Insightface: Test on CPU

[MyraBaba]

Hi,

How we can test pretrained models on CPU ?

it gaves:

RuntimeError: simple_bind error. Arguments:
data: (1, 3L, 68L, 68L)
[23:39:56] src/storage/storage.cc:118: Compile with USE_CUDA=1 to enable GPU usage

[xxllp]

这个需要GPU支持

[Sparkkkk]

You can manually set Mxnet to use CPU. In face_model.py line 46, set ctx = mx.cpu(0) instead of the argument that passed.

[MyraBaba]

Thanks I missed that.

now it gaves below error:

Traceback (most recent call last):
File "test.py", line 23, in
gender, age = model.get_ga(img)
File "face_model.py", line 104, in get_ga
self.ga_model.forward(db, is_train=False)
AttributeError: 'NoneType' object has no attribute 'forward'

[Wisgon]

Gender and age model haven't been release yet.

[xxllp]

@Wisgon ,你有训练好的模型嘛

[Wisgon]

No, I don't have.

[MyraBaba]

I just want to test face recognition performance. than what is the step that I need to follow ? I am confused

Best

[MyraBaba]

Also have this error in test.py:

line 21 : f1 = model.get_feature(img)

Traceback (most recent call last):
File "/Users/alpullu/Projects/insightface/deploy/test.py", line 21, in
f1 = model.get_feature(img)
File "/Users/alpullu/Projects/insightface/deploy/face_model.py", line 93, in get_feature
self.model.forward(db, is_train=False)
AttributeError: 'NoneType' object has no attribute 'forward'

[Wisgon]

The author had comment ~/deploy/test.py line 21 now. You had better update your repo

[MyraBaba]

I will update and let you know.

[MyraBaba]

Still not updated. I want to basicly test to compare faces to thest model accuracy. Yes we are not expert try to learn now.

How I can do this ? I need just need a simple guide for:

1- load model LResNet50E-IR
2 - load to face images
3-extrac features
4-compare their similarity
5-make decision depens on the number.

and need to try on CPU.

any help ? it could also helpfull for others.

Best

[glorioushedgehog]

Hey MyraBaba,

What arguments are you passing when you run test.py? I got the same error when I wasn't specifying the directory to the model. You also have to give it the epoch number. I got it to work with this:

--model C:\Users\{user}\insightface\models\model-r50-am-lfw\,0

The number at the end specifies the epoch. The model-r50-am-lfw folder contains the .params, symbol.json, and the log. It looks like mxnet assumes a lot about the names of these files. For example, the number in front of "-params" has to be the epoch number. So I have "-0000.params" and I pass an epoch of 0. You might have to change the file names to get it to work.

[glorioushedgehog]

Also, for some reason test.py attempts to do gender and age recognition by default, so that has to be turned off. The end of my test.py looks like this:

`model = face_model.FaceModel(args)
img = cv2.imread('Tom_Hanks_54745.png')
img = model.get_input(img)
f1 = model.get_feature(img)
print(f1[0:10])

gender, age = model.get_ga(img)

print(gender)

print(age)

sys.exit(0)

img = cv2.imread('/raid5data/dplearn/megaface/facescrubr/112x112/Tom_Hanks/Tom_Hanks_54733.png')

f2 = model.get_feature(img)
dist = np.sum(np.square(f1-f2))
print(dist)
sim = np.dot(f1, f2.T)
print(sim)

diff = np.subtract(source_feature, target_feature)

dist = np.sum(np.square(diff),1)`

[MyraBaba]

I got below error:

loading /Users/alpullu/Projects/insightface/models/model-r50-am-lfw/ 0
[19:43:26] src/nnvm/legacy_json_util.cc:190: Loading symbol saved by previous version v0.12.1. Attempting to upgrade...
[19:43:26] src/nnvm/legacy_json_util.cc:198: Symbol successfully upgraded!
[19:43:26] src/nnvm/legacy_json_util.cc:190: Loading symbol saved by previous version v0.8.0. Attempting to upgrade...
[19:43:26] src/nnvm/legacy_json_util.cc:198: Symbol successfully upgraded!
/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/mxnet/model.py:927: DeprecationWarning: mxnet.model.FeedForward has been deprecated. Please use mxnet.mod.Module instead.
kwargs)
[19:43:26] src/nnvm/legacy_json_util.cc:190: Loading symbol saved by previous version v0.8.0. Attempting to upgrade...
[19:43:26] src/nnvm/legacy_json_util.cc:198: Symbol successfully upgraded!
[19:43:26] src/nnvm/legacy_json_util.cc:190: Loading symbol saved by previous version v0.8.0. Attempting to upgrade...
[19:43:26] src/nnvm/legacy_json_util.cc:198: Symbol successfully upgraded!
[19:43:26] src/nnvm/legacy_json_util.cc:190: Loading symbol saved by previous version v0.8.0. Attempting to upgrade...
[19:43:26] src/nnvm/legacy_json_util.cc:198: Symbol successfully upgraded!
[ 1.06205717 -3.00269191 84.31859815 118.66388603]
[[35 50]
[72 48]
[61 78]
[36 90]
[66 87]]
Traceback (most recent call last):
File "/Users/alpullu/Projects/insightface/deploy/test.py", line 29, in
f2 = model.get_feature(img)
File "/Users/alpullu/Projects/insightface/deploy/face_model.py", line 93, in get_feature
self.model.forward(db, is_train=False)
File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/mxnet/module/module.py", line 608, in forward
infer_shape error. Arguments:
data: (1,)
self.reshape(new_dshape, new_lshape)
File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/mxnet/module/module.py", line 470, in reshape
self._exec_group.reshape(self._data_shapes, self._label_shapes)
File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/mxnet/module/executor_group.py", line 381, in reshape
self.bind_exec(data_shapes, label_shapes, reshape=True)
File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/mxnet/module/executor_group.py", line 357, in bind_exec
allow_up_sizing=True, *dict(data_shapes_i + label_shapes_i))
File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/mxnet/executor.py", line 402, in reshape
arg_shapes, _, aux_shapes = self._symbol.infer_shape(kwargs)
File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/mxnet/symbol/symbol.py", line 989, in infer_shape
res = self._infer_shape_impl(False, *args, *kwargs)
File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/mxnet/symbol/symbol.py", line 1119, in _infer_shape_impl
ctypes.byref(complete)))
File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/mxnet/base.py", line 146, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: Error in operator conv0: [19:43:27] src/operator/nn/./convolution-inl.h:492: Check failed: dshp.ndim() == 4U (1 vs. 4) Input data should be 4D in batch-num_filter-y-x

Stack trace returned 8 entries:
[bt] (0) 0 libmxnet.so 0x0000000107d707cf libmxnet.so + 59343
[bt] (1) 1 libmxnet.so 0x0000000107d7056f libmxnet.so + 58735
[bt] (2) 2 libmxnet.so 0x0000000107e983ed libmxnet.so + 1270765
[bt] (3) 3 libmxnet.so 0x0000000108e9077e MXNDListFree + 1571518
[bt] (4) 4 libmxnet.so 0x0000000108d5de8a MXNDListFree + 315850
[bt] (5) 5 libmxnet.so 0x0000000108d5682e MXNDListFree + 285550
[bt] (6) 6 libmxnet.so 0x0000000108d02d69 MXSymbolInferShape + 2281
[bt] (7) 7 _ctypes.so 0x0000000104c4d5c7 ffi_call_unix64 + 79

Process finished with exit code 1

---

code is:

`import face_model
import argparse
import cv2
import sys
import numpy as np

parser = argparse.ArgumentParser(description='face model test')
# general
parser.add_argument('--image-size', default='112,112', help='')
parser.add_argument('--model', default='', help='path to load model.')
parser.add_argument('--ga-model', default='', help='path to load model.')
parser.add_argument('--gpu', default=0, type=int, help='gpu id')
parser.add_argument('--det', default=0, type=int, help='mtcnn option, 1 means using R+O, 0 means detect from begining')
parser.add_argument('--flip', default=0, type=int, help='whether do lr flip aug')
parser.add_argument('--threshold', default=1.24, type=float, help='ver dist threshold')
args = parser.parse_args()

model = face_model.FaceModel(args)
img = cv2.imread('Tom_Hanks_54745.png')
img = model.get_input(img)

f1 = model.get_feature(img)
#print(f1[0:10])
# gender, age = model.get_ga(img)
# print(gender)
# print(age)
# sys.exit(0)
img = cv2.imread('/raid5data/dplearn/megaface/facescrubr/112x112/Tom_Hanks/Tom_Hanks_54733.png')
f2 = model.get_feature(img)
dist = np.sum(np.square(f1-f2))
print(dist)
sim = np.dot(f1, f2.T)
print(sim)
#diff = np.subtract(source_feature, target_feature)
#dist = np.sum(np.square(diff),1)
`

models folder has :
model-r50-am-lfw/
-symbol.json
-0000.params

and the run parameter 👍
--model Projects/insightface/models/model-r50-am-lfw/,0

[glorioushedgehog]

just comment out the line img = cv2.imread('/raid5data/dplearn/megaface/facescrubr/112x112/Tom_Hanks/Tom_Hanks_54733.png')
It's replacing img with nothing because that path doesn't exist

[MyraBaba]

I missed that now it worked.
I gave:

[20:05:47] src/nnvm/legacy_json_util.cc:198: Symbol successfully upgraded!
[ 1.06205717 -3.00269191 84.31859815 118.66388603]
[[35 50]
[72 48]
[61 78]
[36 90]
[66 87]]
0.0
1.0

Process finished with exit code 0

So it means the distance is the for the same image.

if I want to compare two image they should be the 122x122 right ? and also face already detected and focused or it si detecting faces for this compare ?

thanks

[glorioushedgehog]

They should be 112x112. It is detecting faces for this compare.

[MyraBaba]

Hi,

Is this dist = np.sum(np.square(f1-f2)) distance for loss-type=0: Softmax . or loss-type=4: ArcFace ?

And the distance between 0 and 1 right ? closing means to 1 more similar or opposite?

[MyraBaba]

I guess :

distance lower means more similar.
similarity close to 1 means more similar

[glorioushedgehog]

Sounds right

[gwsbhqt]

@MyraBaba @xxllp @Sparkkkk @Wisgon @glorioushedgehog @nttstar

Do you have the new ga-model for age and sex classification? I've been looking all over the place, I still can't find the ga-model.

I don't have a good GPU so that I can't train a model myself, so I want to wait for the author's pre training model.

Do you know that the author is sharing this ga-model?

thanks!

[MyraBaba]

I didnt find also . as far as I know there is not.
by the way ; I am trying to convert the some part to the c++ from python for more speed gain. any experience ?

[gwsbhqt]

@MyraBaba
I advise you not to do this because it is too difficult and too much trouble. Spending a lot of time and effort may only get a very small boost. The best way to increase speed is to spend money on GPUs.

[symisc]

@MyraBaba what's the CPU performance result?

[MyraBaba]

Still not able to port %100 to c++ . But I am sure it will be better that python and compilable for the edge.