feat: export for unet_mha

jolibrain · Dec 2, 2022 · b4c3cfd · b4c3cfd
1 parent 67bede6
commit b4c3cfd
Show file tree

Hide file tree

Showing 3 changed files with 57 additions and 33 deletions.
diff --git a/models/base_model.py b/models/base_model.py
@@ -227,7 +227,7 @@ def __init__(self, opt, rank):
 
         if "segformer" in self.opt.G_netG:
             self.onnx_opset_version = 11
-        elif "ittr" in self.opt.G_netG:
+        elif "ittr" in self.opt.G_netG or "unet_mha" in self.opt.G_netG:
             self.onnx_opset_version = 12
         else:
             self.onnx_opset_version = 9
@@ -795,39 +795,63 @@ def export_networks(self, epoch):
 
                 net = getattr(self, "net" + name)
 
-                input_nc = self.opt.model_input_nc
-                if self.opt.model_multimodal:
-                    input_nc += self.opt.train_mm_nz
-
-                dummy_input = torch.randn(
-                    1,
-                    input_nc,
-                    self.opt.data_crop_size,
-                    self.opt.data_crop_size,
-                    device=self.device,
-                )
-
                 # onnx
+
                 if (
                     not "ittr" in self.opt.G_netG
                     and not "unet_mha" in self.opt.G_netG
                     and not "palette" in self.opt.model_type
                 ):
-                    export_path_onnx = save_path.replace(".pth", ".onnx")
-
-                    torch.onnx.export(
-                        net,
-                        dummy_input,
-                        export_path_onnx,
-                        verbose=False,
-                        opset_version=self.onnx_opset_version,
-                    )
+                    input_nc = self.opt.model_input_nc
+                    if self.opt.model_multimodal:
+                        input_nc += self.opt.train_mm_nz
+
+                        dummy_input = torch.randn(
+                            1,
+                            input_nc,
+                            self.opt.data_crop_size,
+                            self.opt.data_crop_size,
+                            device=self.device,
+                        )
+
+                    # onnx
+                    if not "ittr" in self.opt.G_netG:
+                        export_path_onnx = save_path.replace(".pth", ".onnx")
+
+                        export_device = torch.device("cpu")
+                        net = net.to(export_device)
+
+                        torch.onnx.export(
+                            net,
+                            self.get_dummy_input(export_device),
+                            export_path_onnx,
+                            verbose=False,
+                            opset_version=self.onnx_opset_version,
+                        )
+                        net.to(self.device)
+
+                    # jit
+                    if self.opt.train_export_jit and not "segformer" in self.opt.G_netG:
+                        export_path_jit = save_path.replace(".pth", ".pt")
+                        jit_model = torch.jit.trace(net, self.get_dummy_input())
+                        jit_model.save(export_path_jit)
+
+    def get_dummy_input(self, device=None):
+        input_nc = self.opt.model_input_nc
+        if self.opt.model_multimodal:
+            input_nc += self.opt.train_mm_nz
+
+        if device is None:
+            device = self.device
+        dummy_input = torch.randn(
+            1,
+            input_nc,
+            self.opt.data_crop_size,
+            self.opt.data_crop_size,
+            device=device,
+        )
 
-                # jit
-                if self.opt.train_export_jit and not "segformer" in self.opt.G_netG:
-                    export_path_jit = save_path.replace(".pth", ".pt")
-                    jit_model = torch.jit.trace(net, dummy_input)
-                    jit_model.save(export_path_jit)
+        return dummy_input
 
     def __patch_instance_norm_state_dict(self, state_dict, module, keys, i=0):
         """Fix InstanceNorm checkpoints incompatibility (prior to 0.4)"""

diff --git a/models/modules/unet_generator_attn/unet_generator_attn.py b/models/modules/unet_generator_attn/unet_generator_attn.py
@@ -242,7 +242,7 @@ def __init__(
         self.proj_out = zero_module(nn.Conv1d(channels, channels, 1))
 
     def forward(self, x):
-        return checkpoint(self._forward, (x,), self.parameters(), True)
+        return checkpoint(self._forward, (x,), self.parameters(), self.use_checkpoint)
 
     def _forward(self, x):
         b, c, *spatial = x.shape

diff --git a/scripts/export_onnx_model.py b/scripts/export_onnx_model.py
@@ -31,7 +31,7 @@
     help="optional model configuration, e.g /path/to/segformer_config_b0.py",
 )
 parser.add_argument("--img-size", default=256, type=int, help="square image size")
-parser.add_argument("--cpu", action="store_true", help="whether to export for CPU")
+parser.add_argument("--cuda", action="store_true", help="whether to export using gpu")
 parser.add_argument("--bw", action="store_true", help="whether input/output is bw")
 parser.add_argument(
     "--padding-type",
@@ -80,14 +80,14 @@
 model.eval()
 model.load_state_dict(torch.load(args.model_in_file))
 
-if not args.cpu:
+if args.cuda:
     model = model.cuda()
 
 # export to ONNX via tracing
-if args.cpu:
-    device = "cpu"
-else:
+if args.cuda:
     device = "cuda"
+else:
+    device = "cpu"
 dummy_input = torch.randn(1, input_nc, args.img_size, args.img_size, device=device)