This repository has been archived by the owner on Nov 17, 2023. It is now read-only.
[Metric] Accelerate the calculation of F1 #9833
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
szha
reviewed
Feb 20, 2018
python/mxnet/metric.py
Outdated
| self.false_negatives += 1. | ||
| else: | ||
| self.true_negatives += 1. | ||
| self.true_positives += ((pred_label == 1) * (label == 1)).sum() |
There was a problem hiding this comment.
you can cache the computation such as predicted_true = pred_label == 1 and use later.
|
After some investigation, I find using GPU is not faster because the batch_size tested is rather small, i.e, 16, 64, 256, 1024. NDArray + GPU is not as fast as numpy in these cases. import mxnet as mx
import mxnet.ndarray as nd
import numpy as np
import time
# Warm up the GPU
for _ in range(10):
a = nd.ones((100, 100), ctx=mx.gpu())
b = a * 2
b.asnumpy()
N = 100
# Test the speed
for data_shape in [(16,), (64,), (256,), (1024,)]:
dat_npy = np.random.uniform(0, 1, data_shape)
dat_nd_gpu = nd.array(dat_npy, ctx=mx.gpu())
dat_nd_cpu = nd.array(dat_npy, ctx=mx.cpu())
nd.waitall()
start = time.time()
for _ in range(N):
np_ret = np.sum(dat_npy)
end = time.time()
np_time = end - start
start = time.time()
for _ in range(N):
nd_ret = nd.sum(dat_nd_gpu).asscalar()
end = time.time()
nd_gpu_time = end - start
start = time.time()
for _ in range(N):
nd_ret = nd.sum(dat_nd_cpu).asscalar()
end = time.time()
nd_cpu_time = end - start
print('sum, data_shape=%s, numpy time=%g, mxnet gpu time=%g, mxnet cpu time=%g' %(str(data_shape), np_time, nd_gpu_time, nd_cpu_time))Result |
|
Given that those are typical cases, would it make sense to always use cpu? |
|
Yes, I think CPU should be used. |
|
There's no wait in between the for loops. That should be addressed before calling the results conclusive. |
|
We have |
|
Ah, sorry I read the results wrong. Looks like numpy sum is an order of magnitude faster than either nd CPU or GPU. What does it look like if we take into account asnumpy()? |
|
Without import mxnet as mx
import mxnet.ndarray as nd
import numpy as np
import time
# Warm up the GPU
for _ in range(10):
a = nd.ones((100, 100), ctx=mx.gpu())
b = a * 2
b.asnumpy()
N = 100
# Test the speed
for data_shape in [(16,), (64,), (256,), (1024,)]:
dat_npy = np.random.uniform(0, 1, data_shape)
dat_nd_gpu = nd.array(dat_npy, ctx=mx.gpu())
dat_nd_cpu = nd.array(dat_npy, ctx=mx.cpu())
nd.waitall()
start = time.time()
for _ in range(N):
np_ret = np.sum(dat_npy)
end = time.time()
np_time = end - start
start = time.time()
for _ in range(N):
nd_ret = nd.sum(dat_nd_gpu)
nd.waitall()
end = time.time()
nd_gpu_time = end - start
start = time.time()
for _ in range(N):
nd_ret = nd.sum(dat_nd_cpu)
nd.waitall()
end = time.time()
nd_cpu_time = end - start
print('sum, data_shape=%s, numpy time=%g, mxnet gpu time=%g, mxnet cpu time=%g' %(str(data_shape), np_time, nd_gpu_time, nd_cpu_time))Result: |
|
Our implementations are much slower than numpy if the inputs are small. |
|
I've tested the case that calls import mxnet as mx
import mxnet.ndarray as nd
import numpy as np
import time
# Warm up the GPU
for _ in range(10):
a = nd.ones((100, 100), ctx=mx.gpu())
b = a * 2
b.asnumpy()
N = 100
# Test the speed
for data_shape in [(16,), (64,), (256,), (1024,)]:
dat_npy = np.random.uniform(0, 1, data_shape)
dat_nd_gpu = nd.array(dat_npy, ctx=mx.gpu())
dat_nd_cpu = nd.array(dat_npy, ctx=mx.cpu())
nd.waitall()
start = time.time()
for _ in range(N):
np_ret = np.sum(dat_nd_gpu.asnumpy())
end = time.time()
np_copy_from_gpu_time = end - start
start = time.time()
for _ in range(N):
np_ret = np.sum(dat_nd_cpu.asnumpy())
end = time.time()
np_copy_from_cpu_time = end - start
start = time.time()
for _ in range(N):
nd_ret = nd.sum(dat_nd_gpu).asscalar()
end = time.time()
nd_gpu_time = end - start
start = time.time()
for _ in range(N):
nd_ret = nd.sum(dat_nd_cpu).asscalar()
end = time.time()
nd_cpu_time = end - start
print('sum, data_shape=%s, numpy from gpu=%g, numpy from cpu=%g, mxnet gpu=%g, mxnet cpu=%g' %(str(data_shape), np_copy_from_gpu_time, np_copy_from_cpu_time, nd_gpu_time, nd_cpu_time))Result: |
rahul003
pushed a commit
to rahul003/mxnet
that referenced
this pull request
Jun 4, 2018
* Accelerate the calculation of F1 * cache the mid results * trigger CI
zheng-da
pushed a commit
to zheng-da/incubator-mxnet
that referenced
this pull request
Jun 28, 2018
* Accelerate the calculation of F1 * cache the mid results * trigger CI
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
Accelerate the calculation of F1 by removing the for-loop. I find that using GPU will in fact make the code slower 😅.
Checklist
Essentials
make lint)Changes
Results of running the script in #9705
Before:
After