Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 4 years ago.
Improve this question
I'm evaluating the options of using HTM (hierarchical temporal memory) and CNN (convolutional neural network) for object recognition. Which architecture (model) would is most appropriate in this case?
Convolutional Neural Network and its variants are best tool for object recognition .
You can try with AlexNet,VGGNEt, ResNet, Batch Normalization , Dropout etc.
Always prefer using pretrained models and using transfer learning first in these cases. You can check out the implementation of Inception V3 etc. for object detection on tensorflow website and use them for transfer learning for your project.
Related
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 1 year ago.
Improve this question
This is a cross-post to my question in the Pytorch forum.
When using DistributedDataParallel (DDP) from PyTorch on only one node I expect it to be the same as a script without DistributedDataParallel.
I created a simple MNIST training setup with a three-layer feed forward neural network. It gives significantly lower accuracy (around 10%) if trained with the same hyperparameters, same epochs, and generally same code but the usage of the DDP library.
I created a GitHub repository demonstrating my problem.
I hope it is a usage error of the library, but I do not see how there is a problem, also colleges of mine did already audit the code. Also, I tried it on macOS with a CPU and on three different GPU/ubuntu combinations (one with a 1080-Ti, one with a 2080-Ti and a cluster with P100s inside) all giving the same results. Seeds are fixed for reproducibility.
You are using different batch sizes in your two experiments: batch_size=128, and batch_size=32 for mnist-distributed.py and mnist-plain.py respectively. This would indicate that you won't have the same performance result with those two trainings.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 1 year ago.
Improve this question
Do you have any code example or paper that refers to something like the following diagram?
I want to know why we want to stack multiple resnet blocks as opposed to multiple convolutional block as in more traditional architectures? Any code sample or referring to one will be really helpful.
Also, how can I transfer that to something like the following that can contain self-attention module for each resnet block?
Applying self-attention to the outputs of Resnet blocks at the very high resolution of the input image may lead to memory issues: The memory requirements of self-attention blocks grow quadratically with the input size (=resolution). This is why in, e.g., Xiaolong Wang, Ross Girshick, Abhinav Gupta, Kaiming He Non-Local Neural Networks (CVPR 2018) they introduced self-attention only at a very deep layer of the architecture, once the feature map was substantially sub-sampled.
Closed. This question needs details or clarity. It is not currently accepting answers.
Want to improve this question? Add details and clarify the problem by editing this post.
Closed 1 year ago.
Improve this question
my question is:
for example: i have a trained FCC and i want to implement it on Hardware(ASIC). i want to ask how to utilize weights and biases from trained model in verilog ?
Should i make RAM and then store the values in it, or is any other way to be used?
I need this values(weights and biases) to propagate them to MAC units.
The weights and biases need to be converted into specific number format (say Fixed Point) and then stored in RAM.
Then the values should be fetched and given to the MAC units.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 3 years ago.
Improve this question
I already trained a CNN for steering angle prediction with regression output. Also I trained CNN for classification (Two classes: road, no road). But I want to combine it. How to do it? Train a new model is not a problem.
You will need to update your y values to have both the category and the angle, and write a custom loss function that adds the regression loss and the category loss.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 5 years ago.
Improve this question
I have executed a program with image classification and it was running good .I was running the code with conv1D and conv2D . I am getting accuracy of 0.854 for both conv1D and conv2D.
Can i know the exact differences between these two things in detail?
Conv1d is a convolution filter of 1 dimension (imagine it like a one dimension array). Conv2d is a filter with 2 dimensions (like a 2d array) and it is more suitable for data like images where it can retain more spatial information in a data point because it is applied to all the neighbors. You can see what is a kernel to understand why this is better for data like images. For non image data I guess it will not have significant impact whether you use 1d or 2d convolution arrays.
Note: Also this site is for programming problems, maybe you should ask your question in Data Science