i did not find any SVD operations in the code

[yhao-z]

In your paper, u proposed to use low rank proporties in the CNN-extracted domain and also used the singular value thresholding (NOT soft thresholding) in that domain. Indeed, it is an excellent paper.

However, when i check the code in BuildModel.py, i just saw the soft thresholding without any singular value decompositions, which in the paper are for getting the singular values and then thresholding them.

The code is like,

    # rk <-- yk
    rk = tf.subtract(layeryk[-1], tf.scalar_mul(lambdaStep, tf.multiply(PhiT, tf.subtract(tf.multiply(Phi, layeryk[-1]), Yinput))))

    # Drk <-- rk
    weight0 = add_conv2d_weight([filterSize1, filterSize1, nOfModel, convSize1], 0)
    Drk = tf.nn.conv2d(rk, weight0, strides=[1, 1, 1, 1], padding='SAME')

    # Transform Module
    weight1 = add_conv2d_weight([filterSize2, filterSize2, convSize1, convSize2], 1)
    weight2 = add_conv2d_weight([filterSize3, filterSize3, convSize2, convSize3], 2)
    Frk = tf.nn.conv2d(Drk, weight1, strides=[1, 1, 1, 1], padding='SAME')
    Frk = tf.nn.relu(Frk)
    Frk = tf.nn.conv2d(Frk, weight2, strides=[1, 1, 1, 1], padding='SAME')
    field = Frk

    # Soft-thresholding Module
    softFRk = tf.multiply(tf.sign(Frk), tf.nn.relu(tf.subtract(tf.abs(Frk), softThr)))


    # Inverse Transform Module
    weight3 = add_conv2d_weight([filterSize3, filterSize3, convSize3, convSize2], 3)
    weight4 = add_conv2d_weight([filterSize2, filterSize2, convSize2, convSize1], 4)
    FFrk = tf.nn.conv2d(softFRk, weight3, strides=[1, 1, 1, 1], padding='SAME')
    FFrk = tf.nn.relu(FFrk)
    FFrk = tf.nn.conv2d(FFrk, weight4, strides=[1, 1, 1, 1], padding='SAME')

So if it is in this structure, then i think there are no differences between this code and the ISTA-Net.

I'm totally confused. If I ignored anything important for this issue in the code, plz point it out. Thanks a lot!