Update models/pcsr.py
Browse files- models/pcsr.py +197 -196
models/pcsr.py
CHANGED
@@ -1,197 +1,198 @@
|
|
1 |
-
import math
|
2 |
-
import numpy as np
|
3 |
-
import torch
|
4 |
-
import torch.nn as nn
|
5 |
-
import torch.nn.functional as F
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
from
|
10 |
-
from
|
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 |
-
l =
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
self.
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
prob =
|
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 |
-
inp_cell
|
111 |
-
inp_cell[:,:,
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
cost_list =
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
l =
|
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 |
-
l
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
pred_unfold = F.
|
187 |
-
|
188 |
-
|
189 |
-
flag_unfold = F.
|
190 |
-
|
191 |
-
|
192 |
-
cond
|
193 |
-
|
194 |
-
|
195 |
-
pred =
|
196 |
-
|
|
|
197 |
return pred, flag
|
|
|
1 |
+
import math
|
2 |
+
import numpy as np
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
from huggingface_hub import PyTorchModelHubMixin
|
7 |
+
|
8 |
+
import models
|
9 |
+
from models import register
|
10 |
+
from fast_pytorch_kmeans import KMeans
|
11 |
+
from utils import *
|
12 |
+
|
13 |
+
|
14 |
+
@register('pcsr-phase0')
|
15 |
+
class PCSR(nn.Module, PyTorchModelHubMixin):
|
16 |
+
def __init__(self, encoder_spec, heavy_sampler_spec):
|
17 |
+
super().__init__()
|
18 |
+
self.encoder = models.make(encoder_spec)
|
19 |
+
in_dim = self.encoder.out_dim
|
20 |
+
self.heavy_sampler = models.make(heavy_sampler_spec,
|
21 |
+
args={'in_dim': in_dim, 'out_dim': 3})
|
22 |
+
|
23 |
+
def forward(self, lr, coord, cell, **kwargs):
|
24 |
+
if self.training:
|
25 |
+
return self.forward_train(lr, coord, cell)
|
26 |
+
else:
|
27 |
+
return self.forward_test(lr, coord, cell, **kwargs)
|
28 |
+
|
29 |
+
def forward_train(self, lr, coord, cell):
|
30 |
+
feat = self.encoder(lr)
|
31 |
+
res = F.grid_sample(lr, coord.flip(-1).unsqueeze(1), mode='bilinear',
|
32 |
+
padding_mode='border', align_corners=False)[:,:,0,:].permute(0,2,1)
|
33 |
+
pred_heavy = self.heavy_sampler(feat, coord, cell) + res
|
34 |
+
return pred_heavy
|
35 |
+
|
36 |
+
def forward_test(self, lr, coord, cell, pixel_batch_size=None):
|
37 |
+
feat = self.encoder(lr)
|
38 |
+
b,q = coord.shape[:2]
|
39 |
+
tot = b*q
|
40 |
+
if not pixel_batch_size:
|
41 |
+
pixel_batch_size = q
|
42 |
+
|
43 |
+
preds = []
|
44 |
+
for i in range(b): # for each image
|
45 |
+
pred = torch.zeros((q,3), device=lr.device)
|
46 |
+
l = 0
|
47 |
+
while l < q:
|
48 |
+
r = min(q, l+pixel_batch_size)
|
49 |
+
coord_split = coord[i:i+1,l:r,:]
|
50 |
+
cell_split = cell[i:i+1,l:r,:]
|
51 |
+
res = F.grid_sample(lr[i:i+1], coord_split.flip(-1).unsqueeze(1), mode='bilinear',
|
52 |
+
padding_mode='border', align_corners=False)[:,:,0,:].squeeze(0).transpose(0,1)
|
53 |
+
pred[l:r] = self.heavy_sampler(feat[i:i+1], coord_split, cell_split) + res
|
54 |
+
l = r
|
55 |
+
preds.append(pred)
|
56 |
+
pred = torch.stack(preds, dim=0)
|
57 |
+
return pred
|
58 |
+
|
59 |
+
|
60 |
+
@register('pcsr-phase1')
|
61 |
+
class PCSR(nn.Module, PyTorchModelHubMixin):
|
62 |
+
|
63 |
+
def __init__(self, encoder_spec, heavy_sampler_spec, light_sampler_spec, classifier_spec):
|
64 |
+
super().__init__()
|
65 |
+
self.encoder = models.make(encoder_spec)
|
66 |
+
in_dim = self.encoder.out_dim
|
67 |
+
self.heavy_sampler = models.make(heavy_sampler_spec,
|
68 |
+
args={'in_dim': in_dim, 'out_dim': 3})
|
69 |
+
self.light_sampler = models.make(light_sampler_spec,
|
70 |
+
args={'in_dim': in_dim, 'out_dim': 3})
|
71 |
+
self.classifier = models.make(classifier_spec,
|
72 |
+
args={'in_dim': in_dim, 'out_dim': 2})
|
73 |
+
self.kmeans = KMeans(n_clusters=2, max_iter=20, mode='euclidean', verbose=0)
|
74 |
+
self.cost_list = {}
|
75 |
+
|
76 |
+
def forward(self, lr, coord, cell, **kwargs):
|
77 |
+
if self.training:
|
78 |
+
return self.forward_train(lr, coord, cell)
|
79 |
+
else:
|
80 |
+
return self.forward_test(lr, coord, cell, **kwargs)
|
81 |
+
|
82 |
+
def forward_train(self, lr, coord, cell):
|
83 |
+
feat = self.encoder(lr)
|
84 |
+
prob = self.classifier(feat, coord, cell)
|
85 |
+
prob = F.softmax(prob, dim=-1) # (b,q,2)
|
86 |
+
|
87 |
+
pred_heavy = self.heavy_sampler(feat, coord, cell)
|
88 |
+
pred_light = self.light_sampler(feat, coord, cell)
|
89 |
+
pred = prob[:,:,0:1] * pred_light + prob[:,:,1:2] * pred_heavy
|
90 |
+
|
91 |
+
res = F.grid_sample(lr, coord.flip(-1).unsqueeze(1), mode='bilinear',
|
92 |
+
padding_mode='border', align_corners=False)[:,:,0,:].permute(0,2,1)
|
93 |
+
pred = pred + res
|
94 |
+
return pred, prob
|
95 |
+
|
96 |
+
def forward_test(self, lr, coord, cell, scale=None, hr_size=None, k=0., pixel_batch_size=None, adaptive_cluster=False, refinement=True):
|
97 |
+
h,w = lr.shape[-2:]
|
98 |
+
if not scale and hr_size:
|
99 |
+
H,W = hr_size
|
100 |
+
scale = round((H/h + W/w)/2, 1)
|
101 |
+
else:
|
102 |
+
assert scale and not hr_size
|
103 |
+
H,W = round(h*scale), round(w*scale)
|
104 |
+
hr_size = (H,W)
|
105 |
+
|
106 |
+
if scale not in self.cost_list:
|
107 |
+
h0,w0 = 16,16
|
108 |
+
H0,W0 = round(h0*scale), round(w0*scale)
|
109 |
+
inp_coord = make_coord((H0,W0), flatten=True, device='cuda').unsqueeze(0)
|
110 |
+
inp_cell = torch.ones_like(inp_coord)
|
111 |
+
inp_cell[:,:,0] *= 2/H0
|
112 |
+
inp_cell[:,:,1] *= 2/W0
|
113 |
+
inp_encoder = torch.zeros((1,3,h0,w0), device='cuda')
|
114 |
+
flops_encoder = get_model_flops(self.encoder, inp_encoder)
|
115 |
+
inp_sampler = torch.zeros((1,self.encoder.out_dim,h0,w0), device='cuda')
|
116 |
+
x = get_model_flops(self.light_sampler, inp_sampler, coord=inp_coord, cell=inp_cell)
|
117 |
+
y = get_model_flops(self.heavy_sampler, inp_sampler, coord=inp_coord, cell=inp_cell)
|
118 |
+
cost_list = torch.FloatTensor([x,y]).cuda() + flops_encoder
|
119 |
+
cost_list = cost_list / cost_list.sum()
|
120 |
+
self.cost_list[scale] = cost_list
|
121 |
+
print('cost_list calculated (x{}): {}'.format(scale, cost_list))
|
122 |
+
cost_list = self.cost_list[scale]
|
123 |
+
|
124 |
+
feat = self.encoder(lr)
|
125 |
+
b,q = coord.shape[:2]
|
126 |
+
assert H*W == q
|
127 |
+
tot = b*q
|
128 |
+
if not pixel_batch_size:
|
129 |
+
pixel_batch_size = q
|
130 |
+
|
131 |
+
# pre-calculate flag
|
132 |
+
prob = torch.zeros((b,q,2), device=lr.device)
|
133 |
+
pb = pixel_batch_size//b*b
|
134 |
+
assert pb > 0
|
135 |
+
l = 0
|
136 |
+
while l < q:
|
137 |
+
r = min(q, l+pb)
|
138 |
+
coord_split = coord[:,l:r,:]
|
139 |
+
cell_split = cell[:,l:r,:]
|
140 |
+
prob_split = self.classifier(feat, coord_split, cell_split)
|
141 |
+
prob[:,l:r] = F.softmax(prob_split, dim=-1)
|
142 |
+
l = r
|
143 |
+
|
144 |
+
if adaptive_cluster: # auto-decide threshold
|
145 |
+
diff = prob[:,:,1].view(-1,1) # (tot,1)
|
146 |
+
assert diff.max() > diff.min()
|
147 |
+
diff = (diff - diff.min()) / (diff.max() - diff.min())
|
148 |
+
centroids = torch.FloatTensor([[0.5]]).cuda()
|
149 |
+
flag = self.kmeans.fit_predict(diff, centroids=centroids)
|
150 |
+
_, min_index = torch.min(diff.flatten(), dim=0)
|
151 |
+
if flag[min_index] == 1:
|
152 |
+
flag = 1 - flag # (tot,)
|
153 |
+
flag = flag.view(b,q)
|
154 |
+
else:
|
155 |
+
prob = prob / torch.pow(cost_list, k).view(1,1,2)
|
156 |
+
flag = torch.argmax(prob, dim=-1) # (b,q)
|
157 |
+
|
158 |
+
# inference per image
|
159 |
+
# more efficient implementation may exist
|
160 |
+
preds = []
|
161 |
+
for i in range(b):
|
162 |
+
pred = torch.zeros((q,3), device=lr.device)
|
163 |
+
l = 0
|
164 |
+
while l < q:
|
165 |
+
r = min(q, l+pixel_batch_size)
|
166 |
+
coord_split = coord[i:i+1,l:r,:]
|
167 |
+
cell_split = cell[i:i+1,l:r,:]
|
168 |
+
flg = flag[i,l:r]
|
169 |
+
|
170 |
+
idx_easy = torch.where(flg == 0)[0]
|
171 |
+
idx_hard = torch.where(flg == 1)[0]
|
172 |
+
num_easy, num_hard = len(idx_easy), len(idx_hard)
|
173 |
+
if num_easy > 0:
|
174 |
+
pred[l+idx_easy] = self.light_sampler(feat[i:i+1], coord_split[:,idx_easy,:], cell_split[:,idx_easy,:]).squeeze(0)
|
175 |
+
if num_hard > 0:
|
176 |
+
pred[l+idx_hard] = self.heavy_sampler(feat[i:i+1], coord_split[:,idx_hard,:], cell_split[:,idx_hard,:]).squeeze(0)
|
177 |
+
res = F.grid_sample(lr[i:i+1], coord_split.flip(-1).unsqueeze(1), mode='bilinear',
|
178 |
+
padding_mode='border', align_corners=False)[:,:,0,:].squeeze(0).transpose(0,1)
|
179 |
+
pred[l:r] += res
|
180 |
+
l = r
|
181 |
+
preds.append(pred)
|
182 |
+
pred = torch.stack(preds, dim=0) # (b,q,3)
|
183 |
+
|
184 |
+
if refinement:
|
185 |
+
pred = pred.transpose(1,2).view(-1,3,H,W)
|
186 |
+
pred_unfold = F.pad(pred, (1,1,1,1), mode='replicate')
|
187 |
+
pred_unfold = F.unfold(pred_unfold, 3, padding=0).view(-1,3,9,H,W).mean(dim=2) # (b,3,H,W)
|
188 |
+
flag = flag.view(-1,1,H,W)
|
189 |
+
flag_unfold = F.pad(flag.float(), (1,1,1,1), mode='replicate')
|
190 |
+
flag_unfold = F.unfold(flag_unfold, 3, padding=0).view(-1,1,9,H,W).int().sum(dim=2) # (b,1,H,W)
|
191 |
+
|
192 |
+
cond = (flag==0) & (flag_unfold>0) #
|
193 |
+
cond[:,:,[0,-1],:] = cond[:,:,:,[0,-1]] = False
|
194 |
+
#print('refined: {} / {}'.format(cond.sum().item(), tot))
|
195 |
+
pred = torch.where(cond, pred_unfold, pred)
|
196 |
+
pred = pred.view(-1,3,q).transpose(1,2)
|
197 |
+
flag = flag.view(b,q,1)
|
198 |
return pred, flag
|