EasyCV/easycv/models/segmentation/encoder_decoder.py

336 lines
13 KiB
Python

# Copyright (c) Alibaba, Inc. and its affiliates.
import torch
import torch.nn as nn
import torch.nn.functional as F
from easycv.models import builder
from easycv.models.base import BaseModel
from easycv.models.builder import MODELS
from easycv.models.utils.ops import resize_tensor
from easycv.utils.checkpoint import load_checkpoint
from easycv.utils.logger import get_root_logger, print_log
from easycv.utils.misc import add_prefix
# Modified from https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py
@MODELS.register_module()
class EncoderDecoder(BaseModel):
"""Encoder Decoder segmentors.
EncoderDecoder typically consists of backbone, decode_head, auxiliary_head.
Note that auxiliary_head is only used for deep supervision during training,
which could be dumped during inference.
"""
def __init__(
self,
backbone,
decode_head,
neck=None,
auxiliary_head=None,
train_cfg=None,
test_cfg=None,
pretrained=None,
):
super(EncoderDecoder, self).__init__()
self.backbone = builder.build_backbone(backbone)
self.neck = neck
self.auxiliary_head = auxiliary_head
self.pretrained = pretrained
if self.neck is not None:
self.neck = builder.build_neck(self.neck)
self.decode_head = builder.build_head(decode_head)
self.align_corners = self.decode_head.align_corners
self.num_classes = self.decode_head.num_classes
if auxiliary_head is not None:
auxiliary_head = [
auxiliary_head
] if not isinstance(auxiliary_head, list) else auxiliary_head
self.auxiliary_head = nn.ModuleList()
for head_cfg in auxiliary_head:
self.auxiliary_head.append(builder.build_head(head_cfg))
self.train_cfg = train_cfg
self.test_cfg = test_cfg
self.init_weights()
def init_weights(self):
logger = get_root_logger()
if isinstance(self.pretrained, str):
load_checkpoint(
self.backbone, self.pretrained, strict=False, logger=logger)
elif self.pretrained:
if self.backbone.__class__.__name__ == 'PytorchImageModelWrapper':
self.backbone.init_weights(pretrained=self.pretrained)
elif hasattr(self.backbone, 'default_pretrained_model_path'
) and self.backbone.default_pretrained_model_path:
print_log(
'load model from default path: {}'.format(
self.backbone.default_pretrained_model_path), logger)
load_checkpoint(
self.backbone,
self.backbone.default_pretrained_model_path,
strict=False,
logger=logger)
else:
print_log('load model from init weights')
self.backbone.init_weights()
else:
print_log('load model from init weights')
self.backbone.init_weights()
if hasattr(self.decode_head, 'init_weights'):
self.decode_head.init_weights()
if self.auxiliary_head is not None:
for idx in range(len(self.auxiliary_head)):
if hasattr(self.auxiliary_head[idx], 'init_weights'):
self.auxiliary_head[idx].init_weights()
if self.neck is not None and hasattr(self.neck, 'init_weights'):
self.neck.init_weights()
def extract_feat(self, img):
"""Extract features from images."""
x = self.backbone(img)
if self.neck is not None:
x = self.neck(x)
return x
def encode_decode(self, img, img_metas):
"""Encode images with backbone and decode into a semantic segmentation
map of the same size as input."""
x = self.extract_feat(img)
out = self._decode_head_forward_test(x, img_metas)
out = resize_tensor(
input=out,
size=img.shape[2:],
mode='bilinear',
align_corners=self.align_corners)
return out
def _decode_head_forward_train(self, x, img_metas, gt_semantic_seg):
"""Run forward function and calculate loss for decode head in
training."""
losses = dict()
loss_decode = self.decode_head.forward_train(x, img_metas,
gt_semantic_seg,
self.train_cfg)
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."""
seg_logits = self.decode_head.forward_test(x, img_metas, self.test_cfg)
return seg_logits
def _auxiliary_head_forward_train(self, x, img_metas, gt_semantic_seg):
"""Run forward function and calculate loss for auxiliary head in
training."""
losses = dict()
for idx, aux_head in enumerate(self.auxiliary_head):
loss_aux = aux_head.forward_train(x, img_metas, gt_semantic_seg,
self.train_cfg)
losses.update(add_prefix(loss_aux, f'aux_{idx}'))
return losses
def forward_train(self, img, img_metas, gt_semantic_seg):
"""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
`mmseg/datasets/pipelines/formatting.py:Collect`.
gt_semantic_seg (Tensor): Semantic segmentation masks
used if the architecture supports semantic segmentation task.
Returns:
dict[str, Tensor]: a dictionary of loss components
"""
x = self.extract_feat(img)
losses = dict()
loss_decode = self._decode_head_forward_train(x, img_metas,
gt_semantic_seg)
losses.update(loss_decode)
if self.auxiliary_head is not None:
loss_aux = self._auxiliary_head_forward_train(
x, img_metas, gt_semantic_seg)
losses.update(loss_aux)
return losses
def forward_test(self, imgs, img_metas, **kwargs):
"""
Args:
imgs (List[Tensor]): the outer list indicates test-time
augmentations and inner Tensor should have a shape NxCxHxW,
which contains all images in the batch.
img_metas (List[List[dict]]): the outer list indicates test-time
augs (multiscale, flip, etc.) and the inner list indicates
images in a batch.
"""
for var, name in [(imgs, 'imgs'), (img_metas, 'img_metas')]:
if not isinstance(var, list):
raise TypeError(f'{name} must be a list, but got '
f'{type(var)}')
num_augs = len(imgs)
if num_augs != len(img_metas):
raise ValueError(f'num of augmentations ({len(imgs)}) != '
f'num of image meta ({len(img_metas)})')
# all images in the same aug batch all of the same ori_shape and pad shape
for img_meta in img_metas:
ori_shapes = [_['ori_shape'] for _ in img_meta]
assert all(shape == ori_shapes[0] for shape in ori_shapes)
img_shapes = [_['img_shape'] for _ in img_meta]
assert all(shape == img_shapes[0] for shape in img_shapes)
pad_shapes = [_['pad_shape'] for _ in img_meta]
assert all(shape == pad_shapes[0] for shape in pad_shapes)
if num_augs == 1:
return self.simple_test(imgs[0], img_metas[0], **kwargs)
else:
return self.aug_test(imgs, img_metas, **kwargs)
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()
num_classes = self.num_classes
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, num_classes, 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]
crop_seg_logit = self.encode_decode(crop_img, img_meta)
preds += F.pad(crop_seg_logit,
(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
if rescale:
preds = resize_tensor(
preds,
size=img_meta[0]['ori_shape'][:2],
mode='bilinear',
align_corners=self.align_corners,
warning=False)
return preds
def whole_inference(self, img, img_meta, rescale):
"""Inference with full image."""
seg_logit = self.encode_decode(img, img_meta)
if rescale:
# support dynamic shape for onnx
if torch.onnx.is_in_onnx_export():
size = img.shape[2:]
else:
size = img_meta[0]['ori_shape'][:2]
seg_logit = resize_tensor(
seg_logit,
size=size,
mode='bilinear',
align_corners=self.align_corners,
warning=False)
return seg_logit
def inference(self, img, img_meta, rescale):
"""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
`mmseg/datasets/pipelines/formatting.py:Collect`.
rescale (bool): Whether rescale back to original shape.
Returns:
Tensor: The output segmentation 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':
seg_logit = self.slide_inference(img, img_meta, rescale)
else:
seg_logit = self.whole_inference(img, img_meta, rescale)
output = F.softmax(seg_logit, dim=1)
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."""
seg_logit = self.inference(img, img_meta, rescale)
seg_pred = seg_logit.argmax(dim=1)
if torch.onnx.is_in_onnx_export():
# our inference backend only support 4D output
seg_pred = seg_pred.unsqueeze(0)
return seg_pred
seg_pred = seg_pred.cpu().numpy()
# unravel batch dim
seg_pred = list(seg_pred)
return {'seg_pred': seg_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 seg logit inplace
seg_logit = self.inference(imgs[0], img_metas[0], rescale)
for i in range(1, len(imgs)):
cur_seg_logit = self.inference(imgs[i], img_metas[i], rescale)
seg_logit += cur_seg_logit
seg_logit /= len(imgs)
seg_pred = seg_logit.argmax(dim=1)
seg_pred = seg_pred.cpu().numpy()
# unravel batch dim
seg_pred = list(seg_pred)
return {'seg_pred': seg_pred}