Titan x

In our proposed model, there are two layers of LSTM, with 128 hidden nodes for each phase. We apply dropout layer in the end of decoding phase with a rate of 0.2 to prevent overfitting. Our model’s learning rate is set to 0.0025 and uses Adam optimizer for the training process. The maximal epochs for the training are set to 50 for both source field and target field. We built our model in Python using Keras with Tensorflow 2.2 as the backend. Using four GPU machines (NVIDIA Titan X from Taipei, Taiwan), it takes up to 9 h to train both the PPG to PPG translation model and the PPG to ABP translation model.

All trainings ran on a GPU workstation with 16 GB RAM, Intel(R) Xeon(R) CPU E3-1231 v3 @ 3.40GHz and a NVidia TITAN X (Pascal) GPU with 12 GB VRAM.

Model parameters consumed up to 1 GB depending on the number of included components. The training was performed on the Entropy computation cluster (Faculty of Mathematics, Informatics and Mechanics, University of Warsaw, Warsaw, Poland) with the following GPU hardware: RTX 2080 Ti (MSI, Taipei, Taiwan), TITAN V, and TITAN X (NVIDIA, Santa Clara, CA, USA). The performance of these GPUs was enough to run the model.
The categorical cross-entropy cost function was minimised by Adam optimiser [54 ].

The proposed models were implemented using PyTorch version 1.8.0 [29 ], with NVIDIA CUDA version 10.0 (Santa Clara, CA, USA), Single GPU NVIDIA TITAN X and 64 GB RAM. For our algorithm approximately 6 GB should be enough depending on the batch size. The models were trained and evaluated on Linux (Ubuntu) operating system (also compatible with Windows) and were coded in Python 3.9. The performance measurement and statistical analysis were performed using the publicly available libraries including Scikit-learn version 0.20.3.

Our model is a Fully Connected Deep Neural Network (DNN) [39 (link)] having an input layer (of 13,654 dimensions), 18 hidden layers (of 512 dimensions each) and a scalar output layer. We employ the logistic function and log loss at the output layer for binary classification (with 0/1 labels), and use the Scaled Exponential Linear Unit (SeLU) activation function [40 ] at each layer. The model is optimized using the Adam optimizer [41 ], with a mini-batch size of 128 examples. The default learning rate was used (0.001).
Intermediate model snapshots of the model weights were taken every 250 mini-batch iterations, and the snapshot that performed best on the validation test was retroactively selected as the final model. Explicit regularization was not found necessary. The network configuration was reached by extensive hyperparameter search over various network depths (ranging from 2 to 32) and activation functions (tanh, ReLU and SeLU).
The software was implemented using the Python programming language (version 2.7), PyTorch framework [42 ], and the scikit-learn library (version 0.17.1) [43 ]. The training was performed on an NVIDIA TitanX (12 GB RAM) with CUDA version 8.0.

All neural network models were trained in the open-source package PyTorch⁷¹ on four NVIDIA GTX 1080 or Titan X graphics processing units. Our optimized model used a simple convolutional architecture for image classification, consisting of alternating (3 × 3) kernels of stride 1 and padding 1 followed by max pooling (Fig. 3a), followed by two fully connected hidden layers (512 and 100 neurons) and rectified linear units as the nonlinear activation function. All neural network models were trained using backpropagation. The optimized training procedure used the Adam⁷² optimizer with a multi-label soft margin loss function with weight decay (L2 penalty, 0.008) and dropout (probability 0.5 for the first two fully connected layers and probability 0.2 for all convolutional layers). Training proceeded with mini-batches of 64 images with real-time data augmentation including random flips, rotations, zoom, shear, and color jitter. When calculating the classification accuracy, a threshold of 0.91, 0.1, and 0.85 was used for cored plaque, diffuse plaque, and CAA prediction, respectively. Predictions with confidence above the threshold were considered to be positives.

We perform all experiments with TensorFlow 2.4 on an NVIDIA Titan Xp, except for batch sizes of more than 64 total images, where we use an additional NVIDIA Titan X. The CNN architecture uses a batch size of 72 images. We train the LSTM with a batch size of 8 sequences, each sequence holding 9 images, which corresponds to 72 images per iteration. Increasing the batch size in either of the above architectures yields negligible improvements. Nevertheless, using larger batch sizes for the BLSTM shows more significant improvements. Therefore, in the end, the BLSTM is trained with a batch size of 16 sequences, each holding 9 images, for a total of 144 images per batch. It takes approximately 25 min to train the model. This corresponds to 4500 iterations. It iterates through 4500 × batch_size sequences in total. This corresponds to about 16 epochs for the BLSTM and 8 for the CNN and LSTM architectures. Stochastic gradient descent without momentum is used to train the network. In terms of data augmentation, we use colour jittering, which includes adjustments to brightness, contrast, saturation, and hue, random gray-scale conversion, rotations, random left-right as well as up-down flips, and finally introduce a mask with a radius of 128px to remove any possible artefacts near the borders.