Spaces:
Runtime error
Runtime error
File size: 8,925 Bytes
3fad000 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
import torch
import torch.nn.functional as F
from ...models import builder
from ...models.builder import DEPTHER
from ...ops import resize
from .base import BaseDepther
def add_prefix(inputs, prefix):
"""Add prefix for dict.
Args:
inputs (dict): The input dict with str keys.
prefix (str): The prefix to add.
Returns:
dict: The dict with keys updated with ``prefix``.
"""
outputs = dict()
for name, value in inputs.items():
outputs[f"{prefix}.{name}"] = value
return outputs
@DEPTHER.register_module()
class DepthEncoderDecoder(BaseDepther):
"""Encoder Decoder depther.
EncoderDecoder typically consists of backbone, (neck) and decode_head.
"""
def __init__(self, backbone, decode_head, neck=None, train_cfg=None, test_cfg=None, pretrained=None, init_cfg=None):
super(DepthEncoderDecoder, self).__init__(init_cfg)
if pretrained is not None:
assert backbone.get("pretrained") is None, "both backbone and depther set pretrained weight"
backbone.pretrained = pretrained
self.backbone = builder.build_backbone(backbone)
self._init_decode_head(decode_head)
if neck is not None:
self.neck = builder.build_neck(neck)
self.train_cfg = train_cfg
self.test_cfg = test_cfg
assert self.with_decode_head
def _init_decode_head(self, decode_head):
"""Initialize ``decode_head``"""
self.decode_head = builder.build_head(decode_head)
self.align_corners = self.decode_head.align_corners
def extract_feat(self, img):
"""Extract features from images."""
x = self.backbone(img)
if self.with_neck:
x = self.neck(x)
return x
def encode_decode(self, img, img_metas, rescale=True, size=None):
"""Encode images with backbone and decode into a depth estimation
map of the same size as input."""
x = self.extract_feat(img)
out = self._decode_head_forward_test(x, img_metas)
# crop the pred depth to the certain range.
out = torch.clamp(out, min=self.decode_head.min_depth, max=self.decode_head.max_depth)
if rescale:
if size is None:
if img_metas is not None:
size = img_metas[0]["ori_shape"][:2]
else:
size = img.shape[2:]
out = resize(input=out, size=size, mode="bilinear", align_corners=self.align_corners)
return out
def _decode_head_forward_train(self, img, x, img_metas, depth_gt, **kwargs):
"""Run forward function and calculate loss for decode head in
training."""
losses = dict()
loss_decode = self.decode_head.forward_train(img, x, img_metas, depth_gt, self.train_cfg, **kwargs)
losses.update(add_prefix(loss_decode, "decode"))
return losses
def _decode_head_forward_test(self, x, img_metas):
"""Run forward function and calculate loss for decode head in
inference."""
depth_pred = self.decode_head.forward_test(x, img_metas, self.test_cfg)
return depth_pred
def forward_dummy(self, img):
"""Dummy forward function."""
depth = self.encode_decode(img, None)
return depth
def forward_train(self, img, img_metas, depth_gt, **kwargs):
"""Forward function for training.
Args:
img (Tensor): Input images.
img_metas (list[dict]): List of image info dict where each dict
has: 'img_shape', 'scale_factor', 'flip', and may also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
For details on the values of these keys see
`depth/datasets/pipelines/formatting.py:Collect`.
depth_gt (Tensor): Depth gt
used if the architecture supports depth estimation task.
Returns:
dict[str, Tensor]: a dictionary of loss components
"""
x = self.extract_feat(img)
losses = dict()
# the last of x saves the info from neck
loss_decode = self._decode_head_forward_train(img, x, img_metas, depth_gt, **kwargs)
losses.update(loss_decode)
return losses
def whole_inference(self, img, img_meta, rescale, size=None):
"""Inference with full image."""
depth_pred = self.encode_decode(img, img_meta, rescale, size=size)
return depth_pred
def slide_inference(self, img, img_meta, rescale):
"""Inference by sliding-window with overlap.
If h_crop > h_img or w_crop > w_img, the small patch will be used to
decode without padding.
"""
h_stride, w_stride = self.test_cfg.stride
h_crop, w_crop = self.test_cfg.crop_size
batch_size, _, h_img, w_img = img.size()
h_grids = max(h_img - h_crop + h_stride - 1, 0) // h_stride + 1
w_grids = max(w_img - w_crop + w_stride - 1, 0) // w_stride + 1
preds = img.new_zeros((batch_size, 1, h_img, w_img))
count_mat = img.new_zeros((batch_size, 1, h_img, w_img))
for h_idx in range(h_grids):
for w_idx in range(w_grids):
y1 = h_idx * h_stride
x1 = w_idx * w_stride
y2 = min(y1 + h_crop, h_img)
x2 = min(x1 + w_crop, w_img)
y1 = max(y2 - h_crop, 0)
x1 = max(x2 - w_crop, 0)
crop_img = img[:, :, y1:y2, x1:x2]
depth_pred = self.encode_decode(crop_img, img_meta, rescale)
preds += F.pad(depth_pred, (int(x1), int(preds.shape[3] - x2), int(y1), int(preds.shape[2] - y2)))
count_mat[:, :, y1:y2, x1:x2] += 1
assert (count_mat == 0).sum() == 0
if torch.onnx.is_in_onnx_export():
# cast count_mat to constant while exporting to ONNX
count_mat = torch.from_numpy(count_mat.cpu().detach().numpy()).to(device=img.device)
preds = preds / count_mat
return preds
def inference(self, img, img_meta, rescale, size=None):
"""Inference with slide/whole style.
Args:
img (Tensor): The input image of shape (N, 3, H, W).
img_meta (dict): Image info dict where each dict has: 'img_shape',
'scale_factor', 'flip', and may also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
For details on the values of these keys see
`depth/datasets/pipelines/formatting.py:Collect`.
rescale (bool): Whether rescale back to original shape.
Returns:
Tensor: The output depth map.
"""
assert self.test_cfg.mode in ["slide", "whole"]
ori_shape = img_meta[0]["ori_shape"]
assert all(_["ori_shape"] == ori_shape for _ in img_meta)
if self.test_cfg.mode == "slide":
depth_pred = self.slide_inference(img, img_meta, rescale)
else:
depth_pred = self.whole_inference(img, img_meta, rescale, size=size)
output = depth_pred
flip = img_meta[0]["flip"]
if flip:
flip_direction = img_meta[0]["flip_direction"]
assert flip_direction in ["horizontal", "vertical"]
if flip_direction == "horizontal":
output = output.flip(dims=(3,))
elif flip_direction == "vertical":
output = output.flip(dims=(2,))
return output
def simple_test(self, img, img_meta, rescale=True):
"""Simple test with single image."""
depth_pred = self.inference(img, img_meta, rescale)
if torch.onnx.is_in_onnx_export():
# our inference backend only support 4D output
depth_pred = depth_pred.unsqueeze(0)
return depth_pred
depth_pred = depth_pred.cpu().numpy()
# unravel batch dim
depth_pred = list(depth_pred)
return depth_pred
def aug_test(self, imgs, img_metas, rescale=True):
"""Test with augmentations.
Only rescale=True is supported.
"""
# aug_test rescale all imgs back to ori_shape for now
assert rescale
# to save memory, we get augmented depth logit inplace
depth_pred = self.inference(imgs[0], img_metas[0], rescale)
for i in range(1, len(imgs)):
cur_depth_pred = self.inference(imgs[i], img_metas[i], rescale, size=depth_pred.shape[-2:])
depth_pred += cur_depth_pred
depth_pred /= len(imgs)
depth_pred = depth_pred.cpu().numpy()
# unravel batch dim
depth_pred = list(depth_pred)
return depth_pred
|