Last active
December 11, 2017 15:37
-
-
Save wtnb75/80f5a2f1892f9ee349d55b0d865a023e to your computer and use it in GitHub Desktop.
cpu-caffe vs. movidius ncs
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
from mvnc import mvncapi as mvnc | |
import cv2 | |
import numpy | |
import os | |
import time | |
import subprocess | |
import caffe | |
import click | |
import struct | |
class RunByCaffe: | |
def __init__(self, labelfn, meanfn, prototxt, caffemodel): | |
self.dim = (227, 227) | |
self.labels = numpy.loadtxt(labelfn, str, delimiter="\t") | |
self.mean = numpy.load(meanfn).mean(1).mean(1) | |
self.net = caffe.Net(prototxt, caffemodel, caffe.TEST) | |
self.transformer = caffe.io.Transformer( | |
{'data': self.net.blobs['data'].data.shape}) | |
self.transformer.set_transpose('data', (2, 0, 1)) | |
self.transformer.set_mean('data', self.mean) | |
self.transformer.set_raw_scale('data', 255) | |
self.transformer.set_channel_swap('data', (2, 1, 0)) | |
caffe.set_mode_cpu() | |
def loadimg(self, fn): | |
# print("loading", fn) | |
image = caffe.io.load_image(fn) | |
transformed_image = self.transformer.preprocess('data', image) | |
return transformed_image | |
def classify(self, img): | |
# self.net.blobs['data'].reshape(1, 3, self.dim[0], self.dim[1]) | |
self.net.blobs['data'].data[...] = [img] | |
output = self.net.forward() | |
output_prob = output['prob'][0] | |
return output_prob | |
class RunByMovidius(RunByCaffe): | |
def __init__(self, labelfn, meanfn, graph): | |
self.dim = [227, 227] | |
self.channel = 3 | |
with open(graph, "rb") as ifp: | |
ifp.seek(0xfa) | |
cpus, = struct.unpack("h", ifp.read(2)) | |
print("cpus", cpus + 1) | |
ifp.seek(0x178) | |
width, height, channel = struct.unpack("3i", ifp.read(4 * 3)) | |
self.dim[0] = width | |
self.dim[1] = height | |
self.channels = channel | |
print("width", width, "height", height, "channel", channel) | |
self.labels = numpy.loadtxt(labelfn, str, delimiter="\t") | |
self.mean = numpy.load(meanfn).mean(1).mean(1) | |
if False: | |
self.transformer = caffe.io.Transformer( | |
{'data': (1, self.channel, self.dim[0], self.dim[1])}) | |
self.transformer.set_transpose('data', (2, 0, 1)) | |
self.transformer.set_mean('data', self.mean) | |
self.transformer.set_raw_scale('data', 255) | |
self.transformer.set_channel_swap('data', (2, 1, 0)) | |
mvnc.SetGlobalOption(mvnc.GlobalOption.LOG_LEVEL, 2) | |
devices = mvnc.EnumerateDevices() | |
if len(devices) == 0: | |
print('No devices found') | |
quit() | |
self.device = mvnc.Device(devices[0]) | |
self.device.OpenDevice() | |
with open(graph, mode='rb') as f: | |
blob = f.read() | |
self.graph = self.device.AllocateGraph(blob) | |
def loadimg(self, fn): | |
# print("loading", fn) | |
img = cv2.imread(fn) | |
img = cv2.resize(img, tuple(self.dim)) | |
img = img.astype(numpy.float16) | |
img[:, :, 0] = (img[:, :, 0] - self.mean[0]) | |
img[:, :, 1] = (img[:, :, 1] - self.mean[1]) | |
img[:, :, 2] = (img[:, :, 2] - self.mean[2]) | |
return img | |
def classify(self, img): | |
self.graph.LoadTensor(img.astype(numpy.float16), 'user object') | |
output, userobj = self.graph.GetResult() | |
return output | |
def __del__(self): | |
if hasattr(self, "graph"): | |
self.graph.DeallocateGraph() | |
if hasattr(self, "device"): | |
self.device.CloseDevice() | |
class RunByMovidius2(RunByMovidius): | |
def __init__(self, labelfn, meanfn, prototxt, caffemodel): | |
if not os.path.exists("graph"): | |
res = subprocess.run(["mvNCCompile", "-w", caffemodel, "-s", "12", prototxt]) | |
print("result", res) | |
RunByMovidius.__init__(self, labelfn, meanfn, "graph") | |
def getts(ts): | |
return list(map(lambda f: f[1] - f[0], zip(ts, ts[1:]))) | |
@click.group(invoke_without_command=True) | |
@click.pass_context | |
def cli(ctx): | |
if ctx.invoked_subcommand is None: | |
print(ctx.get_help()) | |
else: | |
print('gonna invoke %s' % ctx.invoked_subcommand) | |
def do_solve(solver, images): | |
print("images", images) | |
print("|file|time(load)|time(eval)|id|prob|label|") | |
print("|----|---------:|---------:|--|----|-----|") | |
for i in images: | |
ts = [time.time()] | |
img = solver.loadimg(i) | |
ts.append(time.time()) | |
output = solver.classify(img) | |
ts.append(time.time()) | |
loadtime, evaltime = getts(ts) | |
am = output.argmax() | |
print("|%s|%.3f|%.3f|%s|%.2f|%s|" % (os.path.basename(i), | |
loadtime, evaltime, am, output[am], solver.labels[am])) | |
@cli.command() | |
@click.option("--label") | |
@click.option("--mean") | |
@click.option("--prototxt") | |
@click.option("--caffemodel") | |
@click.argument("images", nargs=-1) | |
def bycaffe(label, mean, prototxt, caffemodel, images): | |
solver = RunByCaffe(label, mean, prototxt, caffemodel) | |
do_solve(solver, images) | |
@cli.command() | |
@click.option("--label") | |
@click.option("--mean") | |
@click.option("--graph") | |
@click.argument("images", nargs=-1) | |
def byncs(label, mean, graph, images): | |
solver = RunByMovidius(label, mean, graph) | |
do_solve(solver, images) | |
@cli.command() | |
@click.option("--label") | |
@click.option("--mean") | |
@click.option("--prototxt") | |
@click.option("--caffemodel") | |
@click.argument("images", nargs=-1) | |
def byncs2(label, mean, prototxt, caffemodel, images): | |
solver = RunByMovidius2(label, mean, prototxt, caffemodel) | |
do_solve(solver, images) | |
def main(): | |
cli() | |
if __name__ == "__main__": | |
main() |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
SqueezeNet, caffe
SqueezeNet, ncs