updated codes and readme with results

justanhduc · Aug 30, 2019 · 9cdfd03 · 9cdfd03
1 parent f323012
commit 9cdfd03
Show file tree

Hide file tree

Showing 17 changed files with 554 additions and 940 deletions.
diff --git a/README.md b/README.md
@@ -1,17  1,29 @@
 # GraphX-Convolution
 
-The official implementation of the ICCV paper "GraphX-convolution for point cloud deformation in 2D-to-3D conversion".
 The official implementation of the ICCV 2019 paper 
 "GraphX-convolution for point cloud deformation in 2D-to-3D conversion".
 
 ![framework](./imgs/framework.jpg)
 
 __Updates:__ We have achieved a better performance with `PointCloudResLowRankGraphXUpDecoder`, 
 which is an upgraded version of the best deformation network in the paper. The new network works 
 comparably with the best network in the paper but consumes much less memory.
 
 ## Prerequisite
 
 [Pytorch](https://pytorch.org/get-started/locally/) (>=1.0.0)
 
 [Neuralnet-pytorch](https://github.com/justanhduc/neuralnet-pytorch) 
 (branch [fancy](https://github.com/justanhduc/neuralnet-pytorch/tree/fancy))
 
 [Gin-config](https://github.com/google/gin-config)
 
 [Matplotlib](https://matplotlib.org/)
 
 [TensorboardX](https://github.com/lanpa/tensorboardX)
 
 The code has been tested on a Linux Ubuntu 16.04 machine.
 
 ## Data
 
 The data we used in our experiment provided by [ShapeNet](https://www.shapenet.org/). 
@@ -33,32  45,62 @@ python split.py path/to/the/downloaded/data/folder
 ### Training
 
 After the database is setup, we are ready to train the model. 
-At the root of the project, execute
 In the [configs](./src/configs) folder, we have prepared the basic training configs for several models
 used in the paper.
 After the data is downloaded, specify the data root in these config files.
 
-```
-cd src
-python train.py path/to/the/downloaded/data/folder
-```
 By default, the script creates a `results` folder containing checkpoint folders 
 corresponding to the model name.
 The model name can be specified in the config files.
 
-To resume a half-done training, simply specify the folder containing the weight file using the checkpoint flag
 Then, simply execute
 
 ```
-python train.py path/to/the/downloaded/data/folder --checkpoint path/to/the/checkpoint/folder
 cd src
 python train.py path/to/a/config/file (--gpu 0)
 ```
 
-For more options, use ```python train.py -h```.
 To resume a half-done training, simply specify the checkpoint folder in the config file.
 Then run the training script as in the training step.
 
 ### Evaluation
 
 Be sure to use the same config file and specify the checkpoint folder in the config file.
 After the model is fully trained, to test the model, use
 
 ```
-python test.py path/to/the/downloaded/data/folder --checkpoint path/to/the/checkpoint/folder
 python test.py path/to/a/config/file (--gpu 0)
 ```
 
-The script calculates the Chamfer distance (CD) scores similar to [Pixel2mesh](https://github.com/nywang16/Pixel2Mesh).
 The script calculates the Chamfer distance (CD) scores similar to 
 [Pixel2mesh](https://github.com/nywang16/Pixel2Mesh).
 
 ## Results
 
 ### Visualizations
 
 <p align='center'>
   <img src='imgs/results/airplane.gif' width="250px">
   <img src='imgs/results/bench.gif' width="250px">
   <img src='imgs/results/car.gif' width="250px">
 </p>
 
 <p align='center'>
   <img src='imgs/results/lamp.gif' width="250px">
   <img src='imgs/results/monitor.gif' width="250px">
   <img src='imgs/results/table.gif' width="250px">
 </p>
 
 ### Numerical results
 
 |Methods|IoU|Chamfer|
 |:-------:|:---:|:-------:|
 |[3D-R2N2](http://3d-r2n2.stanford.edu/)|1.445|0.631|
 |[PSG](https://arxiv.org/abs/1612.00603)|0.593|0.640|
 |[Pixel2mesh](https://github.com/nywang16/Pixel2Mesh)|0.591|-|
 |[GAL](https://link.springer.com/chapter/10.1007/978-3-030-01237-3_49)|-|0.712|
 |Ours (UpResGraphX)|__0.252__|__0.725__|
 
-## References
 ## Citation
 
-In our experiment, we actually used a CUDA implementation of CD available [here](https://github.com/ThibaultGROUEIX/AtlasNet/tree/master/extension), 
-which is much more efficient in terms of memory and speed.
 TBA
diff --git a/imgs/results/airplane.gif b/imgs/results/airplane.gif
diff --git a/imgs/results/bench.gif b/imgs/results/bench.gif
diff --git a/imgs/results/car.gif b/imgs/results/car.gif
diff --git a/imgs/results/lamp.gif b/imgs/results/lamp.gif
diff --git a/imgs/results/monitor.gif b/imgs/results/monitor.gif
diff --git a/imgs/results/table.gif b/imgs/results/table.gif
diff --git a/src/configs/fc-final.gin b/src/configs/fc-final.gin
@@ -0,0  1,24 @@
 GraphX.data_root = '/media/hdd2/csh/ShapeNet7Class2'
 
 GraphX.name = 'ICCV-fc-final'
 GraphX.img_enc = @cnn18_enc
 GraphX.pc_enc = @pc_enc
 GraphX.pc_dec = @pc_dec
 GraphX.adain = True
 GraphX.projection = True
 
 GraphX.color_img = False
 GraphX.n_points = 2000
 GraphX.bs = 4
 
 GraphX.optimizer = @adam
 GraphX.scheduler = @multistep_lr
 GraphX.lr = 5e-5
 GraphX.weight_decay = 1e-5
 GraphX.gamma = .3
 GraphX.milestones = (5, 8)
 GraphX.n_epochs = 10
 GraphX.print_freq = 1000
 GraphX.val_freq = 10000
 
 GraphX.checkpoint_folder = None
diff --git a/src/configs/graphx-up-final.gin b/src/configs/graphx-up-final.gin
@@ -0,0  1,24 @@
 GraphX.data_root = '/media/hdd2/csh/ShapeNet7Class2'
 
 GraphX.name = 'ICCV-graphx-conv-up-final'
 GraphX.img_enc = @cnn18_enc
 GraphX.pc_enc = @pc_enc
 GraphX.pc_dec = @pc_upresgraphxdec
 GraphX.adain = True
 GraphX.projection = True
 
 GraphX.color_img = False
 GraphX.n_points = 250
 GraphX.bs = 4
 
 GraphX.optimizer = @adam
 GraphX.scheduler = @multistep_lr
 GraphX.lr = 5e-5
 GraphX.weight_decay = 1e-5
 GraphX.gamma = .3
 GraphX.milestones = (5, 8)
 GraphX.n_epochs = 10
 GraphX.print_freq = 1000
 GraphX.val_freq = 10000
 
 GraphX.checkpoint_folder = None
diff --git a/src/configs/lowrankgraphx-up-final.gin b/src/configs/lowrankgraphx-up-final.gin
@@ -0,0  1,25 @@
 GraphX.data_root = '/media/hdd2/csh/ShapeNet7Class2'
 
 GraphX.name = 'ICCV-lowrankgraphx-conv-up-final'
 GraphX.img_enc = @cnn18_enc
 GraphX.pc_enc = @pc_enc
 GraphX.pc_dec = @pc_upreslowrankgraphxdec
 GraphX.adain = True
 GraphX.projection = True
 decimation = .5
 
 GraphX.color_img = False
 GraphX.n_points = 250
 GraphX.bs = 4
 
 GraphX.optimizer = @adam
 GraphX.scheduler = @multistep_lr
 GraphX.lr = 5e-5
 GraphX.weight_decay = 1e-5
 GraphX.gamma = .3
 GraphX.milestones = (5, 8)
 GraphX.n_epochs = 10
 GraphX.print_freq = 1000
 GraphX.val_freq = 100000
 
 GraphX.checkpoint_folder = None
diff --git a/src/data_loader.py b/src/data_loader.py
@@ -38,26  38,25 @@ def collate(batch):
 
 
 class ShapeNet(Dataset):
-    def __init__(self, file_list, path, grayscale=None, type='train', n_points=2000, metadata=False, **kwargs):
     def __init__(self, path, grayscale=None, type='train', n_points=2000, **kwargs):
         assert type in ('train', 'valid', 'test')
         self.n_points = n_points
         self.grayscale = grayscale
-        self.file_list = file_list
         self.file_list = os.listdir(path)
         self.path = path
         self.type = type if type in ('train', 'test') else 'test'
-        self.metadata = metadata
         self.num_vals = kwargs.pop('num_vals', 30)
         self.pkl_list = []
         self.sample_weights = []
-        for folder in file_list:
         for folder in self.file_list:
             file_path = os.listdir(os.path.join(path, folder, self.type))
             if type == 'valid':
-                idx = np.random.randint(len(file_path), size=self.num_vals // len(file_list))
                 idx = np.random.randint(len(file_path), size=self.num_vals // len(self.file_list))
                 file_path = [file_path[i] for i in idx]
 
             file_path = [os.path.join(self.path, folder, self.type, f) for f in file_path]
             self.pkl_list.extend(file_path)
-            self.sample_weights.extend([1 / len(file_path)] * len(file_path))
             self.sample_weights.extend([1/len(file_path)] * len(file_path))
 
     def __len__(self):
         return len(self.pkl_list)
@@ -69,9  68,4 @@ def __getitem__(self, idx):
         img = (np.transpose(img / 255.0, (2, 0, 1)) - .5) * 2
         pc = np.array(contents[1], 'float32')[:, :3]
         pc -= np.mean(pc, 0, keepdims=True)
-        item = (init_pointcloud_loader(self.n_points), np.array(img, 'float32'), pc)
-        if self.metadata:
-            metadata = pkl_path.split('\\')[-1][:-4]
-            item  = (metadata,)
-
-        return item
         return init_pointcloud_loader(self.n_points), np.array(img, 'float32'), pc