Obs window size

[SatriaPriambada]

1. Change the frequency from 125Hz to 250Hz
   We change the firing frequency according to dataset.
   This means prediction will happen after 30 sec * 250Hz = 7500 request

2. Change the server to accomodate obs_w_30sec.
   obs_w_30sec is equal to batch size of waveform with observation window of 30 sec.
   For example if we want to do profile_ensemble of 5 minutes of waveform we will send tensor in size of (obs_w_30sec, 1, 7500).

[alindkhare]

why obs_window has to be per query basis? Can it part of http server instantiation?

[SatriaPriambada]

I assume that when profile client make request it can change the obs_window.
If we put it as part of http server instantiation then the client will always have to send the same obs_window.
For example, when the patients are on stable condition client will only look at interval of obs_window 30s, however when the patients are critical condition, the client can send obs_window 30s, 2min, 5 min, etc

[alindkhare]

Yes the client has to be consistent in sending the information. Any reasons for the assumption you are making?

[alindkhare]

I think we need one more profiling :

@ray.remote
def echo_tensor(a : torch.Tensor): 
	return a
>>> %timeit ray.get(echo_tensor.remote(torch.zeros((obs_w_30sec, 1, PREDITICATE_INTERVAL))))

[SatriaPriambada]

This is the result from experimenting with ray profiling:
For context there are two function tested
Normal function means we input a tensor with ray remote and return the tensor back

@ray.remote
def echo_tensor(a : torch.Tensor):
   return a

Simple function means we input a tensor with ray remote and return a boolean whether serialization finished or not

@ray.remote
def echo_tensor_simple(a : torch.Tensor):
	return 1

The results are performed by doing the following:
for obs_w_30sec in range(1,11)

>>> PREDITICATE_INTERVAL = 7500
>>> %timeit ray.get(echo_tensor.remote(torch.zeros((obs_w_30sec, 1, PREDITICATE_INTERVAL))))
>>> %timeit ray.get(echo_tensor_simple.remote(torch.zeros((obs_w_30sec, 1, PREDITICATE_INTERVAL))))

[SatriaPriambada]

This is the latency profile for different observation_window_30sec
Based on my implementation, each observation window will be treated like a batch_size
The experiment uses the new tensor size of 30 sec * 250Hz = 7500

• timeit time means the time taken for %timeit to run the tensor with batch_size
• Tq time means the maximum expected time for request to wait in the queue before profiled
• Ts time means the time request get executed from ensemble
• Total time = Tq + Ts

[SatriaPriambada]

This is the latency profile for different observation_window_30sec after the fix of calculate throughput

The experiment uses tensor size of 30 sec * 250Hz = 7500
Number of client is 1 patient
Model are ResNet1D with base_filters = 128, kernel_size = 16, n_block = 8

• timeit time means the time taken for %timeit to run the tensor with batch_size
• Tq time means the maximum expected time for request to wait in the queue before profiled
• Ts time means the time request get executed from ensemble
• Total time = Tq + Ts