Merge branch 'develop'

README.md

@@ -1,3 +1,7 @@
+**About PyTorch 1.2.0**
+  * Now the master branch supports PyTorch 1.2.0 by default.
+  * Due to the serious version problem (especially torch.utils.data.dataloader), MDSR functions are temporarily disabled. If you have to train/evaluate the MDSR model, please use legacy branches.
+
 # EDSR-PyTorch
 
 **About PyTorch 1.1.0**
@@ -20,7 +24,7 @@ If you find our work useful in your research or publication, please cite our wor
   year = {2017}
 }
 ```
-We provide scripts for reproducing all the results from our paper. You can train your own model from scratch, or use pre-trained model to enlarge your images.
+We provide scripts for reproducing all the results from our paper. You can train your model from scratch, or use a pre-trained model to enlarge your images.
 
 **Differences between Torch version**
 * Codes are much more compact. (Removed all unnecessary parts.)
@@ -46,8 +50,8 @@ git clone https://github.com/thstkdgus35/EDSR-PyTorch
 cd EDSR-PyTorch
 ```
 
-## Quick start (Demo)
-You can test our super-resolution algorithm with your own images. Place your images in ``test`` folder. (like ``test/<your_image>``) We support **png** and **jpeg** files.
+## Quickstart (Demo)
+You can test our super-resolution algorithm with your images. Place your images in ``test`` folder. (like ``test/<your_image>``) We support **png** and **jpeg** files.
 
 Run the script in ``src`` folder. Before you run the demo, please uncomment the appropriate line in ```demo.sh``` that you want to execute.
 ```bash
@@ -123,17 +127,17 @@ sh demo.sh
   * Basically, this function first split a large image to small patches. Those images are merged after super-resolution. I checked this function with 12GB memory, 4000 x 2000 input image in scale 4. (Therefore, the output will be 16000 x 8000.)
 
 * Feb 21, 2018
-  * Fixed the problem when loading pre-trained multi-gpu model.
+  * Fixed the problem when loading pre-trained multi-GPU model.
   * Added pre-trained scale 2 baseline model.
-  * This code now only saves the best-performing model by default. For MDSR, 'the best' can be ambiguous. Use --save_models argument to save all the intermediate models.
+  * This code now only saves the best-performing model by default. For MDSR, 'the best' can be ambiguous. Use --save_models argument to keep all the intermediate models.
   * PyTorch 0.3.1 changed their implementation of DataLoader function. Therefore, I also changed my implementation of MSDataLoader. You can find it on feature/dataloader branch.
 
 * Feb 23, 2018
-  * Now PyTorch 0.3.1 is default. Use legacy/0.3.0 branch if you use the old version.
+  * Now PyTorch 0.3.1 is a default. Use legacy/0.3.0 branch if you use the old version.
   * With a new ``src/data/DIV2K.py`` code, one can easily create new data class for super-resolution.
   * New binary data pack. (Please remove the ``DIV2K_decoded`` folder from your dataset if you have.)
-  * With ``--ext bin``, this code will automatically generates and saves the binary data pack that corresponds to previous ``DIV2K_decoded``. (This requires huge RAM (~45GB, Swap can be used.), so please be careful.)
-  * If you cannot make the binary pack, just use the default setting (``--ext img``).
+  * With ``--ext bin``, this code will automatically generate and saves the binary data pack that corresponds to previous ``DIV2K_decoded``. (This requires huge RAM (~45GB, Swap can be used.), so please be careful.)
+  * If you cannot make the binary pack, use the default setting (``--ext img``).
 
   * Fixed a bug that PSNR in the log and PSNR calculated from the saved images does not match.
   * Now saved images have better quality! (PSNR is ~0.1dB higher than the original code.)
@@ -146,23 +150,23 @@ sh demo.sh
 * Mar 11, 2018
   * Fixed some typos in the code and script.
   * Now --ext img is default setting. Although we recommend you to use --ext bin when training, please use --ext img when you use --test_only.
-  * Skip_batch operation is implemented. Use --skip_threshold argument to skip the batch that you want to ignore. Although this function is not exactly same with that of Torch7 version, it will work as you expected.
+  * Skip_batch operation is implemented. Use --skip_threshold argument to skip the batch that you want to ignore. Although this function is not exactly the same with that of Torch7 version, it will work as you expected.
 
 * Mar 20, 2018
-  * Use ``--ext sep_reset`` to pre-decode large png files. Those decoded files will be saved to the same directory with DIV2K png files. After the first run, you can use ``--ext sep`` to save time.
+  * Use ``--ext sep-reset`` to pre-decode large png files. Those decoded files will be saved to the same directory with DIV2K png files. After the first run, you can use ``--ext sep`` to save time.
   * Now supports various benchmark datasets. For example, try ``--data_test Set5`` to test your model on the Set5 images.
   * Changed the behavior of skip_batch.
 
 * Mar 29, 2018
   * We now provide all models from our paper.
-  * We also provide ``MDSR_baseline_jpeg`` model that suppresses JPEG artifacts in original low-resolution image. Please use it if you have any trouble.
+  * We also provide ``MDSR_baseline_jpeg`` model that suppresses JPEG artifacts in the original low-resolution image. Please use it if you have any trouble.
   * ``MyImage`` dataset is changed to ``Demo`` dataset. Also, it works more efficient than before.
   * Some codes and script are re-written.
 
 * Apr 9, 2018
   * VGG and Adversarial loss is implemented based on [SRGAN](http://openaccess.thecvf.com/content_cvpr_2017/papers/Ledig_Photo-Realistic_Single_Image_CVPR_2017_paper.pdf). [WGAN](https://arxiv.org/abs/1701.07875) and [gradient penalty](https://arxiv.org/abs/1704.00028) are also implemented, but they are not tested yet.
   * Many codes are refactored. If there exists a bug, please report it.
-  * [D-DBPN](https://arxiv.org/abs/1803.02735) is implemented. Default setting is D-DBPN-L.
+  * [D-DBPN](https://arxiv.org/abs/1803.02735) is implemented. The default setting is D-DBPN-L.
 
 * Apr 26, 2018
   * Compatible with PyTorch 0.4.0
@@ -171,9 +175,12 @@ sh demo.sh
 
 * July 22, 2018
   * Thanks for recent commits that contains RDN and RCAN. Please see ``code/demo.sh`` to train/test those models.
-  * Now the dataloader is much stable than the previous version. Please erase ``DIV2K/bin`` folder that is created before this commit. Also, please avoid to use ``--ext bin`` argument. Our code will automatically pre-decode png images before training. If you do not have enough spaces(~10GB) in your disk, we recommend ``--ext img``(But SLOW!).
+  * Now the dataloader is much stable than the previous version. Please erase ``DIV2K/bin`` folder that is created before this commit. Also, please avoid using ``--ext bin`` argument. Our code will automatically pre-decode png images before training. If you do not have enough spaces(~10GB) in your disk, we recommend ``--ext img``(But SLOW!).
 
 * Oct 18, 2018
-  * with ``--pre_train download``, pretrained models will be automatically downloaded from server.
+  * with ``--pre_train download``, pretrained models will be automatically downloaded from the server.
   * Supports video input/output (inference only). Try with ``--data_test video --dir_demo [video file directory]``.
 
+* About PyTorch 1.0.0
+  * We support PyTorch 1.0.0. If you prefer the previous versions of PyTorch, use legacy branches.
+  * ``--ext bin`` is not supported. Also, please erase your bin files with ``--ext sep-reset``. Once you successfully build those bin files, you can remove ``-reset`` from the argument.

src/data/__init__.py

@@ -1,5 +1,6 @@
 from importlib import import_module
-from dataloader import MSDataLoader
+#from dataloader import MSDataLoader
+from torch.utils.data import dataloader
 from torch.utils.data import ConcatDataset
 
 # This is a simple wrapper function for ConcatDataset
@@ -22,12 +23,12 @@ def __init__(self, args):
                 m = import_module('data.' + module_name.lower())
                 datasets.append(getattr(m, module_name)(args, name=d))
 
-            self.loader_train = MSDataLoader(
-                args,
+            self.loader_train = dataloader.DataLoader(
                 MyConcatDataset(datasets),
                 batch_size=args.batch_size,
                 shuffle=True,
-                pin_memory=not args.cpu
+                pin_memory=not args.cpu,
+                num_workers=args.n_threads,
             )
 
         self.loader_test = []
@@ -40,11 +41,12 @@ def __init__(self, args):
                 m = import_module('data.' + module_name.lower())
                 testset = getattr(m, module_name)(args, train=False, name=d)
 
-            self.loader_test.append(MSDataLoader(
-                args,
-                testset,
-                batch_size=1,
-                shuffle=False,
-                pin_memory=not args.cpu
-            ))
-
+            self.loader_test.append(
+                dataloader.DataLoader(
+                    testset,
+                    batch_size=1,
+                    shuffle=False,
+                    pin_memory=not args.cpu,
+                    num_workers=args.n_threads,
+                )
+            )

src/demo.sh

@@ -1,5 +1,5 @@
 # EDSR baseline model (x2) + JPEG augmentation
-#python main.py --model EDSR --scale 2 --patch_size 96 --save edsr_baseline_x2 --reset
+python main.py --model EDSR --scale 2 --patch_size 96 --save edsr_baseline_x2 --reset
 #python main.py --model EDSR --scale 2 --patch_size 96 --save edsr_baseline_x2 --reset --data_train DIV2K+DIV2K-Q75 --data_test DIV2K+DIV2K-Q75
 
 # EDSR baseline model (x3) - from EDSR baseline model (x2)

src/trainer.py

@@ -37,13 +37,15 @@ def train(self):
         self.model.train()
 
         timer_data, timer_model = utility.timer(), utility.timer()
-        for batch, (lr, hr, _, idx_scale) in enumerate(self.loader_train):
+        # TEMP
+        self.loader_train.dataset.set_scale(0)
+        for batch, (lr, hr, _,) in enumerate(self.loader_train):
             lr, hr = self.prepare(lr, hr)
             timer_data.hold()
             timer_model.tic()
 
             self.optimizer.zero_grad()
-            sr = self.model(lr, idx_scale)
+            sr = self.model(lr, 0)
             loss = self.loss(sr, hr)
             loss.backward()
             if self.args.gclip > 0:
@@ -84,7 +86,7 @@ def test(self):
         for idx_data, d in enumerate(self.loader_test):
             for idx_scale, scale in enumerate(self.scale):
                 d.dataset.set_scale(idx_scale)
-                for lr, hr, filename, _ in tqdm(d, ncols=80):
+                for lr, hr, filename in tqdm(d, ncols=80):
                     lr, hr = self.prepare(lr, hr)
                     sr = self.model(lr, idx_scale)
                     sr = utility.quantize(sr, self.args.rgb_range)