Fix CUDA inference

nets/attention/position_encoding.py

@@ -18,7 +18,6 @@ class PositionEncodingSine(nn.Module):
                 We will remove the buggy impl after re-training all variants of our released models.
         """
         super().__init__()
-
         pe = torch.zeros((d_model, *max_shape))
         y_position = torch.ones(max_shape).cumsum(0).float().unsqueeze(0)
         x_position = torch.ones(max_shape).cumsum(1).float().unsqueeze(0)
@@ -39,4 +38,4 @@ class PositionEncodingSine(nn.Module):
         Args:
             x: [N, C, H, W]
         """
-        return x + self.pe[:, :, :x.size(2), :x.size(3)]
+        return x + self.pe[:, :, :x.size(2), :x.size(3)].to(x.device)

test_model.py

@@ -6,6 +6,8 @@ from imread_from_url import imread_from_url
 
 from nets import Model
 
+device = 'cuda'
+
 #Ref: https://github.com/megvii-research/CREStereo/blob/master/test.py
 def inference(left, right, model, n_iter=20):
 
@@ -15,8 +17,8 @@ def inference(left, right, model, n_iter=20):
 	imgL = np.ascontiguousarray(imgL[None, :, :, :])
 	imgR = np.ascontiguousarray(imgR[None, :, :, :])
 
-	imgL = torch.tensor(imgL.astype("float32"))
+	imgL = torch.tensor(imgL.astype("float32")).to(device)
-	imgR = torch.tensor(imgR.astype("float32"))
+	imgR = torch.tensor(imgR.astype("float32")).to(device)
 
 	imgL_dw2 = F.interpolate(
 		imgL,
@@ -35,26 +37,32 @@ def inference(left, right, model, n_iter=20):
 	pred_flow_dw2 = model(imgL_dw2, imgR_dw2, iters=n_iter, flow_init=None)
 
 	pred_flow = model(imgL, imgR, iters=n_iter, flow_init=pred_flow_dw2)
-	pred_disp = torch.squeeze(pred_flow[:, 0, :, :]).detach().numpy()
+	pred_disp = torch.squeeze(pred_flow[:, 0, :, :]).cpu().detach().numpy()
 
 	return pred_disp
 
 if __name__ == '__main__':
 
-	left_img = imread_from_url("https://vision.middlebury.edu/stereo/data/scenes2003/newdata/cones/im2.png")
+	left_img = imread_from_url("https://raw.githubusercontent.com/megvii-research/CREStereo/master/img/test/left.png")
-	right_img = imread_from_url("https://vision.middlebury.edu/stereo/data/scenes2003/newdata/cones/im6.png")
+	right_img = imread_from_url("https://raw.githubusercontent.com/megvii-research/CREStereo/master/img/test/right.png")
+
+	# Resize image in case the GPU memory overflows
+	eval_h, eval_w = (240,426)
+	imgL = cv2.resize(left, (eval_w, eval_h), interpolation=cv2.INTER_LINEAR)
+	imgR = cv2.resize(right, (eval_w, eval_h), interpolation=cv2.INTER_LINEAR)
 
 	model_path = "models/crestereo_eth3d.pth"
 
 	model = Model(max_disp=256, mixed_precision=False, test_mode=True)
 	model.load_state_dict(torch.load(model_path), strict=True)
+	model.to(device)
 	model.eval()
 
-	disp = inference(left_img, right_img, model, n_iter=20)
+	disp = inference(imgL, imgR, model, n_iter=20)
 	disp_vis = (disp - disp.min()) / (disp.max() - disp.min()) * 255.0
 	disp_vis = disp_vis.astype("uint8")
 	disp_vis = cv2.applyColorMap(disp_vis, cv2.COLORMAP_INFERNO)
-	disp_vis = cv2.resize(disp_vis, left_img.shape[1::-1])
+	left_img = cv2.resize(left_img, disp_vis.shape[1::-1])
 
 	combined_img = np.hstack((left_img, disp_vis))
 	cv2.imshow("output", combined_img)