Skip to content
This repository has been archived by the owner on Oct 27, 2021. It is now read-only.

TypeError: 'Convolution2D' object has no attribute '__getitem__' #3

Closed
zjsjack opened this issue Jun 23, 2017 · 22 comments
Closed

TypeError: 'Convolution2D' object has no attribute '__getitem__' #3

zjsjack opened this issue Jun 23, 2017 · 22 comments

Comments

@zjsjack
Copy link

zjsjack commented Jun 23, 2017

I convent the VGG16 model to pkl as instructed by https://github.com/mitmul/chainer-faster-rcnn/issues/15：

import pickle
from chainer.links.caffe import CaffeFunction

vgg16 = CaffeFunction('VGG_ILSVRC_16_layers.caffemodel')
pickle.dump(vgg16, open('VGG16.model', 'wb'))

Got TypeError when run python index.es.py.

2017-06-23 09:23:26.732486: I tensorflow/core/common_runtime/gpu/gpu_device.cc:940] Found device 0 with properties:
name: Tesla K80
major: 3 minor: 7 memoryClockRate (GHz) 0.8235
pciBusID 0000:00:1e.0
Total memory: 11.17GiB
Free memory: 11.11GiB
2017-06-23 09:23:26.732530: I tensorflow/core/common_runtime/gpu/gpu_device.cc:961] DMA: 0
2017-06-23 09:23:26.732539: I tensorflow/core/common_runtime/gpu/gpu_device.cc:971] 0: Y
2017-06-23 09:23:26.732555: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030] Creating TensorFlow device (/gpu:0) -> (device: 0, name: Tesla K80, pci bus id: 0000:00:1e.0)
Namespace(es_host='localhost', es_index='im_data', es_port=9200, es_type='obj', input='tf-faster-rcnn/data/demo/', model_path='./coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_490000.ckpt.data-00000-of-00001', weight='./VGG16.model')
Loading caffe weights...
Done!
Traceback (most recent call last):
File "index_es.py", line 104, in
sess=sess)
File "/home/ubuntu/workspace/visual_search/visual_search/extractor.py", line 53, in init
tag='default', anchor_scales=anchors)
File "/home/ubuntu/workspace/visual_search/visual_search/lib/nets/vgg16.py", line 496, in create_architecture
rois, cls_prob, bbox_pred = self._vgg16_from_imagenet(sess, training)
File "/home/ubuntu/workspace/visual_search/visual_search/lib/nets/vgg16.py", line 303, in _vgg16_from_imagenet
net = self._conv_layer(sess, self._image, "conv1_1", False)
File "/home/ubuntu/workspace/visual_search/visual_search/lib/nets/vgg16.py", line 111, in _conv_layer
filt=self._get_conv_filter(sess, name, trainable=trainable)
File "/home/ubuntu/workspace/visual_search/visual_search/lib/nets/vgg16.py", line 64, in _get_conv_filter
w=self._caffe2tf_filter(name)
File "/home/ubuntu/workspace/visual_search/visual_search/lib/nets/vgg16.py", line 59, in _caffe2tf_filter
f=self._caffe_weights(name)
File "/home/ubuntu/workspace/visual_search/visual_search/lib/nets/vgg16.py", line 52, in _caffe_weights
return layer['weights']
TypeError: 'Convolution2D' object has no attribute 'getitem'
@tuan3w
Copy link
Owner

tuan3w commented Jun 23, 2017

Hi @zjsjack,
You used wrong pretrained-model :D. I use a pretrained model from this repo (https://github.com/endernewton/tf-faster-rcnn). So you can follow the guide in that repo to know how to get it.
Thanks

@zjsjack
Copy link
Author

zjsjack commented Jun 25, 2017 •
edited
Loading

I cannot find out how to get the right pretrained-model and weights. Could you explain more on it or show me where to get the model and weights?
In one of my tests, I used http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz for the weights and coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_490000.ckpt.data-00000-of-00001 for the model.
With this I got the following error.

Loading caffe weights...
Traceback (most recent call last):
File "index_es.py", line 104, in
sess=sess)
File "/home/ubuntu/workspace/visual_search/visual_search/extractor.py", line 53, in init
tag='default', anchor_scales=anchors)
File "/home/ubuntu/workspace/visual_search/visual_search/lib/nets/vgg16.py", line 493, in create_architecture
self._caffe_layers = pickle.load(f)
EOFError

So I tried to convert the wights from VGG_ILSVRC_16_layers.caffemodel and then got the TypeError

@tuan3w
Copy link
Owner

tuan3w commented Jun 25, 2017 •
edited
Loading

Hi @zjsjack ,
My PC is broken, so I can't check the the weights path you downloaded from. I think the problem is that pretrained model is trained with different version of tensorflow. You should use tensorflow==0.12.1.
By the way, you can download the weights path from here https://drive.google.com/drive/folders/0B1_fAEgxdnvJSmF3YUlZcHFqWTQ (file imagenet_weights.tgz)
In the folder that contains the model path should contains all metadata and index files. You can download coco_900k-1190k.tgz (https://drive.google.com/drive/folders/0B1_fAEgxdnvJNEdDUTRSOU11cW8) and change the model path appropriately.

@zjsjack
Copy link
Author

zjsjack commented Jun 26, 2017

Now my configuration is:
tensorflow==0.12.1
WEIGHT_PATH=./imagenet_weights/vgg16.weights
MODEL_PATH=./coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_1190000.ckpt

Namespace(es_host='localhost', es_index='im_data', es_port=9200, es_type='obj', input='tf-faster-rcnn/data/demo/', model_path='./coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_1190000.ckpt', weight='./imagenet_weights/vgg16.weights')
Loading caffe weights...
Done!
Loading model check point from ./coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_1190000.ckpt
W tensorflow/core/framework/op_kernel.cc:975] Not found: Key vgg16_default/conv4_3/bias not found in checkpoint
W tensorflow/core/framework/op_kernel.cc:975] Not found: Key vgg16_default/conv2_2/bias not found in checkpoint
...
NotFoundError (see above for traceback): Key vgg16_default/conv4_3/bias not found in checkpoint

Any suggestions?

@tuan3w
Copy link
Owner

tuan3w commented Jun 26, 2017

As I checked, some variable scopes are changed in the new pretrained model so it will throws errors as you see. So you have tried with old pretrained model ( coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_490000). Otherwise you have to adapt your code to new model.

@zjsjack
Copy link
Author

zjsjack commented Jun 26, 2017

Also tried the 490000, the same error.
I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:0) -> (device: 0, name: Tesla K80, pci bus id: 0000:00:1e.0)
Namespace(es_host='localhost', es_index='im_data', es_port=9200, es_type='obj', input='tf-faster-rcnn/data/demo/', model_path='./coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_490000.ckpt', weight='./imagenet_weights/vgg16.weights')
Loading caffe weights...
Done!
Loading model check point from ./coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_490000.ckpt
W tensorflow/core/framework/op_kernel.cc:975] Not found: Key vgg16_default/cls_score/bias not found in checkpoint

@tuan3w
Copy link
Owner

tuan3w commented Jun 26, 2017

How does ./coco_2014_train+coco_2014_valminusminival/ folder look like ? Does it contains all metadata files ?

@zjsjack
Copy link
Author

zjsjack commented Jun 26, 2017

ubuntu@ip-172-31-18-140:~/workspace/visual_search/visual_search$ tree coco_2014_train+coco_2014_valminusminival/
coco_2014_train+coco_2014_valminusminival/
├── vgg16_faster_rcnn_iter_490000.ckpt.data-00000-of-00001
├── vgg16_faster_rcnn_iter_490000.ckpt.index
├── vgg16_faster_rcnn_iter_490000.ckpt.meta
└── vgg16_faster_rcnn_iter_490000.pkl

@tuan3w
Copy link
Owner

tuan3w commented Jun 26, 2017

This error means that variable bias with name (vgg16_default/cls_score/bias) is not found in the check point path. The model you downloaded is different with the model I used when I work on it. I can't access to my PC right now, so I can't share you the model file that I used.
Can you give me output when you run following command. Maybe I can help fix this error.

cat coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_490000.ckpt.index | grep -a conv4

@zjsjack
Copy link
Author

zjsjack commented Jun 26, 2017

The output is not readable. Just copy&paste as below:

ubuntu@ip-172-31-18-140:~/workspace/visual_search/visual_search$ cat coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_490000.ckpt.index | grep -a conv4
(???5>?? %/Mome? ???
(???5????
4/conv4_1/bias? ???
(?5oI?? /Moment? ???
(?5[5
?%vgg_16/conv4/conv4_1/wei? ???
(???5?Ӆ? %/Mome? ???(???5??N2/bias? ???(?5Y?9 /Moment? ???(?5eS?%wei? ???(???5 խw%/Mome? ?˺(???5??3/bias? ???(?5??? /Moment? ???(?5.ʎ%wei? ???(???5?8?? %/Mome? ??(???5??0?
5/conv5_1/bias? ???"(?5?O? /Moment? ???"(?5?;$#%wei? ???"(???5j mY %/Mome? ???'(???5?9a2/bias? ???+(?5???? /Moment? ???+(?56?|?%vgg_16/conv5/conv5_2/wei? ???+(???52?&d %/Mome? ???0(???5???3/bias? ???4(?5?^? /Moment? ???4(?5̓?z%wei? ???4(???5?tb? %/Mome? ?˻9(???5Pv??

@tuan3w
Copy link
Owner

tuan3w commented Jun 26, 2017

sorry, I point to wrong file.

cat coco_2014_train+coco_2014_valminusminival/vgg16_faster_rcnn_iter_490000.ckpt.meta | grep -a conv4

@zjsjack
Copy link
Author

zjsjack commented Jun 26, 2017

value0B$vgg_16/conv4/conv4_3/biases/Momentum*
save_1/Assign_38Assign$vgg_16/conv4/conv4_3/biases/Momentumsave_1/RestoreV2_38*
" loc:@vgg_16/conv4/conv4_3/biases*
value(Bvgg_16/conv4/conv4_3/weights*

save_1/Assign_39Assignvgg_16/conv4/conv4_3/weightssave_1/RestoreV2_39*
#!loc:@vgg_16/conv4/conv4_3/weights*
value1B%vgg_16/conv4/conv4_3/weights/Momentum*
save_1/Assign_40Assign%vgg_16/conv4/conv4_3/weights/Momentumsave_1/RestoreV2_40*
#!loc:@vgg_16/conv4/conv4_3/weights*
8vgg_16/conv4/conv4_1/kernel/Regularizer/l2_regularizer:0
8vgg_16/conv4/conv4_2/kernel/Regularizer/l2_regularizer:0
8vgg_16/conv4/conv4_3/kernel/Regularizer/l2_regularizer:0
vgg_16/conv4/conv4_1/weights:0#vgg_16/conv4/conv4_1/weights/Assign#vgg_16/conv4/conv4_1/weights/read:0
vgg_16/conv4/conv4_1/biases:0"vgg_16/conv4/conv4_1/biases/Assign"vgg_16/conv4/conv4_1/biases/read:0
vgg_16/conv4/conv4_2/weights:0#vgg_16/conv4/conv4_2/weights/Assign#vgg_16/conv4/conv4_2/weights/read:0
vgg_16/conv4/conv4_2/biases:0"vgg_16/conv4/conv4_2/biases/Assign"vgg_16/conv4/conv4_2/biases/read:0
vgg_16/conv4/conv4_3/weights:0#vgg_16/conv4/conv4_3/weights/Assign#vgg_16/conv4/conv4_3/weights/read:0
vgg_16/conv4/conv4_3/biases:0"vgg_16/conv4/conv4_3/biases/Assign"vgg_16/conv4/conv4_3/biases/read:0
'vgg_16/conv4/conv4_1/weights/Momentum:0,vgg_16/conv4/conv4_1/weights/Momentum/Assign,vgg_16/conv4/conv4_1/weights/Momentum/read:0
&vgg_16/conv4/conv4_1/biases/Momentum:0+vgg_16/conv4/conv4_1/biases/Momentum/Assign+vgg_16/conv4/conv4_1/biases/Momentum/read:0
'vgg_16/conv4/conv4_2/weights/Momentum:0,vgg_16/conv4/conv4_2/weights/Momentum/Assign,vgg_16/conv4/conv4_2/weights/Momentum/read:0
&vgg_16/conv4/conv4_2/biases/Momentum:0+vgg_16/conv4/conv4_2/biases/Momentum/Assign+vgg_16/conv4/conv4_2/biases/Momentum/read:0
'vgg_16/conv4/conv4_3/weights/Momentum:0,vgg_16/conv4/conv4_3/weights/Momentum/Assign,vgg_16/conv4/conv4_3/weights/Momentum/read:0
&vgg_16/conv4/conv4_3/biases/Momentum:0+vgg_16/conv4/conv4_3/biases/Momentum/Assign+vgg_16/conv4/conv4_3/biases/Momentum/read:0
vgg_16/conv4/conv4_1/weights:0#vgg_16/conv4/conv4_1/weights/Assign#vgg_16/conv4/conv4_1/weights/read:0
vgg_16/conv4/conv4_1/biases:0"vgg_16/conv4/conv4_1/biases/Assign"vgg_16/conv4/conv4_1/biases/read:0
vgg_16/conv4/conv4_2/weights:0#vgg_16/conv4/conv4_2/weights/Assign#vgg_16/conv4/conv4_2/weights/read:0
vgg_16/conv4/conv4_2/biases:0"vgg_16/conv4/conv4_2/biases/Assign"vgg_16/conv4/conv4_2/biases/read:0
vgg_16/conv4/conv4_3/weights:0#vgg_16/conv4/conv4_3/weights/Assign#vgg_16/conv4/conv4_3/weights/read:0
vgg_16/conv4/conv4_3/biases:0"vgg_16/conv4/conv4_3/biases/Assign"vgg_16/conv4/conv4_3/biases/read:0
$TRAIN/vgg_16/conv4/conv4_1/weights:0
#TRAIN/vgg_16/conv4/conv4_1/biases:0
$TRAIN/vgg_16/conv4/conv4_2/weights:0
#TRAIN/vgg_16/conv4/conv4_2/biases:0
$TRAIN/vgg_16/conv4/conv4_3/weights:0
#TRAIN/vgg_16/conv4/conv4_3/biases:0
vgg_16/conv4/conv4_1/weights:0
vgg_16/conv4/conv4_1/biases:0
vgg_16/conv4/conv4_2/weights:0
vgg_16/conv4/conv4_2/biases:0
vgg_16/conv4/conv4_3/weights:0
vgg_16/conv4/conv4_3/biases:0

@tuan3w
Copy link
Owner

tuan3w commented Jun 26, 2017 •
edited
Loading

Unluckly, you have to fix all the variable scope names manually in (https://github.com/tuan3w/visual_search/blob/master/visual_search/lib/nets/vgg16.py) to fit names as above.
for example:

 if cfg.TRAIN.BIAS_DECAY:
      bias = tf.get_variable("bias", initializer=phb, dtype=tf.float32, trainable=trainable)
    else:
      bias = tf.get_variable("bias", initializer=phb, regularizer=tf.no_regularizer, dtype=tf.float32, trainable=trainable)

should change to:

 if cfg.TRAIN.BIAS_DECAY:
      bias = tf.get_variable("biases", initializer=phb, dtype=tf.float32, trainable=trainable)
    else:
      bias = tf.get_variable("biases", initializer=phb, regularizer=tf.no_regularizer, dtype=tf.float32, trainable=trainable)

Does anyone here can share the old pretrained file ?

@zjsjack
Copy link
Author

zjsjack commented Jul 3, 2017

After manually change the source code, still got the Error as blew:
NotFoundError (see above for traceback): Key vgg16_default/conv3_2/weights not found in checkpoint

Not sure if you have the old pretrained file to share?

@tuan3w
Copy link
Owner

tuan3w commented Jul 3, 2017 •
edited
Loading

Hi @zjsjack,
I used pre-trained version provided by @endernewton at the time I write it. I don't have any copy of it in my machine now. If you can, try to adapt the code with new version of tf-faster-rnn. I used fc7 as image feature .
I don't have much time right now, so I can't fix the problem for you.
If you want to build own image search, you should consider this repo.
Thanks

@tuan3w
Copy link
Owner

tuan3w commented Jul 26, 2017

Hi @zjsjack,
You can get pre-trained model here.
Thanks

@zjsjack
Copy link
Author

zjsjack commented Jul 28, 2017

Hi, thanks for sharing the model. But one of the .tgz file can not be decompressed.
ubuntu@ip-172-31-18-140:/mnt/train2014$ tar xvf faster_rcnn_models.tgz
faster_rcnn_models/
faster_rcnn_models/coco_2014_train+coco_2014_valminusminival/
faster_rcnn_models/coco_2014_train+coco_2014_valminusminival/default/
faster_rcnn_models/coco_2014_train+coco_2014_valminusminival/default/vgg16_faster_rcnn_iter_490000.ckpt.index
faster_rcnn_models/coco_2014_train+coco_2014_valminusminival/default/vgg16_faster_rcnn_iter_490000.pkl
faster_rcnn_models/coco_2014_train+coco_2014_valminusminival/default/vgg16_faster_rcnn_iter_490000.ckpt.meta
faster_rcnn_models/coco_2014_train+coco_2014_valminusminival/default/vgg16_faster_rcnn_iter_490000.ckpt.data-00000-of-00001

gzip: stdin: invalid compressed data--format violated
tar: Unexpected EOF in archive
tar: Unexpected EOF in archive
tar: Error is not recoverable: exiting now

And with these partial data, I got the below new error msg.
Loading model check point from /mnt/train2014/faster_rcnn_models/coco_2014_train+coco_2014_valminusminival/default/vgg16_faster_rcnn_iter_490000.ckpt
W tensorflow/core/framework/op_kernel.cc:975] Out of range: Read less bytes than requested
W tensorflow/core/framework/op_kernel.cc:975] Out of range: Read less bytes than requested
W tensorflow/core/framework/op_kernel.cc:975] Out of range: Read less bytes than requested
W tensorflow/core/framework/op_kernel.cc:975] Out of range: Read less bytes than requested
W tensorflow/core/framework/op_kernel.cc:975] Out of range: Read less bytes than requested
W tensorflow/core/framework/op_kernel.cc:975] Out of range: Read less bytes than requested
W tensorflow/core/framework/op_kernel.cc:975] Out of range: Read less bytes than requested
W tensorflow/core/framework/op_kernel.cc:975] Out of range: Read less bytes than requested
W tensorflow/core/framework/op_kernel.cc:975] Data loss: Checksum does not match: stored 1475381895 vs. calculated on the restored bytes 4182060358
W tensorflow/core/framework/op_kernel.cc:975] Data loss: Checksum does not match: stored 3119027912 vs. calculated on the restored bytes 4182060358
W tensorflow/core/framework/op_kernel.cc:975] Data loss: Checksum does not match: stored 4197277905 vs. calculated on the restored bytes 2076928609
W tensorflow/core/framework/op_kernel.cc:975] Data loss: Checksum does not match: stored 2777681086 vs. calculated on the restored bytes 3183351207
W tensorflow/core/framework/op_kernel.cc:975] Data loss: Checksum does not match: stored 3304795247 vs. calculated on the restored bytes 2076928609
W tensorflow/core/framework/op_kernel.cc:975] Out of range: Read less bytes than requested

@tuan3w
Copy link
Owner

tuan3w commented Jul 28, 2017

You can get it now.

@xiesibo
Copy link

xiesibo commented Aug 16, 2017

Hi, thanks for sharing the model.But I can not find cython_nms.
[root@134 visual_search]# python index_es.py --weight $WEIGHT_PATH --model_path $MODEL_PATH --input $INPUT
Traceback (most recent call last):
File "index_es.py", line 14, in
from extractor import Extractor
File "/usr/local/visual_search_master/visual_search/extractor.py", line 2, in
from model.test import extract_regions_and_feats
File "/usr/local/visual_search_master/visual_search/lib/model/test.py", line 14, in
from utils.cython_nms import nms, nms_new
ImportError: No module named cython_nms

@tuan3w
Copy link
Owner

tuan3w commented Aug 16, 2017

Hi @xiesibo ,
You have to build cython module first.

$ cd visual_search/visual_search/lib/
$ make

@xiesibo
Copy link

xiesibo commented Aug 16, 2017

Hi @tuan3w ,
Can I run it on platform where there is no GPU? Because I find error when I make it
[root@134 lib]# sudo make
python setup.py build_ext --inplace
running build_ext
skipping 'utils/bbox.c' Cython extension (up-to-date)
skipping 'utils/nms.c' Cython extension (up-to-date)
skipping 'nms/cpu_nms.c' Cython extension (up-to-date)
skipping 'nms/gpu_nms.cpp' Cython extension (up-to-date)
building 'nms.gpu_nms' extension
{'gcc': ['-Wno-unused-function'], 'nvcc': ['-arch=sm_52', '--ptxas-options=-v', '-c', '--compiler-options', "'-fPIC'"]}
/usr/local/cuda/bin/nvcc -I/usr/lib64/python2.7/site-packages/numpy/core/include -I/usr/local/cuda/include -I/usr/include/python2.7 -c nms/nms_kernel.cu -o build/temp.linux-x86_64-2.7/nms/nms_kernel.o -arch=sm_52 --ptxas-options=-v -c --compiler-options '-fPIC'
nvcc fatal : Value 'sm_52' is not defined for option 'gpu-architecture'
error: command '/usr/local/cuda/bin/nvcc' failed with exit status 1

@tuan3w
Copy link
Owner

tuan3w commented Aug 16, 2017

This error means that only extension for gpu cannot be built. So it still works fine with CPU. The author also claim that, see
Can you check if it works:

$ cd visual_search/visual_search/lib/
$ python -c 'from utils.cython_nms import nms, nms_new'

If no exception throws, it's ok

@tuan3w tuan3w closed this as completed Sep 19, 2017
@tuan3w tuan3w mentioned this issue Nov 15, 2017
Closed
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
3 participants
@tuan3w @zjsjack @xiesibo