Titan rtx gpu

For this study, all the CNN architectures were implemented using the Keras framework [35 ] with TensorFlow [36 ] as back-end. We performed our experiment on two powerful NVIDIA Titan RTX GPUs with 4608 CUDA cores and 24 GB GDDR6 SDRAM. The batch size for this experiment was set to 2 for both models and the proposed architecture was optimized with the Adam optimizer. The learning rate to train the model was set to 1 × 10⁻⁵. To reduce the training time and use the GPU efficiently, we use 10 to 20 percent of the 3D CBCT scans for training of model for jaw localization as well as canal segmentation. The size of images while training the jaw localization model is kept to 128 × 128 × 128. After localizing the jaw, the 3D images are cropped and resized to three fixed sizes as mentioned in Section 3.4.2. We used the dice loss function Equation (1) to calculate the loss. Labels are the segmentation annotation of images containing 0 as background and 1 as foreground. We trained the localization model for 50 epochs and the 3D segmentation models for 80 epochs by keeping the learning rate lower in order to train a generalized model. The batch size, epoch and learning rate were reset depending upon the need.
$$\begin{array}{cc}\hfill \mathrm{Dice}\mathrm{Loss}& \hfill =1-\mathrm{Dice}\mathrm{Coefficient}\hfill \end{array}$$
where, dice coefficient is given by Equation (6).

All experiments are conducted on the PyTorch platform with two NVIDIA TITAN RTX GPUs (24GB). We use the ADAM optimizer to optimize all networks. The initial learning rates of the whole-breast and tumor segmentation models are 0.002 and 0.001, respectively. And, the learning rate decays by half for every 50 epochs. A total number of 300 epochs are set for each task. We compute the training loss within 10 epochs to determine the convergence. While using the well-trained segmentation models to test the results, we use sliding windows to crop the overlapping patches, whose stride is half of the patch size. Then, we average the overlapping patches to obtain the final results.

All of the experiments were conducted on a machine with 48 cores Intel Xeon Platinum (Cooper Lake) 8369 processor and 192 GB memory. For training the model, the PyTorch framework was utilized with 4 Nvidia Titan RTX GPUs.

For training and testing, we used the PyTorch⁵¹ deep-learning framework on 8 × NVIDIA TITAN RTX GPUs. The Adam optimizer^{41 (link)} with a weight decay of 0.0001 was used to train the CT-Nets. The initial learning rate was set at 0.0005, and the learning rate decayed by a factor of 10 after the 35th, 40th, and 43rd epochs. All models were trained for 45 epochs. Owing to the restricted GPU memory, the batch sizes on each GPU were set to 16 for the abnormality model and 8 for the disease model.
To train CXR-Nets, an Adam optimizer with a weight decay of 0.0001 was used. The initial learning rate was set at 0.0005, and the learning rate decayed by a factor of ten after the 25th and 35th epochs. All models were trained for 45 epochs. Owing to the restricted GPU capacity, the batch sizes on each GPU were adjusted to 128 for the abnormality model and 64 for the disease model.