FC-DenseNet-Tensorflow
This is a re-implementation of the 100 layer tiramisu, technically a fully convolutional DenseNet, in TensorFlow (Tiramisu). The aim of the repository is to break down the working modules of the network, as presented in the paper, for ease of understanding. To facilitate this, the network is defined in a class, with functions for each block in the network. This promotes a modular view, and an understanding of what each component does individually. I tried to make the model code more readable, and this is the main aim of the this repository.
Network Architecture
Submodules
The "submodules" that build up the Tiramisu are explained here. Note: The graphics are just a redrawing of the ones from the original paper.
The Conv Layer:
The "conv layer" is the most atomic unit of the FC-DenseNet, it is the building block of all other modules. The following image shows the conv layer:
def conv_layer(self, x, training, filters, name):
with tf.name_scope(name):
x = self.batch_norm(x, training, name=name+'_bn')
x = tf.nn.relu(x, name=name+'_relu')
x = tf.layers.conv2d(x,
filters=filters,
kernel_size=[3, 3],
strides=[1, 1],
padding='SAME',
dilation_rate=[1, 1],
activation=None,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name=name+'_conv3x3')
x = tf.layers.dropout(x, rate=0.2, training=training, name=name+'_dropout')
As can be seen, each "convolutional" layer is actually a 4 step procedure of batch normalization -> Relu -> 2D-Convolution -> Dropout.
The Dense Block
The dense block is a sequence of convolutions followed by concatenations. The output of a conv layer is concated depth wise with its input, this forms the input to the next layer, and is repeated for all layers in a dense block. For the final output i.e., the output of the Dense Block, all the outputs of each conv layer in the block are concated, as shown:
In code, it is implemented as:
def dense_block(self, x, training, block_nb, name):
dense_out = []
with tf.name_scope(name):
for i in range(self.layers_per_block[block_nb]):
conv = self.conv_layer(x, training, self.growth_k, name=name+'_layer_'+str(i))
x = tf.concat([conv, x], axis=3)
dense_out.append(conv)
x = tf.concat(dense_out, axis=3)
return x
How to Run
To run the network on your own dataset, do the following:
- Clone this repository.
- Open up your terminal and navigate to the cloned repository
- Type in the following:
python main.py --mode=train --train_data=path/to/train/data --val_data=path/to/validation/data \
--ckpt=path/to/save/checkpoint/model.ckpt --layers_per_block=4,5,7,10,12,15 \
--batch_size=8 --epochs=10 --growth_k=16 --num_classes=2 --learning_rate=0.001
The "layers_per_block" argument is only specified for the downsample path, upto the final bottleneck dense block, the upsample path is then automatically built by mirroring the downsample path.
Run with trained checkpoint
To run the code with a trained checkpoint file on images, use the infer mode in in the command line options, like so:
python main.py --mode=infer --infer_data=path/to/infer/data --batch_size=4 \
--ckpt=models/model.ckpt-20 --output_folder=outputs
Tests
The python files ending with "*_test.py" are unit test files, if you make changes or have just cloned the repo, it is a good idea to run them once in your favorite Python IDE, they should let you know if your changes break anything. Currently, the test coverage is not that high, I plan to keep adding more in the future.
TODOs:
Add some more functionality in the code.- Add more detail into this readme.
Save model graph.Rework command line arguments.- Update with some examples of performance once trained.
Increase test coverage.- Save loss summaries for Tensorboard.
349 Dec 8, 2022
771 Dec 15, 2022
568 Oct 31, 2022
606 Dec 21, 2022
9 Nov 24, 2022
3 Nov 30, 2021
32 Sep 21, 2022
299 Dec 17, 2022
734 Jan 3, 2023
9 Nov 29, 2022
65 Sep 12, 2022
321 Dec 1, 2022
472 Dec 22, 2022
115 Dec 28, 2022
111 Dec 27, 2022
158 Dec 21, 2022
1.6k Jan 7, 2023