File size: 16,347 Bytes
7da7768
 
 
 
 
 
 
 
 
 
 
 
 
 
 
47163f5
 
 
 
 
 
e102afc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
47163f5
 
 
 
7da7768
98fc92a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7da7768
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
47163f5
 
 
 
 
 
 
 
 
28425ed
f016e80
98fc92a
47163f5
 
 
98fc92a
 
 
 
 
f016e80
98fc92a
 
 
 
 
55b8905
f016e80
98fc92a
 
 
 
 
 
 
 
 
 
 
47163f5
 
 
928a499
758cfc5
47163f5
 
7e98f35
7093637
7e98f35
c6e2b5c
5914cbe
7beb833
5914cbe
 
7093637
5914cbe
c6e2b5c
5914cbe
7beb833
47163f5
1eefa67
7093637
47163f5
c6e2b5c
 
47163f5
 
 
 
c6e2b5c
7093637
47163f5
 
03bc7e7
c6e2b5c
 
 
47163f5
 
 
acce61f
e102afc
 
448a2e3
e102afc
 
448a2e3
a976ca4
 
 
 
 
 
 
 
 
47163f5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e102afc
 
47163f5
 
 
aff3d3b
47163f5
 
 
 
448a2e3
47163f5
 
 
 
 
 
 
 
 
 
 
 
e102afc
47163f5
5f8442e
47163f5
 
 
 
c6e2b5c
 
 
 
8b3abf5
 
e102afc
 
 
448a2e3
e102afc
 
448a2e3
c6e2b5c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
98fc92a
 
8b3abf5
98fc92a
 
 
 
 
9022b97
98fc92a
 
 
 
8b3abf5
98fc92a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c6e2b5c
e102afc
 
c6e2b5c
 
 
98fc92a
 
c242bd5
c6e2b5c
 
 
 
448a2e3
c6e2b5c
 
 
 
 
 
 
 
 
2b80ac9
c6e2b5c
 
e102afc
c6e2b5c
 
 
 
 
e2e2967
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
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
import os
import os.path as osp

import cv2
import numpy as np
import torch
from basicsr.utils import img2tensor, tensor2img
from pytorch_lightning import seed_everything
from ldm.models.diffusion.plms import PLMSSampler
from ldm.modules.encoders.adapter import Adapter
from ldm.util import instantiate_from_config
from model_edge import pidinet
import gradio as gr
from omegaconf import OmegaConf

import pathlib
import random
import shlex
import subprocess
import sys

import mmcv
from mmdet.apis import inference_detector, init_detector
from mmpose.apis import (inference_top_down_pose_model, init_pose_model, process_mmdet_results, vis_pose_result)

skeleton = [[15, 13], [13, 11], [16, 14], [14, 12], [11, 12], [5, 11], [6, 12], [5, 6], [5, 7], [6, 8], [7, 9], [8, 10],
            [1, 2], [0, 1], [0, 2], [1, 3], [2, 4], [3, 5], [4, 6]]

pose_kpt_color = [[51, 153, 255], [51, 153, 255], [51, 153, 255], [51, 153, 255], [51, 153, 255], [0, 255, 0],
                  [255, 128, 0], [0, 255, 0], [255, 128, 0], [0, 255, 0], [255, 128, 0], [0, 255, 0], [255, 128, 0],
                  [0, 255, 0], [255, 128, 0], [0, 255, 0], [255, 128, 0]]

pose_link_color = [[0, 255, 0], [0, 255, 0], [255, 128, 0], [255, 128, 0],
                   [51, 153, 255], [51, 153, 255], [51, 153, 255], [51, 153, 255], [0, 255, 0], [255, 128, 0],
                   [0, 255, 0], [255, 128, 0], [51, 153, 255], [51, 153, 255], [51, 153, 255], [51, 153, 255],
                   [51, 153, 255], [51, 153, 255], [51, 153, 255]]


sys.path.append('T2I-Adapter')

config_path =  'https://github.com/TencentARC/T2I-Adapter/raw/main/configs/stable-diffusion/'
model_path = 'https://github.com/TencentARC/T2I-Adapter/raw/main/models/'


def imshow_keypoints(img,
                     pose_result,
                     skeleton=None,
                     kpt_score_thr=0.1,
                     pose_kpt_color=None,
                     pose_link_color=None,
                     radius=4,
                     thickness=1):
    """Draw keypoints and links on an image.
    Args:
            img (ndarry): The image to draw poses on.
            pose_result (list[kpts]): The poses to draw. Each element kpts is
                a set of K keypoints as an Kx3 numpy.ndarray, where each
                keypoint is represented as x, y, score.
            kpt_score_thr (float, optional): Minimum score of keypoints
                to be shown. Default: 0.3.
            pose_kpt_color (np.array[Nx3]`): Color of N keypoints. If None,
                the keypoint will not be drawn.
            pose_link_color (np.array[Mx3]): Color of M links. If None, the
                links will not be drawn.
            thickness (int): Thickness of lines.
    """

    img_h, img_w, _ = img.shape
    img = np.zeros(img.shape)

    for idx, kpts in enumerate(pose_result):
        if idx > 1:
            continue
        kpts = kpts['keypoints']
        # print(kpts)
        kpts = np.array(kpts, copy=False)

        # draw each point on image
        if pose_kpt_color is not None:
            assert len(pose_kpt_color) == len(kpts)

            for kid, kpt in enumerate(kpts):
                x_coord, y_coord, kpt_score = int(kpt[0]), int(kpt[1]), kpt[2]

                if kpt_score < kpt_score_thr or pose_kpt_color[kid] is None:
                    # skip the point that should not be drawn
                    continue

                color = tuple(int(c) for c in pose_kpt_color[kid])
                cv2.circle(img, (int(x_coord), int(y_coord)), radius, color, -1)

        # draw links
        if skeleton is not None and pose_link_color is not None:
            assert len(pose_link_color) == len(skeleton)

            for sk_id, sk in enumerate(skeleton):
                pos1 = (int(kpts[sk[0], 0]), int(kpts[sk[0], 1]))
                pos2 = (int(kpts[sk[1], 0]), int(kpts[sk[1], 1]))

                if (pos1[0] <= 0 or pos1[0] >= img_w or pos1[1] <= 0 or pos1[1] >= img_h or pos2[0] <= 0
                        or pos2[0] >= img_w or pos2[1] <= 0 or pos2[1] >= img_h or kpts[sk[0], 2] < kpt_score_thr
                        or kpts[sk[1], 2] < kpt_score_thr or pose_link_color[sk_id] is None):
                    # skip the link that should not be drawn
                    continue
                color = tuple(int(c) for c in pose_link_color[sk_id])
                cv2.line(img, pos1, pos2, color, thickness=thickness)

    return img


def load_model_from_config(config, ckpt, verbose=False):
    print(f"Loading model from {ckpt}")
    pl_sd = torch.load(ckpt, map_location="cpu")
    if "global_step" in pl_sd:
        print(f"Global Step: {pl_sd['global_step']}")
    if "state_dict" in pl_sd:
        sd = pl_sd["state_dict"]
    else:
        sd = pl_sd
    model = instantiate_from_config(config.model)
    m, u = model.load_state_dict(sd, strict=False)
    # if len(m) > 0 and verbose:
    #     print("missing keys:")
    #     print(m)
    # if len(u) > 0 and verbose:
    #     print("unexpected keys:")
    #     print(u)

    model.cuda()
    model.eval()
    return model

class Model:
    def __init__(self,
                 model_config_path: str = 'ControlNet/models/cldm_v15.yaml',
                 model_dir: str = 'models',
                 use_lightweight: bool = True):
        self.device = torch.device(
            'cuda:0' if torch.cuda.is_available() else 'cpu')    
        self.model_dir = pathlib.Path(model_dir)
        self.model_dir.mkdir(exist_ok=True, parents=True)
        self.download_pose_models()
        self.download_models()


    def download_pose_models(self) -> None:
        ## mmpose
        device = "cuda"
        det_config_file = model_path+"faster_rcnn_r50_fpn_coco.py"
        subprocess.run(shlex.split(f'wget {det_config_file} -O models/faster_rcnn_r50_fpn_coco.py'))
        det_config = 'models/faster_rcnn_r50_fpn_coco.py'
        
        det_checkpoint_file = "https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth"
        subprocess.run(shlex.split(f'wget {det_checkpoint_file} -O models/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth'))
        det_checkpoint = 'models/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth'

        pose_config_file = model_path+"hrnet_w48_coco_256x192.py"
        subprocess.run(shlex.split(f'wget {pose_config_file} -O models/hrnet_w48_coco_256x192.py'))
        pose_config = 'models/hrnet_w48_coco_256x192.py'

        pose_checkpoint_file = "https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w48_coco_256x192-b9e0b3ab_20200708.pth"
        subprocess.run(shlex.split(f'wget {pose_checkpoint_file} -O models/hrnet_w48_coco_256x192-b9e0b3ab_20200708.pth'))
        pose_checkpoint = 'models/hrnet_w48_coco_256x192-b9e0b3ab_20200708.pth'
        
        ## detector
        det_config_mmcv = mmcv.Config.fromfile(det_config)
        self.det_model = init_detector(det_config_mmcv, det_checkpoint, device=device)
        pose_config_mmcv = mmcv.Config.fromfile(pose_config)
        self.pose_model = init_pose_model(pose_config_mmcv, pose_checkpoint, device=device)

    def download_models(self) -> None:
        device = 'cuda'
    
        config = OmegaConf.load("configs/stable-diffusion/test_sketch.yaml")
        config.model.params.cond_stage_config.params.device = device

        base_model_file = "https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt"
        base_model_file_anything = "https://huggingface.co/andite/anything-v4.0/resolve/main/anything-v4.0-pruned.ckpt"
        sketch_adapter_file = "https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_sketch_sd14v1.pth"
        pose_adapter_file = "https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_keypose_sd14v1.pth" 
        pidinet_file = model_path+"table5_pidinet.pth"
        clip_file = "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/*"
        
        subprocess.run(shlex.split(f'wget {base_model_file} -O models/sd-v1-4.ckpt'))
        subprocess.run(shlex.split(f'wget {base_model_file_anything} -O models/anything-v4.0-pruned.ckpt'))
        subprocess.run(shlex.split(f'wget {sketch_adapter_file} -O models/t2iadapter_sketch_sd14v1.pth'))
        subprocess.run(shlex.split(f'wget {pose_adapter_file} -O models/t2iadapter_keypose_sd14v1.pth'))
        subprocess.run(shlex.split(f'wget {pidinet_file} -O models/table5_pidinet.pth'))

        
        self.model = load_model_from_config(config, "models/sd-v1-4.ckpt").to(device)
        self.model_anything = load_model_from_config(config, "models/anything-v4.0-pruned.ckpt").to(device)
        current_base = 'sd-v1-4.ckpt'
        self.model_ad_sketch = Adapter(channels=[320, 640, 1280, 1280][:4], nums_rb=2, ksize=1, sk=True, use_conv=False).to(device)
        self.model_ad_sketch.load_state_dict(torch.load("models/t2iadapter_sketch_sd14v1.pth"))
        net_G = pidinet()
        ckp = torch.load('models/table5_pidinet.pth', map_location='cpu')['state_dict']
        net_G.load_state_dict({k.replace('module.',''):v for k, v in ckp.items()})
        net_G.to(device)
        self.sampler= PLMSSampler(self.model)
        self.sampler_anything= PLMSSampler(self.model_anything)
        save_memory=True

        self.model_ad_pose = Adapter(cin=int(3*64),channels=[320, 640, 1280, 1280][:4], nums_rb=2, ksize=1, sk=True, use_conv=False).to(device)
        self.model_ad_pose.load_state_dict(torch.load("models/t2iadapter_keypose_sd14v1.pth"))


    @torch.inference_mode()
    def process_sketch(self, input_img, type_in, color_back, prompt, neg_prompt, fix_sample, scale, con_strength, base_model):    
        global current_base
        device = 'cuda' 
        if base_model == 'sd-v1-4.ckpt':
            model = self.model
            sampler = self.sampler
        else:
            model = self.model_anything
            sampler = self.sampler_anything
        # if current_base != base_model:
        #     ckpt = os.path.join("models", base_model)
        #     pl_sd = torch.load(ckpt, map_location="cpu")
        #     if "state_dict" in pl_sd:
        #         sd = pl_sd["state_dict"]
        #     else:
        #         sd = pl_sd
        #     model.load_state_dict(sd, strict=False) #load_model_from_config(config, os.path.join("models", base_model)).to(device)
        #     current_base = base_model
        con_strength = int((1-con_strength)*50)
        if fix_sample == 'True':
            seed_everything(42)
        
        im = cv2.resize(input_img,(512,512))
    
        if type_in == 'Sketch':
            # net_G = net_G.cpu()
            if color_back == 'White':
                im = 255-im
            im_edge = im.copy()
            im = img2tensor(im)[0].unsqueeze(0).unsqueeze(0)/255.
            # edge = 1-edge # for white background
            im = im>0.5
            im = im.float()
        elif type_in == 'Image':
            im = img2tensor(im).unsqueeze(0)/255.
            im = net_G(im.to(device))[-1]
            im = im>0.5
            im = im.float()
            im_edge = tensor2img(im)
    
        c = model.get_learned_conditioning([prompt])
        nc = model.get_learned_conditioning([neg_prompt])
        
        with torch.no_grad():
            # extract condition features
            features_adapter = self.model_ad_sketch(im.to(device))
    
        shape = [4, 64, 64]
    
        # sampling
        samples_ddim, _ = sampler.sample(S=50,
                                        conditioning=c,
                                        batch_size=1,
                                        shape=shape,
                                        verbose=False,
                                        unconditional_guidance_scale=scale,
                                        unconditional_conditioning=nc,
                                        eta=0.0,
                                        x_T=None,
                                        features_adapter1=features_adapter,
                                        mode = 'sketch',
                                        con_strength = con_strength)
    
        x_samples_ddim = model.decode_first_stage(samples_ddim)
        x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
        x_samples_ddim = x_samples_ddim.permute(0, 2, 3, 1).cpu().numpy()[0]
        x_samples_ddim = 255.*x_samples_ddim
        x_samples_ddim = x_samples_ddim.astype(np.uint8)
    
        return [im_edge, x_samples_ddim]

    @torch.inference_mode()
    def process_pose(self, input_img, prompt, neg_prompt, fix_sample, scale, con_strength, base_model):    
        global current_base
        det_cat_id = 1
        bbox_thr = 0.2
        device = 'cuda'
        if base_model == 'sd-v1-4.ckpt':
            model = self.model
            sampler = self.sampler
        else:
            model = self.model_anything
            sampler = self.sampler_anything
        # if current_base != base_model:
        #     ckpt = os.path.join("models", base_model)
        #     pl_sd = torch.load(ckpt, map_location="cpu")
        #     if "state_dict" in pl_sd:
        #         sd = pl_sd["state_dict"]
        #     else:
        #         sd = pl_sd
        #     model.load_state_dict(sd, strict=False) #load_model_from_config(config, os.path.join("models", base_model)).to(device)
        #     current_base = base_model
        con_strength = int((1-con_strength)*50)
        if fix_sample == 'True':
            seed_everything(42)
        
        im = cv2.resize(input_img,(512,512))

        image = im.copy()
        im = img2tensor(im).unsqueeze(0)/255.
        mmdet_results = inference_detector(self.det_model, image)
        # keep the person class bounding boxes.
        person_results = process_mmdet_results(mmdet_results, det_cat_id)

        # optional
        return_heatmap = False
        dataset = self.pose_model.cfg.data['test']['type']

        # e.g. use ('backbone', ) to return backbone feature
        output_layer_names = None
        pose_results, returned_outputs = inference_top_down_pose_model(
            self.pose_model,
            image,
            person_results,
            bbox_thr=bbox_thr,
            format='xyxy',
            dataset=dataset,
            dataset_info=None,
            return_heatmap=return_heatmap,
            outputs=output_layer_names)

        # show the results
        im_pose = imshow_keypoints(
            image,
            pose_results,
            skeleton=skeleton,
            pose_kpt_color=pose_kpt_color,
            pose_link_color=pose_link_color,
            radius=2,
            thickness=2)

        im_pose = cv2.resize(im_pose,(512,512))
    
        c = model.get_learned_conditioning([prompt])
        nc = model.get_learned_conditioning([neg_prompt])
        
        with torch.no_grad():
            # extract condition features
            pose = img2tensor(im_pose, bgr2rgb=True, float32=True)/255.
            pose = pose.unsqueeze(0)
            features_adapter = self.model_ad_pose(pose.to(device))
    
        shape = [4, 64, 64]
    
        # sampling
        samples_ddim, _ = sampler.sample(S=50,
                                        conditioning=c,
                                        batch_size=1,
                                        shape=shape,
                                        verbose=False,
                                        unconditional_guidance_scale=scale,
                                        unconditional_conditioning=nc,
                                        eta=0.0,
                                        x_T=None,
                                        features_adapter1=features_adapter,
                                        mode = 'sketch',
                                        con_strength = con_strength)
    
        x_samples_ddim = model.decode_first_stage(samples_ddim)
        x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
        x_samples_ddim = x_samples_ddim.permute(0, 2, 3, 1).cpu().numpy()[0]
        x_samples_ddim = 255.*x_samples_ddim
        x_samples_ddim = x_samples_ddim.astype(np.uint8)
    
        return [im_pose[:,:,::-1].astype(np.uint8), x_samples_ddim]