TensorFlow | TensorFlow tutorials from syntax basics | Learning library

This can be achieved by weight sharing or shared layers. To share layers in different models in keras, you just need to pass the same instance of layer to both of the models.

Example Codes:

I fixed my issue.

Here are the steps of solving that issue.

1 ) Download CUDA from https://developer.nvidia.com/cuda-downloads

2 ) Download CUDNN from https://developer.nvidia.com/rdp/cudnn-download

3 ) Copy bin,include and lastly lib from CUDNN zip file and paste it C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA{version}

4 ) Then run the .py code in PyCharm and it perceives GPU at last.

There are multiple potential places to measure execution time, first might be defining very specifically what is the intended measurement.

1. Measuring the training time of each client as proposed is a great way to get a sense of the variability among clients. This could help identify whether rounds frequently have stragglers. Using tf.timestamp() at the beginning and end of the client_update function seems reasonable. The question correctly notes that this happens in parallel, summing all of these times would be akin to CPU time.

2. Measuring the time it takes to complete all client training in a round would generally be the maximum of the values above. This might not be true when simulating FL in TFF, as TFF maybe decided to run some number of clients sequentially due to system resources constraints. In practice all of these clients would run in parallel.

3. Measuring the time it takes to complete a full round (the maximum time it takes to run a client, plus the time it takes for the server to update) could be done by moving the tf.timestamp calls to the outer training loop. This would be wrapping the call to trainer.next() in the snippet on https://www.tensorflow.org/federated. This would be most similar to elapsed real time (wall clock time).

I had the same problem, man. I want to transplant tflite to the development board of CEVA. There is no problem in compiling. In the process of running, there is also an error in AddBuiltin(full_connect). At present, the only possible situation I guess is that some devices can not support tflite.

Generally, if there are any incompatibility between TF, CUDA and cuDNN version you can observed this behavior.

For GeForce RTX 3060, support starts from CUDA 11.x. Once you upgrade to TF2.4 or TF2.5 your issue will be resolved.

For the benefit of community providing tested built configuration

CUDA Support Matrix

I would suggest you to use conda (Ananconda/Miniconda) to create a separate environment and install tensorflow-gpu, cudnn and cudatoolkit. Miniconda has a much smaller footprint than Anaconda. I would suggest you to install Miniconda if you do not have conda already.

In your problem you are trying to use Sequential API to create the Model. There are Limitations to Sequential API, you can just create a layer by layer model. It can't handle multiple inputs/outputs. It can't be used for Branching also.

Below is the text from Keras official website: https://keras.io/guides/functional_api/

The functional API makes it easy to manipulate multiple inputs and outputs. This cannot be handled with the Sequential API.

Also this stack link will be useful for you: Keras' Sequential vs Functional API for Multi-Task Learning Neural Network

Now you can create a Model using Functional API or Model Sub Classing.

In case of functional API Your Model will be

Assuming Output_1 is classification with 17 classes Output_2 is calssification with 2 classes and Output_3 is regression

I think when you are using the conda environment. The conda list is going to show all the general packages that shared by the same conda environment. And the reason why 'tensorflow-gpu' is listed in 'pip freeze', but not in 'conda list', is because you used pip install to installed 'tensorflow-gpu'(could be you or the IDE). In this case, 'tensorflow-gpu' is only exists under this python project I believe. Actually, there is an official document about this topic.

Issues may arise when using pip and conda together. When combining conda and pip, it is best to use an isolated conda environment. Only after conda has been used to install as many packages as possible should pip be used to install any remaining software. If modifications are needed to the environment, it is best to create a new environment rather than running conda after pip. When appropriate, conda and pip requirements should be stored in text files.

Use pip only after conda Install as many requirements as possible with conda then use pip.

Pip should be run with --upgrade-strategy only-if-needed (the default).

Do not use pip with the --user argument, avoid all users installs.

And here is the link.

for MDN model , the likelihood for each sample has to be calculated with all the Gaussians pdf , to do that I think you have to reshape your matrices ( y_true and mu) and take advantage of the broadcasting operation by adding 1 as the last dimension . e.g: