As creating a neural network for digit classification seems to be a bit outdated, we will create a fictional network that learns to colorize grayscale images. In this case-study, you will learn to do the following using TensorPack.
As creating a neural network for digit classification seems to be a bit outdated, we will create a fictional network that learns to colorize grayscale images. In this case-study, you will learn to do the following using TensorPack.
- Dataflow
- DataFlow
+ create a basic dataflow containing images
+ create a basic dataflow containing images
+ debug you dataflow
+ debug you dataflow
+ add custom manipulation to your data such as converting to Lab-space
+ add custom manipulation to your data such as converting to Lab-space
...
@@ -16,94 +16,86 @@ As creating a neural network for digit classification seems to be a bit outdated
...
@@ -16,94 +16,86 @@ As creating a neural network for digit classification seems to be a bit outdated
- Callbacks
- Callbacks
+ write your own callback to export predicted images after each epoch
+ write your own callback to export predicted images after each epoch
## Dataflow
## DataFlow
The basic idea is to gather a huge amount of images, resizing them to the same size and extract the luminance channel after converting from RGB to Lab. For demonstration purposes, we will split the dataflow definition into single steps, though it might more efficient to combine some steps.
The basic idea is to gather a huge amount of images, resizing them to the same size and extract
the luminance channel after converting from RGB to Lab. For demonstration purposes, we will split
the dataflow definition into separate steps, though it might more efficient to combine some steps.
### Reading data
### Reading data
The first node in the dataflow is the image-reader. There are several ways to read a lot of images:
The first node in the dataflow is the image reader. You can implement the reader however you want, but there are some existing ones we can use, e.g.:
- use lmdb files you probably already used for the Caffe framework
- use the lmdb files you probably already have for the Caffe framework
- collect images from a specific directory
- collect images from a specific directory
- read data from the ImageNet if you have already downloaded these images
- read ImageNet dataset if you have already downloaded these images
We will use simply a directory which consists of many RGB images. This is as simple as:
We will use simply a directory which consists of many RGB images. This is as simple as:
dp 0: is ndarray of shape (1920, 2560, 3) with range [0.0000, 255.0000]
dp 0: is ndarray of shape (1920, 2560, 3) with range [0.0000, 255.0000]
datapoint 1<2 with 1 elements consists of
datapoint 1<2 with 1 components consists of
dp 0: is ndarray of shape (850, 1554, 3) with range [0.0000, 255.0000]
dp 0: is ndarray of shape (850, 1554, 3) with range [0.0000, 255.0000]
````
```
To actually get data you can add
To actually access the datapoints generated by the dataflow, you can use
````python
```python
ds.reset_state()
fordpinds.get_data():
fordpinds.get_data():
printdp[0]# this is RGB data!
printdp[0]# this is an RGB image!
````
```
This kind of iteration is used behind the scenes to feed data for training.
This iteration is used in an additional process. Of course, without the `print` statement. There is some other magic behind the scenes. The dataflow checks of the image is actually an RGB image with 3 channels and skip those gray-scale images.
### Manipulate incoming data
### Manipulate incoming data
Now, training a network which is not fully convolutional requires inputs of fixed size. Let us add this to the dataflow.
Now, training a ConvNet which is not fully convolutional requires images of known shape, but our
directory may contain images of different sizes. Let us add this to the dataflow:
Note that we've also added batch and prefetch, so that the dataflow now generates images of shape (32, 256, 256, 3), and faster.
But wait! The alert reader makes a critical observation! We need the L channel *only* and we should add the RGB image as ground-truth data. Let's fix that.
But wait! The alert reader makes a critical observation! For input we need the L channel *only* and we should add the RGB image as ground-truth data. Let's fix that.
ds=MapData(ds,lambdadp:[dp[0][:,:,0],dp[1]])# get L channel from first entry
ds=BatchData(ds,32)
ds=BatchData(ds,32)
ds=PrefetchData(ds,4)# use queue size 4
ds=PrefetchData(ds,4)# use queue size 4
ds=PrintData(ds,num=2)# only for debug
returnds
returnds
````
```
Here, we simply duplicate the rgb image and only apply the `image augmentors` to the first copy of the datapoint. The output when using `PrintData` should be like
Here, we simply apply a mapping function to the datapoint, transform the single component to two components: the first is the L color space, and the second is just itself.
The output when using `PrintData` should be like:
````
```
datapoint 0<2 with 2 elements consists of
datapoint 0<2 with 2 components consists of
dp 0: is ndarray of shape (256, 256) with range [0, 100.0000]
dp 0: is ndarray of shape (32, 256, 256) with range [0, 100.0000]
dp 1: is ndarray of shape (256, 256, 3) with range [0, 221.6387]
dp 1: is ndarray of shape (32, 256, 256, 3) with range [0, 221.6387]
datapoint 1<2 with 2 elements consists of
datapoint 1<2 with 2 components consists of
dp 0: is ndarray of shape (256, 256) with range [0, 100.0000]
dp 0: is ndarray of shape (32, 256, 256) with range [0, 100.0000]
dp 1: is ndarray of shape (256, 256, 3) with range [0, 249.6030]
dp 1: is ndarray of shape (32, 256, 256, 3) with range [0, 249.6030]
```
````
Again, do not use `PrintData` in combination with `PrefetchData` because the prefetch will be done in another process.
Well, this is probably not the most efficient way to encode this process. But it clearly demonstrates how much flexibility the `dataflow` gives.
Well, this is probably not the most efficient way to encode this process. But it clearly demonstrates how much flexibility the `dataflow` gives.
You can easily insert you own functions, and utilize the pre-defined modules at the same time.
## Network
## Network
If you are surprised how far we already are, you will enjoy how easy it is to define a network model. The most simple model is probably:
If you are surprised how far we already are, you will enjoy how easy it is to define a network model. The most simple model is probably:
````python
```python
classModel(ModelDesc):
classModel(ModelDesc):
def_get_input_vars(self):
def_get_input_vars(self):
...
@@ -163,27 +151,30 @@ class Model(ModelDesc):
...
@@ -163,27 +151,30 @@ class Model(ModelDesc):
def_build_graph(self,input_vars):
def_build_graph(self,input_vars):
self.cost=0
self.cost=0
````
```
The framework expects:
The framework expects:
- a definition of inputs in `_get_input_vars`
- a definition of inputs in `_get_input_vars`
- a computation graph containing the actual network layers in `_build_graph`
- a computation graph containing the actual network layers in `_build_graph`
- a member `self.cost` representing the loss function we would like to minimize.
-In single-cost optimization problem, a member `self.cost` representing the loss function we would like to minimize.
### Define inputs
### Define inputs
Our dataflow produces data which looks like `[(256, 256), (256, 256, 3)]`. The first entry is the luminance channel as input and the latter is the original RGB image with all three channels. So we will write
Our dataflow produces data which looks like `[(32, 256, 256), (32, 256, 256, 3)]`.
The first entry is the luminance channel as input and the latter is the original RGB image with all three channels. So we will write
This is pretty straight forward, isn't it? We defined the shapes of the input and spend each entry a name. This is very generous of us and will us help later to build an inference mechanism.
This is pretty straight forward, isn't it? We defined the shapes of the input and give each entry a name.
You can certainly use 32 instead of `None`, but since the model itself doesn't really need to know
the batch size, using `None` offers the extra flexibility to run inference with a different batch size in the same graph.
From now, the `input_vars` in `_build_graph(self, input_vars)` will have the shapes `[(256, 256), (256, 256, 3)]` because of the completed method `_get_input_vars`. We can therefore write
From now, the `input_vars` in `_build_graph(self, input_vars)` will be the tensors of the defined shapes in the method `_get_input_vars`. We can therefore write
The process of coming up with such a network architecture is usually a soup of experience, a lot of trials and much time laced with magic or simply chance, depending what you prefer. We will use an auto-encoder with a lot of convolutions to squeeze the information through a bottle-neck (encoder) and then upsample from a hopefully meaningful compact representation (decoder).
The process of coming up with such a network architecture is usually a soup of experience,
a lot of trials and much time laced with magic or simply chance, depending what you prefer.
We will use an auto-encoder with a lot of convolutions to squeeze the information through a bottle-neck
(encoder) and then upsample from a hopefully meaningful compact representation (decoder).
Because we are fancy, we will use a U-net layout with skip-connections.
Because we are fancy, we will use a U-net layout with skip-connections.
There are probably many better tutorials about defining your network model. And there are definitely [better models](../../examples/GAN/image2image.py). You should check them later. A good way to understand layers from this library is to play with those examples.
There are probably many better tutorials about defining your network model. And there are definitely [better models](../../examples/GAN/image2image.py). You should check them later. A good way to understand layers from this library is to play with those examples.
It should be noted that you can write your models using [tfSlim](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/slim)
which comes along [architectures and pre-trained models](https://github.com/tensorflow/models/tree/master/slim/nets) for image classification.
It should be noted that you can write your models using [tfSlim](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/slim) which comes along architectures [architectures and pre-trained models](https://github.com/tensorflow/models/tree/master/slim/nets) for image classification. The library automatically handles regularization and batchnorm updates from tfSlim. And you can directly load these pre-trained checkpoints from state-of-the-art models in TensorPack. Isn't this cool?
TensorPack automatically handles regularization and batchnorm updates from tfSlim. And you can directly load these pre-trained checkpoints from state-of-the-art models in TensorPack. Isn't this cool?
The remaining part is a boring L2-loss function given by:
The remaining part is a boring L2-loss function given by:
It is a good idea to track the progress of your training session using TensorBoard. This library provides several functions to simplify the output of summaries and visualization of intermediate states.
It is a good idea to track the progress of your training session using TensorBoard.
TensorPack provides several functions to simplify the output of summaries and visualization of intermediate states.
add a plot of the costs from our loss function and add some intermediate results to the tab of "images" inside TensorBoard. The updates are then triggered after each epoch.
add a plot of the moving average of the cost tensor, and add some intermediate results to the tab of "images" inside TensorBoard. The summary is written after each epoch.
Note that you can certainly use `tf.summary.scalar(self.cost)`, but then you'll only see a single cost value (rather than moving average) which is much less informative.
## Training
## Training
Let's summarize: We have a model and data. The missing piece which stitches these parts together is the training protocol. It is only a [configuration](../../tensorpack/train/config.py#L23-#L29)
Let's summarize: we have a model and data.
The missing piece which stitches these parts together is the training protocol.
It is only a [configuration](http://tensorpack.readthedocs.io/en/latest/modules/tensorpack.train.html#tensorpack.train.TrainConfig)
For the dataflow, we already implemented `get_data` in the first part. Specifying the learning rate is done by
For the dataflow, we already implemented `get_data` in the first part. Specifying the learning rate is done by
This essentially creates a non-trainable variable with initial value `1e-4` and also track this value inside TensorBoard. Let's have a look at the entire code:
This essentially creates a non-trainable variable with initial value `1e-4` and also track this value inside TensorBoard.
Let's have a look at the entire code:
````python
```python
defget_config():
defget_config():
logger.auto_set_dir()
logger.auto_set_dir()
dataset=get_data()
dataset=get_data()
...
@@ -300,18 +302,25 @@ def get_config():
...
@@ -300,18 +302,25 @@ def get_config():
step_per_epoch=dataset.size(),
step_per_epoch=dataset.size(),
max_epoch=100,
max_epoch=100,
)
)
````
```
There is not really new stuff. The model was implemented, and `max_epoch` is set to 100. This means 100 runs over the entire dataset. The alert reader who almost already had gone to sleep makes some noise: "Where is `dataset.size()` coming from?" This values represents all images in one directory and is forwarded by all mappings. If you have 42 images in your directory, then this value is 42. Satisfied with this answer, the alert reader went out of the room. But he will miss the most interesting part: the callback section. We will cover this in the next section.
There is not really new stuff.
The model was implemented, and `max_epoch` is set to 100.
The alert reader who almost already had gone to sleep makes some noise: "Where is `dataset.size()` coming from?"
This method is implemented by `ImageFromFile` and is forwarded by all mappings.
If you have 42 images in your directory, then this value would be 42.
Satisfied with this answer, the alert reader went out of the room.
But he will miss the most interesting part: the callback section. We will cover this in the next section.
## Callbacks
## Callbacks
Until this point, we spoke about all necessary part of deep learning pipelines which are common from GANs, image-recognition and embedding learning. But sometimes you want to add your own code. We will now add a functionality which will export some entries of the tensor `prediction`. Remember, this is the result of the decoder part in our network.
Until this point, we spoke about all necessary parts of deep learning pipelines which are common for GANs, image-recognition and embedding learning.
But sometimes you want to add your own code to do something extra. We will now add a functionality which will export some entries of the tensor `prediction`.
Remember, this tensor is the result of the decoder part in our network.
To not mess up the code, there is a plug-in mechanism with callbacks. Our callback looks like
To modularize the code, there is a plug-in mechanism called callbacks. Our callback looks like
````python
```python
classOnlineExport(Callback):
classOnlineExport(Callback):
def__init__(self):
def__init__(self):
pass
pass
...
@@ -321,61 +330,75 @@ class OnlineExport(Callback):
...
@@ -321,61 +330,75 @@ class OnlineExport(Callback):
def_trigger_epoch(self):
def_trigger_epoch(self):
pass
pass
````
```
So it has 3 methods, although there are some more. TensorPack is really conservative regarding the computation graph. After the network is constructed and all callbacks are initialized the graph is finalized. So once you started training, there is no way of adding nodes to the graph, which we actually want to do for inference.
So it has 3 methods, although there are some more.
TensorPack is conservative regarding the computation graph.
After the network is constructed and all callbacks are initialized the graph is read-only.
So once you started training, there is no way of modifying the graph, which we actually want to do for inference.
You'll need to define the whole graph before training starts.
Let us fill in some parts
Let us fill in some parts:
````python
```python
classOnlineExport(Callback):
classOnlineExport(Callback):
def__init__(self):
def__init__(self):
self.cc=0
self.cc=0
self.example_input=color.rgb2lab(cv2.imread('myimage.jpg')[:,:,[2,1,0]])[:,:,0]# read rgb image and extract luminance
self.example_input=color.rgb2lab(cv2.imread('myimage.jpg')[:,:,::-1])[:,:,0]# read rgb image and extract luminance
Can you remember the method `_get_input_vars` in our model? We used the name `luminance` to identify one input. If not, this is the best time to go back in this text and read how to specify input variables for the network. In the deconvolution step there was also:
Can you remember the method `_get_input_vars` in our model? We used the name `luminance` to identify one input.
If not, this is the best time to go back in this text and read how to specify input variables for the network.
In the deconvolution step there was also:
````python
```python
prediction=Deconv2D('prediction',e2,3,nl=tf.tanh)# name is 'prediction'
prediction=Deconv2D('prediction',e2,3,nl=tf.tanh)# name of the tensor is 'prediction/output'
````
```
Usually the name of the output tensor of a layer is in the API documentation. If you are uncertain,
you can simply `print(prediction)` to find out the name.
These two names allows to build the inference part of the network in
These two names allows us to build the inference part of the network in
