flyyufelix/densenet-keras

DenseNet Implementation in Keras with ImageNet Pretrained Models

DenseNet-Keras with ImageNet Pretrained Models

This is an Keras implementation of DenseNet with ImageNet pretrained weights. The weights are converted from Caffe Models. The implementation supports both Theano and TensorFlow backends.

To know more about how DenseNet works, please refer to the original paper

Densely Connected Convolutional Networks
Gao Huang, Zhuang Liu, Kilian Q. Weinberger, Laurens van der Maaten
arXiv:1608.06993

Pretrained DenseNet Models on ImageNet

The top-1/5 accuracy rates by using single center crop (crop size: 224x224, image size: 256xN)

Network	Top-1	Top-5	Theano	Tensorflow
DenseNet 121 (k=32)	74.91	92.19	model (32 MB)	model (32 MB)
DenseNet 169 (k=32)	76.09	93.14	model (56 MB)	model (56 MB)
DenseNet 161 (k=48)	77.64	93.79	model (112 MB)	model (112 MB)

Usage

First, download the above pretrained weights to the imagenet_models folder.

Run test_inference.py for an example of how to use the pretrained model to make inference.

python test_inference.py

Fine-tuning

Check this out to see example of fine-tuning DenseNet with your own dataset.

Requirements

• Keras 1.2.2 2.0.5
• Theano 0.8.2 or TensorFlow 0.12.0 1.2.1

Updates

• Keras 2.0.5 and TensorFlow 1.2.1 are supported

densenet-keras's Issues

TypeError: 'module' object is not callable

Version

DenseNet-Keras git:(master) ipython
Python 2.7.12 (default, Dec  4 2017, 14:50:18)
Type "copyright", "credits" or "license" for more information.

IPython 5.5.0 -- An enhanced Interactive Python.
?         -> Introduction and overview of IPython's features.
%quickref -> Quick reference.
help      -> Python's own help system.
object?   -> Details about 'object', use 'object??' for extra details.

In [1]: import cv2

In [2]: import tensorflow as tf

In [3]: cv2.__version__
Out[3]: '2.4.13.6'

In [4]: tf.__version__
Out[4]: '1.4.0'

In [5]: import keras
Using TensorFlow backend.

In [6]: keras.__version__
Out[6]: '2.2.2'

DenseNet-Keras git:(master) ls -hl imagenet_models
total 33M
drwxrwxr-x 2 epbox epbox 4.0K 8月  30 15:08 .
drwxrwxr-x 5 epbox epbox 4.0K 8月  30 15:09 ..
-rw-r--r-- 1 epbox epbox  32M 8月  30 15:07 densenet121_weights_tf.h5
-rw-rw-r-- 1 epbox epbox   55 8月  30 14:40 README.md

Error

DenseNet-Keras git:(master) python test_inference.py
Using TensorFlow backend.
/home/epbox/Github/DenseNet-Keras/densenet121.py:45: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(64, (7, 7), strides=(2, 2), use_bias=False, name="conv1")`
  x = Convolution2D(nb_filter, 7, 7, subsample=(2, 2), name='conv1', bias=False)(x)
2018-09-10 14:25:19.720132: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
2018-09-10 14:25:19.821988: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:892] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2018-09-10 14:25:19.822306: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030] Found device 0 with properties:
name: GeForce GTX TITAN X major: 5 minor: 2 memoryClockRate(GHz): 1.076
pciBusID: 0000:01:00.0
totalMemory: 11.92GiB freeMemory: 11.64GiB
2018-09-10 14:25:19.822322: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1120] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: GeForce GTX TITAN X, pci bus id: 0000:01:00.0, compute capability: 5.2)
/home/epbox/Github/DenseNet-Keras/densenet121.py:99: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(128, (1, 1), use_bias=False, name="conv2_1_x1")`
  x = Convolution2D(inter_channel, 1, 1, name=conv_name_base+'_x1', bias=False)(x)
/home/epbox/Github/DenseNet-Keras/densenet121.py:109: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(32, (3, 3), use_bias=False, name="conv2_1_x2")`
  x = Convolution2D(nb_filter, 3, 3, name=conv_name_base+'_x2', bias=False)(x)
Traceback (most recent call last):
  File "test_inference.py", line 34, in <module>
    model = DenseNet(reduction=0.5, classes=1000, weights_path=weights_path)
  File "/home/epbox/Github/DenseNet-Keras/densenet121.py", line 55, in DenseNet
    x, nb_filter = dense_block(x, stage, nb_layers[block_idx], nb_filter, growth_rate, dropout_rate=dropout_rate, weight_decay=weight_decay)
  File "/home/epbox/Github/DenseNet-Keras/densenet121.py", line 165, in dense_block
    concat_feat = merge([concat_feat, x], mode='concat', concat_axis=concat_axis, name='concat_'+str(stage)+'_'+str(branch))
TypeError: 'module' object is not callable

Is there a pre-train module for CIFAR100 on keras?

Dear @flyyufelix

Is there a pre-train module for CIFAR100 on keras?

Thanks a lot.

Step by step implementation of code

Can i get step by step implementation of this code please.

TypeError: 'module' object is not callable

Hi, i try to run the test code ,but i meet the above error. Has anyone encountered the same problem?

DocUNet

Hi, I notice your project at shahsohil/sunets#8. Have you reproduced the paper<DocUNet: Document Image Unwarping via A Stacked U-Net>? I trained the S-UNet for this paper some times,but it doesn't work. Can you give me some suggestions?

Regarding scale

Can you explain why you're using a separate scale function when batch norm scales the output anyway?

AssertionError: AbstractConv shape mismatch: shape of filters does not match given kshp.

AssertionError: AbstractConv shape mismatch: shape of filters does not match given kshp.
Apply node that caused the error: Assert{msg='AbstractConv shape mismatch: shape of filters does not match given kshp.'}(InplaceDimShuffle{3,2,0,1}.0, Elemwise{eq,no_inplace}.0, Elemwise{eq,no_inplace}.0, Elemwise{eq,no_inplace}.0, Elemwise{eq,no_inplace}.0)
Toposort index: 2069
Inputs types: [TensorType(float32, 4D), TensorType(bool, scalar), TensorType(bool, scalar), TensorType(bool, scalar), TensorType(bool, scalar)]
Inputs shapes: [(3L, 64L, 7L, 7L), (), (), (), ()]
Inputs strides: [(196L, 588L, 4L, 28L), (), (), (), ()]
Inputs values: ['not shown', array(False, dtype=bool), array(False, dtype=bool), array(True, dtype=bool), array(True, dtype=bool)]
Outputs clients: [[Subtensor{::, ::, ::int64, ::int64}(Assert{msg='AbstractConv shape mismatch: shape of filters does not match given kshp.'}.0, Constant{-1}, Constant{-1})]]

License?

As in flyyufelix/cnn_finetune#5 please consider applying a license (I suggest MIT, which is at that link).

will you provide no top architecture of the densenet with the pre-trained weigths?

Pretrained DenseNet Models

Hello, thanks for your great work. I am researching some related issues and need your pretrained model for my own experiment. I'll appreciate it if you can give me the access. My google mail address: [email protected]

Not working with keras 2.0.5 and tf 1.2.1

Getting error while loading model 'AttributeError: 'Tensor' object has no attribute 'assign'

Pretrained model

Hello, thanks for your great work. I need your pretrained model for my own experiment. I'll appreciate it if you can give me the access. My google mail address: [email protected]

Code is using growth_rate instead of nb_filter

The iterative updates of nb_filter are never used:

Line 164+: x = conv_block(concat_feat, stage, branch, growth_rate, dropout_rate, weight_decay)

AttributeError: 'Tensor' object has no attribute 'assign'

Hi,

Thanks a lot for the repo, but I hit the following error when trying the python test_inference.py with the densenet169:

Traceback (most recent call last):
  File "test_inference.py", line 27, in <module>
    model = DenseNet(reduction=0.5, classes=1000, weights_path=weights_path)
  File "densenet169.py", line 78, in DenseNet
    model.load_weights(weights_path, by_name=True)
  File "/home/petteri/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 2498, in load_weights
    load_weights_from_hdf5_group_by_name(f, self.layers)
  File "/home/petteri/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 2976, in load_weights_from_hdf5_group_by_name
    K.batch_set_value(weight_value_tuples)
  File "/home/petteri/anaconda3/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 2025, in batch_set_value
    assign_op = x.assign(assign_placeholder)
AttributeError: 'Tensor' object has no attribute 'assign'

So when printing the types of the variables in the saved weights I got the following:

<tf.Variable 'conv1/kernel:0' shape=(7, 7, 3, 64) dtype=float32_ref>
<tf.Variable 'conv1_bn/gamma:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'conv1_bn/beta:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'conv1_bn/moving_mean:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'conv1_bn/moving_variance:0' shape=(64,) dtype=float32_ref>
Tensor("conv1_scale/Const:0", shape=(64,), dtype=float32)

And the batch_set_value(tuples) from the backend hit the Tensor problem with "conv1_scale", is there a fix for this to allow the assigning or are the weights somewhat weirdly saved?:

            if hasattr(x, '_assign_placeholder'):
                assign_placeholder = x._assign_placeholder
                assign_op = x._assign_op
            else:
                assign_placeholder = tf.placeholder(tf_dtype, shape=value.shape)
                assign_op = x.assign(assign_placeholder)
                x._assign_placeholder = assign_placeholder
                x._assign_op = assign_op
            assign_ops.append(assign_op)
            feed_dict[assign_placeholder] = value

Training own model

how do I train my own model ?

about the pre-train weight

Hi,
Did you test the pre-train model on tensorflow backend?
I got an error when use it while the theano backend works.
I need a tensorflow pre-train model.

Thanks

Scale layer vs BatchNormalization in Keras v2

Hi,

At first, thanks for this repo with DenseNet for Keras. As I understand you ported the architecture and weights from Caffe. I've just a question on actual purpose of your custom layer Scale after BatchNormalization in Keras v2 in the sense don't they perform the same work ?

Preprocessing for finetuning the models

do we need to do preprocssing of data like normalization and mean subtraction to fine tune
pretrained densenet models?

face recognition

By leveraging on this denseNet-Keras framework can we undergo face recognition.
if so, what is the accuracy we can achieve ?

conversion tool

I'm curious, what did you use to actually convert the weights from caffe to keras?

Why the 'bias' is set to False?

Hi,

Is it possible to get better results by setting 'bias' to false?

RGB or BGR?

Hi, this is a nice work. And I have some questions about how to use the model.

What tool is used convert the caffe model? And what order should be used for testing, RGB or BGR. Thanks.

Bounding box of predicted class

@flyyufelix, can u plz tell me how to get the bounding box of the detected object class?
What to modify to get the bounding box?