jiny2001 / deeply-recursive-cnn-tf Goto Github PK

As you said, 'I use half num of features (128) to make training faster for those results below.' I am a little confuse. How many of your feature number of result image shown at the page. 64, 128 or 96. Thanks.

Duplicate flags error, keyerror verbosity
on windows 10 platform for spyder

When I test with a trained model, I use the command: python test.py --dataSet14 --inference_depth 9 --feature_num 96 has the following error

Features:96 Inference Depth:9 Initial LR:0.00100 [model_F96_D9_LR0.001000]
(3, 3, 1, 96)-864, (96,)-96, (3, 3, 96, 96)-82944, (96,)-96, (3, 3, 96, 96)-82944, (96,)-96, (3, 3, 96, 96)-82944, (96,)-96, (3, 3, 97, 1)-873, (1,)-1, (9,)-9,
Total 11 variables, 250,963 params
Model restored.
Traceback (most recent call last):
File "test.py", line 82, in
if name == 'main':
File "C:\Users\lenovo\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\platform\app.py", line 125, in run
_sys.exit(main(argv))
File "test.py", line 78, in main
model.init_all_variables(load_initial_data=FLAGS.load_model)
File "E:\程序\deeply-recursive-cnn-tf-master\super_resolution.py", line 446, in do_super_resolution
org_image = util.load_image(file_path)
File "E:\程序\deeply-recursive-cnn-tf-master\super_resolution_utilty.py", line 205, in load_image
raise LoadError("File not found [%s]" % filename)
super_resolution_utilty.LoadError: File not found []

Hi @jiny2001 , thank you for sharing your great work , i have some (1000 x 660 ) photos want to scale it with 2x upscal ( 2000 x 1334 ) , could you please explan to me how to get that , when i try it give me output photo with the same size for input photo .

Thank you for your project. It seems that the BSD 100 4x link in your url provided is broken and I cannot find the link your Urban 100 4x database in your link. Could you please upload to Dropbox? Thanks.

hi, thank you for your code! i want to try your code on gpu ,but i don not know how to add the code, Cloud you tell me where i add the code " tf.device("/gpu:0")"?

Could you please show you progress log of training with scale 4x? I used my custom domain images for training and want to compare with your result? Thanks.

Their resolutions are the same and their quality seem almost the same. Are image_SRF_4 is used for training scale 4x? Thanks.

Hello there! Which is your X4 model? I changed the scale to 4 and the result was very poor.

I run the code you provided in my own computer and finally the PSNR eauqls to about from 5 to 7, not 3* as you showed. What happened to this suitation?

In paper, we find the authors use conv+relu in inference Network. Due to:

The recurrence relation is Hd = g(Hd−1) = max(0; W ∗ Hd−1 + b), Inference net f2 is equivalent to the composition of the same elementary function g: f2(H) = (g ◦ g ◦ · · · ◦)g(H) = gD(H)

But in your code, you have not use relu in the inference Network.
Your code: (the 200th-201th line in super_resolution.py file)

for i in range(0, self.inference_depth):
      self.H_conv[i+1] = util.conv2d_with_bias(self.H_conv[i], self.W_conv, 1, self.B_conv, name="H%d"%(i+1))

I think need to change like this:

for i in range(0, self.inference_depth):
      self.H_conv[i+1] = util.conv2d_with_bias(self.H_conv[i], self.W_conv, 1, self.B_conv, add_relu=True, name="H%d"%(i+1))

Is this a mistake? The reason you can't the achieve the performance like authors?
Looking forward to your reply

i wanna to kown, is that a trick to convert RGB space to Ycrcb space

In the paper, the inference network only contains a single conv layer, and it is conducted with a D-times loop. In this implementation, it seems construct D conv layers with same W and B. Thus, this version cannot accomplish reduction on parameters and memory, with no recursion, right?

Hi, Jin:
I am sorry to bother you again. However, I have met a question recently.

In your super_resolution.py file, 227-230 line:
Your code:

for i in range(0, self.inference_depth + 1):
	self.Y1_conv[i] = util.conv2d_with_bias(self.H_conv[i], self.WD1_conv, self.cnn_stride, self.BD1_conv,
											add_relu=not self.residual, name="Y%d_1" % i)
	self.Y2_conv[i] = util.conv2d_with_bias(self.Y1_conv[i], self.WD2_conv, self.cnn_stride, self.BD2_conv,
											add_relu=not self.residual, name="Y%d_2" % i)

However, self.H_conv[0] is in embeding network. And from the authors' papet Figure 3(c). The skip connection should start self_conv[1].

and in 263-271 line,
Your code:

for i in range(0, self.inference_depth):
	if self.residual:
		self.Y2_conv[i] = self.Y2_conv[i] + self.x
	inference_sub = tf.subtract(self.y, self.Y2_conv[i], name="Loss1_%d_sub" % i)
	inference_square = tf.square(inference_sub, name="Loss1_%d_squ" % i)
	loss1_mse[i] = tf.reduce_mean(inference_square, name="Loss1_%d" % i)

loss1 = loss1_mse[0]
for i in range(1, self.inference_depth):
	if i == self.inference_depth:
		loss1 = tf.add(loss1, loss1_mse[i], name="loss1")
	else:
		loss1 = tf.add(loss1, loss1_mse[i], name="loss1_%d_add" % i)

only calculate loss from H[0] to H[self.inference_depth-1].
In fact, the loss should calculate from H[1] to H[self.inference_depth].

I think it should change like this. for 220-230 lines:

self.Y1_conv = self.inference_depth * [None]
self.Y2_conv = self.inference_depth * [None]
self.W = tf.Variable(
	np.full(fill_value=1.0 / (self.inference_depth), shape=[self.inference_depth], dtype=np.float32),name="layer_weight")
W_sum = tf.reduce_sum(self.W)

for i in range(0, self.inference_depth):
	self.Y1_conv[i] = util.conv2d_with_bias(self.H_conv[i+1], self.WD1_conv, self.cnn_stride, self.BD1_conv,
											add_relu=True, name="Y%d_1" % i)
	self.Y2_conv[i] = util.conv2d_with_bias(self.Y1_conv[i], self.WD2_conv, self.cnn_stride, self.BD2_conv,
											add_relu=not self.residual, name="Y%d_2" % i)

And it doesn't need to change for 263-271 line,.

I have got these results: (Set91 aug x4, with residual learning)
Set5: 37.04
Set14: 32.57
urban100: 29.58
BSD: 31.41

And the converge is faster, I only use 2.5 hours get the results in NV1080 GPU

Hi,

if I understand the paper correctly, we input a small image and output a large image with good resolution.

So I read through your code and I dont know where you apply transposed convolution(upscaling) to get larger outputs.

from the paper it must be in re-construct layer but I just cannot find in your code.

Can you explain a bit about this point?