Newer
Older
from spherical_models import SDPAConv, SphereSkipConnection, SpherePixelShuffle, SphereGDN
from utils import common_function as util_common
from utils import healpix as hp_utils
class SLB_Downsample(torch.nn.Module):
r"""Spherical Layer Block for Downsampling consists of:
one or several convolutions (with desired aggregation of conv outputs) +
optional non-linearity on the output of conv +
optional down sampling
Args:
"""
def __init__(self,
conv_name,
in_channels,
out_channels,
bias=True,
hop=1,
skip_conn_aggr=None,
activation=None,
activation_args=dict(),
pool_func=None,
pool_size_sqrt=1):
super().__init__()
# 1- Setting convolution
self.node_dim = 1
self.list_conv = torch.nn.ModuleList()
num_conv = hop if conv_name in ["GraphConv", "SDPAConv"] else 1
if skip_conn_aggr=='cat':
out_channels //= num_conv
if conv_name == "ChebConv":
self.list_conv.append(conv(in_channels=in_channels, out_channels=out_channels, K=hop+1, bias=bias, node_dim=self.node_dim))
elif conv_name in ['TAGConv', 'SGConv']: # the graph convolutions in torch_geometric which need number of hop as input
self.list_conv.append(conv(in_channels=in_channels, out_channels=out_channels, K=hop, bias=bias, node_dim=self.node_dim))
elif conv_name in ['GraphConv']: # These convolutions don't accept number of hops as input
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
self.list_conv.append(conv(in_channels=in_channels, out_channels=out_channels, aggr='mean', bias=bias, node_dim=self.node_dim))
self.list_conv.extend(torch.nn.ModuleList([conv(in_channels=out_channels, out_channels=out_channels, aggr='mean', bias=bias, node_dim=self.node_dim) for _ in range(num_conv - 1)])) # Maybe later not all of them has aggr as argument
elif conv_name == "SDPAConv":
conv = SDPAConv
n_firstHopNeighbors = 8
n_neighbors = util_common.sumOfAP(a=n_firstHopNeighbors, d=n_firstHopNeighbors, n=1)
self.list_conv.append(conv(in_channels=in_channels, out_channels=out_channels, kernel_size=n_neighbors + 1, bias=bias, node_dim=self.node_dim))
self.list_conv.extend(torch.nn.ModuleList([conv(in_channels=out_channels, out_channels=out_channels, kernel_size=n_neighbors + 1, bias=bias, node_dim=self.node_dim) for _ in range(num_conv - 1)]))
else:
raise ValueError('Convolution is not defined')
self.in_channels = in_channels
assert len(self.list_conv) == num_conv, "list conv must be equal to num_conv"
# Setting aggregation of convolution results
self.out_channels = out_channels
if num_conv > 1 and skip_conn_aggr not in ["non", "none"]:
self.skipconn = SphereSkipConnection(skip_conn_aggr)
if skip_conn_aggr == "cat":
self.out_channels *= num_conv
else:
self.register_parameter('skipconn', None)
self.conv_out_channels = self.out_channels
# 2- Setting nonlinearity
if activation is None:
self.register_parameter('activation', None)
elif activation in ["GDN"]:
self.activation = SphereGDN(self.out_channels, **activation_args)
else:
self.activation = getattr(torch.nn, activation)(**activation_args)
# 3- Setting Downsampling
self.pool_size_sqrt = pool_size_sqrt
self.pool_size = pool_size_sqrt*pool_size_sqrt
assert (self.pool_size==1 and pool_func is None) or (self.pool_size > 1 and pool_func is not None), "pool_func and pool_size must match."
if pool_func is None or self.pool_size==1:
self.register_parameter('pool', None)
elif pool_func == 'max_pool':
self.pool = getattr(torch.nn, "MaxPool3d")(kernel_size=(1, self.pool_size, 1))
elif pool_func == "avg_pool":
self.pool = getattr(torch.nn, "AvgPool3d")(kernel_size=(1, self.pool_size, 1))
elif pool_func == "stride":
self.pool = "stride"
else:
raise ValueError('Pooling is not defined')
def forward(self, x, index, weight, valid_index=None, mapping=None): # x is a tensor of size [batch_size, num_nodes, num_features]
device = x.device
index = index.to(device)
weight = weight.to(device)
valid_index = valid_index.to(device) if valid_index is not None else None
xs = []
for conv in self.list_conv:
if conv.__class__.__name__ == "SDPAConv":
x = conv(x, neighbors_indices=index, neighbors_weights=weight, valid_index=valid_index)
else:
x = conv(x, edge_index=index, edge_weight=weight)
xs += [x] if self.pool!="stride" else [x.index_select(self.node_dim, torch.arange(0, x.size(self.node_dim), step=self.pool_size, device=x.device))]
x = self.skipconn(xs) if self.skipconn is not None else xs[-1]
if mapping is not None:
mapping = mapping.to(device)
x = x.index_select(self.node_dim, mapping)
if self.activation is not None:
x = self.activation(x)
if (self.pool is not None) and (not isinstance(self.pool, str)):
x = torch.squeeze(self.pool(torch.unsqueeze(x, dim=0)), dim=0)
return x
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
def get_conv_input_res_offset(self):
r"""
Show the offset of the healpix resolution of struct data for the "input of the conv".
Returns
-------
Integer that shows the offset resolution for the convolution of
"""
return 0
def get_output_res_offset(self):
r"""
Show the offset of the healpix resolution of struct data for the "output of the module".
Returns
-------
Integer that shows the offset resolution for the convolution of
"""
if self.pool is None:
return 0
# Otherwise the unpooling is Upsampling
return hp_utils.healpix_getResolutionDownsampled(0, self.pool_size_sqrt)
class SLB_Upsample(torch.nn.Module):
r"""Spherical Layer Block for Upsampling sists of:
one or several convolutions (with desired aggregation of conv outputs) +
optional non-linearity on the output of conv +
optional up-sampling
Args:
"""
def __init__(self,
conv_name,
in_channels,
out_channels,
bias=True,
hop=1,
skip_conn_aggr=None,
activation=None,
activation_args=dict(),
unpool_func=None,
unpool_size_sqrt=1):
super().__init__()
self.node_dim = 1
# 1- Setting up upsampling
self.unpool_size_sqrt = unpool_size_sqrt
self.unpool_size = unpool_size_sqrt * unpool_size_sqrt
assert (self.unpool_size == 1 and unpool_func is None) or (self.unpool_size > 1 and unpool_func is not None), "unpool_func and unpool_size must match."
if unpool_func is None or self.unpool_size == 1:
self.register_parameter('unpool', None)
elif unpool_func in ['nearest', 'linear', 'bilinear', 'bicubic', 'trilinear']:
self.unpool = getattr(torch.nn, "Upsample")(scale_factor=(self.unpool_size, 1), mode=unpool_func)
elif unpool_func == "pixel_shuffle":
self.unpool = SpherePixelShuffle(self.unpool_size_sqrt, self.node_dim)
out_channels *= self.unpool_size
else:
raise ValueError('Unpooling is not defined')
# 2- Setting convolution
self.list_conv = torch.nn.ModuleList()
num_conv = hop if conv_name in ["GraphConv", "SDPAConv"] else 1
if skip_conn_aggr == 'cat':
out_channels //= num_conv
if conv_name == "ChebConv":
self.list_conv.append(conv(in_channels=in_channels, out_channels=out_channels, K=hop+1, bias=bias, node_dim=self.node_dim))
elif conv_name in ['TAGConv', 'SGConv']: # the graph convolutions in torch_geometric which need number of hop as input
self.list_conv.append(conv(in_channels=in_channels, out_channels=out_channels, K=hop, bias=bias, node_dim=self.node_dim))
elif conv_name in ['GraphConv']: # These convolutions don't accept number of hops as input
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
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
self.list_conv.append(conv(in_channels=in_channels, out_channels=out_channels, aggr='mean', bias=bias, node_dim=self.node_dim))
self.list_conv.extend(torch.nn.ModuleList([conv(in_channels=out_channels, out_channels=out_channels, aggr='mean', bias=bias, node_dim=self.node_dim) for _ in range(num_conv - 1)])) # Maybe later not all of them has aggr as argument
elif conv_name == "SDPAConv":
conv = SDPAConv
n_firstHopNeighbors = 8
n_neighbors = util_common.sumOfAP(a=n_firstHopNeighbors, d=n_firstHopNeighbors, n=1)
self.list_conv.append(conv(in_channels=in_channels, out_channels=out_channels, kernel_size=n_neighbors + 1, bias=bias, node_dim=self.node_dim))
self.list_conv.extend(torch.nn.ModuleList([conv(in_channels=out_channels, out_channels=out_channels, kernel_size=n_neighbors + 1, bias=bias, node_dim=self.node_dim) for _ in range(num_conv - 1)]))
else:
raise ValueError('Convolution is not defined')
self.in_channels = in_channels
assert len(self.list_conv) == num_conv, "list conv must be equal to num_conv"
# Setting aggregation of convolution results
self.out_channels = out_channels
if num_conv > 1 and skip_conn_aggr not in ["non", "none"]:
self.skipconn = SphereSkipConnection(skip_conn_aggr)
if skip_conn_aggr == "cat":
self.out_channels *= num_conv
else:
self.register_parameter('skipconn', None)
self.conv_out_channels = self.out_channels
if unpool_func == "pixel_shuffle":
self.out_channels //= self.unpool_size
# 3- Setting nonlinearity
if activation is None:
self.register_parameter('activation', None)
elif activation in ["GDN"]:
self.activation = SphereGDN(self.out_channels, **activation_args)
else:
self.activation = getattr(torch.nn, activation)(**activation_args)
def forward(self, x, index, weight, valid_index=None, mapping=None): # x is a tensor of size [batch_size, num_nodes, num_features]
device = x.device
if mapping is not None:
raise NotImplementedError("Not implemented")
# Note for unpooling:
# if unpooling is Upsample the order is: Upsample then Convolution
# if unpooling is SpherePixelShuffle the order is: Convolution then SpherePixelShuffle
if self.unpool is not None and self.unpool.__class__.__name__ == "Upsample":
x = torch.squeeze(self.unpool(torch.unsqueeze(x, dim=0)), dim=0)
index = index.to(device)
weight = weight.to(device)
valid_index = valid_index.to(device) if valid_index is not None else None
xs = []
for conv in self.list_conv:
if conv.__class__.__name__ == "SDPAConv":
x = conv(x, neighbors_indices=index, neighbors_weights=weight, valid_index=valid_index)
else:
x = conv(x, edge_index=index, edge_weight=weight)
xs += [x]
x = self.skipconn(xs) if self.skipconn is not None else xs[-1]
if self.unpool is not None and self.unpool.__class__.__name__ == "SpherePixelShuffle":
x = self.unpool(x)
if self.activation is not None:
x = self.activation(x)
return x
def get_conv_input_res_offset(self):
r"""
Show the offset of the healpix resolution of struct data for the "input of the conv".
For example, if we use Upsampling, since first the upsampling is applied and then convolution, for unpool_size_sqrt=2
it returns 1 because conv is appliad on upsampled data.
For pixel shuffling, since pixel shuffling is applied after convolution, the function return 0 no matter of unpool_size_sqrt
Returns
-------
Integer that shows the offset resolution for the convolution of
"""
if self.unpool is None:
return 0
# There is an unpooling
if self.unpool.__class__.__name__ == "SpherePixelShuffle":
return 0
# Otherwise the unpooling is Upsampling
return hp_utils.healpix_getResolutionUpsampled(0, self.unpool_size_sqrt)
def get_output_res_offset(self):
r"""
Show the offset of the healpix resolution of struct data for the "output of the module".
Returns
-------
Integer that shows the offset resolution for the convolution of
"""
if self.unpool is None:
return 0
# Otherwise the unpooling is Upsampling
return hp_utils.healpix_getResolutionUpsampled(0, self.unpool_size_sqrt)
if __name__ == '__main__':
import healpy as hp
import healpix_graph_loader
import healpix_sdpa_struct_loader
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
resolution = 2
patch_resolution = 2
patch_id = 5
nside = hp.order2nside(resolution) # == 2 ** sampling_resolution
nPix = hp.nside2npix(nside)
use_geodesic = True
folder = "../GraphData"
cutGraphForPatchOutside = True
weight_type = "gaussian"
K = 1 # Number of hops
conv_name = "SDPAConv" # SDPAConv, 'ChebConv', 'TAGConv', 'SGConv', GraphConv
unpool_func = "nearest" # 'nearest', 'linear', 'bilinear', 'bicubic', 'trilinear', pixel_shuffle
scale_factor = 2
if conv_name=="SDPAConv":
loader = healpix_sdpa_struct_loader.HealpixSdpaStructLoader(weight_type=weight_type,
use_geodesic=use_geodesic,
use_4connectivity=False,
normalization_method="sym",
cutGraphForPatchOutside=cutGraphForPatchOutside,
load_save_folder=folder)
struct_data = loader.getStruct(resolution, K, patch_resolution, patch_id)
# struct_sdpa = sdpa_loader.getStruct(resolution, K)
index_downsample = struct_data[0]
weight_downsample = struct_data[1]
nodes = struct_data[3]
if unpool_func=="pixel_shuffle":
index_upsample = index_downsample
weight_upsample = weight_downsample
else:
struct_data = loader.getStruct(hp_utils.healpix_getResolutionUpsampled(resolution, scale_factor), K,
hp_utils.healpix_getResolutionUpsampled(patch_resolution, scale_factor), patch_id)
# struct_graph = graph_loader.getGraph(sampling_res=resolution)
index_upsample = struct_data[0]
weight_upsample = struct_data[1]
else:
loader = healpix_graph_loader.HealpixGraphLoader(weight_type=weight_type,
use_geodesic=use_geodesic,
use_4connectivity=False,
load_save_folder=folder)
n_hop_graph = 0 if cutGraphForPatchOutside else K
struct_data= loader.getGraph(sampling_res=resolution, patch_res=patch_resolution, num_hops=n_hop_graph, patch_id=patch_id)
# struct_graph = graph_loader.getGraph(sampling_res=resolution)
index_downsample = struct_data[0]
weight_downsample = struct_data[1]
nodes = struct_data[2]
if unpool_func=="pixel_shuffle":
index_upsample = index_downsample
weight_upsample = weight_downsample
else:
struct_data = loader.getGraph(sampling_res=hp_utils.healpix_getResolutionUpsampled(resolution, scale_factor),
patch_res=hp_utils.healpix_getResolutionUpsampled(patch_resolution, scale_factor),
num_hops=n_hop_graph, patch_id=patch_id)
# struct_graph = graph_loader.getGraph(sampling_res=resolution)
index_upsample = struct_data[0]
weight_upsample = struct_data[1]
B = 4 # batch size
in_channels = 2
out_channels = 10
data_th = torch.randn(B, nPix, in_channels)
data_th = data_th.index_select(dim=1, index=nodes)
print("data_th.size()=", data_th.size())
slb_down = SLB_Downsample(conv_name, in_channels, out_channels,
bias=True, hop=2,
skip_conn_aggr="sum",
activation="GDN",
pool_func="max_pool", pool_size_sqrt=scale_factor
)
print(slb_down)
out_down = slb_down(data_th, index_downsample, weight_downsample)
print("out_down.size()=", out_down.size())
# TODO: Check the same for SLB_Upsample
slb_up = SLB_Upsample(conv_name, in_channels, out_channels,
bias=True, hop=2,
skip_conn_aggr="sum",
activation="GDN", activation_args={"inverse":True},
unpool_func=unpool_func, unpool_size_sqrt=scale_factor
)
print(slb_up)
out_up = slb_up(data_th, index_upsample, weight_upsample)
print("out_up.size()=", out_up.size())