When building the U-Net, the degrees of freedom (or hyperparameters) that determine the capacity of the network are (1) the number of resolution levels in the U-Net, (2) the number of feature maps in the first layer, (3) the input image size, and (4) the training batch size. We optimized these hyperparameters using subsets of training data from the two datasets used in this paper. We observed that the input image size was the most important parameter, and that by using larger input patches we obtained better results and faster convergence of the training and validation losses. With this, we found an optimal U-Net of 5 levels, 16 feature maps in the first layer, input size of , and training batches containing only one image. This network can fit in a graphical card GPU NVIDIA Quadro P6000 with 24 GB memory during training. This U-Net construction is used for all the experiments undertaken in this paper. The implementation of the network is done using the Pytorch framework39 .
Quadro p6000
Manufactured by NVIDIA
The NVIDIA Quadro P6000 is a high-performance professional graphics card designed for demanding visual computing applications. It features 3,840 CUDA cores, 24GB of GDDR5X memory, and a maximum single-precision floating-point performance of 12 teraflops. The Quadro P6000 is optimized for professional visualization, scientific computing, and advanced rendering workloads.
Automatically generated - may contain errors
Lab products found in correlation
Titan rtx gpu, by NVIDIA (2 mentions)
Quadro m5000, by NVIDIA (2 mentions)
Titan vs, by NVIDIA (2 mentions)
Tesla v100, by NVIDIA (2 mentions)
Geforce rtx 2080 ti, by NVIDIA (2 mentions)
Xeon e5 2620 v4, by NVIDIA (1 mentions)
Titan xp, by NVIDIA (1 mentions)
E5 2697, by NVIDIA (1 mentions)
Titan x, by NVIDIA (1 mentions)
19 protocols using quadro p6000
1
Optimizing U-Net Hyperparameters for Efficient Training
2
Large-Scale 3D Imaging Data Processing
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
The data were automatically transferred every day from the acquisition computer to a Lustre server for storage. The processing with ClearMap was done on local workstations, either Dell Precision T7920 or HP Z840. Each workstation was equipped with 2 Intel Xeon Gold 6128 3.4G 6C/12T CPUs, 512Gb of 2666MHz DDR4 RAM, 4x1Tb NVMe Class 40 Solid State Drives in a RAID0 array (plus a separate system disk), and an NVIDIA Quadro P6000, 24Gb VRAM video card. The workstations were operated by Linux Ubuntu 20.04LTS. ClearMap 2.0 was used on Anaconda Python 3.7 environment.
3
Efficient Deep Learning Model Training
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
We have trained our models running on a single NVidia Quadro P6000 with stochastic gradient descent with momentum in pytorch. PyTorch is a handy deep learning library that extends Python. The training step used momentum with a decay of 0.98, a learning rate of 0.002, and decayed every epoch using an exponential rate of 0.97. We also used a mini-batch size of 128 samples and trained the model for 100 iterations. Each iteration taked about one minute. The well trained model size was about 12.5 megabyte, and the number of parameters was 3,077,382.
4
Jet Image Classification and Probability Recovery
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
In this section, we present in detail the methodology used to obtain the presented results. First, we describe the jet generator used to crate the jet images. Next, we present the detailed architecture of the neural network used as the classifier and finally, we detail the algorithm used to recover of the underlying probability distributions.
The computational code used to develop the particle generator, the neural network model and the calculation of the probability distributions is written in the the Python programming language using the Keras module with the TensorFlow backend [11 ]. Both the classifier training and jet generating were performed using a standardized PC setup equipped with an NVIDIA Quadro p6000 graphics processing unit.
The computational code used to develop the particle generator, the neural network model and the calculation of the probability distributions is written in the the Python programming language using the Keras module with the TensorFlow backend [11 ]. Both the classifier training and jet generating were performed using a standardized PC setup equipped with an NVIDIA Quadro p6000 graphics processing unit.
5
DeepHL: Deep Learning Trajectory Analysis
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
The DeepHL system consists of three server computers. The first one is a web server that receives a trajectory data file from a user and provides analysis results to the user (Intel Xeon E5-2620 v4, 16 cores, 32 GB RAM, Ubuntu 14.04). The second one is a storage server that stores data files and analysis results. The third one is a GPU server that analyzes data provided by the user (Intel Xeon E5-2620 v4, 32 cores, 512 GB RAM, four NVIDIA Quadro P6000, Ubuntu 14.04). Supplementary Information , Algorithm, provides a complete description of the DeepHL method. DeepHL is accessible on the Internet through http://www-mmde.ist.osaka-u.ac.jp/maekawa/deephl/ . Supplementary Information , User guide to DeepHL, provides a user guide to DeepHL. In addition, Supplementary Information , Usage of Python-based Software, and Supplementary Software 1 present the Python code of DeepHL.
6
High-Speed Microscope Image Acquisition
7
Robust Plant Segmentation with PlantU-net
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
The PlantU-net was trained using the Keras framework (Fig 1 ) with acceleration from GPUs (NVIDIA Quadro P6000). Five hundred and twelve images were used to train the model. Data expansion is the key to making the network have the required invariance and robustness because this model uses a small number of samples for training. For top-view images of maize shoots, PlantU-net needs to meet the robustness of plant morphology changes and value changes of gray images. Increasing the random elastic deformation of training samples is the key to training segmentation networks with a small number of labeled images. Therefore, during the data reading phase, PlantU-net uses a random displacement vector on the 3 × 3 grid to generate a smooth deformation, where the displacement comes from a Gaussian distribution with a standard deviation of 10 pixels. Because the number of training samples is small, the dropout layer is added to prevent the network from overfitting. Through these "data enhancement" methods, the model performance is improved and overfitting is avoided. In each epoch, the batch size was 1, the initial learning rate was 0.0001, and adam is used as an optimizer to quickly converge the model. PlantU-net was trained until the model converged (the training loss was satisfied and remained nearly unchanged).
8
Statistical Analysis of Experimental Data
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
All data analyses, model training, and experiments were run on a standard workstation (64 GB RAM, 3.70 GHz Intel Core i9 CPU, NVidia Quadro P6000, 56 GB VRAM). The SPSS software (20.0, IBM Corporation, USA) was used to perform statistical analyses. First, a Kolmogorov–Smirnov one‐sample test was used to determine whether the data were normally distributed. The data with normal distribution were expressed as mean differences ± standard deviation (x ¯ ± s
9
Neural Network Particle Generator
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
The code used for the development the particle generator, the neural network models, and the calculations is written in the the Python programming language using the Keras module with the TensorFlow2 backend [5 ] and Numpy modules. The calculations were performed using a standardized PC setup equipped with an NVIDIA Quadro P6000 graphics processing unit.
10
Optimizing 2D CNN and 3D V-Net for Brain Tumor Segmentation
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
Learning rate, kernel size, and network depth were considered for hyper parameter tuning. We varied the learning rate and tested a variable learning rate (cosine annealing) for each network. For 2D CNNs, our experiments included testing 3 × 3 and 5 × 5 kernels. Neither a kernel of 5 × 5 nor an increase in depth from 6 to 7 lead to significant performance gains. We note that almost 90% of the coronal and sagittal slices do not contain tumors; thus, in order to avoid converging to null predictions, we rebalanced the dataset so that approximately 10% of slices did not contain tumors (98,000 training slices).
2D networks were trained on 2 Nvidia Quadro P6000 graphical processing units using the RMSProp optimizer, 25 epochs, and a batch size of 16. We set the learning rate at 10−5 for 13 epochs and divided by 2 after every 3 epochs. The V-Nets were trained using the Adam [21 ] optimizer for 100 epochs with a batch size of 4. The learning rate was set at 10−4 for 50 epochs, 10−4/2 for 25 epochs, and 10−4/4 for 25 epochs.
2D networks were trained on 2 Nvidia Quadro P6000 graphical processing units using the RMSProp optimizer, 25 epochs, and a batch size of 16. We set the learning rate at 10−5 for 13 epochs and divided by 2 after every 3 epochs. The V-Nets were trained using the Adam [21 ] optimizer for 100 epochs with a batch size of 4. The learning rate was set at 10−4 for 50 epochs, 10−4/2 for 25 epochs, and 10−4/4 for 25 epochs.
About PubCompare
Our mission is to provide scientists with the largest repository of trustworthy protocols and intelligent analytical tools, thereby offering them extensive information to design robust protocols aimed at minimizing the risk of failures.
We believe that the most crucial aspect is to grant scientists access to a wide range of reliable sources and new useful tools that surpass human capabilities.
However, we trust in allowing scientists to determine how to construct their own protocols based on this information, as they are the experts in their field.
Ready to get started?
Sign up for free.
Registration takes 20 seconds.
Available from any computer
No download required
Revolutionizing how scientists
search and build protocols!