fix preprocessing
Browse files- .gitignore +3 -0
- exaonepath.py +31 -30
.gitignore
CHANGED
@@ -172,3 +172,6 @@ cython_debug/
|
|
172 |
|
173 |
# PyPI configuration file
|
174 |
.pypirc
|
|
|
|
|
|
|
|
172 |
|
173 |
# PyPI configuration file
|
174 |
.pypirc
|
175 |
+
|
176 |
+
# Project-specific files
|
177 |
+
test.py
|
exaonepath.py
CHANGED
@@ -1,5 +1,6 @@
|
|
1 |
import math
|
2 |
import typing as t
|
|
|
3 |
|
4 |
import torch
|
5 |
import torch.nn as nn
|
@@ -22,7 +23,13 @@ if t.TYPE_CHECKING:
|
|
22 |
from _typeshed import StrPath
|
23 |
|
24 |
|
25 |
-
class
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
def __init__(self, size: int, pad_value: float | None = None):
|
27 |
super().__init__()
|
28 |
self.size = size
|
@@ -44,32 +51,6 @@ class PadToDivisible(T.Transform):
|
|
44 |
return inpt
|
45 |
|
46 |
|
47 |
-
class Preprocessing(T.Transform):
|
48 |
-
def __init__(
|
49 |
-
self, small_tile_size_with_this_mpp: int, small_tile_size_with_target_mpp: int
|
50 |
-
):
|
51 |
-
self.small_tile_size_with_this_mpp = small_tile_size_with_this_mpp
|
52 |
-
self.small_tile_size_with_target_mpp = small_tile_size_with_target_mpp
|
53 |
-
|
54 |
-
def transform(self, inpt, params):
|
55 |
-
assert isinstance(inpt, torch.Tensor) and inpt.ndim >= 3
|
56 |
-
|
57 |
-
# Scale the input tensor to the target MPP
|
58 |
-
if self.small_tile_size_with_this_mpp != self.small_tile_size_with_target_mpp:
|
59 |
-
inpt = TF.resize(
|
60 |
-
inpt,
|
61 |
-
[
|
62 |
-
self.small_tile_size_with_target_mpp,
|
63 |
-
self.small_tile_size_with_target_mpp,
|
64 |
-
],
|
65 |
-
)
|
66 |
-
|
67 |
-
# Normalize the input tensor
|
68 |
-
inpt = scale_and_normalize(inpt)
|
69 |
-
|
70 |
-
return inpt
|
71 |
-
|
72 |
-
|
73 |
class EXAONEPathV20(nn.Module, PyTorchModelHubMixin):
|
74 |
def __init__(
|
75 |
self,
|
@@ -103,7 +84,8 @@ class EXAONEPathV20(nn.Module, PyTorchModelHubMixin):
|
|
103 |
self.model_first_stg,
|
104 |
small_tiles,
|
105 |
batch_size_on_gpu=first_stg_batch_size,
|
106 |
-
preproc_fn=
|
|
|
107 |
small_tile_size_with_this_mpp=small_tile_size,
|
108 |
small_tile_size_with_target_mpp=self.small_tile_size,
|
109 |
),
|
@@ -111,14 +93,14 @@ class EXAONEPathV20(nn.Module, PyTorchModelHubMixin):
|
|
111 |
out_device=self.device,
|
112 |
dtype=torch.bfloat16,
|
113 |
)
|
114 |
-
|
115 |
act1,
|
116 |
height=height,
|
117 |
width=width,
|
118 |
small_tile_size=small_tile_size,
|
119 |
large_tile_size=large_tile_size,
|
120 |
)
|
121 |
-
act2: torch.Tensor = self.model_second_stg(
|
122 |
act2_formatted = format_second_stg_act_as_third_stg_inp(
|
123 |
act2,
|
124 |
height=height,
|
@@ -126,6 +108,7 @@ class EXAONEPathV20(nn.Module, PyTorchModelHubMixin):
|
|
126 |
large_tile_size=large_tile_size,
|
127 |
)
|
128 |
act3: torch.Tensor = self.model_third_stg(act2_formatted)
|
|
|
129 |
return act1[is_tile_valid], act2, act3
|
130 |
|
131 |
def _load_wsi(self, svs_path: "StrPath", target_mpp: float):
|
@@ -163,3 +146,21 @@ class EXAONEPathV20(nn.Module, PyTorchModelHubMixin):
|
|
163 |
is_tile_valid = mask_tile.sum(dim=(1, 2)) > 0
|
164 |
|
165 |
return x, is_tile_valid, padded_size, small_tile_size, large_tile_size
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import math
|
2 |
import typing as t
|
3 |
+
from functools import partial
|
4 |
|
5 |
import torch
|
6 |
import torch.nn as nn
|
|
|
23 |
from _typeshed import StrPath
|
24 |
|
25 |
|
26 |
+
class Transform(T.Transform):
|
27 |
+
# For compatibility with torchvision <= 0.20
|
28 |
+
def _transform(self, inpt, params):
|
29 |
+
return self.transform(inpt, params)
|
30 |
+
|
31 |
+
|
32 |
+
class PadToDivisible(Transform):
|
33 |
def __init__(self, size: int, pad_value: float | None = None):
|
34 |
super().__init__()
|
35 |
self.size = size
|
|
|
51 |
return inpt
|
52 |
|
53 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
class EXAONEPathV20(nn.Module, PyTorchModelHubMixin):
|
55 |
def __init__(
|
56 |
self,
|
|
|
84 |
self.model_first_stg,
|
85 |
small_tiles,
|
86 |
batch_size_on_gpu=first_stg_batch_size,
|
87 |
+
preproc_fn=partial(
|
88 |
+
_preproc,
|
89 |
small_tile_size_with_this_mpp=small_tile_size,
|
90 |
small_tile_size_with_target_mpp=self.small_tile_size,
|
91 |
),
|
|
|
93 |
out_device=self.device,
|
94 |
dtype=torch.bfloat16,
|
95 |
)
|
96 |
+
act1_formatted = format_first_stg_act_as_second_stg_inp(
|
97 |
act1,
|
98 |
height=height,
|
99 |
width=width,
|
100 |
small_tile_size=small_tile_size,
|
101 |
large_tile_size=large_tile_size,
|
102 |
)
|
103 |
+
act2: torch.Tensor = self.model_second_stg(act1_formatted)
|
104 |
act2_formatted = format_second_stg_act_as_third_stg_inp(
|
105 |
act2,
|
106 |
height=height,
|
|
|
108 |
large_tile_size=large_tile_size,
|
109 |
)
|
110 |
act3: torch.Tensor = self.model_third_stg(act2_formatted)
|
111 |
+
|
112 |
return act1[is_tile_valid], act2, act3
|
113 |
|
114 |
def _load_wsi(self, svs_path: "StrPath", target_mpp: float):
|
|
|
146 |
is_tile_valid = mask_tile.sum(dim=(1, 2)) > 0
|
147 |
|
148 |
return x, is_tile_valid, padded_size, small_tile_size, large_tile_size
|
149 |
+
|
150 |
+
|
151 |
+
def _preproc(
|
152 |
+
x: torch.Tensor,
|
153 |
+
small_tile_size_with_this_mpp: int,
|
154 |
+
small_tile_size_with_target_mpp: int,
|
155 |
+
):
|
156 |
+
# Scale the input tensor to the target MPP
|
157 |
+
if small_tile_size_with_this_mpp != small_tile_size_with_target_mpp:
|
158 |
+
x = TF.resize(
|
159 |
+
x,
|
160 |
+
[small_tile_size_with_target_mpp, small_tile_size_with_target_mpp],
|
161 |
+
)
|
162 |
+
|
163 |
+
# Normalize the input tensor
|
164 |
+
x = scale_and_normalize(x)
|
165 |
+
|
166 |
+
return x
|