OOM solved

.gitignore

@@ -13,3 +13,4 @@ kaolin/
 neural_voxelization_layer/
 pytorch3d/
 force_push.sh
+results/

app.py

@@ -117,7 +117,7 @@ with gr.Blocks() as demo:
 
             gr.Examples(examples=examples,
                         inputs=[inp, radio_choice],
-                        cache_examples=False,
+                        cache_examples=True,
                         fn=generate_model,
                         outputs=out_lst)
 

apps/infer.py

@@ -14,10 +14,9 @@
 #
 # Contact: ps-license@tuebingen.mpg.de
 
-import os
+import os, gc
 
 import logging
-from lib.common.render import query_color
 from lib.common.config import cfg
 from lib.dataset.mesh_util import (
     load_checkpoint,
@@ -403,7 +402,7 @@ def generate_model(in_path, model_type):
                 loop_cloth.set_description(pbar_desc)
 
                 # update params
-                cloth_loss.backward(retain_graph=True)
+                cloth_loss.backward()
                 optimizer_cloth.step()
                 scheduler_cloth.step(cloth_loss)
 
@@ -414,13 +413,6 @@ def generate_model(in_path, model_type):
                 process=False, maintains_order=True
             )
             
-            # with front texture
-            # final_colors = query_color(
-            #     mesh_pr.verts_packed().detach().squeeze(0).cpu(),
-            #     mesh_pr.faces_packed().detach().squeeze(0).cpu(),
-            #     in_tensor["image"],
-            #     device=device,
-            # )
             
             # without front texture
             final_colors = (mesh_pr.verts_normals_padded().squeeze(0).detach().cpu() + 1.0) * 0.5 * 255.0
@@ -458,7 +450,7 @@ def generate_model(in_path, model_type):
     for element in dir():
         if 'path' not in element:
             del locals()[element]
-            
+    gc.collect()
     torch.cuda.empty_cache()
     
     return [smpl_path, smpl_path, smpl_npy_path, recon_path, recon_path, refine_path, refine_path, video_path, overlap_path]

gradio_cached_examples/13/log.csv

@@ -0,0 +1,9 @@
+'flag','username','timestamp'
+'','','2022-08-01 22:47:24.908073'
+'','','2022-08-01 22:48:20.753663'
+'','','2022-08-01 22:49:15.504871'
+'','','2022-08-01 22:50:16.762017'
+'','','2022-08-01 22:51:09.444531'
+'','','2022-08-01 22:52:09.357219'
+'','','2022-08-01 22:53:02.217347'
+'','','2022-08-01 22:54:11.545178'

lib/net/FBNet.py

@@ -81,7 +81,8 @@ def define_G(input_nc,
     # print(netG)
     if len(gpu_ids) > 0:
         assert (torch.cuda.is_available())
-        netG.cuda(gpu_ids[0])
+        device=torch.device(f"cuda:{gpu_ids[0]}")
+        netG = netG.to(device)
     netG.apply(weights_init)
     return netG