afiliot commited on
Commit
377f9d4
1 Parent(s): 2a1dd93

Upload 4 files

Browse files
Files changed (4) hide show
  1. data.py +46 -0
  2. module.py +484 -0
  3. trainer.py +440 -0
  4. utils.py +123 -0
data.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Owkin, Inc.
2
+ # All rights reserved.
3
+ #
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ # Copyright (c) Owkin, Inc.
8
+ # All rights reserved.
9
+ #
10
+ # This source code is licensed under the license found in the
11
+ # LICENSE file in the root directory of this source tree.
12
+
13
+ from typing import Tuple
14
+ import numpy as np
15
+ import torch
16
+
17
+ import datasets
18
+
19
+ class SlideFeaturesDataset:
20
+ """Slide features dataset."""
21
+ def __init__(self, *args, **kwargs):
22
+ self.hf_dataset = datasets.oad_dataset(*args, **kwargs).with_format("torch")
23
+
24
+ def __getitem__(self, item: np.int64) -> Tuple[torch.Tensor, torch.Tensor]:
25
+ """
26
+ Parameters
27
+ ----------
28
+ item: np.int64
29
+ Index of item, will be converted to int.
30
+
31
+ Returns
32
+ -------
33
+ Tuple[torch.Tensor, torch.Tensor]
34
+ (1000, 768), ()
35
+ """
36
+ return (
37
+ self.hf_dataset[int(item)]["features"],
38
+ self.hf_dataset[int(item)]["label"].unsqueeze(0).float()
39
+ )
40
+
41
+ def __len__(self) -> int:
42
+ return len(self.hf_dataset)
43
+
44
+ @property
45
+ def labels(self):
46
+ return self.hf_dataset["label"]
module.py ADDED
@@ -0,0 +1,484 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Owkin, Inc.
2
+ # All rights reserved.
3
+ #
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ # Copyright (c) Owkin, Inc.
8
+ # All rights reserved.
9
+ #
10
+ # This source code is licensed under the license found in the
11
+ # LICENSE file in the root directory of this source tree.
12
+
13
+
14
+
15
+ import warnings
16
+ from typing import List, Optional, Tuple, Union
17
+
18
+ import torch
19
+ from torch import nn
20
+
21
+
22
+ class MLP(torch.nn.Sequential):
23
+ """MLP Module.
24
+
25
+ Parameters
26
+ ----------
27
+ in_features: int
28
+ Features (model input) dimension.
29
+ out_features: int = 1
30
+ Prediction (model output) dimension.
31
+ hidden: Optional[List[int]] = None
32
+ Dimension of hidden layer(s).
33
+ dropout: Optional[List[float]] = None
34
+ Dropout rate(s).
35
+ activation: Optional[torch.nn.Module] = torch.nn.Sigmoid
36
+ MLP activation.
37
+ bias: bool = True
38
+ Add bias to MLP hidden layers.
39
+
40
+ Raises
41
+ ------
42
+ ValueError
43
+ If ``hidden`` and ``dropout`` do not share the same length.
44
+ """
45
+
46
+ def __init__(
47
+ self,
48
+ in_features: int,
49
+ out_features: int,
50
+ hidden: Optional[List[int]] = None,
51
+ dropout: Optional[List[float]] = None,
52
+ activation: Optional[torch.nn.Module] = torch.nn.Sigmoid(),
53
+ bias: bool = True,
54
+ ):
55
+ if dropout is not None:
56
+ if hidden is not None:
57
+ assert len(hidden) == len(
58
+ dropout
59
+ ), "hidden and dropout must have the same length"
60
+ else:
61
+ raise ValueError(
62
+ "hidden must have a value and have the same length as dropout if dropout is given."
63
+ )
64
+
65
+ d_model = in_features
66
+ layers = []
67
+
68
+ if hidden is not None:
69
+ for i, h in enumerate(hidden):
70
+ seq = [torch.nn.Linear(d_model, h, bias=bias)]
71
+ d_model = h
72
+
73
+ if activation is not None:
74
+ seq.append(activation)
75
+
76
+ if dropout is not None:
77
+ seq.append(torch.nn.Dropout(dropout[i]))
78
+
79
+ layers.append(torch.nn.Sequential(*seq))
80
+
81
+ layers.append(torch.nn.Linear(d_model, out_features))
82
+
83
+ super(MLP, self).__init__(*layers)
84
+
85
+ class MaskedLinear(torch.nn.Linear):
86
+ """
87
+ Linear layer to be applied tile wise.
88
+ This layer can be used in combination with a mask
89
+ to prevent padding tiles from influencing the values of a subsequent
90
+ activation.
91
+ Example:
92
+ >>> module = Linear(in_features=128, out_features=1) # With Linear
93
+ >>> out = module(slide)
94
+ >>> wrong_value = torch.sigmoid(out) # Value is influenced by padding
95
+ >>> module = MaskedLinear(in_features=128, out_features=1, mask_value='-inf') # With MaskedLinear
96
+ >>> out = module(slide, mask) # Padding now has the '-inf' value
97
+ >>> correct_value = torch.sigmoid(out) # Value is not influenced by padding as sigmoid('-inf') = 0
98
+ Parameters
99
+ ----------
100
+ in_features: int
101
+ size of each input sample
102
+ out_features: int
103
+ size of each output sample
104
+ mask_value: Union[str, int]
105
+ value to give to the mask
106
+ bias: bool = True
107
+ If set to ``False``, the layer will not learn an additive bias.
108
+ """
109
+
110
+ def __init__(
111
+ self,
112
+ in_features: int,
113
+ out_features: int,
114
+ mask_value: Union[str, float],
115
+ bias: bool = True,
116
+ ):
117
+ super(MaskedLinear, self).__init__(
118
+ in_features=in_features, out_features=out_features, bias=bias
119
+ )
120
+ self.mask_value = mask_value
121
+
122
+ def forward(
123
+ self, x: torch.Tensor, mask: Optional[torch.BoolTensor] = None
124
+ ): # pylint: disable=arguments-renamed
125
+ """Forward pass.
126
+
127
+ Parameters
128
+ ----------
129
+ x: torch.Tensor
130
+ Input tensor, shape (B, SEQ_LEN, IN_FEATURES).
131
+ mask: Optional[torch.BoolTensor] = None
132
+ True for values that were padded, shape (B, SEQ_LEN, 1),
133
+
134
+ Returns
135
+ -------
136
+ x: torch.Tensor
137
+ (B, SEQ_LEN, OUT_FEATURES)
138
+ """
139
+ x = super(MaskedLinear, self).forward(x)
140
+ if mask is not None:
141
+ x = x.masked_fill(mask, float(self.mask_value))
142
+ return x
143
+
144
+ def extra_repr(self):
145
+ return (
146
+ f"in_features={self.in_features}, out_features={self.out_features}, "
147
+ f"mask_value={self.mask_value}, bias={self.bias is not None}"
148
+ )
149
+
150
+ class TilesMLP(torch.nn.Module):
151
+ """MLP to be applied to tiles to compute scores.
152
+ This module can be used in combination of a mask
153
+ to prevent padding from influencing the scores values.
154
+ Parameters
155
+ ----------
156
+ in_features: int
157
+ size of each input sample
158
+ out_features: int
159
+ size of each output sample
160
+ hidden: Optional[List[int]] = None
161
+ Number of hidden layers and their respective number of features.
162
+ bias: bool = True
163
+ If set to ``False``, the layer will not learn an additive bias.
164
+ activation: torch.nn.Module = torch.nn.Sigmoid()
165
+ MLP activation function
166
+ dropout: Optional[torch.nn.Module] = None
167
+ Optional dropout module. Will be interlaced with the linear layers.
168
+ """
169
+
170
+ def __init__(
171
+ self,
172
+ in_features: int,
173
+ out_features: int = 1,
174
+ hidden: Optional[List[int]] = None,
175
+ bias: bool = True,
176
+ activation: torch.nn.Module = torch.nn.Sigmoid(),
177
+ dropout: Optional[torch.nn.Module] = None,
178
+ ):
179
+ super(TilesMLP, self).__init__()
180
+
181
+ self.hidden_layers = torch.nn.ModuleList()
182
+ if hidden is not None:
183
+ for h in hidden:
184
+ self.hidden_layers.append(
185
+ MaskedLinear(in_features, h, bias=bias, mask_value="-inf")
186
+ )
187
+ self.hidden_layers.append(activation)
188
+ if dropout:
189
+ self.hidden_layers.append(dropout)
190
+ in_features = h
191
+
192
+ self.hidden_layers.append(
193
+ torch.nn.Linear(in_features, out_features, bias=bias)
194
+ )
195
+
196
+ def forward(
197
+ self, x: torch.Tensor, mask: Optional[torch.BoolTensor] = None
198
+ ):
199
+ """Forward pass.
200
+
201
+ Parameters
202
+ ----------
203
+ x: torch.Tensor
204
+ (B, N_TILES, IN_FEATURES)
205
+ mask: Optional[torch.BoolTensor] = None
206
+ (B, N_TILES), True for values that were padded.
207
+
208
+ Returns
209
+ -------
210
+ x: torch.Tensor
211
+ (B, N_TILES, OUT_FEATURES)
212
+ """
213
+ for layer in self.hidden_layers:
214
+ if isinstance(layer, MaskedLinear):
215
+ x = layer(x, mask)
216
+ else:
217
+ x = layer(x)
218
+ return x
219
+
220
+ class ExtremeLayer(torch.nn.Module):
221
+ """Extreme layer.
222
+ Returns concatenation of n_top top tiles and n_bottom bottom tiles
223
+ .. warning::
224
+ If top tiles or bottom tiles is superior to the true number of
225
+ tiles in the input then padded tiles will be selected and their value
226
+ will be 0.
227
+ Parameters
228
+ ----------
229
+ n_top: Optional[int] = None
230
+ Number of top tiles to select
231
+ n_bottom: Optional[int] = None
232
+ Number of bottom tiles to select
233
+ dim: int = 1
234
+ Dimension to select top/bottom tiles from
235
+ return_indices: bool = False
236
+ Whether to return the indices of the extreme tiles
237
+
238
+ Raises
239
+ ------
240
+ ValueError
241
+ If ``n_top`` and ``n_bottom`` are set to ``None`` or both are 0.
242
+ """
243
+
244
+ def __init__(
245
+ self,
246
+ n_top: Optional[int] = None,
247
+ n_bottom: Optional[int] = None,
248
+ dim: int = 1,
249
+ return_indices: bool = False,
250
+ ):
251
+ super(ExtremeLayer, self).__init__()
252
+
253
+ if not (n_top is not None or n_bottom is not None):
254
+ raise ValueError("one of n_top or n_bottom must have a value.")
255
+
256
+ if not (
257
+ (n_top is not None and n_top > 0)
258
+ or (n_bottom is not None and n_bottom > 0)
259
+ ):
260
+ raise ValueError("one of n_top or n_bottom must have a value > 0.")
261
+
262
+ self.n_top = n_top
263
+ self.n_bottom = n_bottom
264
+ self.dim = dim
265
+ self.return_indices = return_indices
266
+
267
+ def forward(
268
+ self, x: torch.Tensor, mask: Optional[torch.BoolTensor] = None
269
+ ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
270
+ """Forward pass.
271
+ Parameters
272
+ ----------
273
+ x: torch.Tensor
274
+ Input tensor, shape (B, N_TILES, IN_FEATURES).
275
+ mask: Optional[torch.BoolTensor]
276
+ True for values that were padded, shape (B, N_TILES, 1).
277
+
278
+ Warnings
279
+ --------
280
+ If top tiles or bottom tiles is superior to the true number of tiles in
281
+ the input then padded tiles will be selected and their value will be 0.
282
+
283
+ Returns
284
+ -------
285
+ values: torch.Tensor
286
+ Extreme tiles, shape (B, N_TOP + N_BOTTOM).
287
+ indices: torch.Tensor
288
+ If ``self.return_indices=True``, return extreme tiles' indices.
289
+ """
290
+
291
+ if (
292
+ self.n_top
293
+ and self.n_bottom
294
+ and ((self.n_top + self.n_bottom) > x.shape[self.dim])
295
+ ):
296
+ warnings.warn(
297
+ f"Sum of tops is larger than the input tensor shape for dimension {self.dim}: "
298
+ f"{self.n_top + self.n_bottom} > {x.shape[self.dim]}. "
299
+ f"Values will appear twice (in top and in bottom)"
300
+ )
301
+
302
+ top, bottom = None, None
303
+ top_idx, bottom_idx = None, None
304
+ if mask is not None:
305
+ if self.n_top:
306
+ top, top_idx = x.masked_fill(mask, float("-inf")).topk(
307
+ k=self.n_top, sorted=True, dim=self.dim
308
+ )
309
+ top_mask = top.eq(float("-inf"))
310
+ if top_mask.any():
311
+ warnings.warn(
312
+ "The top tiles contain masked values, they will be set to zero."
313
+ )
314
+ top[top_mask] = 0
315
+
316
+ if self.n_bottom:
317
+ bottom, bottom_idx = x.masked_fill(mask, float("inf")).topk(
318
+ k=self.n_bottom, largest=False, sorted=True, dim=self.dim
319
+ )
320
+ bottom_mask = bottom.eq(float("inf"))
321
+ if bottom_mask.any():
322
+ warnings.warn(
323
+ "The bottom tiles contain masked values, they will be set to zero."
324
+ )
325
+ bottom[bottom_mask] = 0
326
+ else:
327
+ if self.n_top:
328
+ top, top_idx = x.topk(k=self.n_top, sorted=True, dim=self.dim)
329
+ if self.n_bottom:
330
+ bottom, bottom_idx = x.topk(
331
+ k=self.n_bottom, largest=False, sorted=True, dim=self.dim
332
+ )
333
+
334
+ if top is not None and bottom is not None:
335
+ values = torch.cat([top, bottom], dim=self.dim)
336
+ indices = torch.cat([top_idx, bottom_idx], dim=self.dim)
337
+ elif top is not None:
338
+ values = top
339
+ indices = top_idx
340
+ elif bottom is not None:
341
+ values = bottom
342
+ indices = bottom_idx
343
+ else:
344
+ raise ValueError
345
+
346
+ if self.return_indices:
347
+ return values, indices
348
+ else:
349
+ return values
350
+
351
+ def extra_repr(self) -> str:
352
+ """Format representation."""
353
+ return f"n_top={self.n_top}, n_bottom={self.n_bottom}"
354
+
355
+
356
+ class Chowder(nn.Module):
357
+ """Chowder MIL model (See [1]_).
358
+
359
+ Example:
360
+ >>> module = Chowder(in_features=128, out_features=1, n_top=5, n_bottom=5)
361
+ >>> logits, extreme_scores = module(slide, mask=mask)
362
+ >>> scores = module.score_model(slide, mask=mask)
363
+
364
+ Parameters
365
+ ----------
366
+ in_features: int
367
+ Features (model input) dimension.
368
+ out_features: int
369
+ Controls the number of scores and, by extension, the number of out_features.
370
+ n_top: int
371
+ Number of tiles with hightest scores that are selected and fed to the MLP.
372
+ n_bottom: int
373
+ Number of tiles with lowest scores that are selected and fed to the MLP.
374
+ tiles_mlp_hidden: Optional[List[int]] = None
375
+ Number of units for layers in the first MLP applied tile wise to compute
376
+ a score for each tiles from the tile features.
377
+ If `None`, a linear layer is used to compute tile scores.
378
+ If e.g. `[128, 64]`, the tile scores are computed with a MLP of dimension
379
+ features_dim -> 128 -> 64 -> 1.
380
+ mlp_hidden: Optional[List[int]] = None
381
+ Number of units for layers of the second MLP that combine top and bottom
382
+ scores and outputs a final prediction at the slide-level. If `None`, a
383
+ linear layer is used to compute the prediction from the extreme scores.
384
+ If e.g. `[128, 64]`, the prediction is computed
385
+ with a MLP n_top + n_bottom -> 128 -> 64 -> 1.
386
+ mlp_dropout: Optional[List[float]] = None
387
+ Dropout that is used for each layer of the MLP. If `None`, no dropout
388
+ is used.
389
+ mlp_activation: Optional[torch.nn.Module] = torch.nn.Sigmoid
390
+ Activation that is used after each layer of the MLP.
391
+ bias: bool = True
392
+ Whether to add bias for layers of the tiles MLP.
393
+
394
+ References
395
+ ----------
396
+ .. [1] Pierre Courtiol, Eric W. Tramel, Marc Sanselme, and Gilles Wainrib. Classification
397
+ and disease localization in histopathology using only global labels: A weakly-supervised
398
+ approach. CoRR, abs/1802.02212, 2018.
399
+ """
400
+
401
+ def __init__(
402
+ self,
403
+ in_features: int,
404
+ out_features: int,
405
+ n_top: Optional[int] = None,
406
+ n_bottom: Optional[int] = None,
407
+ tiles_mlp_hidden: Optional[List[int]] = None,
408
+ mlp_hidden: Optional[List[int]] = None,
409
+ mlp_dropout: Optional[List[float]] = None,
410
+ mlp_activation: Optional[torch.nn.Module] = torch.nn.Sigmoid(),
411
+ bias: bool = True,
412
+ ) -> None:
413
+ super(Chowder, self).__init__()
414
+ if n_top is None and n_bottom is None:
415
+ raise ValueError(
416
+ "At least one of `n_top` or `n_bottom` must not be None."
417
+ )
418
+
419
+ if mlp_dropout is not None:
420
+ if mlp_hidden is not None:
421
+ assert len(mlp_hidden) == len(
422
+ mlp_dropout
423
+ ), "mlp_hidden and mlp_dropout must have the same length"
424
+ else:
425
+ raise ValueError(
426
+ "mlp_hidden must have a value and have the same length as mlp_dropout if mlp_dropout is given."
427
+ )
428
+
429
+ self.score_model = TilesMLP(
430
+ in_features,
431
+ hidden=tiles_mlp_hidden,
432
+ bias=bias,
433
+ out_features=out_features,
434
+ )
435
+ self.score_model.apply(self.weight_initialization)
436
+
437
+ self.extreme_layer = ExtremeLayer(n_top=n_top, n_bottom=n_bottom)
438
+
439
+ mlp_in_features = n_top + n_bottom
440
+ self.mlp = MLP(
441
+ mlp_in_features,
442
+ 1,
443
+ hidden=mlp_hidden,
444
+ dropout=mlp_dropout,
445
+ activation=mlp_activation,
446
+ )
447
+ self.mlp.apply(self.weight_initialization)
448
+
449
+ @staticmethod
450
+ def weight_initialization(module: torch.nn.Module) -> None:
451
+ """Initialize weights for the module using Xavier initialization method,
452
+ "Understanding the difficulty of training deep feedforward neural networks",
453
+ Glorot, X. & Bengio, Y. (2010)."""
454
+ if isinstance(module, torch.nn.Linear):
455
+ torch.nn.init.xavier_uniform_(module.weight)
456
+
457
+ if module.bias is not None:
458
+ module.bias.data.fill_(0.0)
459
+
460
+ def forward(
461
+ self, features: torch.Tensor, mask: Optional[torch.BoolTensor] = None
462
+ ) -> torch.Tensor:
463
+ """
464
+ Parameters
465
+ ----------
466
+ features: torch.Tensor
467
+ (B, N_TILES, IN_FEATURES)
468
+ mask: Optional[torch.BoolTensor] = None
469
+ (B, N_TILES, 1), True for values that were padded.
470
+
471
+ Returns
472
+ -------
473
+ logits, extreme_scores: Tuple[torch.Tensor, torch.Tensor]:
474
+ (B, OUT_FEATURES), (B, N_TOP + N_BOTTOM, OUT_FEATURES)
475
+ """
476
+ scores = self.score_model(x=features[..., 3:], mask=mask)
477
+ extreme_scores = self.extreme_layer(
478
+ x=scores, mask=mask
479
+ ) # (B, N_TOP + N_BOTTOM, OUT_FEATURES)
480
+
481
+ # Apply MLP to the N_TOP + N_BOTTOM scores.
482
+ y = self.mlp(extreme_scores.transpose(1, 2)) # (B, OUT_FEATURES, 1)
483
+
484
+ return y.squeeze(2)
trainer.py ADDED
@@ -0,0 +1,440 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Owkin, Inc.
2
+ # All rights reserved.
3
+ #
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ # Copyright (c) Owkin, Inc.
8
+ # All rights reserved.
9
+ #
10
+ # This source code is licensed under the license found in the
11
+ # LICENSE file in the root directory of this source tree.
12
+
13
+ import pickle
14
+ from pathlib import Path
15
+ from typing import Callable, Dict, List, Optional, Tuple, Union
16
+
17
+ import numpy as np
18
+ import torch
19
+ from torch import nn
20
+ from torch.optim import Adam
21
+ from torch.utils.data import DataLoader, Subset
22
+
23
+
24
+ def slide_level_train_step(
25
+ model: torch.nn.Module,
26
+ train_dataloader: torch.utils.data.DataLoader,
27
+ criterion: torch.nn.Module,
28
+ optimizer: torch.optim.Optimizer,
29
+ device: str = "cpu",
30
+ ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
31
+ """Training step for slide-level experiments. This will serve as the
32
+ ``train_step`` in ``TorchTrainer``printclass.
33
+
34
+ Parameters
35
+ ----------
36
+ model: nn.Module
37
+ The PyTorch model to be trained.
38
+ train_dataloader: torch.utils.data.DataLoader
39
+ Training data loader.
40
+ criterion: nn.Module
41
+ The loss criterion used for training.
42
+ optimizer: Callable = Adam
43
+ The optimizer class to use.
44
+ device : str = "cpu"
45
+ The device to use for training and evaluation.
46
+ """
47
+ model.train()
48
+
49
+ _epoch_loss, _epoch_logits, _epoch_labels = [], [], []
50
+
51
+ for batch in train_dataloader:
52
+ # Get data.
53
+ features, mask, labels = batch
54
+
55
+ # Put on device.
56
+ features = features.to(device)
57
+ mask = mask.to(device)
58
+ labels = labels.to(device)
59
+
60
+ # Compute logits and loss.
61
+ logits = model(features, mask)
62
+ loss = criterion(logits, labels)
63
+ loss.backward()
64
+ optimizer.step()
65
+ optimizer.zero_grad()
66
+
67
+ # Stack logits & labels to compute epoch metrics.
68
+ _epoch_loss.append(loss.detach().cpu().numpy())
69
+ _epoch_logits.append(logits.detach())
70
+ _epoch_labels.append(labels.detach())
71
+
72
+ _epoch_loss = np.mean(_epoch_loss)
73
+ _epoch_logits = torch.cat(_epoch_logits, dim=0).cpu().numpy()
74
+ _epoch_labels = torch.cat(_epoch_labels, dim=0).cpu().numpy()
75
+
76
+ return _epoch_loss, _epoch_logits, _epoch_labels
77
+
78
+
79
+ def slide_level_val_step(
80
+ model: torch.nn.Module,
81
+ val_dataloader: torch.utils.data.DataLoader,
82
+ criterion: torch.nn.Module,
83
+ device: str,
84
+ ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
85
+ """Inference step for slide-level experiments. This will serve as the
86
+ ``val_step`` in ``TorchTrainer``class.
87
+
88
+ Parameters
89
+ ----------
90
+ model: nn.Module
91
+ The PyTorch model to be trained.
92
+ val_dataloader: torch.utils.data.DataLoader
93
+ Inference data loader.
94
+ criterion: nn.Module
95
+ The loss criterion used for training.
96
+ device : str = "cpu"
97
+ The device to use for training and evaluation.
98
+ """
99
+ model.eval()
100
+
101
+ with torch.no_grad():
102
+ _epoch_loss, _epoch_logits, _epoch_labels = [], [], []
103
+
104
+ for batch in val_dataloader:
105
+ # Get data.
106
+ features, mask, labels = batch
107
+
108
+ # Put on device.
109
+ features = features.to(device)
110
+ mask = mask.to(device)
111
+ labels = labels.to(device)
112
+
113
+ # Compute logits and loss.
114
+ logits = model(features, mask)
115
+ loss = criterion(logits, labels)
116
+
117
+ # Stack logits & labels to compute epoch metrics.
118
+ _epoch_loss.append(loss.detach().cpu().numpy())
119
+ _epoch_logits.append(logits.detach())
120
+ _epoch_labels.append(labels.detach())
121
+
122
+ _epoch_loss = np.mean(_epoch_loss)
123
+ _epoch_logits = torch.cat(_epoch_logits, dim=0).cpu().numpy()
124
+ _epoch_labels = torch.cat(_epoch_labels, dim=0).cpu().numpy()
125
+
126
+ return _epoch_loss, _epoch_logits, _epoch_labels
127
+
128
+
129
+ class BaseTrainer:
130
+ """Base trainer class with ``train``, ``evaluate``, ``save`` and ``load``
131
+ methods. ``train`` and ``evaluate`` methods should be overriden."""
132
+
133
+ def __init__(self):
134
+ pass
135
+
136
+ def train(
137
+ self,
138
+ train_set: Subset,
139
+ val_set: Subset,
140
+ ) -> Tuple[Dict[str, float], Dict[str, float]]:
141
+ """Training function."""
142
+ raise NotImplementedError
143
+
144
+ def evaluate(
145
+ self,
146
+ test_set: Subset,
147
+ ) -> Dict[str, float]:
148
+ """Inference function."""
149
+ raise NotImplementedError
150
+
151
+ def save(self, filepath: Union[Path, str]):
152
+ """Model serialization."""
153
+ filepath = Path(filepath).with_suffix(".pkl")
154
+ with filepath.open("wb") as p:
155
+ pickle.dump(self, p)
156
+
157
+ @classmethod
158
+ def load(cls, filepath: Union[Path, str]):
159
+ """Model loading."""
160
+ del cls
161
+ filepath = Path(filepath).with_suffix(".pkl")
162
+ with filepath.open("rb") as p:
163
+ obj = pickle.load(p)
164
+ return obj
165
+
166
+
167
+ class TorchTrainer(BaseTrainer):
168
+ """Trainer class for training and evaluating PyTorch models.
169
+
170
+ Parameters
171
+ ----------
172
+ model: nn.Module
173
+ The PyTorch model to be trained.
174
+ criterion: nn.Module
175
+ The loss criterion used for training.
176
+ metrics: Dict[str, Callable]
177
+ Dictionary of metrics functions to evaluate the model's performance.
178
+ batch_size: int = 16
179
+ The batch size for training and evaluation
180
+ num_epochs : int = 10
181
+ The number of training epochs.
182
+ learning_rate: float = 1.0e-3
183
+ The learning rate for the optimizer.
184
+ weight_decay: float = 0.0
185
+ The weight decay for the optimizer.
186
+ device : str = "cpu"
187
+ The device to use for training and evaluation.
188
+ optimizer: Callable = Adam
189
+ The optimizer class to use.
190
+ train_step: Callable = slide_level_train_step
191
+ The function for training step.
192
+ val_step: Callable = slide_level_val_step
193
+ The function for validation step.
194
+ collator: Optional[Callable] = None
195
+ The collator function for data preprocessing.
196
+ """
197
+
198
+ def __init__(
199
+ self,
200
+ model: nn.Module,
201
+ criterion: nn.Module,
202
+ metrics: Dict[str, Callable],
203
+ batch_size: int = 16,
204
+ num_epochs: int = 10,
205
+ learning_rate: float = 1.0e-3,
206
+ weight_decay: float = 0.0,
207
+ device: str = "cpu",
208
+ optimizer: Callable = Adam,
209
+ train_step: Callable = slide_level_train_step,
210
+ val_step: Callable = slide_level_val_step,
211
+ collator: Optional[Callable] = None,
212
+ ):
213
+ super().__init__()
214
+ self.model = model
215
+ self.criterion = criterion
216
+ self.optimizer = optimizer
217
+ self.metrics = metrics
218
+
219
+ self.train_step = train_step
220
+ self.val_step = val_step
221
+
222
+ self.num_epochs = num_epochs
223
+ self.batch_size = batch_size
224
+ self.learning_rate = learning_rate
225
+ self.weight_decay = weight_decay
226
+
227
+ self.collator = collator
228
+ self.device = device
229
+
230
+ self.train_losses: List[float]
231
+ self.val_losses: List[float]
232
+ self.train_metrics: Dict[str, List[float]]
233
+ self.val_metrics: Dict[str, List[float]]
234
+
235
+ def train(
236
+ self,
237
+ train_set: Subset,
238
+ val_set: Subset,
239
+ ) -> Tuple[Dict[str, List[float]], Dict[str, List[float]]]:
240
+ """
241
+ Train the model using the provided training and validation datasets.
242
+
243
+ Parameters
244
+ ----------
245
+ train_set: Subset
246
+ The training dataset.
247
+ val_set: Subset
248
+ The validation dataset.
249
+
250
+ Returns
251
+ -------
252
+ Tuple[Dict[str, List[float]], Dict[str, List[float]]]
253
+ 2 dictionaries containing the training and validation metrics for each epoch.
254
+ """
255
+ # Dataloaders.
256
+ train_dataloader = DataLoader(
257
+ dataset=train_set,
258
+ shuffle=True,
259
+ batch_size=self.batch_size,
260
+ pin_memory=True,
261
+ collate_fn=self.collator,
262
+ drop_last=True,
263
+ )
264
+ val_dataloader = DataLoader(
265
+ dataset=val_set,
266
+ shuffle=False,
267
+ batch_size=self.batch_size,
268
+ pin_memory=True,
269
+ collate_fn=self.collator,
270
+ drop_last=False,
271
+ )
272
+
273
+ # Prepare modules.
274
+ model = self.model.to(self.device)
275
+ criterion = self.criterion.to(self.device)
276
+ optimizer = self.optimizer(
277
+ params=model.parameters(),
278
+ lr=self.learning_rate,
279
+ weight_decay=self.weight_decay,
280
+ )
281
+
282
+ # Training.
283
+ train_losses, val_losses = [], []
284
+ train_metrics: Dict[str, List[float]] = {
285
+ k: [] for k in self.metrics.keys()
286
+ }
287
+ val_metrics: Dict[str, List[float]] = {
288
+ k: [] for k in self.metrics.keys()
289
+ }
290
+ for ep in range(self.num_epochs):
291
+ # Train step.
292
+ (
293
+ train_epoch_loss,
294
+ train_epoch_logits,
295
+ train_epoch_labels,
296
+ ) = self.train_step(
297
+ model=model,
298
+ train_dataloader=train_dataloader,
299
+ criterion=criterion,
300
+ optimizer=optimizer,
301
+ device=self.device,
302
+ )
303
+
304
+ # Inference step.
305
+ val_epoch_loss, val_epoch_logits, val_epoch_labels = self.val_step(
306
+ model=model,
307
+ val_dataloader=val_dataloader,
308
+ criterion=criterion,
309
+ device=self.device,
310
+ )
311
+
312
+ # Compute metrics.
313
+ for k, m in self.metrics.items():
314
+ train_metric = m(train_epoch_labels, train_epoch_logits)
315
+ val_metric = m(val_epoch_labels, val_epoch_logits)
316
+
317
+ train_metrics[k].append(train_metric)
318
+ val_metrics[k].append(val_metric)
319
+
320
+ print(
321
+ f"Epoch {ep+1}: train_loss={train_epoch_loss:.5f}, train_{k}={train_metric:.4f}, val_loss={val_epoch_loss:.5f}, val_{k}={val_metric:.4f}"
322
+ )
323
+
324
+ train_losses.append(train_epoch_loss)
325
+ val_losses.append(val_epoch_loss)
326
+
327
+ self.train_losses = train_losses
328
+ self.val_losses = val_losses
329
+ self.train_metrics = train_metrics
330
+ self.val_metrics = val_metrics
331
+
332
+ return train_metrics, val_metrics
333
+
334
+ def evaluate(
335
+ self,
336
+ test_set: Subset,
337
+ ) -> Dict[str, float]:
338
+ """Evaluate the model using the provided test dataset.
339
+
340
+ Parameters
341
+ ----------
342
+ test_set: Subset
343
+ The test dataset.
344
+
345
+ Returns
346
+ -------
347
+ Dict[str, float]
348
+ A dictionary containing the test metrics.
349
+ """
350
+ # Dataloader.
351
+ test_dataloader = DataLoader(
352
+ dataset=test_set,
353
+ shuffle=False,
354
+ batch_size=self.batch_size,
355
+ pin_memory=True,
356
+ collate_fn=self.collator,
357
+ drop_last=False,
358
+ )
359
+
360
+ # Prepare modules.
361
+ model = self.model.to(self.device)
362
+ criterion = self.criterion.to(self.device)
363
+
364
+ # Inference step.
365
+ _, test_epoch_logits, test_epoch_labels = self.val_step(
366
+ model=model,
367
+ val_dataloader=test_dataloader,
368
+ criterion=criterion,
369
+ device=self.device,
370
+ )
371
+
372
+ # Compute metrics.
373
+ test_metrics = {
374
+ k: m(test_epoch_labels, test_epoch_logits)
375
+ for k, m in self.metrics.items()
376
+ }
377
+
378
+ return test_metrics
379
+
380
+ def predict(
381
+ self,
382
+ test_set: Subset,
383
+ ) -> Tuple[np.array, np.array]:
384
+ """Make predictions using the provided test dataset.
385
+
386
+ Parameters
387
+ ----------
388
+ test_set: Subset
389
+ The test dataset.
390
+
391
+ Returns
392
+ --------
393
+ Tuple[np.array, np.array]
394
+ A tuple containing the test labels and logits.
395
+ """
396
+ # Dataloader
397
+ test_dataloader = DataLoader(
398
+ dataset=test_set,
399
+ shuffle=False,
400
+ batch_size=self.batch_size,
401
+ pin_memory=True,
402
+ collate_fn=self.collator,
403
+ drop_last=False,
404
+ )
405
+
406
+ # Prepare modules
407
+ model = self.model.to(self.device)
408
+ criterion = self.criterion.to(self.device)
409
+
410
+ # Val step
411
+ _, test_epoch_logits, test_epoch_labels = self.val_step(
412
+ model=model,
413
+ val_dataloader=test_dataloader,
414
+ criterion=criterion,
415
+ device=self.device,
416
+ )
417
+
418
+ return test_epoch_labels, test_epoch_logits
419
+
420
+ def compute_metrics(
421
+ self, labels: np.array, logits: np.array
422
+ ) -> Dict[str, float]:
423
+ """Compute metrics using the provided labels and logits.
424
+
425
+ Parameters
426
+ ----------
427
+ labels: np.ndarray
428
+ The ground truth labels.
429
+ logits: np.ndarray
430
+ The predicted logits.
431
+
432
+ Returns:
433
+ Dict[str, float]
434
+ A dictionary containing the computed metrics.
435
+ """
436
+ test_metrics = {
437
+ k: metric(labels, logits) for k, metric in self.metrics.items()
438
+ }
439
+ return test_metrics
440
+
utils.py ADDED
@@ -0,0 +1,123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Owkin, Inc.
2
+ # All rights reserved.
3
+ #
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ # Copyright (c) Owkin, Inc.
8
+ # All rights reserved.
9
+ #
10
+ # This source code is licensed under the license found in the
11
+ # LICENSE file in the root directory of this source tree.
12
+
13
+
14
+ import pickle
15
+ from pathlib import Path
16
+ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
17
+
18
+ import datasets
19
+ import numpy as np
20
+ import torch
21
+ from sklearn.metrics import roc_auc_score
22
+ from torch.utils.data.dataloader import default_collate
23
+
24
+
25
+ def pad_collate_fn(
26
+ batch: List[Tuple[torch.Tensor, Any]],
27
+ batch_first: bool = True,
28
+ max_len: Optional[int] = None,
29
+ ) -> Tuple[torch.Tensor, torch.BoolTensor, Any]:
30
+ """Pad together sequences of arbitrary lengths.
31
+ Add a mask of the padding to the samples that can later be used
32
+ to ignore padding in activation functions.
33
+
34
+ Expected to be used in combination of a torch.utils.datasets.DataLoader.
35
+
36
+ Expect the sequences to be padded to be the first one in the sample tuples.
37
+ Others members will be batched using ``torch.utils.data.dataloader.default_collate``.
38
+
39
+ Parameters
40
+ ----------
41
+ batch: List[Tuple[torch.Tensor, Any]]
42
+ List of tuples (features, Any). Features have shape (N_slides_tiles, F)
43
+ with ``N_slides_tiles`` being specific to each slide depending on the
44
+ number of extractable tiles in the tissue matter. ``F`` is the feature
45
+ extractor output dimension.
46
+ batch_first: bool = True
47
+ Either return (B, N_TILES, F) or (N_TILES, B, F)
48
+ max_len: Optional[int] = None
49
+ Pre-defined maximum length for elements inside a batch.
50
+
51
+ Returns
52
+ -------
53
+ padded_sequences, masks, Any: Tuple[torch.Tensor, torch.BoolTensor, Any]
54
+ - if batch_first: Tuple[(B, N_TILES, F), (B, N_TILES, 1), ...]
55
+ - else: Tuple[(N_TILES, B, F), (N_TILES, B, 1), ...]
56
+
57
+ with N_TILES = max_len if max_len is not None
58
+ or N_TILES = max length of the training samples.
59
+
60
+ """
61
+ # Expect the sequences to be the first one in the sample tuples
62
+ sequences = []
63
+ others = []
64
+ for sample in batch:
65
+ sequences.append(sample[0])
66
+ others.append(sample[1:])
67
+
68
+ if max_len is None:
69
+ max_len = max([s.size(0) for s in sequences])
70
+
71
+ trailing_dims = sequences[0].size()[1:]
72
+
73
+ if batch_first:
74
+ padded_dims = (len(sequences), max_len) + trailing_dims
75
+ masks_dims = (len(sequences), max_len, 1)
76
+ else:
77
+ padded_dims = (max_len, len(sequences)) + trailing_dims
78
+ masks_dims = (max_len, len(sequences), 1)
79
+
80
+ padded_sequences = sequences[0].data.new(*padded_dims).fill_(0.0)
81
+ masks = torch.ones(*masks_dims, dtype=torch.bool)
82
+
83
+ for i, tensor in enumerate(sequences):
84
+ length = tensor.size(0)
85
+ # use index notation to prevent duplicate references to the tensor
86
+ if batch_first:
87
+ padded_sequences[i, :length, ...] = tensor[:max_len, ...]
88
+ masks[i, :length, ...] = False
89
+ else:
90
+ padded_sequences[:length, i, ...] = tensor[:max_len, ...]
91
+ masks[:length, i, ...] = False
92
+
93
+ # Batching other members of the tuple using default_collate
94
+ others = default_collate(others)
95
+
96
+ return (padded_sequences, masks, *others)
97
+
98
+
99
+ def auc(labels: np.array, logits: np.array) -> float:
100
+ """ROC AUC score for binary classification.
101
+ Parameters
102
+ ----------
103
+ labels: np.array
104
+ Labels of the outcome.
105
+ logits: np.array
106
+ Probabilities.
107
+ """
108
+ preds = 1.0 / (1.0 + np.exp(-logits))
109
+ return roc_auc_score(labels, preds)
110
+
111
+
112
+ def get_cv_metrics(
113
+ cv_metrics: List[Dict[str, float]], epoch: int = -1
114
+ ) -> Dict[str, float]:
115
+ """Get mean and std from cross-validation metrics at a given epoch."""
116
+ cv_mean_metrics = {}
117
+ metrics_names = cv_metrics[0].keys()
118
+ for m_name in metrics_names:
119
+ values = [fold_metrics[m_name][epoch] for fold_metrics in cv_metrics]
120
+ mean_metric, std_metric = np.mean(values), np.std(values)
121
+ cv_mean_metrics[m_name] = f"{mean_metric:.4f} ± {std_metric:.4f}"
122
+ return cv_mean_metrics
123
+