EEGAN_pretrained_doesn't work #6

ajitaru · 2020-11-17T21:34:50Z

Is there a new pretrained EEGAN weights file?
I got the following error message>
Key generator/conv_m2/weight not found in checkpoint

System:
tensorflow 1.10.0
python 3.6

paugalles · 2021-01-27T10:53:28Z

Same problem here.
I have also trained for some epochs and used my trained weights for testing. Got the same problem

paugalles · 2021-01-28T08:58:09Z

@kuihua if you could please double check. I think you might have pushed the wrong version of the file. Is that possible?

It seems the TESTGAN.py is wrong.
The layer named conv_m2 does not exist in the trained weights.
Instead of this layer, all this block seems to be missing:

            res_in = x_f
            # frame
            for i in range(3):
                with tf.variable_scope('block{}ex2'.format(i+1)):
                    x1=x2=x3=x_f
                    for j in range(3):
                        with tf.variable_scope('block{}_{}ex1'.format(i+1,j+1)):
                            with tf.variable_scope('ud1'):
                                a1 = lrelu(deconv_layer(x1, [3, 3, 64, 64], [self.batch_size, self.weight, self.height, 64], 1))
                                #a1 = batch_normalize(a1, is_training)
                            with tf.variable_scope('ud2'):
                                b1 = lrelu(deconv_layer(x2, [3, 3, 64, 64], [self.batch_size, self.weight, self.height, 64], 1))
                                #b1 = batch_normalize(b1, is_training)
                            with tf.variable_scope('ud3'):
                                c1 = lrelu(deconv_layer(x3, [3, 3, 64, 64], [self.batch_size, self.weight, self.height, 64], 1))
                                #c1 = batch_normalize(c1, is_training)
                            sum = tf.concat([a1,b1,c1],3)
                            #sum = batch_normalize(sum, is_training)
                            with tf.variable_scope('ud4'):
                                x1 = lrelu(deconv_layer(tf.concat([sum,x1],3), [1, 1, 64, 256], [self.batch_size, self.weight, self.height, 64], 1))
                                #x1 = batch_normalize(x1, is_training)
                            with tf.variable_scope('ud5'):
                                x2 = lrelu(deconv_layer(tf.concat([sum,x2],3), [1, 1, 64, 256], [self.batch_size, self.weight, self.height, 64], 1))
                                #x2 = batch_normalize(x2, is_training)
                            with tf.variable_scope('ud6'):
                                x3 = lrelu(deconv_layer(tf.concat([sum,x3],3), [1, 1, 64, 256], [self.batch_size, self.weight, self.height, 64], 1))
                                #x3 = batch_normalize(x3, is_training)
                    with tf.variable_scope('ud7'):
                        block_out = lrelu(deconv_layer(tf.concat([x1, x2, x3],3), [3, 3, 64, 192], [self.batch_size, self.weight, self.height, 64], 1))
                    #x = x1+x2+x3+x
                    x_f+=block_out
            with tf.variable_scope('conv_e8'):
                x_f = conv_layer(x_f, [3, 3, 64, 256], 1)
                x_f = lrelu(x_f)
            #res_in = x_f
            # mask

Then in test.py comment this problematic lines:

            fake = sess.run(
                [model.ZConv_VDSR],
                feed_dict={x: input_, is_training: False})

And change it with pass

This fixes seem to work for me. I hope they are correct

hqmf8104 · 2021-02-15T15:51:59Z

@kuihua if you could please double check. I think you might have pushed the wrong version of the file. Is that possible?

It seems the TESTGAN.py is wrong.
The layer named conv_m2 does not exist in the trained weights.
Instead of this layer, all this block seems to be missing:

            res_in = x_f
            # frame
            for i in range(3):
                with tf.variable_scope('block{}ex2'.format(i+1)):
                    x1=x2=x3=x_f
                    for j in range(3):
                        with tf.variable_scope('block{}_{}ex1'.format(i+1,j+1)):
                            with tf.variable_scope('ud1'):
                                a1 = lrelu(deconv_layer(x1, [3, 3, 64, 64], [self.batch_size, self.weight, self.height, 64], 1))
                                #a1 = batch_normalize(a1, is_training)
                            with tf.variable_scope('ud2'):
                                b1 = lrelu(deconv_layer(x2, [3, 3, 64, 64], [self.batch_size, self.weight, self.height, 64], 1))
                                #b1 = batch_normalize(b1, is_training)
                            with tf.variable_scope('ud3'):
                                c1 = lrelu(deconv_layer(x3, [3, 3, 64, 64], [self.batch_size, self.weight, self.height, 64], 1))
                                #c1 = batch_normalize(c1, is_training)
                            sum = tf.concat([a1,b1,c1],3)
                            #sum = batch_normalize(sum, is_training)
                            with tf.variable_scope('ud4'):
                                x1 = lrelu(deconv_layer(tf.concat([sum,x1],3), [1, 1, 64, 256], [self.batch_size, self.weight, self.height, 64], 1))
                                #x1 = batch_normalize(x1, is_training)
                            with tf.variable_scope('ud5'):
                                x2 = lrelu(deconv_layer(tf.concat([sum,x2],3), [1, 1, 64, 256], [self.batch_size, self.weight, self.height, 64], 1))
                                #x2 = batch_normalize(x2, is_training)
                            with tf.variable_scope('ud6'):
                                x3 = lrelu(deconv_layer(tf.concat([sum,x3],3), [1, 1, 64, 256], [self.batch_size, self.weight, self.height, 64], 1))
                                #x3 = batch_normalize(x3, is_training)
                    with tf.variable_scope('ud7'):
                        block_out = lrelu(deconv_layer(tf.concat([x1, x2, x3],3), [3, 3, 64, 192], [self.batch_size, self.weight, self.height, 64], 1))
                    #x = x1+x2+x3+x
                    x_f+=block_out
            with tf.variable_scope('conv_e8'):
                x_f = conv_layer(x_f, [3, 3, 64, 256], 1)
                x_f = lrelu(x_f)
            #res_in = x_f
            # mask

Then in test.py comment this problematic lines:

            fake = sess.run(
                [model.ZConv_VDSR],
                feed_dict={x: input_, is_training: False})

And change it with pass

This fixes seem to work for me. I hope they are correct

Hi! Thanks for this. This is a bit of stupid question, but where do you need to put the pretrained weights? Also, are you saying that :

        with tf.variable_scope('conv_m2'):
            x_f = conv_layer(x_f, [3, 3, 64, 256], 1)
            x_f = lrelu(x_f)
        res_in = x_f

should be replaced by:

        res_in = x_f
        # frame
        for i in range(3):
            with tf.variable_scope('block{}ex2'.format(i+1)):
                x1=x2=x3=x_f
                for j in range(3):
                    with tf.variable_scope('block{}_{}ex1'.format(i+1,j+1)):
                        with tf.variable_scope('ud1'):
                            a1 = lrelu(deconv_layer(x1, [3, 3, 64, 64], [self.batch_size, self.weight, self.height, 64], 1))
                            #a1 = batch_normalize(a1, is_training)
                        with tf.variable_scope('ud2'):
                            b1 = lrelu(deconv_layer(x2, [3, 3, 64, 64], [self.batch_size, self.weight, self.height, 64], 1))
                            #b1 = batch_normalize(b1, is_training)
                        with tf.variable_scope('ud3'):
                            c1 = lrelu(deconv_layer(x3, [3, 3, 64, 64], [self.batch_size, self.weight, self.height, 64], 1))
                            #c1 = batch_normalize(c1, is_training)
                        sum = tf.concat([a1,b1,c1],3)
                        #sum = batch_normalize(sum, is_training)
                        with tf.variable_scope('ud4'):
                            x1 = lrelu(deconv_layer(tf.concat([sum,x1],3), [1, 1, 64, 256], [self.batch_size, self.weight, self.height, 64], 1))
                            #x1 = batch_normalize(x1, is_training)
                        with tf.variable_scope('ud5'):
                            x2 = lrelu(deconv_layer(tf.concat([sum,x2],3), [1, 1, 64, 256], [self.batch_size, self.weight, self.height, 64], 1))
                            #x2 = batch_normalize(x2, is_training)
                        with tf.variable_scope('ud6'):
                            x3 = lrelu(deconv_layer(tf.concat([sum,x3],3), [1, 1, 64, 256], [self.batch_size, self.weight, self.height, 64], 1))
                            #x3 = batch_normalize(x3, is_training)
                with tf.variable_scope('ud7'):
                    block_out = lrelu(deconv_layer(tf.concat([x1, x2, x3],3), [3, 3, 64, 192], [self.batch_size, self.weight, self.height, 64], 1))
                #x = x1+x2+x3+x
                x_f+=block_out
        with tf.variable_scope('conv_e8'):
            x_f = conv_layer(x_f, [3, 3, 64, 256], 1)
            x_f = lrelu(x_f)
        #res_in = x_f
        # mask

kuijiang94 · 2021-02-16T03:06:09Z

I am sorry for this issue. the model.ZConv_VDSR should be [model.imitation_sr, model.base_sr, model.frame_sr]. I have corrected it.

hqmf8104 · 2021-02-16T17:29:07Z

I am sorry for this issue. the model.ZConv_VDSR should be [model.imitation_sr, model.base_sr, model.frame_sr]. I have corrected it.

I'm afraid this generated the following error for me:

File "test.py", line 67, in
[model.model.imitation_sr, model.base_sr, model.frame_sr],
AttributeError: 'Model' object has no attribute 'model'

kuijiang94 · 2021-02-22T08:04:46Z

hi, the first item has two model. Only one is necessary. kuijiang 邮箱：[email protected] 签名由网易邮箱大师定制 On 02/17/2021 01:29, hqmf8104 wrote: I am sorry for this issue. the model.ZConv_VDSR should be [model.imitation_sr, model.base_sr, model.frame_sr]. I have corrected it. I'm afraid this generated the following error for me: File "test.py", line 67, in [model.model.imitation_sr, model.base_sr, model.frame_sr], AttributeError: 'Model' object has no attribute 'model' — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

zensenlon · 2021-06-23T08:29:30Z

Same problem here.
I have also trained for some epochs and used my trained weights for testing. Got the same problem

Did you solved that, I also got this problem but it seems that the hqmf8104's solution doesn't work for me.

paugalles mentioned this issue Jan 28, 2021

what does it mean by 'model.ZConv_VDSR'? #2

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

EEGAN_pretrained_doesn't work #6

EEGAN_pretrained_doesn't work #6

ajitaru commented Nov 17, 2020

paugalles commented Jan 27, 2021 •

edited

Loading

paugalles commented Jan 28, 2021 •

edited

Loading

hqmf8104 commented Feb 15, 2021

kuijiang94 commented Feb 16, 2021

hqmf8104 commented Feb 16, 2021

kuijiang94 commented Feb 22, 2021 via email

zensenlon commented Jun 23, 2021 •

edited

Loading

EEGAN_pretrained_doesn't work #6

EEGAN_pretrained_doesn't work #6

Comments

ajitaru commented Nov 17, 2020

paugalles commented Jan 27, 2021 • edited Loading

paugalles commented Jan 28, 2021 • edited Loading

hqmf8104 commented Feb 15, 2021

kuijiang94 commented Feb 16, 2021

hqmf8104 commented Feb 16, 2021

kuijiang94 commented Feb 22, 2021 via email

zensenlon commented Jun 23, 2021 • edited Loading

paugalles commented Jan 27, 2021 •

edited

Loading

paugalles commented Jan 28, 2021 •

edited

Loading

zensenlon commented Jun 23, 2021 •

edited

Loading