WNJXYK/CMA

A Python Implementation of 'Continuous Manifold-based Adaptation'

CMA

A Python Implementation of 'Continuous Manifold-based Adaptation', the official MatLab version: jhoffman/cma.

Judy Hoffman, Trevor Darrell, Kate Saenko: Continuous Manifold Based Adaptation for Evolving Visual Domains. CVPR 2014: 867-874

How to use

from sklearn.svm import LinearSVC
from cma import CMA

# Define a CMA module with Linear SVM
# Mode is 'cgfk' (cgfk / csa)
# Alpha is 1.5 - Forgetting parameter for online subspace learning
# Dim is 10
cma = CMA(LinearSVC(), **{'alpha': 1.5, 'dim': 10, 'mode': 'cgfk'})

# Init on source domain
cma.fit(Xs, ys.ravel())

# Envolves on data stream
for Xt in data_steam:
    yt = cma.predict(Xt)

We provide a Notebook to reproduce the default experiment in the official Matlab code.

Experiments

Here is the experiment setting and hyper-parameters.

Dataset: caltran_gist
Norm_type: L1 Zscore
Size of Source Domain: 50
Size of Streaming: 480
Block Size: 2
Alpha: 1.5
Dim: 10

Original Matlab Version

StartIdx	KNN	SVM	KNN_cgfk	KNN_csa	SVM_cgfk	SVM_csa
350	65.49	77.75	64.66	64.45	83.99	83.58
400	65.70	71.93	66.53	66.32	73.39	73.80
450	55.30	70.48	55.30	54.89	72.77	72.56
500	54.89	71.93	55.51	55.51	67.98	67.98
550	67.57	71.52	62.99	63.41	79.21	79.21
Mean	61.79	72.72	61.00	60.91	75.47	75.43

This Python Implementation

StartIdx	KNN	SVM	KNN_cgfk	KNN_csa	SVM_cgfk	SVM_csa
350	63.96	77.50	66.46	69.17	84.79	84.79
400	65.21	72.08	64.17	64.17	73.96	74.17
450	56.46	69.58	56.67	56.88	72.50	72.71
500	56.04	71.88	52.92	53.54	66.25	67.92
550	55.00	71.67	55.00	53.96	76.25	79.38
Mean	59.33	72.54	59.04	59.54	74.75	75.79