Embedded Files
In this section we will discuss about the below things:
In this section we will discuss about the below things:
1) Lab cluster setup, which is used to run multiple experiments.
2) Web based interface for both between graph replication and in graph replication.
3) And later We will also present the image pre-processing capabilities of tensorFlow.
The Distributed cluster set up in the lab for testing, details of which are presented below:
The Distributed cluster set up in the lab for testing, details of which are presented below:
In this cluster tensorflow is running on 5 different machine.
In this cluster tensorflow is running on 5 different machine.
1) Machine 1: Raspberry pi 3 (4×1.2GHz +1GB RAM)
2) Machine 2: Linux (8×3.4GHz +15.6GB RAM)
3) Machine 3: Mac (4×2.8GHz +8GB RAM)
4) Machine 4: Raspberry pi 3 (4×1.2GHz +1GB RAM)
5) Machine 5: Raspberry pi 3 (4×1.2GHz +1GB RAM)
Consider the scenario for machine 4. Machine 4 has a camera attached to it. This camera is used to take the picture which can be used as input to the model. We have developed a demo which schedules the different jobs on different machine. The purpose of the camera is to serve as input to one of the device, which can decide at runtime how to schedule different tasks on different available devices. An important thing to note here is that, in this cluster different devices have different processing powers and are also available at different places in the network. Some of the devices are near, whereas some of the devices are far.
Consider the scenario for machine 4. Machine 4 has a camera attached to it. This camera is used to take the picture which can be used as input to the model. We have developed a demo which schedules the different jobs on different machine. The purpose of the camera is to serve as input to one of the device, which can decide at runtime how to schedule different tasks on different available devices. An important thing to note here is that, in this cluster different devices have different processing powers and are also available at different places in the network. Some of the devices are near, whereas some of the devices are far.
Machine 4, has machine 2, machine 3 and machine 5 which are available at one hop distance whereas the machine 1 is available at 2 hop distance. Different scheduling policies has different effect on the running timings.
Machine 4, has machine 2, machine 3 and machine 5 which are available at one hop distance whereas the machine 1 is available at 2 hop distance. Different scheduling policies has different effect on the running timings.
1. Scheduling Tensorflow Computations
1. Scheduling Tensorflow Computations
We have developed a web based interface and server allows user to select the number of computation it want to schedule on different devices and then schedules thoses computations on the devices and report back the result to the user. The details about the computations, sceduling and results is discussed in the subsequent sections.
We have developed a web based interface and server allows user to select the number of computation it want to schedule on different devices and then schedules thoses computations on the devices and report back the result to the user. The details about the computations, sceduling and results is discussed in the subsequent sections.
2. Web Based Interface for : In-graph replication
2. Web Based Interface for : In-graph replication
The web based interface is shown above. The results returned by the running the job are shown below.
The web based interface is shown above. The results returned by the running the job are shown below.
In the web based interface, we have created 15 jobs. The details of the job assignment is given in the input box. 1 job on machine 1, 2 jobs on machine 2 , and so-on. As seen from the output, Machine 1 has local IP 192.168.2.4, and so-on.
In the web based interface, we have created 15 jobs. The details of the job assignment is given in the input box. 1 job on machine 1, 2 jobs on machine 2 , and so-on. As seen from the output, Machine 1 has local IP 192.168.2.4, and so-on.
3. Details About Computations
3. Details About Computations
Every computation running on the cluster is scheduled by the machine 4. Machine 4 is serving as the client which has input source image from the camera. Each computation it schedules on the cluster is a softmax regression to recognize the image. A trained model build using the MNIST dataset is used to recognize the image at runtime. From the user input from the web based interface, it decides how many softmax computations it should schedule on the cluster and on which machines. For example in our current assignment, we are scheduling 15 different softmax computations on the clusters. The total time to run these computations on the cluster is 6.42 seconds.
Every computation running on the cluster is scheduled by the machine 4. Machine 4 is serving as the client which has input source image from the camera. Each computation it schedules on the cluster is a softmax regression to recognize the image. A trained model build using the MNIST dataset is used to recognize the image at runtime. From the user input from the web based interface, it decides how many softmax computations it should schedule on the cluster and on which machines. For example in our current assignment, we are scheduling 15 different softmax computations on the clusters. The total time to run these computations on the cluster is 6.42 seconds.
The Computation Graph schedule on each device is shown by figure in the right. The computation graph is shown below.
The Computation Graph schedule on each device is shown by figure in the right. The computation graph is shown below.
4. Web Based Interface for : Between-graph replication
4. Web Based Interface for : Between-graph replication
The node naming is shown above in the graph, which is tested for the scheduing different nodes on different devices.
The node naming is shown above in the graph, which is tested for the scheduing different nodes on different devices.
5. Image Preprocessing in TensorFlow
5. Image Preprocessing in TensorFlow
During our implementation, we move all the data preprocessing work into TensorFlow framework. No other outer library need for our system.
During our implementation, we move all the data preprocessing work into TensorFlow framework. No other outer library need for our system.
Actually TensorFlow support only a few image processing operation, like Encoding/Decoding, Resizing/Cropping and image adjustments. So we also use the TF math operation to build our own preprocessing operations, e.g. the image binarization.
Actually TensorFlow support only a few image processing operation, like Encoding/Decoding, Resizing/Cropping and image adjustments. So we also use the TF math operation to build our own preprocessing operations, e.g. the image binarization.
Feature of Data Preprocessing inside TensorFlow:
Feature of Data Preprocessing inside TensorFlow:
—Everything represented in Tensors
—Everything represented in Tensors
(symbolic abstract representation, hard to interact)
(symbolic abstract representation, hard to interact)
—Can cause additional overhead.
—Can cause additional overhead.
Google Sites
Report abuse