mxnet - docker performance results

bulk.py

import argparse, os, cv2
import mxnet as mx, numpy as np
from collections import namedtuple

IMAGE_SIZE=224


def resize_min(img, msize = 224):
    old_size = img.shape[:2] # old_size is in (height, width) format
    ratio = float(msize)/max(old_size)
    new_size = tuple([int(x*ratio) for x in old_size])

    # new_size should be in (width, height) format
    img = cv2.resize(img, (new_size[1], new_size[0]))

    delta_w = msize - new_size[1]
    delta_h = msize - new_size[0]
    top, bottom = delta_h//2, delta_h-(delta_h//2)
    left, right = delta_w//2, delta_w-(delta_w//2)

    color = [0, 0, 0]
    img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)
    return img


def load_model_for_batch_prediction(prefix, checkpoint_epoch, shape, device):
    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, checkpoint_epoch)
    all_layers = sym.get_internals()
    devs = mx.gpu() if device == 'gpu' else mx.cpu()
    fe_sym = all_layers['flatten0_output']
    mod = mx.mod.Module(symbol=fe_sym, label_names=None, context=devs)
    mod.bind(for_training=False, data_shapes=[('data', shape)])  # (batch_size,3,224,224)
    mod.set_params(arg_params, aux_params)
    return mod


def prepare_nd_array(filename):
    img = cv2.imread(filename)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = resize_min(img, IMAGE_SIZE)
    img = np.swapaxes(img, 0, 2)
    img = np.swapaxes(img, 1, 2)
    img = img[np.newaxis, :]
    return img


def predict_mxnet(model, dimentions):
    Batch = namedtuple('Batch', ['data'])
    model.forward(Batch([dimentions]))
    prob = model.get_outputs()[0].asnumpy()
    return prob
    #return np.squeeze(prob)


def get_minimal_files_list_size(batch_size, files_list_size):
    size = batch_size
    while size + batch_size <= files_list_size:
        size += batch_size
    if files_list_size > size:
        size += batch_size
    return size


def get_files_list(input_dir, files_num, batch_size):
    files_list = os.listdir(input_dir)[:files_num]
    min_list_size = get_minimal_files_list_size(batch_size, len(files_list))
    last_image = files_list[len(files_list) - 1]
    files_list += (min_list_size - len(files_list))*[last_image]
    return files_list


parser = argparse.ArgumentParser(description='Files generator')
parser.add_argument('--input-dir')
parser.add_argument('--files-num', type=int)
parser.add_argument('--model-prefix')
parser.add_argument('--model-epoch', type=int, default=0)
parser.add_argument('--device', default='cpu')
parser.add_argument('--batch-size', type=int, default=32)

# python mxnet_bulk.py --input-dir=/media/olga/war_and_terror/war_and_terror_manual_filtered --files-num=200 \
#   --model-prefix=/home/olga/IdeaProjects/video-data-science/mxnet_experiments/resnet-152 --model-epoch=0 \
#   --device=cpu --batch-size=32
print("start")
args = parser.parse_args()

batch_size=args.batch_size
mxnet_obj = load_model_for_batch_prediction(args.model_prefix, args.model_epoch, (batch_size,3,IMAGE_SIZE,IMAGE_SIZE), args.device)
files_list = get_files_list(args.input_dir, args.files_num, batch_size)
batches_num = round(len(files_list)/batch_size)
Batch = namedtuple('Batch', ['data'])
#for file_name in files_list:
for i in range(0, batches_num):
    print(i)
    # tic = time.time()
    idx = range(i*batch_size, (i+1)*batch_size)
    batch_files_list = [files_list[j] for j in idx]
    img_batch = [prepare_nd_array(os.path.join(args.input_dir, file)) for file in batch_files_list]
    img = np.concatenate(img_batch)
    mxnet_obj.forward(Batch([mx.nd.array(img)]))
    prob = mxnet_obj.get_outputs()[0].asnumpy()

print("done")

bulk_parallel.py

import argparse, os, cv2
import mxnet as mx, numpy as np, multiprocessing as mproc
from collections import namedtuple

IMAGE_SIZE=224


def resize_min(img, msize = 224):
    old_size = img.shape[:2] # old_size is in (height, width) format
    ratio = float(msize)/max(old_size)
    new_size = tuple([int(x*ratio) for x in old_size])

    # new_size should be in (width, height) format
    img = cv2.resize(img, (new_size[1], new_size[0]))

    delta_w = msize - new_size[1]
    delta_h = msize - new_size[0]
    top, bottom = delta_h//2, delta_h-(delta_h//2)
    left, right = delta_w//2, delta_w-(delta_w//2)

    color = [0, 0, 0]
    img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)
    return img


def load_model_for_batch_prediction(prefix, checkpoint_epoch, shape, device):
    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, checkpoint_epoch)
    all_layers = sym.get_internals()
    devs = mx.gpu() if device == 'gpu' else mx.cpu()
    fe_sym = all_layers['flatten0_output']
    mod = mx.mod.Module(symbol=fe_sym, label_names=None, context=devs)
    mod.bind(for_training=False, data_shapes=[('data', shape)])  # (batch_size,3,224,224)
    mod.set_params(arg_params, aux_params)
    return mod


def prepare_nd_array(filename):
    img = cv2.imread(filename)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = resize_min(img, IMAGE_SIZE)
    img = np.swapaxes(img, 0, 2)
    img = np.swapaxes(img, 1, 2)
    img = img[np.newaxis, :]
    return img


def predict_mxnet(model, dimentions):
    Batch = namedtuple('Batch', ['data'])
    model.forward(Batch([dimentions]))
    prob = model.get_outputs()[0].asnumpy()
    return prob
    #return np.squeeze(prob)


def get_minimal_files_list_size(batch_size, files_list_size):
    size = batch_size
    while size + batch_size <= files_list_size:
        size += batch_size
    if files_list_size > size:
        size += batch_size
    return size


def get_files_list(input_dir, files_num, batch_size):
    files_list = os.listdir(input_dir)[:files_num]
    min_list_size = get_minimal_files_list_size(batch_size, len(files_list))
    last_image = files_list[len(files_list) - 1]
    files_list += (min_list_size - len(files_list))*[last_image]
    return files_list


def process_bulk(files_list, i, batch_size, input_dir):
    print(i)
    # tic = time.time()
    idx = range(i*batch_size, (i+1)*batch_size)
    batch_files_list = [files_list[j] for j in idx]
    img_batch = [prepare_nd_array(os.path.join(input_dir, file)) for file in batch_files_list]
    img = np.concatenate(img_batch)
    mxnet_obj.forward(Batch([mx.nd.array(img)]))
    prob = mxnet_obj.get_outputs()[0].asnumpy()
    return prob


parser = argparse.ArgumentParser(description='Files generator')
parser.add_argument('--input-dir')
parser.add_argument('--files-num', type=int)
parser.add_argument('--model-prefix')
parser.add_argument('--model-epoch', type=int, default=0)
parser.add_argument('--device', default='cpu')
parser.add_argument('--batch-size', type=int, default=32)
parser.add_argument('--pool-size', type=int)

# python mxnet_bulk_parallel.py --input-dir=/media/olga/war_and_terror/war_and_terror_manual_filtered --files-num=200 \
#   --model-prefix=/home/olga/IdeaProjects/video-data-science/mxnet_experiments/resnet-152 --model-epoch=0 \
#   --device=cpu --batch-size=32
print("start")
args = parser.parse_args()

batch_size=args.batch_size
mxnet_obj = load_model_for_batch_prediction(args.model_prefix, args.model_epoch, (batch_size,3,IMAGE_SIZE,IMAGE_SIZE), args.device)
files_list = get_files_list(args.input_dir, args.files_num, batch_size)
batches_num = round(len(files_list)/batch_size)
Batch = namedtuple('Batch', ['data'])
#for file_name in files_list:
# for i in range(0, batches_num):
#     print(i)
#     # tic = time.time()
#     idx = range(i*batch_size, (i+1)*batch_size)
#     batch_files_list = [files_list[j] for j in idx]
#     img_batch = [prepare_nd_array(os.path.join(args.input_dir, file)) for file in batch_files_list]
#     img = np.concatenate(img_batch)
#     mxnet_obj.forward(Batch([mx.nd.array(img)]))
#     prob = mxnet_obj.get_outputs()[0].asnumpy()
#
# print("done")

if __name__ == '__main__':
    print('started')

    files_list = get_files_list(args.input_dir, args.files_num, batch_size)
    iterator = [(files_list, i, batch_size, args.input_dir)  for i in range(0, batches_num)]#os.path.join(args.input_dir, file_name) for file_name in files_list]
    with mproc.Pool(processes=min(args.pool_size, len(iterator))) as pool:
        res = pool.starmap(process_bulk, iterator)

    print("done")

docker_diagnoze_results.txt

root@da4fe28faf8a:/# python3 diagnose.py 
----------Python Info----------
Version      : 3.4.3
Compiler     : GCC 4.8.4
Build        : ('default', 'Nov 28 2017 16:41:13')
Arch         : ('64bit', 'ELF')
------------Pip Info-----------
Version      : 9.0.1
Directory    : /usr/local/lib/python3.4/dist-packages/pip
----------MXNet Info-----------
libdc1394 error: Failed to initialize libdc1394
Version      : 1.1.0
Directory    : /mxnet/python/mxnet
Hashtag not found. Not installed from pre-built package.
----------System Info----------
Platform     : Linux-4.4.0-1055-aws-x86_64-with-Ubuntu-14.04-trusty
system       : Linux
node         : da4fe28faf8a
release      : 4.4.0-1055-aws
version      : #64-Ubuntu SMP Thu Apr 5 17:06:36 UTC 2018
----------Hardware Info----------
machine      : x86_64
processor    : x86_64
Architecture:          x86_64
CPU op-mode(s):        32-bit, 64-bit
Byte Order:            Little Endian
CPU(s):                4
On-line CPU(s) list:   0-3
Thread(s) per core:    2
Core(s) per socket:    2
Socket(s):             1
NUMA node(s):          1
Vendor ID:             GenuineIntel
CPU family:            6
Model:                 85
Stepping:              3
CPU MHz:               3000.000
BogoMIPS:              6000.00
Hypervisor vendor:     KVM
Virtualization type:   full
L1d cache:             32K
L1i cache:             32K
L2 cache:              1024K
L3 cache:              25344K
NUMA node0 CPU(s):     0-3
----------Network Test----------
Setting timeout: 10
Timing for Gluon Tutorial(cn): https://zh.gluon.ai, DNS: 0.0604 sec, LOAD: 0.9503 sec.
Timing for Conda: https://repo.continuum.io/pkgs/free/, DNS: 0.0036 sec, LOAD: 0.0553 sec.
Timing for PYPI: https://pypi.python.org/pypi/pip, DNS: 0.0011 sec, LOAD: 0.0956 sec.
Timing for FashionMNIST: https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/fashion-mnist/train-labels-idx1-ubyte.gz, DNS: 0.0039 sec, LOAD: 0.3542 sec.
Timing for Gluon Tutorial(en): http://gluon.mxnet.io, DNS: 0.0824 sec, LOAD: 0.1584 sec.
Timing for MXNet: https://github.com/apache/incubator-mxnet, DNS: 0.0011 sec, LOAD: 0.7186 sec.

one_by_one.py

import argparse, os, cv2
import mxnet as mx, numpy as np
from collections import namedtuple

IMAGE_SIZE=224


def resize_min(img, msize = 224):
    old_size = img.shape[:2] # old_size is in (height, width) format
    ratio = float(msize)/min(old_size)
    new_size = tuple([int(x*ratio) for x in old_size])

    # new_size should be in (width, height) format
    img = cv2.resize(img, (new_size[1], new_size[0]))
    return img


def load_model_for_one(prefix, checkpoint_epoch, device):
    print('load_model_for_one ', prefix, checkpoint_epoch)
    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, checkpoint_epoch)
    all_layers = sym.get_internals()
    devs = mx.gpu() if device == 'gpu' else mx.cpu()
    fe_sym = all_layers['flatten0_output']
    mod = mx.mod.Module(symbol=fe_sym,  context=devs, label_names=None)
    mod.bind(for_training=False, data_shapes=[('data', (1, 3, IMAGE_SIZE, IMAGE_SIZE))])
    mod.set_params(arg_params, aux_params)
    return mod


def prepare_nd_array(filename):
    img = cv2.imread(filename)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = resize_min(img, IMAGE_SIZE)
    img = np.swapaxes(img, 0, 2)
    img = np.swapaxes(img, 1, 2)
    img = img[np.newaxis, :]
    return mx.nd.array(img)


def predict_mxnet(model, dimentions):
    Batch = namedtuple('Batch', ['data'])
    model.forward(Batch([dimentions]))
    prob = model.get_outputs()[0].asnumpy()
    return prob
    #return np.squeeze(prob)

parser = argparse.ArgumentParser(description='Files generator')
parser.add_argument('--input-dir')
parser.add_argument('--files-num', type=int)
parser.add_argument('--model-prefix')
parser.add_argument('--model-epoch', type=int, default=0)
parser.add_argument('--device', default='cpu')

# python mxnet_one_by_one.py --input-dir=/media/olga/war_and_terror/war_and_terror_manual_filtered --files-num=200 --model-prefix=/home/olga/IdeaProjects/video-data-science/mxnet_experiments/resnet-152 --model-epoch=0
print("start")
args = parser.parse_args()

mxnet_obj = load_model_for_one(args.model_prefix, args.model_epoch, args.device)
files_list = os.listdir(args.input_dir)[:args.files_num]
for file_name in files_list:
    full_path = os.path.join(args.input_dir, file_name)
    array = prepare_nd_array(full_path)
    mxnet_results = predict_mxnet(mxnet_obj, array)

print("done")

parallel.py

import argparse, os, cv2
import mxnet as mx, numpy as np, multiprocessing as mproc
from collections import namedtuple

IMAGE_SIZE=224


def resize_min(img, msize = 224):
    old_size = img.shape[:2] # old_size is in (height, width) format
    ratio = float(msize)/min(old_size)
    new_size = tuple([int(x*ratio) for x in old_size])

    # new_size should be in (width, height) format
    img = cv2.resize(img, (new_size[1], new_size[0]))
    return img


def load_model_for_one(prefix, checkpoint_epoch, device):
    print('load_model_for_one ', prefix, checkpoint_epoch)
    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, checkpoint_epoch)
    devs = mx.gpu() if device == 'gpu' else mx.cpu()
    all_layers = sym.get_internals()
    fe_sym = all_layers['flatten0_output']
    mod = mx.mod.Module(symbol=fe_sym,  context=devs, label_names=None)
    mod.bind(for_training=False, data_shapes=[('data', (1, 3, IMAGE_SIZE, IMAGE_SIZE))])
    mod.set_params(arg_params, aux_params)
    return mod


def prepare_nd_array(filename):
    img = cv2.imread(filename)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = resize_min(img, IMAGE_SIZE)
    img = np.swapaxes(img, 0, 2)
    img = np.swapaxes(img, 1, 2)
    img = img[np.newaxis, :]
    return mx.nd.array(img)


def predict_mxnet(model, dimentions):
    Batch = namedtuple('Batch', ['data'])
    model.forward(Batch([dimentions]))
    prob = model.get_outputs()[0].asnumpy()
    return prob
    #return np.squeeze(prob)


def process_file(full_path):
    array = prepare_nd_array(full_path)
    result = predict_mxnet(mxnet_obj, array)
    # print(full_path, result)
    return result


parser = argparse.ArgumentParser(description='Files generator')
parser.add_argument('--input-dir')
parser.add_argument('--files-num', type=int)
parser.add_argument('--model-prefix')
parser.add_argument('--model-epoch', type=int, default=0)
parser.add_argument('--pool-size', type=int)
parser.add_argument('--device', default='cpu')

# python mxnet_parallel.py --input-dir=/media/olga/war_and_terror/war_and_terror_manual_filtered --files-num=200 \
#   --model-prefix=/home/olga/IdeaProjects/video-data-science/mxnet_experiments/resnet-152 --model-epoch=0 \
#   --pool-size=7 --device=cpu
args = parser.parse_args()

mxnet_obj = load_model_for_one(args.model_prefix, args.model_epoch, args.device)
ncpu = mproc.cpu_count()
print('ncpu - ' + str(ncpu))


if __name__ == '__main__':
    print('started')

    files_list = os.listdir(args.input_dir)[:args.files_num]
    iterator = [os.path.join(args.input_dir, file_name) for file_name in files_list]

    with mproc.Pool(processes=min(args.pool_size, len(iterator))) as pool:
        res = pool.map(process_file, iterator)

    print("done")

pure_os_diagnoze_results.txt

ubuntu@ip-192-168-33-11:~$ python3 diagnose.py 
----------Python Info----------
Version      : 3.5.2
Compiler     : GCC 5.4.0 20160609
Build        : ('default', 'Nov 23 2017 16:37:01')
Arch         : ('64bit', 'ELF')
------------Pip Info-----------
Version      : 8.1.1
Directory    : /usr/lib/python3/dist-packages/pip
----------MXNet Info-----------
Version      : 1.1.0
Directory    : /home/ubuntu/.local/lib/python3.5/site-packages/mxnet
Commit Hash   : 07a83a0325a3d782513a04f47d711710972cb144
----------System Info----------
Platform     : Linux-4.4.0-1055-aws-x86_64-with-Ubuntu-16.04-xenial
system       : Linux
node         : ip-192-168-33-11
release      : 4.4.0-1055-aws
version      : #64-Ubuntu SMP Thu Apr 5 17:06:36 UTC 2018
----------Hardware Info----------
machine      : x86_64
processor    : x86_64
Architecture:          x86_64
CPU op-mode(s):        32-bit, 64-bit
Byte Order:            Little Endian
CPU(s):                4
On-line CPU(s) list:   0-3
Thread(s) per core:    2
Core(s) per socket:    2
Socket(s):             1
NUMA node(s):          1
Vendor ID:             GenuineIntel
CPU family:            6
Model:                 85
Model name:            Intel(R) Xeon(R) Platinum 8124M CPU @ 3.00GHz
Stepping:              3
CPU MHz:               3000.000
BogoMIPS:              6000.00
Hypervisor vendor:     KVM
Virtualization type:   full
L1d cache:             32K
L1i cache:             32K
L2 cache:              1024K
L3 cache:              25344K
NUMA node0 CPU(s):     0-3
Flags:                 fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc aperfmperf eagerfpu pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single rsb_ctxsw retpoline kaiser fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm mpx avx512f rdseed adx smap clflushopt clwb avx512cd xsaveopt xsavec xgetbv1 ida arat
----------Network Test----------
Setting timeout: 10
Timing for MXNet: https://github.com/apache/incubator-mxnet, DNS: 0.0029 sec, LOAD: 0.8403 sec.
Timing for Conda: https://repo.continuum.io/pkgs/free/, DNS: 0.0034 sec, LOAD: 0.0303 sec.
Timing for Gluon Tutorial(en): http://gluon.mxnet.io, DNS: 0.0368 sec, LOAD: 0.6461 sec.
Timing for PYPI: https://pypi.python.org/pypi/pip, DNS: 0.0104 sec, LOAD: 0.0911 sec.
Timing for Gluon Tutorial(cn): https://zh.gluon.ai, DNS: 0.0485 sec, LOAD: 0.4284 sec.
Timing for FashionMNIST: https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/fashion-mnist/train-labels-idx1-ubyte.gz, DNS: 0.0125 sec, LOAD: 0.3838 sec.

def prepare_nd_array(filename):
    img = mx.image.imread(filename)
    img = mx.image.resize_short(img, IMAGE_SIZE)
    return img.transpose((2, 0, 1)).expand_dims(axis=0)

is a bit more concise and does not have external dependencies