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.4762 - lr: 0.0100
Epoch 25/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.5786 - lr: 0.0100
Epoch 26/100
1146/1146 [==============================] - 5s 4ms/step - loss: -9.6250 - lr: 0.0100
Epoch 27/100
1146/1146 [==============================] - 5s 4ms/step - loss: -9.3385 - lr: 0.0100
Epoch 28/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.3302 - lr: 0.0100
Epoch 29/100
1140/1146 [============================>.] - ETA: 0s - loss: -9.5342
Epoch 29: ReduceLROnPlateau reducing learning rate to 0.0038199999146163463.
1146/1146 [==============================] - 4s 4ms/step - loss: -9.5332 - lr: 0.0100
Epoch 30/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.8781 - lr: 0.0038
Epoch 31/100
1146/1146 [==============================] - 4s 4ms/step - loss: -9.9837 - lr: 0.0038
Epoch 32/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.1294 - lr: 0.0038
Epoch 33/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.1395 - lr: 0.0038
Epoch 34/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.1378 - lr: 0.0038
Epoch 35/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.0610 - lr: 0.0038
Epoch 36/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.1958 - lr: 0.0038
Epoch 37/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.2177 - lr: 0.0038
Epoch 38/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.1821 - lr: 0.0038
Epoch 39/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.2419 - lr: 0.0038
Epoch 40/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.1798 - lr: 0.0038
Epoch 41/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.2267 - lr: 0.0038
Epoch 42/100
1139/1146 [============================>.] - ETA: 0s - loss: -10.2399
Epoch 42: ReduceLROnPlateau reducing learning rate to 0.0014592399937100708.
1146/1146 [==============================] - 4s 4ms/step - loss: -10.2379 - lr: 0.0038
Epoch 43/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.2871 - lr: 0.0015
Epoch 44/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.4394 - lr: 0.0015
Epoch 45/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.4516 - lr: 0.0015
Epoch 46/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.4511 - lr: 0.0015
Epoch 47/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.4310 - lr: 0.0015
Epoch 48/100
1137/1146 [============================>.] - ETA: 0s - loss: -10.4450
Epoch 48: ReduceLROnPlateau reducing learning rate to 0.0005574296731501818.
1146/1146 [==============================] - 4s 4ms/step - loss: -10.4506 - lr: 0.0015
Epoch 49/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5040 - lr: 5.5743e-04
Epoch 50/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5365 - lr: 5.5743e-04
Epoch 51/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.4818 - lr: 5.5743e-04
Epoch 52/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.4441 - lr: 5.5743e-04
Epoch 53/100
1146/1146 [==============================] - ETA: 0s - loss: -10.4994
Epoch 53: ReduceLROnPlateau reducing learning rate to 0.00021293813816737384.
1146/1146 [==============================] - 4s 4ms/step - loss: -10.4994 - lr: 5.5743e-04
Epoch 54/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5423 - lr: 2.1294e-04
Epoch 55/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5513 - lr: 2.1294e-04
Epoch 56/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5156 - lr: 2.1294e-04
Epoch 57/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5144 - lr: 2.1294e-04
Epoch 58/100
1136/1146 [============================>.] - ETA: 0s - loss: -10.5268
Epoch 58: ReduceLROnPlateau reducing learning rate to 8.134236637852155e-05.
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5221 - lr: 2.1294e-04
Epoch 59/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5352 - lr: 8.1342e-05
Epoch 60/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5340 - lr: 8.1342e-05
Epoch 61/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5820 - lr: 8.1342e-05
Epoch 62/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5997 - lr: 8.1342e-05
Epoch 63/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5458 - lr: 8.1342e-05
Epoch 64/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5407 - lr: 8.1342e-05
Epoch 65/100
1145/1146 [============================>.] - ETA: 0s - loss: -10.5950
Epoch 65: ReduceLROnPlateau reducing learning rate to 3.107278476818465e-05.
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5960 - lr: 8.1342e-05
Epoch 66/100
1146/1146 [==============================] - 5s 4ms/step - loss: -10.5197 - lr: 3.1073e-05
Epoch 67/100
1146/1146 [==============================] - 5s 4ms/step - loss: -10.5252 - lr: 3.1073e-05
Epoch 68/100
1146/1146 [==============================] - 5s 4ms/step - loss: -10.6115 - lr: 3.1073e-05
Epoch 69/100
1146/1146 [==============================] - 5s 4ms/step - loss: -10.5617 - lr: 3.1073e-05
Epoch 70/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.6030 - lr: 3.1073e-05
Epoch 71/100
1146/1146 [==============================] - ETA: 0s - loss: -10.5464
Epoch 71: ReduceLROnPlateau reducing learning rate to 1.1869803725858219e-05.
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5464 - lr: 3.1073e-05
Epoch 72/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.6003 - lr: 1.1870e-05
Epoch 73/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5594 - lr: 1.1870e-05
Epoch 74/100
1133/1146 [============================>.] - ETA: 0s - loss: -10.5983
Epoch 74: ReduceLROnPlateau reducing learning rate to 4.534264948233613e-06.
1146/1146 [==============================] - 4s 4ms/step - loss: -10.6071 - lr: 1.1870e-05
Epoch 75/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5452 - lr: 4.5343e-06
Epoch 76/100
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5684 - lr: 4.5343e-06
Epoch 77/100
1143/1146 [============================>.] - ETA: 0s - loss: -10.5494
Epoch 77: ReduceLROnPlateau reducing learning rate to 1.7320891984127228e-06.
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5420 - lr: 4.5343e-06
Epoch 78/100
1137/1146 [============================>.] - ETA: 0s - loss: -10.5708Restoring model weights from the end of the best epoch: 68.
1146/1146 [==============================] - 4s 4ms/step - loss: -10.5738 - lr: 1.7321e-06
Epoch 78: early stopping
##################################################
compute the R2 for the fitted localizer.
##################################################
2292/2292 [==============================] - 3s 1ms/step - loss: 0.0483 - accuracy: 0.9861 - auc: 0.9990
2292/2292 [==============================] - 2s 945us/step
2292/2292 [==============================] - 3s 1ms/step - loss: 11.0031 - accuracy: 0.1729 - auc: 0.4940
1/1 [==============================] - 0s 92ms/step
early stop in tau = 10.000, R2: 0.983; target R2: 0.953 is reached
[5]:
## print R_square for Test set
cue.R_square(X_test, y_test)
##################################################
compute the R2 for the fitted localizer.
##################################################
1129/1129 [==============================] - 2s 2ms/step - loss: 0.0455 - accuracy: 0.9862 - auc: 0.9992
1129/1129 [==============================] - 1s 939us/step
1129/1129 [==============================] - 2s 1ms/step - loss: 11.0295 - accuracy: 0.1713 - auc: 0.4927
[5]:
0.9833611752340135
[6]:
## Plot the localization results by the fitted network for novel instances
import seaborn as sns
import matplotlib
import matplotlib.pyplot as plt
n_label = y.shape[1]
cmap = ["Oranges", "Purples", "Reds", "Blues", "Greens"]
c1 = ['darkorange', 'darkslateblue', 'darkred', 'darkblue', 'darkgreen']
c2 = ['darkgreen', 'darkred', 'darkslateblue', 'darkred', 'darkorange']
n_demo = 3
timepoint = list(range(input_shape[0]))
for k in range(n_label):
demo_ind = np.array([np.random.choice(np.where(y_test[:,k] == 1)[0]) for i in range(n_demo)])
X_demo = X_test[demo_ind]
X_demo_detect = cue.localizer.predict(X_demo)
X_demo_hl = cue.locate(X_demo)
sns.set_theme(style= 'white', palette=None)
for i in range(len(X_demo)):
X_tmp, X_detect_tmp, X_hl_tmp = X_demo[i], X_demo_detect[i], X_demo_hl[i]
plt.figure(figsize=(16, 8), dpi=80)
plt.title('detect results for a random sample from class %s' %k)
plt.imshow(X_hl_tmp[np.newaxis,:], cmap=cmap[k], aspect='auto', alpha=0.3,
extent = (0, 187, 0, 1))
plt.colorbar()
plt.plot(timepoint, X_tmp, linewidth=2.5, alpha=.7, color=c1[k],
label='Extracted ECG Beat')
plt.plot(timepoint, X_detect_tmp, linewidth=1.5, alpha=.7, color=c2[k], linestyle='--',
label='Extracted ECG Beat (AFTER removing detected feats)')
plt.legend(loc='best')
plt.tight_layout()
plt.show()
1/1 [==============================] - 0s 101ms/step
1/1 [==============================] - 0s 12ms/step
1/1 [==============================] - 0s 13ms/step
1/1 [==============================] - 0s 12ms/step
1/1 [==============================] - 0s 13ms/step
1/1 [==============================] - 0s 12ms/step
1/1 [==============================] - 0s 14ms/step
1/1 [==============================] - 0s 12ms/step
1/1 [==============================] - 0s 12ms/step
1/1 [==============================] - 0s 12ms/step
