huyng / tensorflow-vgg

Re-implementation of VGG Network in tensorflow

tensorflow-vgg

Re-implementation of VGG Network in tensorflow

setup

pip install pyyaml skimage skdata tensorflow-gpu

training

python train_model_simple.py experiment.yaml

training on multiple gpus

python train_model_parallel.py experiment.yaml

prediction

python predict.py dog.jpg

tensorflow-vgg's Issues

tensorflow new version problems

in vgg.py, line 21,
net = tf.sub(input_tensor, mu, name="input_mean_centered")
latest version of tensorflow (1.3) should be
net = tf.subtract(input_tensor, mu, name="input_mean_centered")

bug fixed on tf1.5

1. tf.pack ---> tf.stack
2. tf.softmax_cross_entropy_with_logits(logits, onehot_labels, name='xentropy') ---> tf.softmax_cross_entropy_with_logits(logits=logits, labels=onehot_labels, name='xentropy')
3. add conv3_3 : net = L.conv(net, name="conv3_3", kh=3, kw=3, n_out=256)

MetricsLogger

The function MetricsLogger is missing from tools.py

ReLU applied to logits

you're applying a relu to your logits

https://github.com/huyng/tensorflow-vgg/blob/master/vgg.py#L59

VGG15? I think in vgg.py, a conv3-256 is missing.

miss out one convolutional layer in VGG16

In vgg.py, block 3 only contains 2 conv layers and 1 pool layer. However, from the paper, it shows there should be 3 conv layers and 1 pool layer in block 3.

Ref: https://arxiv.org/pdf/1409.1556v6.pdf

Accuracy not that great

When I ran the code as it is, the accuracy (loss) wasn't that great. Anyone seeing such issues. I changed the batch size to 50. Apart from that no other change.

Cannot resize raw image when training.

Problem :
The bug

Cause:
size: A 1-D int32 Tensor of 2 elements: new_height, new_width. The new size for the images.

Tensorflow version

I hope you describe the environment that you checked.
I guess some of the functions that you used are not supported by the current version anymore.

predict picture error

The result is error when predict using model trainning from cifar-10.

ImportError: No module named 'sklearn.cross_validation' in python 3.5.6

'function' object has no attribute 'config_file'

(windows7+Python 3.6.4+TensorFlow 1.7)

When I run python train_model_simple.py experiment.yaml , an error occur.

Error:
'function' object has no attribute 'config_file'

they are happens in

parser = argparse.ArgumentParser()
parser.add_argument('config_file', help='YAML formatted config file')
args = parser.parse_args()
with open(args.config_file) as fp:
    config.update(yaml.load(fp))

How should I do,help me. Thanks

Python3.6.4 : No module named 'skdata'

(windows7+Python 3.6.4+TensorFlow 1.7)

When I run python train_model_simple.py experiment.yaml , an error occur.
Error:

from skdata.cifar10.dataset import CIFAR10

ModuleNotFoundError: No module named 'skdata'

How should I do,help me. Thanks

MetricsLogger references instead of StatLogger

Line 77 in train_model_simple.py:
log = tools.MetricsLogger(train_log_fpath)

Also line 57 in train_model_parallel.py:
log = tools.MetricsLogger(train_log_fpath)

TOP1 error always 0 after several training steps

As the title indicated.

Code in train_model_simple.py is cannot work

the last few lines of train_model_simple.py has problems:

args = parser.parse_args()
with open(args.config_file) as fp:
    config.update(yaml.load(fp))

My environment : windows7+Anaconda3(python 3.6.4+Tensorflow 1.7)
Please help me,Thanks !