Cannot specify no cropping

[gwern]

(This is related to but not identical to issue #2 .)

Cropping is particularly undesirable on very small images like 64x64 where it may delete a lot of the image (especially when the images come pre-centered and cropped already). Currently, you cannot run dcgan.torch with no cropping despite the configurable arguments like loadSize=64 fineSize=64 suggesting that should be possible. This is not due to design but a bug in the cropping code in data/donkey_folder.lua, it seems; said in issue #2:

Right now, loadSize has to be greater than fineSize (because of a bug in the cropping logic). So it's okay to have loadSize=65 fineSize=64 th main.lua

---

I messed around some with the responsible trainHook and I think the bug can be fixed by simply checking for the case where the original H/W are greater than the fineSize value and if they aren't, feeding 0s into the crop function, so the new version would look like this:

-- do random crop if fineSize/sampleSize is configured to be smaller than NN's input dimensions, loadSize
local iW = input:size(3)
local iH = input:size(2)
local oW = sampleSize[2]
local oH = sampleSize[2]
if (iW > oW) then
 w1 = math.ceil(torch.uniform(1e-2, iW-oW))
else
 w1 = 0
end
if (iH > oH) then
 h1 = math.ceil(torch.uniform(1e-2, iH-oH))
else
 h1 = 0
end
local out = image.crop(input, w1, h1, w1 + oW, h1 + oH)
assert(out:size(2) == oW)
assert(out:size(3) == oH)

Or to diff it:

diff --git a/data/donkey_folder.lua b/data/donkey_folder.lua
index 3a82393..5248f4e 100644
--- a/data/donkey_folder.lua
+++ b/data/donkey_folder.lua
@@ -52,17 +52,26 @@ local mean,std
 local trainHook = function(self, path)
    collectgarbage()
    local input = loadImage(path)
+
+   -- do random crop if fineSize/sampleSize is configured to be smaller than NN's input dimensions, loadSize
    local iW = input:size(3)
    local iH = input:size(2)
-
-   -- do random crop
-   local oW = sampleSize[2];
+   local oW = sampleSize[2]
    local oH = sampleSize[2]
-   local h1 = math.ceil(torch.uniform(1e-2, iH-oH))
-   local w1 = math.ceil(torch.uniform(1e-2, iW-oW))
+   if (iW > oW) then
+    w1 = math.ceil(torch.uniform(1e-2, iW-oW))
+   else
+    w1 = 0
+   end
+   if (iH > oH) then
+    h1 = math.ceil(torch.uniform(1e-2, iH-oH))
+   else
+    h1 = 0
+   end
    local out = image.crop(input, w1, h1, w1 + oW, h1 + oH)
    assert(out:size(2) == oW)
    assert(out:size(3) == oH)
+
    -- do hflip with probability 0.5
    if torch.uniform() > 0.5 then out = image.hflip(out); end
    out:mul(2):add(-1) -- make it [0, 1] -> [-1, 1]

This seems to work both in the 64x64px default version and the 128x128px fork, eg

$ nThreads=1 DATA_ROOT=myimages dataset=folder batchSize=2 loadSize=128 fineSize=128 nz=75 ngf=106 ndf=48 gpu=0 th main-128.lua
{
  ntrain : inf
  beta1 : 0.5
  name : "experiment1"
  niter : 25
  batchSize : 2
  ndf : 48
  fineSize : 128
  nz : 75
  loadSize : 128
  gpu : 0
  ngf : 106
  dataset : "folder"
  lr : 0.0002
  noise : "normal"
  nThreads : 1
  display_id : 10
  display : 1
}
Random Seed: 5143   
Starting donkey with id: 1 seed: 5144
table: 0x406af6b8
Loading train metadata from cache
Dataset: folder  Size:  442215  
Epoch: [1][       0 /   221107]  Time: 8.181  DataTime: 0.003    Err_G: 1.1998  Err_D: 1.1637   
Epoch: [1][       1 /   221107]  Time: 5.115  DataTime: 0.001    Err_G: 0.3660  Err_D: 1.5839   
Epoch: [1][       2 /   221107]  Time: 5.965  DataTime: 0.001    Err_G: 2.8597  Err_D: 1.7219   
Epoch: [1][       3 /   221107]  Time: 6.163  DataTime: 0.001    Err_G: 0.1956  Err_D: 2.2080   
Epoch: [1][       4 /   221107]  Time: 5.537  DataTime: 0.001    Err_G: 0.7360  Err_D: 1.9527   
Epoch: [1][       5 /   221107]  Time: 6.300  DataTime: 0.001    Err_G: 6.8542  Err_D: 3.6255   
...

And looking at the displayed training sample images in the display server, they don't look cropped like before. So although I haven't run anything to completion, I think that fix works.