Update logit scaling method (#29)

mmlearn/modules/layers/logit_scaling.py

@@ -22,15 +22,15 @@ class LearnableLogitScaling(torch.nn.Module):
 
     def __init__(
         self,
-        logit_scale_init: float = 1 / 0.07,
-        learnable: bool = True,
+        init_logit_scale: float = 1 / 0.07,
         max_logit_scale: float = 100,
+        learnable: bool = True,
     ) -> None:
         super().__init__()
         self.max_logit_scale = max_logit_scale
-        self.logit_scale_init = logit_scale_init
+        self.init_logit_scale = init_logit_scale
         self.learnable = learnable
-        log_logit_scale = torch.ones([]) * np.log(self.logit_scale_init)
+        log_logit_scale = torch.ones([]) * np.log(self.init_logit_scale)
         if learnable:
             self.log_logit_scale = torch.nn.Parameter(log_logit_scale)
         else:
@@ -49,6 +49,6 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
     def extra_repr(self) -> str:
         """Return the string representation of the layer."""
         return (
-            f"logit_scale_init={self.logit_scale_init},learnable={self.learnable},"
+            f"logit_scale_init={self.init_logit_scale},learnable={self.learnable},"
             f" max_logit_scale={self.max_logit_scale}"
         )

mmlearn/modules/losses/contrastive.py

@@ -83,6 +83,7 @@ def _get_logits(
         self,
         features_1: torch.Tensor,
         features_2: torch.Tensor,
+        logit_scale: torch.Tensor,
         rank: int,
         world_size: int,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
@@ -93,7 +94,9 @@ def _get_logits(
         features_1 : torch.Tensor
             First feature tensor.
         features_2 : torch.Tensor
-            Second feature tensor
+            Second feature tensor.
+        logit_scale : torch.Tensor
+            Logit scale.
         rank : int
             Rank of the current process.
         world_size : int
@@ -114,19 +117,28 @@ def _get_logits(
             )
 
             if self.local_loss:
-                logits_per_feature_1 = _safe_matmul(features_1, all_features_2)
-                logits_per_feature_2 = _safe_matmul(features_2, all_features_1)
+                logits_per_feature_1 = logit_scale * _safe_matmul(
+                    features_1, all_features_2
+                )
+                logits_per_feature_2 = logit_scale * _safe_matmul(
+                    features_2, all_features_1
+                )
             else:
-                logits_per_feature_1 = _safe_matmul(all_features_1, all_features_2)
+                logits_per_feature_1 = logit_scale * _safe_matmul(
+                    all_features_1, all_features_2
+                )
                 logits_per_feature_2 = logits_per_feature_1.T
         else:
-            logits_per_feature_1 = _safe_matmul(features_1, features_2)
-            logits_per_feature_2 = _safe_matmul(features_2, features_1)
+            logits_per_feature_1 = logit_scale * _safe_matmul(features_1, features_2)
+            logits_per_feature_2 = logit_scale * _safe_matmul(features_2, features_1)
 
         return logits_per_feature_1, logits_per_feature_2
 
     def forward(
-        self, features_1: torch.Tensor, features_2: torch.Tensor
+        self,
+        features_1: torch.Tensor,
+        features_2: torch.Tensor,
+        logit_scale: torch.Tensor,
     ) -> torch.Tensor:
         """Calculate the CLIP-style loss between two sets of features.
 
@@ -136,6 +148,8 @@ def forward(
             First set of features.
         features_2 : torch.Tensor
             Second set of features.
+        logit_scale : torch.Tensor
+            Logit scale.
 
         Returns
         -------
@@ -150,7 +164,7 @@ def forward(
             features_2 = F.normalize(features_2, p=2, dim=-1)
 
         logits_per_feat1, logits_per_feat2 = self._get_logits(
-            features_1, features_2, rank=rank, world_size=world_size
+            features_1, features_2, logit_scale, rank=rank, world_size=world_size
         )
         labels = self._get_ground_truth(
             features_1.device,

mmlearn/tasks/contrastive_pretraining.py

@@ -2,11 +2,13 @@
 
 import inspect
 import itertools
+import math
 from dataclasses import dataclass
 from functools import partial
 from typing import Any, Dict, List, Literal, Mapping, Optional, Tuple, Union
 
 import lightning as L  # noqa: N812
+import numpy as np
 import torch
 import torch.distributed
 import torch.distributed.nn
@@ -151,6 +153,9 @@ def __init__(  # noqa: PLR0912, PLR0915
                 partial[torch.optim.lr_scheduler.LRScheduler],
             ]
         ] = None,
+        init_logit_scale: float = 1 / 0.07,
+        max_logit_scale: float = 100,
+        learnable_logit_scale: bool = True,
         loss: Optional[CLIPLoss] = None,
         modality_loss_pairs: Optional[List[LossPairSpec]] = None,
         auxiliary_tasks: Optional[Dict[str, AuxiliaryTaskSpec]] = None,
@@ -259,6 +264,19 @@ def __init__(  # noqa: PLR0912, PLR0915
                 }
             )
 
+        # set up logit scaling
+        log_logit_scale = torch.ones([]) * np.log(init_logit_scale)
+        self.max_logit_scale = max_logit_scale
+        self.learnable_logit_scale = learnable_logit_scale
+
+        if self.learnable_logit_scale:
+            self.log_logit_scale = torch.nn.Parameter(
+                log_logit_scale, requires_grad=True
+            )
+        else:
+            self.register_buffer("log_logit_scale", log_logit_scale)
+
+        # set up contrastive loss pairs
         if modality_loss_pairs is None:
             modality_loss_pairs = [
                 LossPairSpec(modalities=(m1.name, m2.name))
@@ -277,6 +295,7 @@ def __init__(  # noqa: PLR0912, PLR0915
                 )
         self.modality_loss_pairs = modality_loss_pairs
 
+        # set up auxiliary tasks
         self.aux_task_specs = auxiliary_tasks or {}
         self.auxiliary_tasks: Dict[str, L.LightningModule] = {}
         for task_name, task_spec in self.aux_task_specs.items():
@@ -313,10 +332,11 @@ def __init__(  # noqa: PLR0912, PLR0915
                         f"Expected {eval_task_spec.task} to be an instance of `EvaluationHooks` "
                         f"but got {type(eval_task_spec.task)}."
                     )
-
         self.evaluation_tasks = evaluation_tasks
 
-    def encode(self, inputs: Dict[str, Any], modality: Modality) -> torch.Tensor:
+    def encode(
+        self, inputs: Dict[str, Any], modality: Modality, normalize: bool = False
+    ) -> torch.Tensor:
         """Encode the input values for the given modality.
 
         Parameters
@@ -325,6 +345,9 @@ def encode(self, inputs: Dict[str, Any], modality: Modality) -> torch.Tensor:
             Input values.
         modality : Modality
             The modality to encode.
+        normalize : bool, optional, default=False
+            Whether to apply L2 normalization to the output (after the head and
+            postprocessor layers, if present).
 
         Returns
         -------
@@ -339,6 +362,9 @@ def encode(self, inputs: Dict[str, Any], modality: Modality) -> torch.Tensor:
         if self.postprocessors and modality.name in self.postprocessors:
             output = self.postprocessors[modality.name](output)
 
+        if normalize:
+            output = torch.nn.functional.normalize(output, p=2, dim=-1)
+
         return output
 
     def forward(self, inputs: Dict[str, Any]) -> Dict[str, torch.Tensor]:
@@ -355,7 +381,7 @@ def forward(self, inputs: Dict[str, Any]) -> Dict[str, torch.Tensor]:
             The encodings for each modality.
         """
         outputs = {
-            modality.embedding: self.encode(inputs, modality)
+            modality.embedding: self.encode(inputs, modality, normalize=True)
             for modality in self._available_modalities
         }
 
@@ -373,6 +399,16 @@ def _compute_loss(
         if self.loss_fn is None:
             return None
 
+        with torch.no_grad():
+            self.log_logit_scale.clamp_(0, math.log(self.max_logit_scale))
+        self.log(
+            "train/logit_scale",
+            self.log_logit_scale.exp(),
+            prog_bar=True,
+            on_step=True,
+            on_epoch=False,
+        )
+
         contrastive_losses: list[torch.Tensor] = []
         for loss_pair in self.modality_loss_pairs:
             modality_a = Modalities.get_modality(loss_pair.modalities[0])
@@ -389,6 +425,7 @@ def _compute_loss(
                 self.loss_fn(
                     outputs[modality_a.embedding][indices_a],
                     outputs[modality_b.embedding][indices_b],
+                    self.log_logit_scale.exp(),
                 )
                 * loss_pair.weight
             )

projects/bioscan_clip/configs/experiment/bioscan_1m.yaml

@@ -12,8 +12,6 @@ defaults:
   - /modules/[email protected]: timm-vit-lora
   - /modules/[email protected]: barcode-bert-lora
   - /modules/[email protected]: MLP # the other modalities have projection heads in their encoders
-  - /modules/[email protected]_and_logit_scale.norm: L2Norm
-  - /modules/[email protected]_and_logit_scale.logit_scale: LearnableLogitScaling
   - /modules/[email protected]: CLIPLoss
   - /modules/[email protected]: AdamW
   - /modules/[email protected]_scheduler.scheduler: OneCycleLR
@@ -67,19 +65,6 @@ task:
     text:
       in_dim: 512
       out_dim: ${task.encoders.rgb.projection_dim}
-  postprocessors:
-    norm_and_logit_scale:
-      norm:
-        dim: -1
-      logit_scale:
-        learnable: True
-  modality_module_mapping:
-    text:
-      postprocessor_key: norm_and_logit_scale
-    rgb:
-      postprocessor_key: norm_and_logit_scale
-    dna:
-      postprocessor_key: norm_and_logit_scale
   optimizer:
     lr: 1.0e-3
     eps: 1.0e-6

projects/med_benchmarking/configs/experiment/baseline.yaml

@@ -15,8 +15,6 @@ defaults:
   - /datasets/[email protected]_fn.batch_processors.text: HFCLIPTokenizer
   - /modules/[email protected]: HFCLIPTextEncoderWithProjection
   - /modules/[email protected]: HFCLIPVisionEncoderWithProjection
-  - /modules/[email protected]_and_logit_scale.norm: L2Norm
-  - /modules/[email protected]_and_logit_scale.logit_scale: LearnableLogitScaling
   - /modules/[email protected]: CLIPLoss
   - /modules/[email protected]: AdamW
   - /modules/[email protected]_scheduler.scheduler: CosineAnnealingLR
@@ -47,17 +45,6 @@ dataloader:
     num_workers: 4
 
 task:
-  postprocessors:
-    norm_and_logit_scale:
-      norm:
-        dim: -1
-      logit_scale:
-        learnable: True
-  modality_module_mapping:
-    text:
-      postprocessor_key: norm_and_logit_scale
-    rgb:
-      postprocessor_key: norm_and_logit_scale
   optimizer:
     betas:
     - 0.9

projects/med_benchmarking/datasets/pad_ufes_20.py

@@ -43,13 +43,13 @@ def __init__(
         self.split = split
 
         # Load cached data if available
-        cache_path = f"cache/PadUfes20_{split}.pkl"
+        cache_path = f".cache/PadUfes20_{split}.pkl"
         if os.path.exists(cache_path):
             print(f"!!! Using cached dataset for {split}")
             with open(cache_path, "rb") as f:
                 self.metadata = pickle.load(f)
         else:
-            os.makedirs("cache/", exist_ok=True)
+            os.makedirs(".cache/", exist_ok=True)
             self.metadata = self._load_and_process_metadata()
             with open(cache_path, "wb") as f:
                 pickle.dump(self.metadata.to_dict("records"), f)
@@ -68,14 +68,13 @@ def _load_and_process_metadata(self) -> pd.DataFrame:
         df["path"] = df["img_id"].apply(
             lambda imgid: os.path.join(self.root_dir, "Dataset", imgid)
         )
-        df.drop(columns=["img_id", "diagnostic"], inplace=True).reset_index(
-            drop=True, inplace=True
-        )
+        df.drop(columns=["img_id", "diagnostic"], inplace=True)
+        df.reset_index(drop=True, inplace=True)
 
         # Split into train and test
         dataset = {}
-        dataset["test"] = df.sample(frac=0.2)
-        dataset["train"] = df.drop(dataset["test"].index)
+        dataset["test"] = df.sample(frac=0.2, ignore_index=True)
+        dataset["train"] = df.drop(dataset["test"].index).reset_index(drop=True)
         return dataset[self.split]
 
     def _build_label(self, str_label: str) -> int: