makes it work

GEOMetrics.py

@@ -47,15 +47,16 @@
 modelA = MeshDeformationBlock(963)
 modelB = MeshDeformationBlock(1155)
 modelC = MeshDeformationBlock(1155)
-encoder_mesh = MeshEncoder(100)
+encoder_mesh = MeshEncoder(50)
 modelA.cuda(), encoder1.cuda(), encoder2.cuda(),encoder3.cuda(), modelB.cuda(), modelC.cuda(), encoder_mesh.cuda()
 params = list(modelA.parameters()) + list(encoder1.parameters()) + list(encoder2.parameters()) + list(encoder3.parameters()) +list(modelB.parameters())  +list(modelC.parameters())
 optimizer = optim.Adam(params,lr=args.lr)
 
 try:
     encoder_mesh.load_state_dict(torch.load('checkpoint/'+ args.object + '/encoder'))
 except:
-    print 'You forgot to load the auto-encoder'
+    print 'You forgot to train the auto-encoder'
+    exit()
 if args.render:
     modelA.load_state_dict(torch.load(checkpoint_dir + 'model_1'))
     modelB.load_state_dict(torch.load(checkpoint_dir + 'model_2'))

Loss_Comparison/loss_comparison.py

@@ -1,9 +1,12 @@
-BASE_DIR = os.path.dirname(os.path.abspath(__file__))
-sys.path.append(BASE_DIR)
-sys.path.append(BASE_DIR + '../scripts/')
-from __future__ import division
+import os
+import sys
+BASEDIR =  os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+sys.path.append(BASEDIR)
+sys.path.append(BASEDIR + '/chamfer_distance/')
+
 from collections import namedtuple
 from utils import * 
+
 from torch.nn.parameter import Parameter
 
 
@@ -34,9 +37,6 @@ def forward(self, points):
 
 
 
-
-
-
 def vert_to_point(p_vert, p_face, t_vert, t_face, num = 3):
 
 	pred_points =  p_vert

README.md

@@ -16,7 +16,13 @@ There are 4 main ideas proposed in this project:
  * An adaptive face splitting procedure which analyses local face curvature to encourage local complexity emerge.
 
 
-
+This project runs with the following dependencies: 
+  * python 2.7 
+  * pytorch 0.4.0
+  * scipy
+  * matplotlib
+  * PIL 
+  * tqdm 
 
 ## Data Production
  To produce the data needed to train and test the methods of this project use the 'data_prep.py' script. This will download CAD models from the core classes of the ShapeNet data set, produce the data require for the latent loss, sample the surface of each ground truth mesh, render the objects as images, and split all the data into training, validation and test sets. This script makes use of the binvoxer executable, so first call
@@ -39,6 +45,13 @@ python data_prep.py --object plane -no 1000
 ## Diffentiable Surface Losses
 We introduce two new losses for reconstructing meshes. These losses are based of the idea of differentiating through the random selection of points on a triangular surface via the reparametrization trick. This allows the adoption of a chamfer loss comparing the samplings of ground truth and predicted mesh surfaces, which does not explicitly penalize the position of vertices. We call this the point-to-point loss. This idea also allows to the adoption of a more accurate loss which compares a sampled set of points to a surface directly, using the "3D point to triangle distance" algorithm. We call this the point-to-surface loss. We compare these two losses and to a loss which directly penalizes vertex position with respect to their ability to reconstruct surfaces, in the Loss_Comparison directory. 
 
+These functions require a python package to be built. To do this call: 
+ ```bash
+python chamfer_distance/build.py
+```
+
+
+
 <p align="center">
   <img  src="images/loss_comparison.png"  >
 </p>

data_prep.py

@@ -24,8 +24,7 @@
 
 parser = argparse.ArgumentParser(description='Dataset prep for image to 3D object super resolution')
 parser.add_argument('-o','--object', default=['chair'], help='List of object classes to be used downloaded and converted.', nargs='+' )
-parser.add_argument('-no','--num_objects', default=2, help='number of objects to be converted', type = int)
-parser.add_argument('-l','--low', default=32, help='Low resolution value', type = int)
+parser.add_argument('-no','--num_objects', default=10000, help='number of objects to be converted', type = int)
 args = parser.parse_args()
 
 
@@ -47,7 +46,7 @@
 		wanted_classes.append(l)
 print wanted_classes
 
-debug_mode = True # change to make all of the called scripts print their errors and warnings
+debug_mode = False # change to make all of the called scripts print their errors and warnings
 if debug_mode:
 	io_redirect = ''
 else:
@@ -186,7 +185,7 @@ def binvox():
 		commands =[]
 		count = 0
 		for d in tqdm(dirs):
-			command = './binvox ' + d  + ' -d ' + str(32)+ ' -pb -cb -c -e'   # this executable can be found at http://www.patrickmin.com/binvox/ ,
+			command = 'scripts/binvox ' + d  + ' -d ' + str(32)+ ' -pb -cb -c -e'   # this executable can be found at http://www.patrickmin.com/binvox/ ,
 			# -d x idicates resoltuion will be x by x by x , -pb is to stop the visualization, the rest of the commnads are to help make the object water tight
 			commands.append(command)
 			if count %20 == 0  and count != 0:
@@ -341,7 +340,7 @@ def calc_surface():
 		commands =[]
 		count = 0
 		for d in tqdm(dirs):
-			command = './binvox ' + d  + ' -d ' + str(dims)+ ' -pb -cb -c -e'   # this executable can be found at http://www.patrickmin.com/binvox/ ,
+			command = 'scripts/binvox ' + d  + ' -d ' + str(dims)+ ' -pb -cb -c -e'   # this executable can be found at http://www.patrickmin.com/binvox/ ,
 			# -d x idicates resoltuion will be x by x by x , -pb is to stop the visualization, the rest of the commnads are to help make the object water tight
 			commands.append(command)
 			if count %20 == 0  and count != 0:

utils.py

@@ -1,4 +1,3 @@
-import time
 import argparse
 import numpy as np
 import torch
@@ -7,8 +6,6 @@
 from torch.autograd import Variable
 import sys
 import os
-import gc
-import scipy.sparse as sp
 import torch
 from glob import glob
 import scipy.io as sio
@@ -24,7 +21,7 @@
 from tqdm import tqdm 
 
 #import chamfer distance 
-sys.path.insert(0,'chamfer_distance/')
+sys.path.insert(0, "chamfer_distance/")
 from modules.nnd import NNDModule
 dist =  NNDModule()
 np.random.seed(2)
@@ -138,7 +135,6 @@ def adj_init(faces,num_verts):
 
 # normalizes symetric, binary adj matrix such that sum of each row is 1 
 def normalize_adj(mx):
-	"""Row-normalize sparse matrix"""
 	rowsum = np.array(mx.sum(1))
 	r_inv = np.power(rowsum, -1).flatten()
 	r_inv[np.isinf(r_inv)] = 0.
@@ -291,13 +287,12 @@ def load_batch(self, size, images = False,  ordered = False):
 			verts = mesh_info['verts']
 			verts= torch.FloatTensor(np.array(verts)).cuda()
 			faces = mesh_info['faces']
-			dist = mesh_info['orig_adj']
-			dist[np.where(dist > 0)] = 1. 
-			dist = calc_dist( verts, dist)
-			dist = normalize_adj(dist)
+			adj = mesh_info['orig_adj']
+			adj[np.where(adj > 0)] = 1. 
+			adj = normalize_adj(adj)
 
-			dist = torch.FloatTensor(dist).cuda()
-			adj_info={'adj': dist}
+			adj = torch.FloatTensor(adj).cuda()
+			adj_info={'adj': adj}
 			batch_verts.append(verts)
 			batch_faces.append(torch.LongTensor(np.array(faces)).cuda())
 			batch_adjs.append(adj_info)
@@ -325,22 +320,6 @@ def load_batch(self, size, images = False,  ordered = False):
 		return batch
 
 
-def calc_dist(verts, adj): 
-	adj_maintained = torch.FloatTensor(adj).cuda()
-	adj = adj_maintained + torch.FloatTensor(np.eye(adj_maintained.shape[0])).cuda()
-
-	positions = torch.FloatTensor(np.array(verts)).cuda()
-	reshape_positions = positions.view(positions.shape[0], positions.shape[1], 1)
-	reshape_positions = reshape_positions.expand(positions.shape[0], positions.shape[1], positions.shape[0]).permute(0,2,1)
-	cross_positions = reshape_positions.permute(1,0,2)
-	dist = torch.sum((reshape_positions- cross_positions)**2, dim = 2)
-
-	dist = 1/dist
-	for k in range(verts.shape[0]): 
-		dist[k,k] = torch.min(dist[k,dist[k].nonzero()])
-	dist = dist*adj
-	dist = (dist).data.cpu().numpy()
-	return dist 
 
 
 # graphs the training and validation information