change a bunch of stuff, add wip lightning implementation

api_server.py

@@ -1,3 +1,4 @@
+import os
 import json
 from datetime import datetime
 from typing import Union, Literal
@@ -7,32 +8,149 @@ import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+import torch_tensorrt
 from cv2 import cv2
 from fastapi import FastAPI, File, UploadFile
 from PIL import Image
 
+
 from nets import Model
 
+from train import inference as ctd_inference
+
+
 app = FastAPI()
 
-reference_pattern_path = '/home/nils/kinect_reference_cropped.png'
+# reference_pattern_path = '/home/nils/kinect_reference_cropped.png'
+reference_pattern_path = '/home/nils/kinect_reference_far.png'
+# reference_pattern_path = '/home/nils/kinect_diff_ref.png'
+print(reference_pattern_path)
 reference_pattern = cv2.imread(reference_pattern_path)
+
+# shift reference pattern a few pixels to the left to simulate further backdrop
+trans_mat = np.float32([[1, 0, 0], [0, 1, 0]])
+reference_pattern = cv2.warpAffine(
+    reference_pattern, trans_mat, reference_pattern.shape[1::-1], flags=cv2.INTER_LINEAR
+)
+
 iters = 20
-minimal_data = False
+minimal_data = True
+temporal_init = False
+last_img = None
 device = torch.device('cuda:0')
 
 
-def load_model(epoch):
+def downsize(img):
+    diff = (512 - 480) // 2
+    downsampled = cv2.pyrDown(img)
+    img = downsampled[diff:downsampled.shape[0] - diff, 0:downsampled.shape[1]]
+    return img
+
+
+if 1024 in reference_pattern.shape:
+    reference_pattern = downsize(reference_pattern)
+
+
+def ghetto_lcn(img):
+    # gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+    gray = img
+
+    float_gray = gray.astype(np.float32) / 255.0
+
+    blur = cv2.GaussianBlur(float_gray, (0, 0), sigmaX=2, sigmaY=2)
+    num = float_gray - blur
+
+    blur = cv2.GaussianBlur(num * num, (0, 0), sigmaX=20, sigmaY=20)
+    den = cv2.pow(blur, 0.5)
+
+    gray = num / den
+
+    # cv2.normalize(gray, dst=gray, alpha=0.0, beta=1.0, norm_type=cv2.NORM_MINMAX)
+    cv2.normalize(gray, dst=gray, alpha=0.0, beta=255.0, norm_type=cv2.NORM_MINMAX)
+    return gray
+
+
+# reference_pattern = ghetto_lcn(reference_pattern)
+
+
+def load_model(epoch, use_tensorrt=False):
     global model
     epoch = f'epoch-{epoch}.pth' if epoch != 'latest' else 'latest.pth'
     model_path = f"train_log/models/{epoch}"
     model = Model(max_disp=256, mixed_precision=False, test_mode=True)
+    # FIXME WIP Workaround Dataparallel TensorRT incompatibility
     model = nn.DataParallel(model, device_ids=[device])
     # model.load_state_dict(torch.load(model_path), strict=False)
     state_dict = torch.load(model_path)['state_dict']
     model.load_state_dict(state_dict, strict=True)
     model.to(device)
     model.eval()
+    if use_tensorrt:
+        np_model = Model(max_disp=256, mixed_precision=False, test_mode=True)
+        np_model.load_state_dict(model.module.state_dict(), strict=True)
+        np_model.to(device)
+        np_model.eval()
+
+        spec_dict = {
+                    "inputs": [
+                        torch_tensorrt.Input(
+                            min_shape=[1, 2, 240, 320], 
+                            max_shape=[1, 2, 480, 640], 
+                            opt_shape=[1, 2, 480, 640], 
+                            dtype=torch.int32,
+                            ), 
+                        torch_tensorrt.Input(
+                            min_shape=[1, 2, 240, 320], 
+                            max_shape=[1, 2, 480, 640], 
+                            opt_shape=[1, 2, 480, 640], 
+                            dtype=torch.int32,
+                            ), 
+                        ],
+                    "enabled_precisions": {torch.float, torch.half},
+                    "refit": False,
+                    "debug": False,
+                    "device": {
+                        "device_type": torch_tensorrt.DeviceType.GPU,
+                        "gpu_id": 0,
+                        "dla_core": 0,
+                        "allow_gpu_fallback": True
+                    },
+                    "capability": torch_tensorrt.EngineCapability.default,
+                    "num_min_timing_iters": 2,
+                    "num_avg_timing_iters": 1,
+                }
+        spec = {
+            "forward":
+                torch_tensorrt.ts.TensorRTCompileSpec(**spec_dict)
+            }
+
+        # trt_model = torch_tensorrt.compile(np_model ,
+                        # inputs=torch_tensorrt.Input(
+                            # min_shape=[1, 2, 240, 320], 
+                            # max_shape=[1, 2, 480, 640], 
+                            # opt_shape=[1, 2, 480, 640], 
+                            # dtype=torch.int32,
+            # inputs = [torch_tensorrt.Input((1, 2, 480, 640)), torch_tensorrt.Input((1, 2, 480, 640))], # input shape
+            # enabled_precisions = {torch_tensorrt.dtype.half} # Run with FP16
+        # )
+        # trt_dw_model = torch_tensorrt.compile(np_model ,
+            # inputs = [torch_tensorrt.Input((1, 2, 240, 320)), torch_tensorrt.Input((1, 2, 240, 320))], # input shape
+            # enabled_precisions = {torch_tensorrt.dtype.half} # Run with FP16
+        # )
+
+        script_model = torch.jit.script(np_model.eval())
+        # script_dw_model = torch.jit.script(trt_dw_model.eval())
+
+        # save the TensorRT embedded Torchscript
+        # torch.jit.save(trt_model, 'trt_torchscript_module.ts')
+        # torch.jit.save(trt_dw_model, 'trt_torchscript_dw_module.ts')
+        print(script_model)
+        print(script_model.forward)
+        print(script_model.forward())
+        print(dir(script_model))
+
+        model = torch._C._jit_to_backend("tensorrt", script_model, spec)
+
     print(f'loaded model {epoch}')
     return model
 
@@ -74,14 +192,37 @@ def inference(left, right, model, n_iter=20):
     )
 
     with torch.inference_mode():
-        pred_flow_dw2 = model(image1=imgL_dw2, image2=imgR_dw2, iters=n_iter, flow_init=None)
-        pred_flow = model(imgL, imgR, iters=n_iter, flow_init=pred_flow_dw2)
+        # pred_flow_dw2 = model(image1=imgL_dw2, image2=imgR_dw2, iters=n_iter, flow_init=None)
+        # pred_flow = model(imgL, imgR, iters=n_iter, flow_init=pred_flow_dw2)
+        pred_flow_dw2 = model.forward(image1=imgL_dw2, image2=imgR_dw2, iters=n_iter, flow_init=None)
+        pred_flow = model.forward(imgL, imgR, iters=n_iter, flow_init=pred_flow_dw2)
 
     pred_disp = torch.squeeze(pred_flow[:, 0, :, :]).cpu().detach().numpy()
 
     return pred_disp
 
 
+def get_reference():
+    refs = [ref.path for ref in os.scandir('/home/nils/references/')]
+    for ref in refs:
+        reference =  cv2.imread(ref)
+        yield reference
+
+references = get_reference()
+
+
+@app.post('/params/update_reference')
+async def update_reference():
+    global references, reference_pattern
+    try:
+        reference_pattern = downsize(next(references))
+        print(reference_pattern.shape)
+        return {'status': 'success'}
+    except StopIteration:
+        references = get_reference()
+        return {'status': 'finished'}
+
+
 @app.post("/model/update/{epoch}")
 async def change_model(epoch: Union[int, Literal['latest']]):
     global model
@@ -103,8 +244,15 @@ async def set_minimal_data(enable: bool):
     minimal_data = enable
 
 
+@app.post("/params/temporal_init/{enable}")
+async def set_temporal_init(enable: bool):
+    global temporal_init
+    temporal_init = enable
+
+
 @app.put("/ir")
 async def read_ir_input(file: UploadFile = File(...)):
+    global last_img, minimal_data
     try:
         img = np.array(Image.open(BytesIO(await file.read())))
     except Exception as e:
@@ -114,24 +262,35 @@ async def read_ir_input(file: UploadFile = File(...)):
     if len(img.shape) == 2:
         img = cv2.merge([img for _ in range(3)])
     if img.shape == (1024, 1280, 3):
-        diff = (512 - 480) // 2
-        downsampled = cv2.pyrDown(img)
-        img = downsampled[diff:downsampled.shape[0] - diff, 0:downsampled.shape[1]]
+        img = downsize(img)
 
-    img = img.transpose((1, 2, 0))
-    ref_pat = reference_pattern.transpose((1, 2, 0))
+    # img = img.transpose((1, 2, 0))
+    # ref_pat = reference_pattern.transpose((1, 2, 0))
+    ref_pat = reference_pattern
 
     start = datetime.now()
-    pred_disp = inference(img, ref_pat, model, iters)
+    if temporal_init:
+        pred_disp = ctd_inference(img, ref_pat, None, None, model, None, iters, False, last_img)
+        last_img = pred_disp
+    else:
+        pred_disp = ctd_inference(img, ref_pat, None, None, model, None, iters, False)
+    # pred_disp = inference(img, ref_pat, model, iters)
     duration = (datetime.now() - start).total_seconds()
 
     if minimal_data:
         return json.dumps({'disp': pred_disp, 'duration': duration}, cls=NumpyEncoder)
     else:
+        # return json.dumps({'disp': pred_disp, 'input': img, 'duration': duration},
         return json.dumps({'disp': pred_disp, 'reference': ref_pat, 'input': img, 'duration': duration},
                           cls=NumpyEncoder)
 
 
+@app.get('/temporal_init')
+def get_temporal_init():
+    return {'status': 'enabled' if temporal_init else 'disabled'}
+
+
+
 @app.get('/')
 def main():
     return {'test': 'abc'}

cfgs/train.yaml

@@ -4,18 +4,22 @@ base_lr: 4.0e-4
 
 nr_gpus: 3
 batch_size: 4
-n_total_epoch: 600
+n_total_epoch: 300
 minibatch_per_epoch: 500
 
 loadmodel: ~
 log_dir: "./train_log"
+log_dir_lightning: "./train_log_lightning"
 model_save_freq_epoch: 1
 
 max_disp: 256
 image_width: 640
 image_height: 480
 # training_data_path: "./stereo_trainset/crestereo"
-training_data_path: "/media/Data1/connecting_the_dots_data/ctd_data/"
+pattern_attention: true
+dataset: "blender"
+# training_data_path: "/media/Data1/connecting_the_dots_data/ctd_data/"
+training_data_path: "/media/Data1/connecting_the_dots_data/blender_renders/data"
 
 log_level: "logging.INFO"
 

dataset.py

@@ -234,12 +234,16 @@ class CREStereoDataset(Dataset):
 
 
 class CTDDataset(Dataset):
-    def __init__(self, root, pattern_path: str, data_type: str = 'syn', augment=True, resize_pattern=True, blur=False):
+    def __init__(self, root, pattern_path: str, data_type: str = 'syn', augment=True, resize_pattern=True, blur=False, split=0.9, test_set=False):
         super().__init__()
         self.rng = np.random.RandomState(0)
         self.augment = augment
         self.blur = blur
-        self.imgs = glob.glob(os.path.join(root, f"{data_type}/*/im0_*.npy"), recursive=True)
+        imgs = glob.glob(os.path.join(root, f"{data_type}/*/im0_*.npy"), recursive=True)
+        if test_set:
+            self.imgs = imgs[:int(split * len(imgs))]
+        else:
+            self.imgs = imgs[int(split * len(imgs)):]
         self.pattern = cv2.imread(pattern_path)#, cv2.IMREAD_GRAYSCALE)
 
         if resize_pattern and self.pattern.shape != (480, 640, 3):
@@ -271,6 +275,10 @@ class CTDDataset(Dataset):
 
         # read img, disp
         left_img = np.load(left_path)
+
+        if left_img.dtype == 'float32':
+            left_img = (left_img * 255).astype('uint8')
+
         left_img = cv2.merge([left_img, left_img, left_img]).reshape((480, 640, 3))
 
         right_img = self.pattern
@@ -307,3 +315,100 @@ class CTDDataset(Dataset):
 
     def __len__(self):
         return len(self.imgs)
+
+
+class BlenderDataset(CTDDataset):
+    def __init__(self, root, pattern_path: str, data_type: str = 'syn', augment=True, resize_pattern=True, blur=False, split=0.9, test_set=False, use_lightning=False):
+        super().__init__(root, pattern_path)
+        self.use_lightning = use_lightning
+        imgs = [f for f in glob.glob(f"{root}/im_*.png", recursive=True) if not 'depth0001' in f]
+        if test_set:
+            self.imgs = imgs[:int(split * len(imgs))]
+        else:
+            self.imgs = imgs[int(split * len(imgs)):]
+        self.pattern = cv2.imread(pattern_path)#, cv2.IMREAD_GRAYSCALE)
+
+        if resize_pattern and self.pattern.shape != (480, 640, 3):
+            downsampled = cv2.pyrDown(self.pattern)
+            diff = (downsampled.shape[0] - 480) // 2
+            self.pattern = downsampled[diff:downsampled.shape[0]-diff, 0:downsampled.shape[1]]
+
+        self.augmentor = Augmentor(
+            image_height=480,
+            image_width=640,
+            max_disp=256,
+            scale_min=0.6,
+            scale_max=1.0,
+            seed=0,
+        )
+
+    def __getitem__(self, index):
+        # find path
+        left_path = self.imgs[index]
+        left_disp_path = left_path.split('.')[0] + '_depth0001.png'
+
+        # read img, disp
+        left_img = cv2.imread(left_path)
+
+        if left_img.dtype == 'float32':
+            left_img = (left_img * 255).astype('uint8')
+
+        if left_img.shape != (480, 640, 3):
+            downsampled = cv2.pyrDown(left_img)
+            diff = (downsampled.shape[0] - 480) // 2
+            left_img = downsampled[diff:downsampled.shape[0]-diff, 0:downsampled.shape[1]]
+            if left_img.shape[-1] != 3:
+                left_img = cv2.merge([left_img, left_img, left_img]).reshape((480, 640, 3))
+
+            right_img = self.pattern
+        left_disp = self.get_disp(left_disp_path)
+
+        if False: # self.rng.binomial(1, 0.5):
+            left_img, right_img = np.fliplr(right_img), np.fliplr(left_img)
+            left_disp, right_disp = np.fliplr(right_disp), np.fliplr(left_disp)
+        left_disp[left_disp == np.inf] = 0
+
+        if self.blur:
+            kernel_size = random.sample([1,3,5,7,9], 1)[0]
+            kernel = (kernel_size, kernel_size)
+            left_img = cv2.GaussianBlur(left_img, kernel, 0)
+
+        # augmentation
+        if not self.augment:
+            _left_img, _right_img, _left_disp, disp_mask = self.augmentor(
+                left_img, right_img, left_disp
+            )
+        else:
+            left_img, right_img, left_disp, disp_mask = self.augmentor(
+                left_img, right_img, left_disp
+            )
+
+        if not self.use_lightning:
+            right_img = right_img.transpose((2, 0, 1)).astype("uint8")
+            return {
+                "left": left_img,
+                "right": right_img,
+                "disparity": left_disp,
+                "mask": disp_mask,
+            }
+
+        right_img = right_img.transpose((2, 0, 1)).astype("uint8")
+        left_img = left_img.transpose((2, 0, 1)).astype("uint8")
+        return left_img, right_img, left_disp, disp_mask
+
+    def get_disp(self, path):
+        baseline = 0.075    # meters
+        fl = 560.           # as per CTD
+        depth = cv2.imread(path, cv2.IMREAD_UNCHANGED)
+        downsampled = cv2.pyrDown(depth)
+        diff = (downsampled.shape[0] - 480) // 2
+        depth = downsampled[diff:downsampled.shape[0]-diff, 0:downsampled.shape[1]]
+        # disp = np.load(path).transpose(1,2,0)
+        # disp = baseline * fl / depth
+        # return disp.astype(np.float32) / 32
+        # FIXME temporarily increase disparity until new data with better depth values is generated
+        #       higher values seem to speedup convergence, but introduce much stronger artifacting
+        # mystery_factor = 150
+        mystery_factor = 1
+        disp = (baseline * fl * mystery_factor) / depth
+        return disp.astype(np.float32)

nets/attention/transformer.py

@@ -86,8 +86,15 @@ class LocalFeatureTransformer(nn.Module):
         """
         assert self.d_model == feat0.size(2), "the feature number of src and transformer must be equal"
 
-        for layer, name in zip(self.layers, self.layer_names):
-
+        # NOTE Workaround for non statically determinable zip
+        # for layer, name in zip(self.layers, self.layer_names):
+        # layer_zip = ((layer, self.layer_names[i]) for i, layer in enumerate(self.layers))
+        # layer_zip = []
+        # for i, layer in enumerate(self.layers):
+            # layer_zip.append((layer, self.layer_names[i]))
+            
+        for i, layer in enumerate(self.layers):
+            name = self.layer_names[i]
             if name == 'self':
                 feat0 = layer(feat0, feat0, mask0, mask0)
                 feat1 = layer(feat1, feat1, mask1, mask1)

nets/crestereo.py

@@ -36,6 +36,12 @@ class CREStereo(nn.Module):
         self.fnet = BasicEncoder(output_dim=256, norm_fn='instance', dropout=self.dropout)  
         self.update_block = BasicUpdateBlock(hidden_dim=self.hidden_dim, cor_planes=4 * 9, mask_size=4)
 
+        # # NOTE Position_encoding as workaround for TensorRt
+        # image1_shape = [1, 2, 480, 640]
+        # self.pos_encoding_fn_small = PositionEncodingSine(
+            # d_model=256, max_shape=(image1_shape[2] // 16, image1_shape[3] // 16)
+        # )
+
         # loftr
         self.self_att_fn = LocalFeatureTransformer(
             d_model=256, nhead=8, layer_names=["self"] * 1, attention="linear"
@@ -81,7 +87,7 @@ class CREStereo(nn.Module):
         zero_flow = torch.cat((_x, _y), dim=1).to(fmap.device)
         return zero_flow
 
-    def forward(self, image1, image2, flow_init=None, iters=10, upsample=True, test_mode=False):
+    def forward(self, image1, image2, flow_init=None, iters=10, upsample=True, test_mode=False, self_attend_right=True):
         """ Estimate optical flow between pair of frames """
 
         image1 = 2 * (image1 / 255.0) - 1.0
@@ -130,17 +136,22 @@ class CREStereo(nn.Module):
             inp_dw16 = F.avg_pool2d(inp, 4, stride=4)
 
             # positional encoding and self-attention
-            pos_encoding_fn_small = PositionEncodingSine(
-                d_model=256, max_shape=(image1.shape[2] // 16, image1.shape[3] // 16)
-            )
+            # pos_encoding_fn_small = PositionEncodingSine(
+                # d_model=256, max_shape=(image1.shape[2] // 16, image1.shape[3] // 16)
+            # )
             # 'n c h w -> n (h w) c'
-            x_tmp = pos_encoding_fn_small(fmap1_dw16)
+            x_tmp = self.pos_encoding_fn_small(fmap1_dw16)
             fmap1_dw16 = x_tmp.permute(0, 2, 3, 1).reshape(x_tmp.shape[0], x_tmp.shape[2] * x_tmp.shape[3], x_tmp.shape[1])
             # 'n c h w -> n (h w) c'
-            x_tmp = pos_encoding_fn_small(fmap2_dw16)
+            x_tmp = self.pos_encoding_fn_small(fmap2_dw16)
             fmap2_dw16 = x_tmp.permute(0, 2, 3, 1).reshape(x_tmp.shape[0], x_tmp.shape[2] * x_tmp.shape[3], x_tmp.shape[1])
 
-            fmap1_dw16, fmap2_dw16 = self.self_att_fn(fmap1_dw16, fmap2_dw16)
+            # FIXME experimental ! no self-attention for pattern
+            if not self_attend_right:
+                fmap1_dw16, _ = self.self_att_fn(fmap1_dw16, fmap2_dw16)
+            else:
+                fmap1_dw16, fmap2_dw16 = self.self_att_fn(fmap1_dw16, fmap2_dw16)
+
             fmap1_dw16, fmap2_dw16 = [
                 x.reshape(x.shape[0], image1.shape[2] // 16, -1, x.shape[2]).permute(0, 3, 1, 2)
                 for x in [fmap1_dw16, fmap2_dw16]
@@ -258,3 +269,4 @@ class CREStereo(nn.Module):
             return flow_up
 
         return predictions
+

nets/extractor.py

@@ -1,6 +1,8 @@
+from typing import List
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from torch import Tensor
 
 # Ref: https://github.com/princeton-vl/RAFT/blob/master/core/extractor.py
 class ResidualBlock(nn.Module):
@@ -96,28 +98,43 @@ class BasicEncoder(nn.Module):
         self.in_planes = dim
         return nn.Sequential(*layers)
 
-    def forward(self, x):
+    def forward(self, x: List[Tensor]):
+        # NOTE always assume list, otherwise TensorRT is sad
+        # batch_dim = x[0].shape[0]
+        # x_tensor = torch.cat(list(x), dim=0)
 
         # if input is list, combine batch dimension
         is_list = isinstance(x, tuple) or isinstance(x, list)
         if is_list:
             batch_dim = x[0].shape[0]
-            x = torch.cat(x, dim=0)
+            x_tensor = torch.cat(x, dim=0)
+        else:
+            x_tensor = x
+
+        print()
+        print()
+        print(x_tensor.shape)
+        print()
+        print()
 
-        x = self.conv1(x)
-        x = self.norm1(x)
-        x = self.relu1(x)
+        x_tensor = self.conv1(x_tensor)
+        x_tensor = self.norm1(x_tensor)
+        x_tensor = self.relu1(x_tensor)
 
-        x = self.layer1(x)
-        x = self.layer2(x)
-        x = self.layer3(x)
+        x_tensor = self.layer1(x_tensor)
+        x_tensor = self.layer2(x_tensor)
+        x_tensor = self.layer3(x_tensor)
 
-        x = self.conv2(x)
+        x_tensor = self.conv2(x_tensor)
 
         if self.dropout is not None:
-            x = self.dropout(x)
+            x_tensor = self.dropout(x_tensor)
 
         if is_list:
-            x = torch.split(x, x.shape[0]//2, dim=0)
+            x_list = torch.split(x_tensor, x_tensor.shape[0]//2, dim=0)
+            return x_list
+
+        x_list = torch.split(x_tensor, x_tensor.shape[0]//2, dim=0)
+        return x_list
 
-        return x
+        # return list(x)

nets/update.py

@@ -77,7 +77,7 @@ class BasicUpdateBlock(nn.Module):
             nn.ReLU(inplace=True),
             nn.Conv2d(256, mask_size**2 *9, 1, padding=0))
 
-    def forward(self, net, inp, corr, flow, upsample=True):
+    def forward(self, net, inp, corr, flow, upsample: bool=True):
         # print(inp.shape, corr.shape, flow.shape)
         motion_features = self.encoder(flow, corr)
         # print(motion_features.shape, inp.shape)

test_model.py

@@ -3,6 +3,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 import numpy as np
 import cv2
+import os
 
 from nets import Model
 
@@ -16,17 +17,20 @@ device = 'cuda'
 wandb.init(project="crestereo", entity="cpt-captain")
 
 
-def do_infer(left_img, right_img, gt_disp, model):
-    disp = ctd_inference(left_img, right_img, gt_disp, None, model, None, n_iter=20, wandb_log=False)
+def do_infer(left_img, right_img, gt_disp, model, attend_pattern=True):
+    disp = ctd_inference(left_img, right_img, gt_disp, None, model, None, n_iter=20, wandb_log=False, attend_pattern=attend_pattern)
 
     disp_vis = normalize_and_colormap(disp)
-    gt_disp_vis = normalize_and_colormap(gt_disp)
-    if gt_disp.shape != disp.shape:
-        gt_disp = gt_disp.reshape(disp.shape)
-    disp_err = gt_disp - disp
-    disp_err = normalize_and_colormap(disp_err.abs())
-
-    wandb.log({
+    # gt_disp_vis = normalize_and_colormap(gt_disp)
+    # if gt_disp.shape != disp.shape:
+        # gt_disp = gt_disp.reshape(disp.shape)
+    # disp_err = gt_disp - disp
+    # disp_err = normalize_and_colormap(disp_err.abs())
+    if isinstance(left_img, torch.Tensor):
+        left_img = left_img.cpu().detach().numpy().astype('uint8')
+        right_img = right_img.cpu().detach().numpy().astype('uint8')
+
+    results = {
         'disp': wandb.Image(
             disp,
             caption=f"Pred. Disparity \n{disp.min():.{2}f}/{disp.max():.{2}f}",
@@ -35,41 +39,60 @@ def do_infer(left_img, right_img, gt_disp, model):
             disp_vis,
             caption=f"Pred. Disparity \n{disp.min():.{2}f}/{disp.max():.{2}f}",
         ),
-        'gt_disp_vis': wandb.Image(
-            gt_disp_vis,
-            caption=f"GT Disparity \n{gt_disp.min():.{2}f}/{gt_disp.max():.{2}f}",
-        ),
-        'disp_err': wandb.Image(
-            disp_err,
-            caption=f"Disparity Error\n{disp_err.min():.{2}f}/{disp_err.max():.{2}f}",
-        ),
+        # 'disp_err': wandb.Image(
+            # disp_err,
+            # caption=f"Disparity Error\n{disp_err.min():.{2}f}/{disp_err.max():.{2}f}",
+        # ),
         'input_left': wandb.Image(
-            left_img.cpu().detach().numpy().astype('uint8'),
+            left_img,
             caption=f"Input left",
         ),
         'input_right': wandb.Image(
-            right_img.cpu().detach().numpy().astype('uint8'),
+            right_img,
             caption=f"Input right",
         ),
-    })
+    }
+
+    if gt_disp is not None:
+        print('logging gt')
+        print(f'gt: {gt_disp.max()}/{gt_disp.min()}/{gt_disp.mean()}')
+        gt_disp_vis = normalize_and_colormap(gt_disp)
+        results.update({
+            'gt_disp_vis': wandb.Image(
+                gt_disp_vis,
+                caption=f"GT Disparity \n{gt_disp.min():.{2}f}/{gt_disp.max():.{2}f}",
+            )})
+    wandb.log(results)
+
+
+def downsample(img, half_height_out=480):
+    downsampled = cv2.pyrDown(img)
+    diff = (downsampled.shape[0] - half_height_out) // 2
+    return downsampled[diff:downsampled.shape[0] - diff, 0:downsampled.shape[1]]
 
 
 if __name__ == '__main__':
     # model_path = "models/crestereo_eth3d.pth"
     model_path = "train_log/models/latest.pth"
+    # model_path = "train_log/models/epoch-120.pth"
+    # model_path = "train_log/models/epoch-250.pth"
 
+    print(model_path)
     # reference_pattern_path = '/home/nils/kinect_reference_high_res_scaled_down.png'
-    reference_pattern_path = '/home/nils/kinect_reference_cropped.png'
     # reference_pattern_path = '/home/nils/new_reference.png'
     # reference_pattern_path = '/home/nils/kinect_reference_high_res.png'
+    reference_pattern_path = '/home/nils/miniprojekt/kinect_high_res_thresh_denoised.png'
+    # reference_pattern_path = '/home/nils/miniprojekt/kinect_syn_ref.png'
+    # reference_pattern_path = '/home/nils/kinect_reference_cropped.png'
     # reference_pattern_path = '/home/nils/orig_ctd/connecting_the_dots/data/kinect_pattern.png'
 
-    data_type = 'kinect'
+    # data_type = 'kinect'
+    data_type = 'blender'
     augment = False
 
     args = parse_yaml("cfgs/train.yaml")
 
-    wandb.config.update({'model_path': model_path, 'reference_pattern': reference_pattern_path, 'augment': augment})
+    wandb.config.update({'model_path': model_path, 'reference_pattern': reference_pattern_path, 'augment': augment, 'data_type': data_type, 'pattern_self_attention': args.pattern_attention})
 
     model = Model(max_disp=256, mixed_precision=False, test_mode=True)
     model = nn.DataParallel(model, device_ids=[device])
@@ -78,16 +101,32 @@ if __name__ == '__main__':
     model.to(device)
     model.eval()
 
-    dataset = CTDDataset('/media/Data1/connecting_the_dots_data/ctd_data/', data_type=data_type,
-                         pattern_path=reference_pattern_path, augment=augment)
-    dataloader = DataLoader(dataset, args.batch_size, shuffle=True,
-                            num_workers=0, drop_last=False, persistent_workers=False, pin_memory=True)
-    for batch in dataloader:
-        for left, right, disparity in zip(batch['left'], batch['right'], batch['disparity']):
-            right = right.transpose(0, 2).transpose(0, 1)
-            left_img = left
-            imgL = left.cpu().detach().numpy()
-            right_img = right
-            imgR = right.cpu().detach().numpy()
-            gt_disp = disparity
-            do_infer(left_img, right_img, gt_disp, model)
+    # dataset = CTDDataset('/media/Data1/connecting_the_dots_data/ctd_data/', data_type=data_type,
+                         # pattern_path=reference_pattern_path, augment=augment)
+    # dataloader = DataLoader(dataset, args.batch_size, shuffle=True,
+                            # num_workers=0, drop_last=False, persistent_workers=False, pin_memory=True)
+    
+    # for batch in dataloader:
+    gt_disp = None
+    right = downsample(cv2.imread(reference_pattern_path))
+
+    if data_type == 'blender':
+        img_path = '/media/Data1/connecting_the_dots_data/blender_renders/data/'
+    elif data_type == 'kinect':
+        img_path = '/home/nils/kinect_pngs/ir/'
+
+    for img in sorted(list(entry for entry in os.scandir(img_path) if 'depth' not in entry.name), key=lambda x:x.name)[:25]:
+        print(img.path)
+        if data_type == 'blender':
+            baseline = 0.075    # meters
+            fl = 560.           # as per CTD
+
+            gt_path = img.path.rsplit('.')[0] + '_depth0001.png'
+            gt_depth = downsample(cv2.imread(gt_path))
+
+            mystery_factor = 35     # we don't get reasonable disparities due to incorrect depth scaling (or something like that)
+            gt_disp = (baseline * fl * mystery_factor) / gt_depth
+
+        left = downsample(cv2.imread(img.path))
+
+        do_infer(left, right, gt_disp, model, attend_pattern=args.pattern_attention)

train.py

@@ -9,7 +9,7 @@ import yaml
 
 from nets import Model
 # from dataset import CREStereoDataset
-from dataset import CREStereoDataset, CTDDataset
+from dataset import BlenderDataset, CREStereoDataset, CTDDataset
 
 import torch
 import torch.nn as nn
@@ -32,15 +32,19 @@ def normalize_and_colormap(img):
     return ret
 
 
-def inference(left, right, gt_disp, mask, model, epoch, n_iter=20, wandb_log=True):
+def inference(left, right, gt_disp, mask, model, epoch, n_iter=20, wandb_log=True, last_img=None, test=False, attend_pattern=True):
 
     print("Model Forwarding...")
-    left = left.cpu().detach().numpy()
+    if isinstance(left, torch.Tensor):
+        left = left.cpu().detach().numpy()
+        imgR = right.cpu().detach().numpy()
     imgL = left
-    imgR = right.cpu().detach().numpy()
+    imgR = right
     imgL = np.ascontiguousarray(imgL[None, :, :, :])
     imgR = np.ascontiguousarray(imgR[None, :, :, :])
 
+    flow_init = None
+    
     # chosen for convenience
     device = torch.device('cuda:0')
 
@@ -55,19 +59,54 @@ def inference(left, right, gt_disp, mask, model, epoch, n_iter=20, wandb_log=Tru
             size=(imgL.shape[2] // 2, imgL.shape[3] // 2),
             mode="bilinear",
             align_corners=True,
-            )
+            ).clamp(min=0, max=255)
     imgR_dw2 = F.interpolate(
             imgR,
             size=(imgL.shape[2] // 2, imgL.shape[3] // 2),
             mode="bilinear",
             align_corners=True,
-            )
+            ).clamp(min=0, max=255)
+    if last_img is not None:
+        print('using flow_initialization')
+        print(last_img.shape)
+        # FIXME trying 8bit normalization because sometimes we get really high values that don't seem to help
+        print(last_img.max(), last_img.min())
+        if last_img.min() < 0:
+            # print('Negative disparity detected. shifting...')
+            last_img = last_img - last_img.min()
+        if last_img.max() > 255:
+        # print('Excessive disparity detected. scaling...')
+            last_img = last_img / (last_img.max() / 255)
+
+
+        last_img = np.dstack([last_img, last_img])
+        # last_img = np.dstack([last_img, last_img, last_img])
+        last_img = np.dstack([last_img])
+        last_img = last_img.reshape((1, 2, 480, 640))
+        # print(last_img.shape)
+        # print(last_img.dtype)
+        # print(last_img.max(), last_img.min())
+        flow_init = torch.tensor(last_img.astype("float32")).to(device)
+        # flow_init = F.interpolate(
+            # last_img,
+            # size=(last_img.shape[0] // 2, last_img.shape[1] // 2),
+            # mode="bilinear",
+            # align_corners=True,
+            # )
     with torch.inference_mode():
-        pred_flow_dw2 = model(image1=imgL_dw2, image2=imgR_dw2, iters=n_iter, flow_init=None)
-        pred_flow = model(imgL, imgR, iters=n_iter, flow_init=pred_flow_dw2)
+        pred_flow_dw2 = model(image1=imgL_dw2, image2=imgR_dw2, iters=n_iter, flow_init=flow_init, self_attend_right=attend_pattern)
+        pred_flow = model(imgL, imgR, iters=n_iter, flow_init=pred_flow_dw2, self_attend_right=attend_pattern)
+        pf_base = pred_flow
+        if isinstance(pf_base, list):
+            pf_base = pred_flow[0]
+        pf = torch.squeeze(pf_base[:, 0, :, :]).cpu().detach().numpy()
+        print('pred_flow max min')
+        print(pf.max(), pf.min())
 
 
     if not wandb_log:
+        if test:
+            return pred_flow
         return torch.squeeze(pred_flow[:, 0, :, :]).cpu().detach().numpy()
 
     log = {}
@@ -96,30 +135,36 @@ def inference(left, right, gt_disp, mask, model, epoch, n_iter=20, wandb_log=Tru
                 np.array([pred_disp.reshape(480, 640)]), 
                 caption=f"Pred. Disp. It {i}\n{pred_disp.min():.{2}f}/{pred_disp.max():.{2}f}",
                 )
-        log[f'pred_norm_{i}'] = wandb.Image(
-                np.array([pred_disp_norm.reshape(480, 640)]), 
-                caption=f"Pred. Disp. It {i}\n{pred_disp.min():.{2}f}/{pred_disp.max():.{2}f}",
-                )
+        # log[f'pred_norm_{i}'] = wandb.Image(
+                # np.array([pred_disp_norm.reshape(480, 640)]), 
+                # caption=f"Pred. Disp. It {i}\n{pred_disp.min():.{2}f}/{pred_disp.max():.{2}f}",
+                # )
 
-        log[f'pred_dw2_{i}'] = wandb.Image(
-                np.array([pred_disp_dw2.reshape(240, 320)]), 
-                caption=f"Pred. Disp. Dw2 It {i}\n{pred_disp_dw2.min():.{2}f}/{pred_disp_dw2.max():.{2}f}",
-                )
-        log[f'pred_dw2_norm_{i}'] = wandb.Image(
-                np.array([pred_disp_dw2_norm.reshape(240, 320)]), 
-                caption=f"Pred. Disp. Dw2 It {i}\n{pred_disp_dw2.min():.{2}f}/{pred_disp_dw2.max():.{2}f}",
-                )
+        # log[f'pred_dw2_{i}'] = wandb.Image(
+                # np.array([pred_disp_dw2.reshape(240, 320)]), 
+                # caption=f"Pred. Disp. Dw2 It {i}\n{pred_disp_dw2.min():.{2}f}/{pred_disp_dw2.max():.{2}f}",
+                # )
+        # log[f'pred_dw2_norm_{i}'] = wandb.Image(
+                # np.array([pred_disp_dw2_norm.reshape(240, 320)]), 
+                # caption=f"Pred. Disp. Dw2 It {i}\n{pred_disp_dw2.min():.{2}f}/{pred_disp_dw2.max():.{2}f}",
+                # )
 
 
     log['input_left'] = wandb.Image(left.astype('uint8'), caption="Input Left")
-    log['input_right'] = wandb.Image(right.cpu().detach().numpy().transpose(1,2,0).astype('uint8'), caption="Input Right")
+    input_right = right.cpu().detach().numpy() if isinstance(right, torch.Tensor) else right
+    if input_right.shape != (480, 640, 3):
+        input_right.transpose(1,2,0)
+    log['input_right'] = wandb.Image(input_right.astype('uint8'), caption="Input Right")
 
     log['gt_disp'] = wandb.Image(gt_disp, caption=f"GT Disparity\n{gt_disp.min():.{2}f}/{gt_disp.max():.{2}f}")
 
+    gt_disp = gt_disp.cpu().detach().numpy() if isinstance(gt_disp, torch.Tensor) else gt_disp
+    disp = disp.cpu().detach().numpy() if isinstance(disp, torch.Tensor) else disp
+
     disp_error = gt_disp - disp
     log['disp_error'] = wandb.Image(
-            normalize_and_colormap(disp_error.abs()),
-            caption=f"Disp. Error\n{disp_error.min():.{2}f}/{disp_error.max():.{2}f}\n{disp_error.abs().mean():.{2}f}",
+            normalize_and_colormap(abs(disp_error)),
+            caption=f"Disp. Error\n{disp_error.min():.{2}f}/{disp_error.max():.{2}f}\n{abs(disp_error).mean():.{2}f}",
             )
 
 
@@ -129,6 +174,7 @@ def inference(left, right, gt_disp, mask, model, epoch, n_iter=20, wandb_log=Tru
         )
 
     wandb.log(log)
+    return pred_flow
 
 
 def parse_yaml(file_path: str) -> namedtuple:
@@ -172,12 +218,25 @@ def adjust_learning_rate(optimizer, epoch):
     for param_group in optimizer.param_groups:
         param_group['lr'] = lr
 
-def sequence_loss(flow_preds, flow_gt, valid, gamma=0.8):
+def sequence_loss(flow_preds, flow_gt, valid, gamma=0.8, test=False):
     '''
     valid: (2, 384, 512) (B, H, W) -> (B, 1, H, W)
     flow_preds[0]: (B, 2, H, W)
     flow_gt: (B, 2, H, W)
     '''
+    if test:
+        # print('sequence loss')
+        if valid.shape != (2, 480, 640):
+            valid = torch.stack([valid, valid])#.transpose(0,1)#.transpose(1,2)
+            # print(valid.shape)
+            #valid = torch.stack([valid, valid])
+            # print(valid.shape)
+            if valid.shape != (2, 480, 640):
+                valid = valid.transpose(0,1)
+                # print(valid.shape)
+    # print(valid.shape)
+    # print(flow_preds[0].shape)
+    # print(flow_gt.shape)
     n_predictions = len(flow_preds)
     flow_loss = 0.0
     
@@ -260,20 +319,41 @@ def main(args):
         start_iters = 0
 
     # pattern_path = '/home/nils/kinect_reference_high_res_scaled_down.png'
-    pattern_path = '/home/nils/kinect_reference_cropped.png'
+    # pattern_path = '/home/nils/kinect_reference_cropped.png'
+    # pattern_path = '/home/nils/kinect_reference_far.png'
+    # pattern_path = '/home/nils/miniprojekt/kinect_syn_ref.png'
+    pattern_path = '/home/nils/miniprojekt/kinect_high_res_thresh_denoised.png'
     # pattern_path = '/home/nils/orig_ctd/connecting_the_dots/data/kinect_pattern.png'
     # datasets
     # dataset = CREStereoDataset(args.training_data_path)
-    dataset = CTDDataset(args.training_data_path, pattern_path=pattern_path)
+    # dataset = CTDDataset(args.training_data_path, pattern_path=pattern_path)
+    # test_dataset = CTDDataset(args.training_data_path, pattern_path=pattern_path, test_set=True)
+    if args.dataset == 'blender':
+        pattern_path = '/home/nils/miniprojekt/kinect_syn_ref.png'
+        dataset = BlenderDataset(root=args.training_data_path, pattern_path=pattern_path)
+        test_dataset = BlenderDataset(root=args.training_data_path, pattern_path=pattern_path, test_set=True)
+    elif args.dataset == 'ctd':
+        dataset = CTDDataset(args.training_data_path, pattern_path=pattern_path)
+        test_dataset = CTDDataset(args.training_data_path, pattern_path=pattern_path, test_set=True)
+    else:
+        print('unrecognized dataset')
+        quit()
+
+    test_data_iter = iter(test_dataset)
     # if rank == 0:
     worklog.info(f"Dataset size: {len(dataset)}")
+    print(args.batch_size)
     dataloader = DataLoader(dataset, args.batch_size, shuffle=True,
-                            num_workers=0, drop_last=True, persistent_workers=False, pin_memory=True)
+                            num_workers=0, drop_last=False, persistent_workers=False, pin_memory=True)
+                            # num_workers=0, drop_last=True, persistent_workers=False, pin_memory=True)
+    test_dataloader = DataLoader(test_dataset, args.batch_size, shuffle=False,
+                            num_workers=0, drop_last=False, persistent_workers=False, pin_memory=True)
 
     # counter
     cur_iters = start_iters
     total_iters = args.minibatch_per_epoch * args.n_total_epoch
     t0 = time.perf_counter()
+    test_idx = 0
     for epoch_idx in range(start_epoch_idx, args.n_total_epoch + 1):
 
         # adjust learning rate
@@ -310,24 +390,59 @@ def main(args):
             # forward
             # left = left.transpose(1, 2).transpose(2, 3)
             left = left.transpose(1, 3).transpose(2, 3)
-            right = right.transpose(2, 3).transpose(2, 3)#.transpose(1, 2)
-            flow_predictions = model(left.cuda(), right.cuda())
+            # right = right.transpose(2, 3).transpose(2, 3)#.transpose(1, 2)
+            flow_predictions = model(left.cuda(), right.cuda(), self_attend_right=args.pattern_attention)
 
             # loss & backword
             loss = sequence_loss(
                 flow_predictions, gt_flow, valid_mask, gamma=0.8
             )
 
-            if batch_idx % 128 == 0: 
-                inference(
-                    mini_batch_data['left'][0],
-                    mini_batch_data['right'][0],
-                    mini_batch_data['disparity'][0],
-                    mini_batch_data['mask'][0],
-                    model,
-                    batch_idx,
-                )
-
+            if batch_idx % 512 == 0: 
+                test_idx = 0
+                test_loss = 0
+                for i, test_batch in enumerate(test_dataset):
+                    # test_batch = next(test_data_iter)
+                    if i >= 24:
+                        break
+
+                    # TODO refactor, DRY
+                    left, right, gt_disp, valid_mask = (
+                        test_batch['left'],
+                        test_batch['right'],
+                        torch.Tensor(test_batch['disparity']).cuda(),
+                        torch.Tensor(test_batch['mask']).cuda(),
+                    )
+                    gt_disp = torch.dstack([gt_disp, gt_disp]).transpose(2,0).transpose(1,2)
+                    gt_disp = torch.unsqueeze(gt_disp, dim=1)  # [2, 384, 512] -> [2, 1, 384, 512]
+                    gt_flow = torch.cat([gt_disp, gt_disp * 0], dim=1)  # [2, 2, 384, 512]
+                    # print(f'left {left.shape}, right {right.shape}')
+                    # left = left.transpose([2, 0, 1])
+                    right = right.transpose([1, 2, 0])
+                    # right = right.transpose(2, 3).transpose(2, 3)#.transpose(1, 2)
+                    # print(f'left {left.shape}, right {right.shape}')
+
+                    model.eval()
+                    flow_predictions = inference(
+                        left, 
+                        right,
+                        # gt_disp,
+                        torch.Tensor(test_batch['disparity']).cuda(),
+                        valid_mask,
+                        model,
+                        test_idx,
+                        wandb_log=i % 4 == 0,
+                        test=True,
+                    )
+                    test_idx += 1
+                    test_loss += sequence_loss(
+                        flow_predictions, gt_flow, valid_mask, gamma=0.8, test=True
+                    ).data.item()
+                    model.train()
+
+                avg_test_loss = test_loss / test_idx
+                print(f'test_loss: {test_loss}\nlen test: {test_idx}\navg. loss: {avg_test_loss}')
+                metrics['test/loss'] = avg_test_loss
             # loss stats
             loss_item = loss.data.item()
             epoch_total_train_loss += loss_item

train_lightning.py

@@ -0,0 +1,346 @@
+import os
+import sys
+import time
+import logging
+from collections import namedtuple
+
+import yaml
+# from tensorboardX import SummaryWriter
+
+from nets import Model
+# from dataset import CREStereoDataset
+from dataset import BlenderDataset, CREStereoDataset, CTDDataset
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.optim as optim
+from torch.utils.data import DataLoader
+from pytorch_lightning.lite import LightningLite
+from pytorch_lightning import LightningDataModule, LightningModule, Trainer
+from pytorch_lightning import Trainer, seed_everything
+from pytorch_lightning.loggers import WandbLogger
+
+seed_everything(42, workers=True)
+
+
+import wandb
+
+import numpy as np
+import cv2
+
+
+def normalize_and_colormap(img):
+    ret = (img - img.min()) / (img.max() - img.min()) * 255.0
+    if isinstance(ret, torch.Tensor):
+        ret = ret.cpu().detach().numpy()
+    ret = ret.astype("uint8")
+    ret = cv2.applyColorMap(ret, cv2.COLORMAP_INFERNO)
+    return ret
+
+
+def inference(left, right, gt_disp, mask, model, epoch, n_iter=20, wandb_log=True, last_img=None, test=False, attend_pattern=True):
+
+    print("Model Forwarding...")
+    if isinstance(left, torch.Tensor):
+        left = left# .cpu().detach().numpy()
+        imgR = right# .cpu().detach().numpy()
+    imgL = left
+    imgR = right
+    imgL = np.ascontiguousarray(imgL[None, :, :, :])
+    imgR = np.ascontiguousarray(imgR[None, :, :, :])
+
+    flow_init = None
+    
+    # chosen for convenience
+
+    imgL = torch.tensor(imgL.astype("float32"))
+    imgR = torch.tensor(imgR.astype("float32"))
+    imgL = imgL.transpose(2,3).transpose(1,2)
+    if imgL.shape != imgR.shape:
+        imgR = imgR.transpose(2,3).transpose(1,2)
+
+    imgL_dw2 = F.interpolate(
+            imgL,
+            size=(imgL.shape[2] // 2, imgL.shape[3] // 2),
+            mode="bilinear",
+            align_corners=True,
+            ).clamp(min=0, max=255)
+    imgR_dw2 = F.interpolate(
+            imgR,
+            size=(imgL.shape[2] // 2, imgL.shape[3] // 2),
+            mode="bilinear",
+            align_corners=True,
+            ).clamp(min=0, max=255)
+    if last_img is not None:
+        print('using flow_initialization')
+        print(last_img.shape)
+        # FIXME trying 8bit normalization because sometimes we get really high values that don't seem to help
+        print(last_img.max(), last_img.min())
+        if last_img.min() < 0:
+            # print('Negative disparity detected. shifting...')
+            last_img = last_img - last_img.min()
+        if last_img.max() > 255:
+        # print('Excessive disparity detected. scaling...')
+            last_img = last_img / (last_img.max() / 255)
+
+
+        last_img = np.dstack([last_img, last_img])
+        # last_img = np.dstack([last_img, last_img, last_img])
+        last_img = np.dstack([last_img])
+        last_img = last_img.reshape((1, 2, 480, 640))
+        # print(last_img.shape)
+        # print(last_img.dtype)
+        # print(last_img.max(), last_img.min())
+        flow_init = torch.tensor(last_img.astype("float32"))
+        # flow_init = F.interpolate(
+            # last_img,
+            # size=(last_img.shape[0] // 2, last_img.shape[1] // 2),
+            # mode="bilinear",
+            # align_corners=True,
+            # )
+    with torch.inference_mode():
+        pred_flow_dw2 = model(image1=imgL_dw2, image2=imgR_dw2, iters=n_iter, flow_init=flow_init, self_attend_right=attend_pattern)
+        pred_flow = model(imgL, imgR, iters=n_iter, flow_init=pred_flow_dw2, self_attend_right=attend_pattern)
+        pf_base = pred_flow
+        if isinstance(pf_base, list):
+            pf_base = pred_flow[0]
+        pf = torch.squeeze(pf_base[:, 0, :, :])# .cpu().detach().numpy()
+        print('pred_flow max min')
+        print(pf.max(), pf.min())
+
+
+    if not wandb_log:
+        if test:
+            return pred_flow
+        return torch.squeeze(pred_flow[:, 0, :, :])# .cpu().detach().numpy()
+
+    log = {}
+    in_h, in_w = left.shape[:2]
+
+    # Resize image in case the GPU memory overflows
+    eval_h, eval_w = (in_h,in_w)
+
+    for i, (pf, pf_dw2) in enumerate(zip(pred_flow, pred_flow_dw2)):
+        pred_disp = torch.squeeze(pf[:, 0, :, :])# .cpu().detach().numpy()
+        pred_disp_dw2 = torch.squeeze(pf_dw2[:, 0, :, :])# .cpu().detach().numpy()
+        
+        # pred_disp_norm = cv2.normalize(pred_disp, None, 0, 255, cv2.NORM_MINMAX, cv2.CV_8U)
+        # pred_disp_dw2_norm = cv2.normalize(pred_disp_dw2, None, 0, 255, cv2.NORM_MINMAX, cv2.CV_8U)
+
+        if i == n_iter-1:
+            t = float(in_w) / float(eval_w)
+            disp = cv2.resize(pred_disp.cpu().detach().numpy(), (in_w, in_h), interpolation=cv2.INTER_LINEAR) * t
+
+            log[f'disp_vis'] = wandb.Image(
+                    normalize_and_colormap(disp),
+                    caption=f"Disparity \n{pred_disp.min():.{2}f}/{pred_disp.max():.{2}f}",
+                    )
+
+        log[f'pred_{i}'] = wandb.Image(
+                np.array([pred_disp.cpu().detach().numpy().reshape(480, 640)]), 
+                caption=f"Pred. Disp. It {i}\n{pred_disp.min():.{2}f}/{pred_disp.max():.{2}f}",
+                )
+        # log[f'pred_norm_{i}'] = wandb.Image(
+                # np.array([pred_disp_norm.reshape(480, 640)]), 
+                # caption=f"Pred. Disp. It {i}\n{pred_disp.min():.{2}f}/{pred_disp.max():.{2}f}",
+                # )
+
+        # log[f'pred_dw2_{i}'] = wandb.Image(
+                # np.array([pred_disp_dw2.reshape(240, 320)]), 
+                # caption=f"Pred. Disp. Dw2 It {i}\n{pred_disp_dw2.min():.{2}f}/{pred_disp_dw2.max():.{2}f}",
+                # )
+        # log[f'pred_dw2_norm_{i}'] = wandb.Image(
+                # np.array([pred_disp_dw2_norm.reshape(240, 320)]), 
+                # caption=f"Pred. Disp. Dw2 It {i}\n{pred_disp_dw2.min():.{2}f}/{pred_disp_dw2.max():.{2}f}",
+                # )
+
+
+    log['input_left'] = wandb.Image(left.astype('uint8'), caption="Input Left")
+    input_right = right# .cpu().detach().numpy() if isinstance(right, torch.Tensor) else right
+    if input_right.shape != (480, 640, 3):
+        input_right.transpose(1,2,0)
+    log['input_right'] = wandb.Image(input_right.astype('uint8'), caption="Input Right")
+
+    log['gt_disp'] = wandb.Image(gt_disp, caption=f"GT Disparity\n{gt_disp.min():.{2}f}/{gt_disp.max():.{2}f}")
+
+    gt_disp = gt_disp# .cpu().detach().numpy() if isinstance(gt_disp, torch.Tensor) else gt_disp
+    disp = disp# .cpu().detach().numpy() if isinstance(disp, torch.Tensor) else disp
+
+    disp_error = gt_disp - disp
+    log['disp_error'] = wandb.Image(
+            normalize_and_colormap(abs(disp_error)),
+            caption=f"Disp. Error\n{disp_error.min():.{2}f}/{disp_error.max():.{2}f}\n{abs(disp_error).mean():.{2}f}",
+            )
+
+
+    log[f'gt_disp_vis'] = wandb.Image(
+        normalize_and_colormap(gt_disp),
+        caption=f"GT Disp Vis \n{gt_disp.min():.{2}f}/{gt_disp.max():.{2}f}",
+        )
+
+    wandb.log(log)
+    return pred_flow
+
+
+def parse_yaml(file_path: str) -> namedtuple:
+    """Parse yaml configuration file and return the object in `namedtuple`."""
+    with open(file_path, "rb") as f:
+        cfg: dict = yaml.safe_load(f)
+    args = namedtuple("train_args", cfg.keys())(*cfg.values())
+    return args
+
+
+def format_time(elapse):
+    elapse = int(elapse)
+    hour = elapse // 3600
+    minute = elapse % 3600 // 60
+    seconds = elapse % 60
+    return "{:02d}:{:02d}:{:02d}".format(hour, minute, seconds)
+
+
+def ensure_dir(path):
+    if not os.path.exists(path):
+        os.makedirs(path, exist_ok=True)
+
+
+def adjust_learning_rate(optimizer, epoch):
+
+    warm_up = 0.02
+    const_range = 0.6
+    min_lr_rate = 0.05
+
+    if epoch <= args.n_total_epoch * warm_up:
+        lr = (1 - min_lr_rate) * args.base_lr / (
+            args.n_total_epoch * warm_up
+        ) * epoch + min_lr_rate * args.base_lr
+    elif args.n_total_epoch * warm_up < epoch <= args.n_total_epoch * const_range:
+        lr = args.base_lr
+    else:
+        lr = (min_lr_rate - 1) * args.base_lr / (
+            (1 - const_range) * args.n_total_epoch
+        ) * epoch + (1 - min_lr_rate * const_range) / (1 - const_range) * args.base_lr
+
+    for param_group in optimizer.param_groups:
+        param_group['lr'] = lr
+
+def sequence_loss(flow_preds, flow_gt, valid, gamma=0.8, test=False):
+    '''
+    valid: (2, 384, 512) (B, H, W) -> (B, 1, H, W)
+    flow_preds[0]: (B, 2, H, W)
+    flow_gt: (B, 2, H, W)
+    '''
+    if test:
+        # print('sequence loss')
+        if valid.shape != (2, 480, 640):
+            valid = torch.stack([valid, valid])#.transpose(0,1)#.transpose(1,2)
+            # print(valid.shape)
+            #valid = torch.stack([valid, valid])
+            # print(valid.shape)
+            if valid.shape != (2, 480, 640):
+                valid = valid.transpose(0,1)
+                # print(valid.shape)
+    # print(valid.shape)
+    # print(flow_preds[0].shape)
+    # print(flow_gt.shape)
+    n_predictions = len(flow_preds)
+    flow_loss = 0.0
+    
+    # TEST
+    flow_gt = torch.squeeze(flow_gt, dim=-1)
+
+    for i in range(n_predictions):
+        i_weight = gamma ** (n_predictions - i - 1)
+        i_loss = torch.abs(flow_preds[i] - flow_gt)
+        flow_loss += i_weight * (valid.unsqueeze(1) * i_loss).mean()
+
+    return flow_loss
+
+
+
+class CREStereoLightning(LightningModule):
+    def __init__(self, args):
+        super().__init__()
+        self.batch_size = args.batch_size
+        self.model = Model(
+            max_disp=args.max_disp, mixed_precision=args.mixed_precision, test_mode=False
+        )
+
+    def forward(self, image1, image2, flow_init=None, iters=10, upsample=True, test_mode=False, self_attend_right=True):
+        return self.model(image1, image2, flow_init, iters, upsample, test_mode, self_attend_right)
+
+    def training_step(self, batch, batch_idx):
+        # loss = self(batch)
+        left, right, gt_disp, valid_mask = batch
+        left = torch.Tensor(left).to(self.device)
+        right = torch.Tensor(right).to(self.device)
+        left = left
+        right = right
+        flow_predictions = self.forward(left, right)
+        loss = sequence_loss(
+            flow_predictions, gt_flow, valid_mask, gamma=0.8
+        )
+        self.log("train_loss", loss)
+        return loss
+
+    def validation_step(self, batch, batch_idx):
+        left, right, gt_disp, valid_mask = batch
+        left = torch.Tensor(left).to(self.device)
+        right = torch.Tensor(right).to(self.device)
+        print(left.shape)
+        print(right.shape)
+        flow_predictions = self.forward(left, right)
+        val_loss = sequence_loss(
+            flow_predictions, gt_flow, valid_mask, gamma=0.8
+        )
+        self.log("val_loss", val_loss)
+
+    def test_step(self, batch, batch_idx):
+        left, right, gt_disp, valid_mask = batch
+        # left, right, gt_disp, valid_mask = (
+            # batch["left"],
+            # batch["right"],
+            # batch["disparity"],
+            # batch["mask"],
+        # )
+        left = torch.Tensor(left).to(self.device)
+        right = torch.Tensor(right).to(self.device)
+        flow_predictions = self.forward(left, right)
+        test_loss = sequence_loss(
+            flow_predictions, gt_flow, valid_mask, gamma=0.8
+        )
+        self.log("test_loss", test_loss)
+    
+    def configure_optimizers(self):
+        return optim.Adam(self.model.parameters(), lr=0.01, betas=(0.9, 0.999))
+
+
+if __name__ == "__main__":
+    # train configuration
+    args = parse_yaml("cfgs/train.yaml")
+    # wandb.init(project="crestereo-lightning", entity="cpt-captain")
+    # Lite(strategy='dp', accelerator='gpu', devices=2).run(args)
+    pattern_path = '/home/nils/miniprojekt/kinect_syn_ref.png'
+    model = CREStereoLightning(args)
+    dataset = BlenderDataset(root=args.training_data_path, pattern_path=pattern_path, use_lightning=True)
+    test_dataset = BlenderDataset(root=args.training_data_path, pattern_path=pattern_path, test_set=True, use_lightning=True)
+    print(len(dataset))
+    print(len(test_dataset))
+    wandb_logger = WandbLogger(project="crestereo-lightning")
+    wandb.config.update(args._asdict())
+
+    trainer = Trainer(
+            max_epochs=args.n_total_epoch, 
+            accelerator='gpu', 
+            devices=2,
+            # auto_scale_batch_size='binsearch',
+            # strategy='ddp',
+            deterministic=True,
+            check_val_every_n_epoch=1,
+            limit_val_batches=24,
+            limit_test_batches=24,
+            logger=wandb_logger,
+            default_root_dir=args.log_dir_lightning,
+            )
+    # trainer.tune(model)
+    trainer.fit(model, dataset, test_dataset)