Incubator-mxnet: asnumpy() on cpu is 80 times slower than gpu

[djaym7]

mxnet 1.7.0 latest cu101 and cu101mkl aws p3.2x instance.

code pasted below to reproduce:
my results :

from mxnet.gluon.model_zoo.vision import get_model
from gluoncv.data.transforms.presets.ssd import transform_test
import mxnet as mx
import numpy as np
import gluoncv 
from mxnet.gluon.data.vision import transforms
import cv2
import time

from matplotlib.pyplot import imshow
import glob


categories = ['Person','non-Moving'] 




# model = self.load_model(best=True)
model = gluoncv.model_zoo.get_model('faster_rcnn_resnet50_v1b_voc')
#print(model)
model.load_parameters('/home/ec2-user/SageMaker/PICV/Segmentation Job/OD/params/faster_rcnn_finalparams.params')

img = mx.random.randn(1,3,608,608,ctx = mx.gpu())
model.collect_params().reset_ctx(mx.gpu())


model.hybridize()

#img,image = transform_test(mx.nd.array(img),short=420, max_size=640)


#warmup
for i in range(10):
    ids,scores,boxes = model(img)
    ids,scores,boxes = ids.asnumpy(),scores.asnumpy(),boxes.asnumpy()
    mx.nd.waitall()

%%time 
#this is very low time
# start_time5 = time.time()
ids,scores,boxes = model(img)
ids,scores,boxes = ids.asnumpy(),scores.asnumpy(),boxes.asnumpy()
mx.nd.waitall()
# end_time5 = time.time()-start_time5
# print("detection time :{:0.2f}".format(end_time5)) 

%%timeit 

ids,scores,boxes = model(img)
ids,scores,boxes = ids.asnumpy(),scores.asnumpy(),boxes.asnumpy()
mx.nd.waitall()


#cpu test
for i in range(10):
    start_time5 = time.time()
    ids,scores,boxes = model(img)
    ids,scores,boxes = ids.asnumpy(),scores.asnumpy(),boxes.asnumpy()
    mx.nd.waitall()
    print(time.time()-start_time5) 

model.collect_params().reset_ctx(mx.cpu())
img = img.as_in_context(mx.cpu())

#cpu test
for i in range(10):
    start_time5 = time.time()
    ids,scores,boxes = model(img)
    ids,scores,boxes = ids.asnumpy(),scores.asnumpy(),boxes.asnumpy()
    mx.nd.waitall()
    print(time.time()-start_time5) 

[samskalicky]

Cleaned up the code a bit:

import mxnet as mx
import numpy as np
import gluoncv 
import time

# model = self.load_model(best=True)
model = gluoncv.model_zoo.get_model('faster_rcnn_resnet50_v1b_voc',pretrained=True)

img = mx.random.randn(1,3,608,608,ctx = mx.gpu())
model.collect_params().reset_ctx(mx.gpu())
model.hybridize(static_shape=True, static_alloc=True)

#warmup
print('gpu warmup')
for i in range(10):
    ids,scores,boxes = model(img)
    ids,scores,boxes = ids.asnumpy(),scores.asnumpy(),boxes.asnumpy()
    mx.nd.waitall()

#gpu test    
for i in range(10):
    start_time5 = time.time()
    ids,scores,boxes = model(img)
    ids,scores,boxes = ids.asnumpy(),scores.asnumpy(),boxes.asnumpy()
    mx.nd.waitall()
    print('gpu: %4.3f seconds' % (time.time()-start_time5))

model.collect_params().reset_ctx(mx.cpu())
img = img.as_in_context(mx.cpu())

#warmup
print('cpu warmup')
for i in range(10):
    ids,scores,boxes = model(img)
    ids,scores,boxes = ids.asnumpy(),scores.asnumpy(),boxes.asnumpy()
    mx.nd.waitall()

#cpu test
for i in range(10):
    start_time5 = time.time()
    ids,scores,boxes = model(img)
    ids,scores,boxes = ids.asnumpy(),scores.asnumpy(),boxes.asnumpy()
    mx.nd.waitall()
    print('cpu: %4.3f seconds' % (time.time()-start_time5))

Heres the output I get on P3.2xl with mxnet-cu102 (version 1.7.0):

gpu warmup
[01:37:43] src/operator/nn/./cudnn/./cudnn_algoreg-inl.h:97: Running performance tests to find the best convolution algorithm, this can take a while... (set the environment variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)
gpu: 0.052 seconds
gpu: 0.052 seconds
gpu: 0.052 seconds
gpu: 0.052 seconds
gpu: 0.052 seconds
gpu: 0.052 seconds
gpu: 0.052 seconds
gpu: 0.052 seconds
gpu: 0.052 seconds
gpu: 0.052 seconds
cpu warmup
cpu: 2.827 seconds
cpu: 2.837 seconds
cpu: 2.829 seconds
cpu: 2.826 seconds
cpu: 2.828 seconds
cpu: 2.830 seconds
cpu: 2.826 seconds
cpu: 2.830 seconds
cpu: 2.832 seconds
cpu: 2.824 seconds

I think this performance is expected. This doesnt seem to be extraneous.