Vincentqyw
commited on
Commit
•
a9b8ec2
1
Parent(s):
bfaa19c
update: roma
Browse files
third_party/Roma/roma/models/encoders.py
CHANGED
@@ -38,10 +38,13 @@ class ResNet50(nn.Module):
|
|
38 |
self.freeze_bn = freeze_bn
|
39 |
self.early_exit = early_exit
|
40 |
self.amp = amp
|
41 |
-
if torch.cuda.is_available()
|
42 |
-
|
|
|
|
|
|
|
43 |
else:
|
44 |
-
self.amp_dtype = torch.
|
45 |
|
46 |
def forward(self, x, **kwargs):
|
47 |
with torch.autocast("cuda", enabled=self.amp, dtype=self.amp_dtype):
|
@@ -78,10 +81,13 @@ class VGG19(nn.Module):
|
|
78 |
super().__init__()
|
79 |
self.layers = nn.ModuleList(tvm.vgg19_bn(pretrained=pretrained).features[:40])
|
80 |
self.amp = amp
|
81 |
-
if torch.cuda.is_available()
|
82 |
-
|
|
|
|
|
|
|
83 |
else:
|
84 |
-
self.amp_dtype = torch.
|
85 |
|
86 |
def forward(self, x, **kwargs):
|
87 |
with torch.autocast("cuda", enabled=self.amp, dtype=self.amp_dtype):
|
@@ -121,10 +127,13 @@ class CNNandDinov2(nn.Module):
|
|
121 |
else:
|
122 |
self.cnn = VGG19(**cnn_kwargs)
|
123 |
self.amp = amp
|
124 |
-
if torch.cuda.is_available()
|
125 |
-
|
|
|
|
|
|
|
126 |
else:
|
127 |
-
self.amp_dtype = torch.
|
128 |
if self.amp:
|
129 |
dinov2_vitl14 = dinov2_vitl14.to(self.amp_dtype)
|
130 |
self.dinov2_vitl14 = [dinov2_vitl14] # ugly hack to not show parameters to DDP
|
|
|
38 |
self.freeze_bn = freeze_bn
|
39 |
self.early_exit = early_exit
|
40 |
self.amp = amp
|
41 |
+
if torch.cuda.is_available():
|
42 |
+
if torch.cuda.is_bf16_supported():
|
43 |
+
self.amp_dtype = torch.bfloat16
|
44 |
+
else:
|
45 |
+
self.amp_dtype = torch.float16
|
46 |
else:
|
47 |
+
self.amp_dtype = torch.float32
|
48 |
|
49 |
def forward(self, x, **kwargs):
|
50 |
with torch.autocast("cuda", enabled=self.amp, dtype=self.amp_dtype):
|
|
|
81 |
super().__init__()
|
82 |
self.layers = nn.ModuleList(tvm.vgg19_bn(pretrained=pretrained).features[:40])
|
83 |
self.amp = amp
|
84 |
+
if torch.cuda.is_available():
|
85 |
+
if torch.cuda.is_bf16_supported():
|
86 |
+
self.amp_dtype = torch.bfloat16
|
87 |
+
else:
|
88 |
+
self.amp_dtype = torch.float16
|
89 |
else:
|
90 |
+
self.amp_dtype = torch.float32
|
91 |
|
92 |
def forward(self, x, **kwargs):
|
93 |
with torch.autocast("cuda", enabled=self.amp, dtype=self.amp_dtype):
|
|
|
127 |
else:
|
128 |
self.cnn = VGG19(**cnn_kwargs)
|
129 |
self.amp = amp
|
130 |
+
if torch.cuda.is_available():
|
131 |
+
if torch.cuda.is_bf16_supported():
|
132 |
+
self.amp_dtype = torch.bfloat16
|
133 |
+
else:
|
134 |
+
self.amp_dtype = torch.float16
|
135 |
else:
|
136 |
+
self.amp_dtype = torch.float32
|
137 |
if self.amp:
|
138 |
dinov2_vitl14 = dinov2_vitl14.to(self.amp_dtype)
|
139 |
self.dinov2_vitl14 = [dinov2_vitl14] # ugly hack to not show parameters to DDP
|
third_party/Roma/roma/models/matcher.py
CHANGED
@@ -76,10 +76,13 @@ class ConvRefiner(nn.Module):
|
|
76 |
self.disable_local_corr_grad = disable_local_corr_grad
|
77 |
self.is_classifier = is_classifier
|
78 |
self.sample_mode = sample_mode
|
79 |
-
if torch.cuda.is_available()
|
80 |
-
|
|
|
|
|
|
|
81 |
else:
|
82 |
-
self.amp_dtype = torch.
|
83 |
|
84 |
def create_block(
|
85 |
self,
|
@@ -337,10 +340,13 @@ class Decoder(nn.Module):
|
|
337 |
self.displacement_dropout_p = displacement_dropout_p
|
338 |
self.gm_warp_dropout_p = gm_warp_dropout_p
|
339 |
self.flow_upsample_mode = flow_upsample_mode
|
340 |
-
if torch.cuda.is_available()
|
341 |
-
|
|
|
|
|
|
|
342 |
else:
|
343 |
-
self.amp_dtype = torch.
|
344 |
|
345 |
def get_placeholder_flow(self, b, h, w, device):
|
346 |
coarse_coords = torch.meshgrid(
|
|
|
76 |
self.disable_local_corr_grad = disable_local_corr_grad
|
77 |
self.is_classifier = is_classifier
|
78 |
self.sample_mode = sample_mode
|
79 |
+
if torch.cuda.is_available():
|
80 |
+
if torch.cuda.is_bf16_supported():
|
81 |
+
self.amp_dtype = torch.bfloat16
|
82 |
+
else:
|
83 |
+
self.amp_dtype = torch.float16
|
84 |
else:
|
85 |
+
self.amp_dtype = torch.float32
|
86 |
|
87 |
def create_block(
|
88 |
self,
|
|
|
340 |
self.displacement_dropout_p = displacement_dropout_p
|
341 |
self.gm_warp_dropout_p = gm_warp_dropout_p
|
342 |
self.flow_upsample_mode = flow_upsample_mode
|
343 |
+
if torch.cuda.is_available():
|
344 |
+
if torch.cuda.is_bf16_supported():
|
345 |
+
self.amp_dtype = torch.bfloat16
|
346 |
+
else:
|
347 |
+
self.amp_dtype = torch.float16
|
348 |
else:
|
349 |
+
self.amp_dtype = torch.float32
|
350 |
|
351 |
def get_placeholder_flow(self, b, h, w, device):
|
352 |
coarse_coords = torch.meshgrid(
|
third_party/Roma/roma/models/transformer/__init__.py
CHANGED
@@ -30,10 +30,14 @@ class TransformerDecoder(nn.Module):
|
|
30 |
self._scales = [16]
|
31 |
self.is_classifier = is_classifier
|
32 |
self.amp = amp
|
33 |
-
if torch.cuda.is_available()
|
34 |
-
|
|
|
|
|
|
|
35 |
else:
|
36 |
-
self.amp_dtype = torch.
|
|
|
37 |
self.pos_enc = pos_enc
|
38 |
self.learned_embeddings = learned_embeddings
|
39 |
if self.learned_embeddings:
|
|
|
30 |
self._scales = [16]
|
31 |
self.is_classifier = is_classifier
|
32 |
self.amp = amp
|
33 |
+
if torch.cuda.is_available():
|
34 |
+
if torch.cuda.is_bf16_supported():
|
35 |
+
self.amp_dtype = torch.bfloat16
|
36 |
+
else:
|
37 |
+
self.amp_dtype = torch.float16
|
38 |
else:
|
39 |
+
self.amp_dtype = torch.float32
|
40 |
+
|
41 |
self.pos_enc = pos_enc
|
42 |
self.learned_embeddings = learned_embeddings
|
43 |
if self.learned_embeddings:
|