PCA in TensorFlow

tensorflow_pca.py

import numpy as np
import tensorflow as tf

# N, size of matrix. R, rank of data
N = 100
R = 5

# generate data
W_true = np.random.randn(N,R)
C_true = np.random.randn(R,N)
Y_true = np.dot(W_true, C_true)
Y_tf = tf.constant(Y_true.astype(np.float32))

W = tf.Variable(np.random.randn(N,R).astype(np.float32))
C = tf.Variable(np.random.randn(R,N).astype(np.float32))
Y_est = tf.matmul(W,C)
loss = tf.reduce_sum((Y_tf-Y_est)**2)

# regularization 
alpha = tf.constant(1e-4)
regW = alpha*tf.reduce_sum(W**2)
regC = alpha*tf.reduce_sum(C**2)

# full objective
objective = loss + regW + regC

# optimization setup
train_step = tf.train.AdamOptimizer(0.001).minimize(objective)

# fit the model
init_op = tf.initialize_all_variables()
with tf.Session() as sess:
    sess.run(init_op)
    for n in range(10000):
        sess.run(train_step)
        if (n+1) % 1000 == 0:
            print('iter %i, %f' % (n+1, sess.run(objective)))

Couldn't one just use tf.svd? The SVD is how PCA is performed in the majority of implementations.

Yeah sure, the point of this was just for demonstration. You could extend this for PCA models that can't be solved in closed form (e.g. sparse PCA).

I have knowledge of the linear algebraic implementation for PCA using SVD. But I see that you are trying to minimize a loss function here. Can you explain how this relates to PCA? Or share resources on the above.

What does this op stand for ？
regW = alphatf.reduce_sum(W**2)
regC = alphatf.reduce_sum(C**2)
It seems to make regularization of W or C？
I think It is not need to square the W or T, I guess.

My implementation: https://gist.github.com/N-McA/bbbaed9d1a4b7c316f5d28cef1b96bdd