csailvision / places365 Goto Github PK

Hello，everyone！ Running run_placesCNN_unified.py in this repository I run into this following error:

Traceback (most recent call last):
File "run_placesCNN_unified.py", line 124, in
model = load_model()
File "run_placesCNN_unified.py", line 94, in load_model
model = wideresnet.resnet18(num_classes=365)
File "/home/user/zou/code/places365/wideresnet.py", line 164, in resnet18
model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
File "/home/user/zou/code/places365/wideresnet.py", line 120, in init
nn.init.constant_(m.weight, 1)
AttributeError: 'module' object has no attribute 'constant_'

Any advice about that? Thanks in advance!

Hi, can you share the labels.pkl file for the for the hybrid model with 1365 categories. I want to build a docker container with resnet hybrid model in that. but I didn't find any labels.pkl file for it. It will be very helpful if you can share the file.

Hi all! Running run_placesCNN_basic.py in this repository I run into this following error:
Traceback (most recent call last):
File "", line 4, in
File "/usr/local/lib/python3.5/dist-packages/torch/serialization.py", line 231, in load
return _load(f, map_location, pickle_module)
File "/usr/local/lib/python3.5/dist-packages/torch/serialization.py", line 379, in _load
result = unpickler.load()
UnicodeDecodeError: 'ascii' codec can't decode byte 0xc3 in position 875: ordinal not in range(128)

Any advice? Thanks in advance!

A Keras (2.0) implementation of the pre-trained VGG 16 convolutional-layer CNN models on Places365-Standard

Hi, is it possible to release the train_val.prototxt / solver.prototxt that you used to train your networks? I could not find them.

Thanks!

why run the run_placesCNN_unified.py show this error.

Traceback (most recent call last):
File "/home/xyx/Downloads/places365 new/run_placesCNN_unified.py", line 132, in
input_img = V(tf(img).unsqueeze(0), volatile=True)
File "/home/xyx/anaconda2/lib/python2.7/site-packages/torchvision/transforms.py", line 29, in call
img = t(img)
File "/home/xyx/anaconda2/lib/python2.7/site-packages/torchvision/transforms.py", line 139, in call
ow = int(self.size * w / h)
TypeError: unsupported operand type(s) for /: 'tuple' and 'int'

I'm attempting to use the mean file with the vgg16 network and the caffe Classifier class and have encountered a shape issue.
Do I even need the mean file here? According to #3 it appears maybe not.

Example:

import caffe
MODEL_PROTOTXT = 'deploy_vgg16_places365.prototxt'
MODEL_TRAINED = 'vgg16_places365.caffemodel'
MEAN_FN = 'places365CNN_mean.binaryproto'

def load_mean():
    mean_fh = open(MEAN_FN, 'rb')
    blob = caffe.proto.caffe_pb2.BlobProto()
    mean_string = mean_fh.read()
    blob.ParseFromString(mean_string)
    mean_fh.close()
    mean = caffe.io.blobproto_to_array(blob)
    return mean

mean = load_mean()
c = caffe.Classifier(MODEL_PROTOTXT, MODEL_TRAINED) # works
c = caffe.Classifier(MODEL_PROTOTXT, MODEL_TRAINED, mean=mean) # raises ValueError: Mean shape invalid

I downloaded the pytorch pretrained models, but failed to unarchive the '.tar' files, and I have tried any method I can think of.
It seems that there is something wrong with the '.tar' files.

Can someone familiar with how the models were trained verify that the preprocessing in run_placesCNN_basic.py and run_placesCNN_unified.py are both correct?

In run_placesCNN_basic.py it's:

centre_crop = trn.Compose([
        trn.Scale(256),
        trn.CenterCrop(224),
        trn.ToTensor(),
        trn.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

That is, rescaling the smaller edge of the image to 256 and then taking the center 224 pix?

In run_placesCNN_unified.py, it's:

    tf = trn.Compose([
        trn.Scale((224,224)),
        trn.ToTensor(),
        trn.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ])

That is, squashing the image to 224x224. Presumably the appropriate preprocessing should be the one that's closes to what was used for training.

Thanks

Hello guys,
I could not find the prototxt file for vgg16_hybrid1365. Where can I find it?
thanks

Hi,
I want to use your models as pre-trained weight for other tasks. However, I am not sure how you processed your input images, as I saw you have :

1. ResNet152-places365 fine-tuned from ResNet152-ImageNet
2. ResNet152-places365 trained from scratch using Torch

As the author of ResNet152-ImageNet, mentioned that he used BGR, which is expected if using caffe... So i guess the model in 1, should also use BGR? While the model in 2 is not clear to me, as it was trained from scratch and used Torch...

It would be nice if you can tell me it directly. Thanks!

When I run

docker build -t places365_container .

I get this error:

---> Running in db1ece13a39a
--2017-01-31 18:04:36--  http://places2.csail.mit.edu/models_places365/alexnet_places365.caffemodel
Resolving places2.csail.mit.edu (places2.csail.mit.edu)... failed: Name or service not known.
wget: unable to resolve host address 'places2.csail.mit.edu'
INFO[0045] The command [/bin/sh -c wget http://places2.csail.mit.edu/models_places365/alexnet_places365.caffemodel] returned a non-zero code: 4 

Any advice?

Thanks!

Hi all! I am dealing with a very annoying issue and as I am new at Python I don't know how to go thought.
Shortly I am running places365 over many images, and it looks that it can't handle with black and white images, since I got the following error:
RuntimeError: Need input of dimension 4 and input.size[1] == 3 but got input to be of shape: [1 x 1 x 224 x 224] at /pytorch/torch/lib/THNN/generic/SpatialConvolutionMM.c:47
As I would not want to delete all the black and white images, could someone suggest me few lines of code that would solve the problem?
Thanks!

I run" CUDA_VISIBLE_DEVICES=0 python run_placesCNN_basic.py" on my server .

but I found the program runs on cpu.

How can I use GPU to do test?

Would it be possible to release the single-crop validation performance? This would make it easier to verify models that have been converted from caffe [without writing a complicated data loader].

Thanks!

Hi, I am sorry for reopening this issue, but I did not get response.

Hi guys,

this is an excellent work. I really like to unified version of prediction multiple attributes. Do you also have the whole_wideresnet18_places365 model used in the unified version in Caffe? Or can you give me a hint how to convert it? Can I also use a different network in the unified prediction?

Does anyone know why the demo here: http://places2.csail.mit.edu/demo.html and the run_placesCNN_unified.py . I tried testing with the same images mentiones in the README.md file, and the result differs a lot:

According to the README file and places demo website, the result is as follows:

RESULT ON http://places.csail.mit.edu/demo/6.jpg
--TYPE: indoor
--SCENE CATEGORIES:
0.690 -> food_court
0.163 -> cafeteria
0.033 -> dining_hall
0.022 -> fastfood_restaurant
0.016 -> restaurant
--SCENE ATTRIBUTES:
no horizon, enclosed area, man-made, socializing, indoor lighting, cloth, congregating, eating, working
Class activation map is output as cam.jpg

But when I run run_placesCNN_unified.py on my server, I get the following result:

RESULT ON http://places.csail.mit.edu/demo/6.jpg
--TYPE OF ENVIRONMENT: indoor
--SCENE CATEGORIES:
0.511 -> food_court
0.085 -> fastfood_restaurant
0.083 -> cafeteria
0.040 -> dining_hall
0.021 -> flea_market/indoor
--SCENE ATTRIBUTES:
no horizon, enclosed area, man-made, socializing, indoor lighting, cloth, congregating, eating, working
Class activation map is saved as cam.jpg

Does anyone knows which model the current demo website is using ?

Hi guys,

this is an excellent work. I really like to unified version of prediction multiple attributes. Do you also have the whole_wideresnet18_places365 model used in the unified version in Caffe? Or can you give me a hint how to convert it? Can I also use a different network in the unified prediction?

I wanted to reproduce the results in the paper. Where can I find the ground truth results for the test data set of Places365 Standard? I am currently trying it on small images http://places2.csail.mit.edu/download.html.

I have downloaded the image list and annotation file of Places365 as given in Places365 dev kit, but it only has the file names of the test data set, and not the ground truth.

I downloaded the 'Places365 Development kit', and I found that there were only the 'categories_places365.txt' and 'categories_hybrid1365.txt' for labels file of training data. So, is there a labels file for validating data?

The repo contains this mean file: places365CNN_mean.binaryproto

But in /docker/run_scene.py this line of code seems to imply using the ILSVRC 2012 mean file.

transformer.set_mean('data', np.load(
    'python/caffe/imagenet/ilsvrc_2012_mean.npy').mean(1).mean(1))

Which is the right one to use?

Hi, What's the inputs' range for the provided Torch models? 0-1 or 0-255?

Hello,

For indoor scene classification, I've run 'run_placeCNN_unified.py' code and used 'wideresnet18' model you attached.

However, I do not know the difference between 'wideresnet18_places365' and 'resnet18_places365' except for maxpooling layer before convolution layer. So, I do not know if it is really wide residual model.

I want to know why you have used 'wideresnet18' model in 'run_placesCNN_unified.py' and named 'wideresnet'.

I would be grateful if you answer my question.
Thanks.

If we want to test the performance of the deep features of Places365-CNNs, how to set up new training for new dataset such as Scene15?

I'm using pytorch0.4 with python2.7 and had some problems when loading the pre-trained model. I had change the model download link by delete '_python36' and download the model for python2.7.

The log is below:
Traceback (most recent call last): File "run_placesCNN_basic.py", line 77, in <module> logit = model.forward(input_img) File "/home/public/anaconda2/lib/python2.7/site-packages/torchvision/models/resnet.py", line 140, in forward x = self.bn1(x) File "/home/public/anaconda2/lib/python2.7/site-packages/torch/nn/modules/module.py", line 491, in __call__ result = self.forward(*input, **kwargs) File "/home/public/anaconda2/lib/python2.7/site-packages/torch/nn/modules/batchnorm.py", line 49, in forward self.training or not self.track_running_stats, self.momentum, self.eps) File "/home/public/anaconda2/lib/python2.7/site-packages/torch/nn/modules/module.py", line 532, in __getattr__ type(self).__name__, name)) AttributeError: 'BatchNorm2d' object has no attribute 'track_running_stats'

How to solve this problem without downgrade the pytorch?

when i run the command:
docker build -t places365_container .

I got the errors as the follow:
error pulling image configuration: Get https://dseasb33srnrn.cloudfront.net/registry-v2/docker/registry/v2/blobs/sha256/0b/0b9fc622f1b7840a160ddb2377fcb085109e26983fe2dcca4ca5d52902f6a65d/data?Expires=1523203761&Signature=hFnXCgLk0ivjhqwl4oXsC~kWnHEkyFCSR3Oy8N2S41MkQc1m~SmoxdW7ZwupbHGLpGljQBvHkGVkgqL9TpcttH8Wt4ug20dqtrE8V5NJ6RmzWXgTFsE-hOXQMJKuy56iV9zsgnAWUAOl1lXGBmyhst7nFiRi21K1I7QgUaITlkM_&Key-Pair-Id=APKAJECH5M7VWIS5YZ6Q: net/http: TLS handshake timeout

I've try to access the image file, and https://dseasb33srnrn.cloudfront.net/ can't be accessed
Thanks

Hi all!
while running run_placesCNN_basic.py, this error comes up
Traceback (most recent call last):
File "run_placesCNN_basic.py", line 31, in
checkpoint = torch.load(model_file, map_location=lambda storage, loc: storage)
File "/home/pytorch_py35/lib/python3.5/site-packages/torch/serialization.py", line 267, in load
return _load(f, map_location, pickle_module)
File "/home/pytorch_py35/lib/python3.5/site-packages/torch/serialization.py", line 412, in _load
magic_number = pickle_module.load(f)
_pickle.UnpicklingError: invalid load key, '<'.
Any advice? Thanks!

I have tried the run_placesCNN_basic.py script with the densenet161 architecture. It works well when I choose a ResNet architecture. I have attached the error log

Do you have any suggestions why this particular model fails to load?

Thank you

Hi,
For those pre-trained models, where can we find the tran_val.prototxt so that we get to know the transform_param ?

I need it to do preprocessing for input images

thanks

I think the training data should also do the scale operation which would help.

Is it possible to describe what preprocessing was done on the data for use in Caffe please?

I modify the code in run_placesCNN_unified.py to input more than one pictures and just need to load the model once, I write a function which input is picture_path, and output is the results.But is seems that SCENE ATTRIBUTES always the same, cause the features_blobs is used in load_model(), so it doesn't modify when the picture is varying.Does any body else know how to fix this and don't need to load the model every time?

For places205 a mapping was provided from the canonical label to indoor/outdoor.

I was wondering if that also exists for places365, as I can't seem to find it.

The pretrained model by mxnet ,resnet-50' accuracy is only 0.3112 and resnet-152' accuracy is only 0.3355.
the site is https://github.com/apache/incubator-mxnet/tree/master/example/image-classification.
So, I want use pretrained model by pytorch ,but ,how to convert pytorch's model to mxnet's model.
Thank you

I'm trying to run the train places py code provided and after a whole weekend of waiting, I got this error:

./train_places_cnn.py: line 23: syntax error near unexpected token `('
./train_places_cnn.py: line 23: 'model_names = sorted(name for name in models.dict'

I have not modified train_places_cnn.py.

I was hoping to get a snapshot.caffemodel file generated so I can plug this model into the nvidia GRE. I have had success testing with the places 205 model, but I was not able to locate a places365 snapshot file on the github or some direct route to downloading it.

It occur Apache Error

Not Found
The requested URL /models_places365/densenet161_places365_python36.pth.tar was not found on this server.

What is problem?

It appears the pre-trained pytorch model consistently predicts a label different from what it should. For example, instead of predicting 2, the model predicts 10; instead of predicting 3, it predicts 100.

Here is some code to reproduce this issue.

import argparse
import torch
import torch.backends.cudnn as cudnn
import torch.nn.functional as F
import torchvision.datasets as dset
import torchvision.transforms as transforms
import torchvision.models as models
from torch.autograd import Variable
import numpy as np

parser = argparse.ArgumentParser(description='Demo of bug',
                                 formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Optimization options
parser.add_argument('--batch_size', '-b', type=int, default=100, help='Batch size.')
parser.add_argument('--test_bs', type=int, default=100)
# Acceleration
parser.add_argument('--ngpu', type=int, default=1, help='0 = CPU.')
parser.add_argument('--prefetch', type=int, default=5, help='Pre-fetching threads.')
args = parser.parse_args()

mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
transform = transforms.Compose(
    [transforms.Scale(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize(mean, std)])

test_data = dset.ImageFolder(root="/share/data/lang/users/dan/datasets/places365/val", transform=transform)
test_loader = torch.utils.data.DataLoader(test_data, batch_size=args.test_bs, shuffle=False,
                                          num_workers=args.prefetch, pin_memory=True)

net = models.resnet50(num_classes=365)
checkpoint = torch.load('/share/data/lang/users/dan/.torch/models/resnet50_places365.pth')
state_dict = {str.replace(k,'module.',''): v for k,v in checkpoint['state_dict'].items()}
net.load_state_dict(state_dict)
net.eval()

for p in net.parameters():
    p.volatile = True

if args.ngpu > 1:
    net = torch.nn.DataParallel(net, device_ids=list(range(args.ngpu)))
if args.ngpu > 0:
    net.cuda()

np.random.seed(1)
torch.manual_seed(1)
if args.ngpu > 0:
    torch.cuda.manual_seed(1)

cudnn.benchmark = True  # fire on all cylinders

to_np = lambda x: x.data.cpu().numpy()

for batch_idx, (data, target) in enumerate(test_loader):
    data = Variable(data.cuda(), volatile=True)

    output = net(data)
    smax = to_np(F.softmax(output))
    # batch_idx = target since batch size = 100, the size of each val folder
    print(np.argmax(smax, axis=1), batch_idx)

Classes 0 and 1 are correctly predicted. After that, 2 -> 10, 3 -> 100, 4 -> 101, 5 -> 102, 6 -> 103, ...

[  0   0   0 170 174   0  18 174 293 170   0 170 293   0 293 174   0   0
 174   0   0   0   0 174 174 293   0   0 169 170   0 293   0   0 293 293
 186   0 174   0 174   0   0 207 293   0 293   0 192   0   0   0   0   0
   0   0   0   0   0 174   0   0   0   0   0 140   0   0   0   0 293 293
   0   0   0   0 293   0 174 293 174 174 293   0 293   0   0   0 293 140
   0 293   0 174   0 293 293 293 174   0] 0
[ 38   1   1 346   1   1  58   1   1   1   1  98   1   1  98   1   1   1
   1   1   1   1  27   1   1   1   1   1   1   1   1   1   1   1   1   1
   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1
   1   1 336   1   1   1   1   1 336   1   1   1   1   1   1   1  98   1
   1   1   1   1   1  55   1  98   1   1   1   1   1   1   1   1   1   2
 135  55   1   1   1   1 298   1   1  82] 1
[ 10 113  10  10 278  84  84 109 347  10 347 347 113  66 218  10  10  10
  10  10 347 218  10  10  10  10 347  10 292 229  10  10 347  10  78 347
 347  10  10  10  66  10  10  10 347  10 349  12  10  10  10  10  10  10
  10 283  10  66  10  10  10 218  10  81  10  10  66 347 347  33  10  10
 292  10 292  10 347 307 127  10  10 347 347  66  10 347  10  10  10  10
  10  85 273  10 347  10  10  10  10  10] 2
[100 238 100 176 176 100 246 100 100 176  56 100 100 235 176 244 182 100
 298 100 264 176 176 244 100 177 264 100 210 246 100 176 210 246 100 100
 244 246 100 212 100 244 102 100 100  63 244 238 244 244 202 100 100 100
 176 246 211 177 100 100 100 177 176 246 100  54 176 100 100 100 100 100
 100 246 244  33 100 238 100 100  19 100 176  20 176 100 100 246 244 246
 100 100 176 100 100 100 100 244 100 100] 3
[ 38 210 101 177 210 101 101 102 210 280 210  16 101 102  27  38 101 211
 101 101 101  27 101 210 101 211 211 101  92 101 101 101  27 102 211  38
 101 101 101 101 101 211 101 101 101 210  27 241 102 211  38 102 101 235
 101 101 211 211 102 211 235 101 211 101 101 210 101 211  27 211 101  18
 101 211 101 244 101  27 101 211 101 101  27  38 101 101 101 101 210 101
 210 211  27 101 211 235  38 101 101 101] 4
[102 102 102 102 101  38 211 102 102 211 102 121 211  38 102 102 102 102
 102 219 102 211 102 246  27 121 102 102  27 121 102 102 102  38 102 244
 102 101 102 102 244 120 210 102 102 102 102 211 100 102 102 102 235 212
 101 102  20 210 102 102 102 128 102 101 102 102 102 101 102 102 102 102
 102 211 102 102 210 102 210 101 211 102 102 102 102 102 102 102 102 102
 210 121 102 102 102 248 244  38 101 102] 5
[103  13 308 192 216 218 192  66 103 136 206 136 226 103 103 294 307 245
 255 103 125 103 103 192   8 216 308 216 125 103 143 165 103 340 256 103
  76 127 125 256 339 103 103 247 103 103 127 103 103 136 103 103 103 103
 103 143  13 103 103   7 103 171 103 338 307 103 136 103 103 103 131 103
 340 103 318 127 228 103  41 103  13  71 103 103 192 131 103 143  19 103
 331 166 339  43 103   5 221 223 103 169] 6

With these commands I got the validation data suitable for pytorch's ImageFolder.

import os
for i in range(365):
    os.mkdir('./val/' + str(i))
with open('places365_val.txt') as f:
    for line in f:
        line = line.split()
        os.rename('./val/images/' + line[0], "./val/"+str(line[1])+'/'+line[0])

Hi, is there a ground truths file for Hybrid models?

Hi,
To be able to use the vgg16_places365.caffemodel, I need the corresponding mean data file (used for pre-processing normalization). Where can I find the places365CNN_mean.binaryproto file?

Regards,
Abhishek

Excuse me,when i download tar from website, the downloaded file can not be untar. How can i fix this problem? Thanks!

The models are downloaded from websites below.

PyTorch Places365 models: AlexNet, ResNet18, ResNet50, DenseNet161. The models are trained in Python2.7+PyTorch 0.2, when the models are being loaded in python3, you might encounter UnicodeDecodeError, see this issue. Run basic code to get the scene prediction from PlacesCNN:

how to save a pytorch model to xxx.pth.tar ? like you

Hi,

I downloaded the two ResNet model from you:

1. ResNet152-places365 trained from scratch using Torch
2. ResNet50-places365 trained from scratch using Torch

As I use tensorflow, so I have to parse the model first...(I use pytorch torch.utils.serialization.load_lua()).
And I was confused by the structure of ResNet50, because I saw it does batch normalization after element-wise addition of original and shortcut.

As an example, I pasted the representation of the 2nd block of conv2_x, which corresponds to part(in your code):

layer {
	bottom: "res2a_branch1"
	bottom: "res2a_branch2c"
	top: "res2a"
	name: "res2a"
	type: "Eltwise"
}

nn.Sequential {
  [input -> (0) -> (1) -> output]
  (0): torch.legacy.nn.ConcatTable.ConcatTable {
    input
      |`-> (0): nn.Sequential {
      |      [input -> (0) -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
      |      (0): nn.SpatialBatchNormalization  
      |      (1): TorchObject(cudnn.ReLU, {'output': [torch.FloatTensor with no dimension]
      |      , '_type': 'torch.FloatTensor', 'train': True, 'inplace': True, 'gradInput': [torch.FloatTensor with no dimension]
      |      })
      |      (2): TorchObject(cudnn.SpatialConvolution, {'groups': 1, '_type': 'torch.FloatTensor', 'weight': 

... (3) - (5)
      |      (6): nn.SpatialBatchNormalization
      |      (7): TorchObject(cudnn.ReLU, {'output': [torch.FloatTensor with no dimension]
      |      , '_type': 'torch.FloatTensor', 'train': True, 'inplace': True, 'gradInput': [torch.FloatTensor with no dimension]
      |      })
      |      (8): TorchObject(cudnn.SpatialConvolution, {'groups': 1, '_type': 'torch.FloatTensor', 
... 
      |      [torch.FloatTensor of size 256x64x1x1]
      |      , 'train': True, 'kW': 1, 'padW': 0, 'output': [torch.FloatTensor with no dimension]
      |      , 'dW': 1, 'nInputPlane': 64})
      |    }
      |`-> (1): nn.Identity
       +. -> output
  }
  (1): nn.CAddTable
}

But according to your code, (0): nn.SpatialBatchNormalization should be before addition, right? After checking model ResNet152,
I found its implementation seems more consistent with the code in deploy_resnet152_places365.prototxt, as following:

nn.Sequential {
  [input -> (0) -> (1) -> (2) -> output]
  (0): torch.legacy.nn.ConcatTable.ConcatTable {
    input
      |`-> (0): nn.Sequential {
      |      [input -> (0) -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> output]
      |      (0): TorchObject(cudnn.SpatialConvolution, {'groups': 1, '_type': 'torch.FloatTensor', 'weight': 

... (1) - (6)

      |      , 'dW': 1, 'nInputPlane': 64})
      |      (7): nn.SpatialBatchNormalization
      |    }
      |`-> (1): nn.Identity
       +. -> output
  }
  (1): nn.CAddTable
  (2): TorchObject(cudnn.ReLU, {'output': [torch.FloatTensor with no dimension]

The confirmation is important for me to reuse your model, thanks!

why in run_placesCNN_unified.py the cv2 is not import.

Could you provide the Torch version of following models, please? I couldn't load and convert them. Thanks

AlexNet-places365
GoogLeNet-places365
VGG16-places365
VGG16-hybrid1365
ResNet152-hybrid1365

Hi,
When I run the run_placesCNN_basic.py in Python3 [with PyTorch 0.2] the following error occurred:
$ python3 run_placesCNN_basic.py

Traceback (most recent call last):
  File "run_placesCNN_basic.py", line 26, in <module>
    model = torch.load(model_weight, map_location=lambda storage, loc: storage) # model trained in GPU could be deployed in CPU machine like this!
  File "/home/karami/anaconda3/lib/python3.6/site-packages/torch/serialization.py", line 231, in load
    return _load(f, map_location, pickle_module)
  File "/home/karami/anaconda3/lib/python3.6/site-packages/torch/serialization.py", line 379, in _load
    result = unpickler.load()
UnicodeDecodeError: 'ascii' codec can't decode byte 0xc3 in position 875: ordinal not in range(128)

would you please help me to fix this issue.

only find places365CNN_mean.binaryproto, is there a mean.binaryproto file for vgg16_hybrid.model?

Hi Would it be possible for you to post the parameters used for training alexnet/caffenet? I'm trying to fine my own network to reach your accuracy but so far only got to 0.47 for one crop validation set.
Thank you!

There is currently no indication of the license for the contents of this repository. Creative commons perhaps?

Hello!
I use python 2.7 and compiled distribution of caffe.
I try to run:
net = caffe.Net('places_model/deploy_googlenet_places365.prototxt', caffe.TEST)
But I get error:

[libprotobuf ERROR google/protobuf/text_format.cc:274] Error parsing text-format caffe.NetParameter: 7:1: Expected identifier. F0912 12:11:03.998572 19221 upgrade_proto.cpp:88] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: places_model/deploy_googlenet_places365.prototxt
*** Check failure stack trace: ***

I open proto file and dont found some syntax error.
What I can do to fix this? Thank you.

Traceback (most recent call last):                                                                                
File "train.py", line 155, in <module>                                                                          
main()                                                                                                        
File "train.py", line 150, in main                                                                              
trainer.train()                                                                                               
File "C:\myFile\code\image_scene_classification\CH\model\trainer.py", line 314, in train                        
self.train_epoch()                                                                                            
File "C:\myFile\code\image_scene_classification\CH\model\trainer.py", line 228, in train_epoch                  
self.validate()                                                                                               
File "C:\myFile\code\image_scene_classification\CH\model\trainer.py", line 123, in validate                     
output = self.model(inputs)                                                                                   
File "C:\Users\chmtt\Anaconda3\envs\SCENE\lib\site-packages\torch\nn\modules\module.py", line 491, in __call__  
result = self.forward(*input, **kwargs)                                                                       
File "C:\Users\chmtt\Anaconda3\envs\SCENE\lib\site-packages\torchvision\models\alexnet.py", line 43, in forward 
x = self.features(x)                                                                                          
File "C:\Users\chmtt\Anaconda3\envs\SCENE\lib\site-packages\torch\nn\modules\module.py", line 491, in __call__  
result = self.forward(*input, **kwargs)                                                                       
File "C:\Users\chmtt\Anaconda3\envs\SCENE\lib\site-packages\torch\nn\modules\container.py", line 91, in forward 
input = module(input)                                                                                         
File "C:\Users\chmtt\Anaconda3\envs\SCENE\lib\site-packages\torch\nn\modules\module.py", line 491, in __call__  
result = self.forward(*input, **kwargs)                                                                       
File "C:\Users\chmtt\Anaconda3\envs\SCENE\lib\site-packages\torch\nn\modules\conv.py", line 301, in forward     
self.padding, self.dilation, self.groups)                                                                     
RuntimeError: thnn_conv2d_forward is not implemented for type torch.ByteTensor                                    

and my code is
// this is for get the file_path
model_path = load_model_from_name('alexnet')

model= torch.load(model_path)

the output is gotten when call model(inputs)