File size: 5,390 Bytes
797142e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
# based on https://github.com/isl-org/MiDaS

import cv2
import torch
import torch.nn as nn
from torchvision.transforms import Compose

from ldm.modules.midas.midas.dpt_depth import DPTDepthModel
from ldm.modules.midas.midas.midas_net import MidasNet
from ldm.modules.midas.midas.midas_net_custom import MidasNet_small
from ldm.modules.midas.midas.transforms import Resize, NormalizeImage, PrepareForNet


ISL_PATHS = {
    "dpt_large": "/fsx/robin/midas_models/dpt_large-midas-2f21e586.pt",  # TODO: adapt
    "dpt_hybrid": "/fsx/robin/midas_models/dpt_hybrid-midas-501f0c75.pt",  # TODO: adapt
    "midas_v21": "",
    "midas_v21_small": "",
}


def disabled_train(self, mode=True):
    """Overwrite model.train with this function to make sure train/eval mode
    does not change anymore."""
    return self


def load_midas_transform(model_type):
    # https://github.com/isl-org/MiDaS/blob/master/run.py
    # load transform only
    if model_type == "dpt_large":  # DPT-Large
        net_w, net_h = 384, 384
        resize_mode = "minimal"
        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])

    elif model_type == "dpt_hybrid":  # DPT-Hybrid
        net_w, net_h = 384, 384
        resize_mode = "minimal"
        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])

    elif model_type == "midas_v21":
        net_w, net_h = 384, 384
        resize_mode = "upper_bound"
        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

    elif model_type == "midas_v21_small":
        net_w, net_h = 256, 256
        resize_mode = "upper_bound"
        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

    else:
        assert False, f"model_type '{model_type}' not implemented, use: --model_type large"

    transform = Compose(
        [
            Resize(
                net_w,
                net_h,
                resize_target=None,
                keep_aspect_ratio=True,
                ensure_multiple_of=32,
                resize_method=resize_mode,
                image_interpolation_method=cv2.INTER_CUBIC,
            ),
            normalization,
            PrepareForNet(),
        ]
    )

    return transform


def load_model(model_type):
    # https://github.com/isl-org/MiDaS/blob/master/run.py
    # load network
    model_path = ISL_PATHS[model_type]
    if model_type == "dpt_large":  # DPT-Large
        model = DPTDepthModel(
            path=model_path,
            backbone="vitl16_384",
            non_negative=True,
        )
        net_w, net_h = 384, 384
        resize_mode = "minimal"
        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])

    elif model_type == "dpt_hybrid":  # DPT-Hybrid
        model = DPTDepthModel(
            path=model_path,
            backbone="vitb_rn50_384",
            non_negative=True,
        )
        net_w, net_h = 384, 384
        resize_mode = "minimal"
        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])

    elif model_type == "midas_v21":
        model = MidasNet(model_path, non_negative=True)
        net_w, net_h = 384, 384
        resize_mode = "upper_bound"
        normalization = NormalizeImage(
            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
        )

    elif model_type == "midas_v21_small":
        model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True,
                               non_negative=True, blocks={'expand': True})
        net_w, net_h = 256, 256
        resize_mode = "upper_bound"
        normalization = NormalizeImage(
            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
        )

    else:
        print(f"model_type '{model_type}' not implemented, use: --model_type large")
        assert False

    transform = Compose(
        [
            Resize(
                net_w,
                net_h,
                resize_target=None,
                keep_aspect_ratio=True,
                ensure_multiple_of=32,
                resize_method=resize_mode,
                image_interpolation_method=cv2.INTER_CUBIC,
            ),
            normalization,
            PrepareForNet(),
        ]
    )

    return model.eval(), transform


class MiDaSInference(nn.Module):
    MODEL_TYPES_TORCH_HUB = [
        "DPT_Large",
        "DPT_Hybrid",
        "MiDaS_small"
    ]
    MODEL_TYPES_ISL = [
        "dpt_large",
        "dpt_hybrid",
        "midas_v21",
        "midas_v21_small",
    ]

    def __init__(self, model_type):
        super().__init__()
        assert (model_type in self.MODEL_TYPES_ISL)
        model, _ = load_model(model_type)
        self.model = model
        self.model.train = disabled_train

    def forward(self, x):
        # x in 0..1 as produced by calling self.transform on a 0..1 float64 numpy array
        # NOTE: we expect that the correct transform has been called during dataloading.
        with torch.no_grad():
            prediction = self.model(x)
            prediction = torch.nn.functional.interpolate(
                prediction.unsqueeze(1),
                size=x.shape[2:],
                mode="bicubic",
                align_corners=False,
            )
        assert prediction.shape == (x.shape[0], 1, x.shape[2], x.shape[3])
        return prediction