Transfer Learning
What is Transfer Learning?
Transfer learning is the improvement of learning in a new task through the transfer of knowledge from a related task that has already been learned.
http://pages.cs.wisc.edu/~shavlik/abstracts/torrey.handbook09.abstract.html
Why we care about this?
- We rarely have a dataset big enough for CNN
- We can take advantage of ImageNet (1.2 million images and 1000 classes)
- Reduced risk of overfitting
- Less training time
The two types
- CNN as fixed feature extractor
- Fine-tuning the CNN
Feature Extractor
VGG16
Top Model
Feature Extractor
from keras.applications.vgg16 import VGG16
from keras.preprocessing import image
from keras.applications.vgg16 import preprocess_input
import numpy as np
model = VGG16(weights='imagenet', include_top=False)
img_path = 'elephant.jpg'
img = image.load_img(img_path, target_size=(224, 224))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
features = model.predict(x)SVM(features)
XGBoost(features)
RandomForest(features)
or
or
FIne Tuning
Steps for FIne TUNING
- Initialize CNN with ImageNet weights and freeze Convolution layers (Without the top model)
- Create top model (Fully Connected Layers and output layer)
- Train top model with the dataset
- Attach top model to the CNN
- Fine Tune the network using a very small learning rate
Resources for this
- Using the CPU (Default)
- Using the GPU (Requires Nvidia graphics card and CUDA drivers)
- Using Amazon p2.xlarge instances (0.9 cents/hour and 30x faster than a traditional CPU)