This tutorial will guide through the steps to create a simple custom layer for Caffe using python. By the end of it, there are some examples of custom layers.
Usually you would create a custom layer to implement a funcionality that isn't available in Caffe, tuning it for your requirements.
Probably just Python and Caffe installed.
Creating a python custom layer adds some overhead to your network and probably isn't as efficient as a C++ custom layer. However, this way, you won't have to compile the whole caffe with your new layer.
import caffe
class My_Custom_Layer(caffe.Layer):
def setup(self, bottom, top):
pass
def forward(self, bottom, top):
pass
def reshape(self, bottom, top):
pass
def backward(self, bottom, top):
pass
So important things to remember:
- Your custom layer has to inherit from caffe.Layer (so don't forget to import caffe);
- You must define the four following methods: setup, forward, reshape and backward;
- All methods have a top and a bottom parameters, which are the blobs that store the input and the output passed to your layer. You can access it using top[i].data or bottom[i].data, where i is the index of the blob in case you have more than one upper or lower blob.
The Setup method is called once during the lifetime of the execution, when Caffe is instantiating all layers. This is where you will read parameters, instantiate fixed-size buffers.
Use the reshape method for initialization/setup that depends on the bottom blob (layer input) size (for example top blob size and internal buffers). It is called before every forward.
The Forward method is called for each input batch and is where most of your logic will be.
The Backward method is called during the backward pass of the network. For example, in a convolution-like layer, this would be where you would calculate the gradients. This is optional (a layer can be forward-only).
Ok, so now you have your layer designed! This is how you define it in your .prototxt file:
layer {
name: "LayerName"
type: "Python"
top: "TopBlobName"
bottom: "BottomBlobName"
python_param {
module: "My_Custom_Layer_File"
layer: "My_Custom_Layer_Class"
param_str: '{"param1": 1,"param2":True, "param3":"some string"}'
}
include{
phase: TRAIN
}
}
Important remarks:
- type must be Python;
- You must have a python_param dictionary with at least the module and layer parameters;
- module refers to the file where you implemented your layer (without the .py);
- layer refers to the name of your class;
- You can pass parameters to the layer using param_str (more on accessing them bellow);
- Just like any other layer, you can define in which phase you want it to be active (see the examples to see how you can check the current phase);
You can define the layer parameters in the prototxt by using param_str. Once you've done it, here is an example on how you access these paremeters inside the layer class:
def setup(self, bottom, top):
params = eval(self.param_str)
param1 = params["param1"]
param2 = params.get('param2', False) #I usually use this when fetching a bool
param3 = params["param3"]
#Continue with the setup
# ...
You have two options (at least that I know of). Either you can save the custom layer file in the same folder as you are going to run the caffe command (probably where your prototxt files would be). Another way, also my favorite one, is to save all your custom layers in a folder and adding this folder to your PYTHONPATH.
Bellow are two examples of layers. One of them is a "measure" layer, that outputs the accuracy and a confusion matrix for a binary problem during training and the accuracy, false positive rate and false negative rate during test/validation. Although Caffe already has a Accuracy layer, sometimes you want something more, like a F-measure.
The other is a custom data layer, that receives a text file with image paths, loads a batch of images and preprocesses them. Just a quick tip, Caffe already has a big range of data layers and probably a custom layer is not the most efficient way if you just want something simple.
This is my measureLayer.py with my class definition:
#Remark: This class is designed for a binary problem, where the first class would be the 'negative'
# and the second class would be 'positive'
import caffe
TRAIN = 0
TEST = 1
class Measure_Layer(caffe.Layer):
#Setup method
def setup(self, bottom, top):
#We want two bottom blobs, the labels and the predictions
if len(bottom) != 2:
raise Exception("Wrong number of bottom blobs (prediction and label)")
#And some top blobs, depending on the phase
if self.phase == TEST and len(top) != 3:
raise Exception("Wrong number of top blobs (acc, FPR, FNR)")
if self.phase == TRAIN and len(top) != 5:
raise Exception("Wrong number of top blobs (acc, tp, tn, fp and fn)")
#Initialize some attributes
self.TPs = 0.0
self.TNs = 0.0
self.FPs = 0.0
self.FNs = 0.0
self.totalImgs = 0
#Forward method
def forward(self, bottom, top):
#The order of these depends on the prototxt definition
predictions = bottom[0].data
labels = bottom[1].data
self.totalImgs += len(labels)
for i in range(len(labels)): #len(labels) is equal to the batch size
pred = predictions[i] #pred is a tuple with the normalized probability
#of a sample i.r.t. two classes
lab = labels[i]
if pred[0] > pred[1]:
if lab == 1.0:
self.FNs += 1.0
else:
self.TNs += 1.0
else:
if lab == 1.0:
self.TPs += 1.0
else:
self.FPs += 1.0
acc = (self.TPs + self.TNs) / self.totalImgs
try: #just assuring we don't divide by 0
fpr = self.FPs / (self.FPs + self.TNs)
except:
fpr = -1.0
try: #just assuring we don't divide by 0
fnr = self.FNs / (self.FNs + self.TPs)
except:
fnr = -1.0
#output data to top blob
top[0].data = acc
if self.phase == TRAIN:
top[1].data = self.TPs
top[2].data = self.TNs
top[3].data = self.FPs
top[4].data = self.FNs
elif self.phase == TEST:
top[1].data = fpr
top[2].data = fnr
def reshape(self, bottom, top):
"""
We don't need to reshape or instantiate anything that is input-size sensitive
"""
pass
def backward(self, bottom, top):
"""
This layer does not back propagate
"""
pass
And this is an example of a prototxt with it:
layer {
name: "metrics"
type: "Python"
top: "Acc"
top: "TPs"
top: "TNs"
top: "FPs"
top: "FNs"
bottom: "prediction" #let's supose we have these two bottom blobs
bottom: "label"
python_param {
module: "measureLayer"
layer: "Measure_Layer"
}
include {
phase: TRAIN
}
}
layer {
name: "metrics"
type: "Python"
top: "Acc"
top: "FPR"
top: "FNR"
bottom: "prediction" #let's supose we have these two bottom blobs
bottom: "label"
python_param {
module: "measureLayer"
layer: "Measure_Layer"
}
include {
phase: TEST
}
}
My dataLayer.py could be something like:
import caffe
class Custom_Data_Layer(caffe.Layer):
def setup(self, bottom, top):
# Check top shape
if len(top) != 2:
raise Exception("Need to define top blobs (data and label)")
#Check bottom shape
if len(bottom) != 0:
raise Exception("Do not define a bottom.")
#Read parameters
params = eval(self.param_str)
src_file = params["src_file"]
self.batch_size = params["batch_size"]
self.im_shape = params["im_shape"]
self.crop_size = params.get("crop_size", False)
###### Reshape top ######
#This could also be done in Reshape method, but since it is a one-time-only
#adjustment, we decided to do it on Setup
if self.crop_size:
top[0].reshape(self.batch_size, 3, self.crop_size, self.crop_size)
else:
top[0].reshape(self.batch_size, 3, self.im_shape, self.im_shape)
top[1].reshape(self.batch_size)
#Read source file
#I'm just assuming we have this method that reads the source file
#and returns a list of tuples in the form of (img, label)
self.imgTuples = readSrcFile(src_file)
self._cur = 0 #use this to check if we need to restart the list of imgs
def forward(self, bottom, top):
for itt in range(self.batch_size):
# Use the batch loader to load the next image.
im, label = self.load_next_image()
#Here we could preprocess the image
# ...
# Add directly to the top blob
top[0].data[itt, ...] = im
top[1].data[itt, ...] = label
def load_next_img(self):
#If we have finished forwarding all images, then an epoch has finished
#and it is time to start a new one
if self._cur == len(self.imgTuples):
self._cur = 0
shuffle(self.imgTuples)
im, label = self.imgTuples[self._cur]
self._cur += 1
return im, label
def reshape(self, bottom, top):
"""
There is no need to reshape the data, since the input is of fixed size
(img shape and batch size)
If we were processing a fixed-sized number of images (for example in Testing)
and their number wasn't a multiple of the batch size, we would need to
reshape the top blob for a smaller batch.
"""
pass
def backward(self, bottom, top):
"""
This layer does not back propagate
"""
pass
And the prototxt would be like:
layer {
name: "Data"
type: "Python"
top: "data"
top: "label"
python_param {
module: "dataLayer"
layer: "Custom_Data_Layer"
param_str: '{"batch_size": 126,"im_shape":256, "crop_size":224, "src_file": "path_to_TRAIN_file.txt"}'
}
}
Sorry everybody, I've just seen your comments. For some reason, I didn't receive a notification/email when you commented or mentioned me. (Edit: I've just found out Gist doesn't support notifications. That's too bad :( )
@danzeng1990, as @Noiredd said, you shouldn't need to comment anything in
.cpp
files. As far as I remember, I only altered the MakeFile. In case you still weren't able to figure out what is it, I suggest you use Docker with an image that already has all caffe dependencies set up.@ujsyehao, thanks! I've fixed that!
@Noiredd, I'm glad that you liked! Thanks! Regarding the backward method, I'm not sure how the python wrapper is implemented, so this is only a guess, but I think that when you implement the backward method, you should "pass" data from top to bottom, i.e. use
top[...].data
as input andbottom[...].data
as output. But once again, I'm not sure about it.@AlexTS1980, that is one way to do it. Indeed it adds overhead to the whole process, making it a bit slower. I had two alternatives for that:
The first alternative seems to be faster (considering only training time), but you need to be able to fit and process all your data in disk (in my case this wasn't possible).
Thank you all for your comments!