Vgg16: Problems of Loading Pretrained Weights without fcs

[LynnHo]

@taehoonlee I tried to use sess.run(net.pretrained()) for vgg16, but it failed because of shape mismatch (I did not use the default input size). I read the source code and I am confused by line 320 in tensornets/utils.py,

is the code at line 321 designed for tackling the shape mismatch of fcs? If yes, I think it has a problem on vgg16 because it has three fc and this code seems to be only suitable for networks with 1 fc.

[taehoonlee]

Thank you for trying TensorNets, @LynnHo. Can you share a TensorNets version and your snippets to reproduce the error? The different input shape is not related to the weight shape mismatching. Maybe, I guess that you use different classes. And the lines 321-324 are to handle the different classes as you expected.

[LynnHo]

@taehoonlee I use a code like below,

import tensorflow as tf
import tensornets as nets

input_256 = tf.placeholder(dtype=tf.float32, shape=[None, 256, 256, 3])
network = nets.VGG16(input_256)
sess = tf.Session()

sess.run(network.pretrained())

and the error is,

ValueError: Dimension 0 in both shapes must be equal, but are 32768 and 25088. Shapes are [32768,4096] and [25088,4096]. for 'Assign_26' (op: 'Assign') with input shapes: [32768,4096], [25088,4096].

If the shape is [None, 224, 224, 3], the code works.

[taehoonlee]

@LynnHo, I see. I misunderstood the problem. The VGG16 and the VGG19 accept only a size of 224x224 because the fc6 layer produces a 4096-dimensional vector from a 25088(=77512)-dim vector. The error is from VGG's non-flexible architecture (accepting only fixed input sizes), not the line 320.

[LynnHo]

@taehoonlee But if I just want to load the pretrained weights of Convs ignoring the FCs‘, line 320 is unsuitable. I modified the code at line 320-324 to blow, it works for me

ops = []
for w, v in zip(weights, values):
    if w.shape != v.shape:
        ops.append(w.initializer)
    else:
        ops.append(w.assign(v))

[taehoonlee]

@LynnHo, Thank you for the idea. But the proposed one will load all the weights except the middle layer fc6. I think it is error-prone. For the users who want to exploit all the weights, I think it is better to show the mismatch error rather than a silent error due to the random middle layer fc6.

[LynnHo]

@taehoonlee You are right, the shape mismatch should be shown. Or an argument to control the ignoring of the shape mismatch would be better. Anyway, thanks a lot!

[taehoonlee]

@LynnHo, Yes I will considering the argument to control the shape mismatch. Thank you for the idea.