adapt params to ijepa paper

configs/wilds_vitb16-448_ep600.yaml

@@ -1,5 +1,5 @@
 data:
-  batch_size: 16
+  batch_size: 256
   color_jitter_strength: 0.0
   crop_scale:
     - 1.0
@@ -13,7 +13,7 @@ data:
   use_horizontal_flip: false
   use_random_resized_crop: false
 logging:
-  folder: experiment_logs/vitb16.448-bs.16-ep.600
+  folder: experiment_logs/vitb16.448-bs.16-ep.600-paper
   write_tag: jepa
 mask:
   allow_overlap: false
@@ -48,6 +48,7 @@ optimization:
   final_weight_decay: 0.4
   ipe_scale: 1.0
   lr: 0.001
-  start_lr: 0.0002
-  warmup: 40
+  start_lr: 0.0001
+  warmup: 15
+  wd_schedule: linear
   weight_decay: 0.04

configs/wilds_vith14_ep300.yaml

@@ -1,5 +1,5 @@
 data:
-  batch_size: 128
+  batch_size: 256
   color_jitter_strength: 0.0
   crop_scale:
     - 0.3
@@ -46,6 +46,7 @@ optimization:
   final_weight_decay: 0.4
   ipe_scale: 1.0
   lr: 0.001
-  start_lr: 0.0002
-  warmup: 40
+  start_lr: 0.0001
+  warmup: 15
+  wd_schedule: linear
   weight_decay: 0.04

configs/wilds_vith16-448_ep300.yaml

@@ -1,5 +1,5 @@
 data:
-  batch_size: 128
+  batch_size: 256
   color_jitter_strength: 0.0
   crop_scale:
     - 1.0
@@ -47,6 +47,7 @@ optimization:
   final_weight_decay: 0.4
   ipe_scale: 1.0
   lr: 0.001
-  start_lr: 0.0002
-  warmup: 40
+  start_lr: 0.0001
+  warmup: 15
+  wd_schedule: linear
   weight_decay: 0.04

src/helper.py

@@ -13,7 +13,8 @@ import torch
 import src.models.vision_transformer as vit
 from src.utils.schedulers import (
     WarmupCosineSchedule,
-    CosineWDSchedule)
+    CosineWDSchedule,
+    LinearWDSchedule)
 from src.utils.tensors import trunc_normal_
 
 logging.basicConfig(stream=sys.stdout, level=logging.INFO)
@@ -116,7 +117,8 @@ def init_opt(
     final_wd=1e-6,
     final_lr=0.0,
     use_bfloat16=False,
-    ipe_scale=1.25
+    ipe_scale=1.25,
+    wd_schedule='cosine'
 ):
     param_groups = [
         {
@@ -147,10 +149,20 @@ def init_opt(
         ref_lr=ref_lr,
         final_lr=final_lr,
         T_max=int(ipe_scale*num_epochs*iterations_per_epoch))
-    wd_scheduler = CosineWDSchedule(
-        optimizer,
-        ref_wd=wd,
-        final_wd=final_wd,
-        T_max=int(ipe_scale*num_epochs*iterations_per_epoch))
+    wd_schedule = str(wd_schedule).lower()
+    if wd_schedule == 'linear':
+        logger.info('Using linear weight-decay schedule')
+        wd_scheduler = LinearWDSchedule(
+            optimizer,
+            ref_wd=wd,
+            final_wd=final_wd,
+            T_max=int(ipe_scale*num_epochs*iterations_per_epoch))
+    else:
+        logger.info('Using cosine weight-decay schedule')
+        wd_scheduler = CosineWDSchedule(
+            optimizer,
+            ref_wd=wd,
+            final_wd=final_wd,
+            T_max=int(ipe_scale*num_epochs*iterations_per_epoch))
     scaler = torch.cuda.amp.GradScaler() if use_bfloat16 else None
     return optimizer, scaler, scheduler, wd_scheduler

src/train.py

@@ -107,6 +107,7 @@ def main(args, resume_preempt=False):
     ipe_scale = args["optimization"]["ipe_scale"]  # scheduler scale factor (def: 1.0)
     wd = float(args["optimization"]["weight_decay"])
     final_wd = float(args["optimization"]["final_weight_decay"])
+    wd_schedule = args["optimization"].get("wd_schedule", "cosine")
     num_epochs = args["optimization"]["epochs"]
     warmup = args["optimization"]["warmup"]
     start_lr = args["optimization"]["start_lr"]
@@ -214,6 +215,7 @@ def main(args, resume_preempt=False):
         num_epochs=num_epochs,
         ipe_scale=ipe_scale,
         use_bfloat16=use_bfloat16,
+        wd_schedule=wd_schedule,
     )
     if dist.is_available() and dist.is_initialized() and world_size > 1:
         encoder = DistributedDataParallel(encoder, static_graph=True)

src/utils/schedulers.py

@@ -74,3 +74,30 @@ class CosineWDSchedule(object):
             if ('WD_exclude' not in group) or not group['WD_exclude']:
                 group['weight_decay'] = new_wd
         return new_wd
+
+
+class LinearWDSchedule(object):
+
+    def __init__(
+        self,
+        optimizer,
+        ref_wd,
+        T_max,
+        final_wd=0.
+    ):
+        self.optimizer = optimizer
+        self.ref_wd = ref_wd
+        self.final_wd = final_wd
+        self.T_max = T_max
+        self._step = 0.
+
+    def step(self):
+        self._step += 1
+        progress = self._step / self.T_max
+        progress = min(max(progress, 0.0), 1.0)
+        new_wd = self.ref_wd + progress * (self.final_wd - self.ref_wd)
+
+        for group in self.optimizer.param_groups:
+            if ('WD_exclude' not in group) or not group['WD_exclude']:
+                group['weight_decay'] = new_wd
+        return new_wd