Add ssl ci

configs/byol/byol_clas_r50.yaml

@@ -1,7 +1,10 @@
-epochs: 100
+epochs: 80
+use_simclr_iters: True
+global_batch_size: 1024
 output_dir: output_dir
 seed: 0
 device: gpu
+total_iters: 51
 
 model:
   name: ByolClassification
@@ -58,13 +61,9 @@ dataloader:
         - name: Clip
 
 lr_scheduler:
-  name: ByolLRScheduler
-  total_image: 1281167
-  total_batch: 1024
-  total_steps: 100
-  warmup_steps: 0
-  start_lr: 0.0001
-  end_lr: 1.6
+  name: Cosinesimclr
+  learning_rate: 0.4
+  T_max: 80
 
 optimizer:
   name: Momentum
@@ -74,5 +73,9 @@ log_config:
     name: LogHook
     interval: 10
 
+lr_config:
+  name: LRSchedulerHook
+  unit: epoch
+
 custom_config:
   - name: EvaluateHook

configs/byol/byol_r50_IM.yaml

@@ -1,10 +1,11 @@
 epochs: 300
-use_byol_iters: True
+use_simclr_iters: True
 total_images: 1281167
 global_batch_size: 4096
 output_dir: output_dir
 seed: 0
 device: gpu
+total_iters: 51
 
 model:
   name: BYOL
@@ -84,12 +85,13 @@ dataloader:
 
 
 lr_scheduler:
-  name: CosineWarmup
-  learning_rate: 4.8
-  T_max: 93835
-  warmup_steps: 3127
-  start_lr: 0.0001
+  name: simclrCosineWarmup
+  learning_rate_scaling: linear
+  total_images: 1281167
+  warmup_epochs: 10
+  start_lr: 0
   end_lr: 4.8
+  T_max: 300
 
 
 optimizer:

configs/simclr/simclr_clas_r50.yaml

@@ -0,0 +1,93 @@
+epochs: 90
+output_dir: cls_output_dir
+use_simclr_iters: True
+seed: 0
+device: gpu
+# used for static mode and model export
+image_shape: [3, 224, 224]
+save_inference_dir: ./cls_inference
+total_iters: 51
+
+model:
+  name: Classification
+  backbone:
+    name: ResNet
+    depth: 50
+    frozen_stages: 4
+  head:
+    name: ClasHead
+    with_avg_pool: true
+    in_channels: 2048
+
+dataloader:
+  train:
+    loader:
+      num_workers: 8
+      use_shared_memory: True
+    sampler:
+      batch_size: 512
+      shuffle: true
+      drop_last: true
+    dataset:
+      name: ImageNet
+      dataroot: data/ILSVRC2012/train
+      return_label: True
+      transforms:
+        - name: RandomResizedCrop
+          size: 224
+        - name: RandomHorizontalFlip
+        - name: Transpose
+        - name: NormalizeImage
+          scale: 1.0/255.0
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
+  val:
+    loader:
+      num_workers: 8
+      use_shared_memory: True
+    sampler:
+      batch_size: 512
+      shuffle: false
+      drop_last: false
+    dataset:
+      name: ImageNet
+      dataroot: data/ILSVRC2012/val
+      return_label: True
+      transforms:
+        - name: Resize
+          size: 256
+        - name: CenterCrop
+          size: 224
+        - name: Transpose
+        - name: NormalizeImage
+          scale: 1.0/255.0
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
+
+
+lr_scheduler:
+  name: Cosinesimclr
+  learning_rate: 1.6
+  T_max: 90
+
+lr_config:
+  name: LRSchedulerHook
+  unit: epoch
+
+optimizer:
+  name: LarsMomentumOptimizer
+  momentum: 0.9
+  lars_weight_decay: 0
+  exclude_from_weight_decay: []
+
+
+log_config:
+    name: LogHook
+    interval: 10
+
+custom_config:
+  - name: EvaluateHook
+
+checkpoint:
+    name: CheckpointHook
+    interval: 10

configs/simclr/simclr_r50_IM.yaml

@@ -2,6 +2,8 @@ epochs: 100
 use_simclr_iters: True
 global_batch_size: 4096
 output_dir: output_dir
+device: gpu
+total_iters: 51
 
 model:
   name: SimCLR
@@ -21,7 +23,9 @@ model:
 
 dataloader:
   train:
-    num_workers: 6
+    loader:
+      num_workers: 4
+      use_shared_memory: True
     sampler:
       batch_size: 32
       shuffle: true
@@ -55,7 +59,7 @@ dataloader:
           p: 0.5
         - name: RandomHorizontalFlip
         - name: Transpose
-        - name: NormalizeImage                                                                                                                                                                                                                                                 
+        - name: NormalizeImage
           scale: 1.0/255.0
           mean: [0.485, 0.456, 0.406]
           std: [0.229, 0.224, 0.225]
@@ -77,15 +81,17 @@ dataloader:
           p: 0.5
         - name: RandomHorizontalFlip
         - name: Transpose
-        - name: NormalizeImage                                                                                                                                                                                                                                                 
+        - name: NormalizeImage
           scale: 1.0/255.0
           mean: [0.485, 0.456, 0.406]
           std: [0.229, 0.224, 0.225]
 
   val:
-    num_workers: 4
+    loader:
+      num_workers: 4
+      use_shared_memory: True
     sampler:
-      batch_size: 512
+      batch_size: 256
       shuffle: false
       drop_last: false
     dataset:
@@ -98,29 +104,27 @@ dataloader:
         - name: CenterCrop
           size: 224
         - name: Transpose
-        - name: NormalizeImage                                                                                                                                                                                                                                                 
+        - name: NormalizeImage
           scale: 1.0/255.0
           mean: [0.485, 0.456, 0.406]
           std: [0.229, 0.224, 0.225]
 
 lr_scheduler:
   name: simclrCosineWarmup
-  learning_rate_scaling: sqrt
+  learning_rate_scaling: linear
   total_images: 1281167
   warmup_epochs: 10
   start_lr: 0
-  end_lr: 1.0
+  end_lr: 0.3
   T_max: 200
 
 
 optimizer:
   name: LarsMomentumOptimizer
   momentum: 0.9
-  lars_weight_decay: 0.0001
+  lars_weight_decay: 1e-6
   exclude_from_weight_decay: ["scale","offset",".bias"]
-  
+
 log_config:
     name: LogHook
     interval: 10
-
-

configs/simsiam/simsiam_clas_r50.yaml

@@ -4,6 +4,7 @@ global_batch_size: 4096
 output_dir: output_dir
 seed: 0
 device: gpu
+total_iters: 51
 
 # used for static mode and model export
 image_shape: [3, 224, 224]
@@ -23,7 +24,7 @@ model:
 dataloader:
   train:
     loader:
-      num_workers: 16
+      num_workers: 8
       use_shared_memory: True
     sampler:
       batch_size: 512
@@ -44,7 +45,7 @@ dataloader:
           std: [0.229, 0.224, 0.225]
   val:
     loader:
-      num_workers: 16
+      num_workers: 8
       use_shared_memory: True
     sampler:
       batch_size: 512
@@ -77,7 +78,7 @@ optimizer:
 
 log_config:
     name: LogHook
-    interval: 50
+    interval: 10
 
 lr_config:
   name: LRSchedulerHook

configs/simsiam/simsiam_r50.yaml

@@ -4,6 +4,7 @@ global_batch_size: 512
 output_dir: output_dir
 seed: 0
 device: gpu
+total_iters: 51
 
 model:
   name: SimSiam

configs/swav/swav_clas_r50.yaml

@@ -2,6 +2,7 @@ epochs: 100
 output_dir: output_dir
 seed: 0
 device: gpu
+total_iters: 51
 
 # used for static mode and model export
 image_shape: [3, 224, 224]
@@ -75,7 +76,7 @@ optimizer:
 
 log_config:
     name: LogHook
-    interval: 50
+    interval: 10
 
 custom_config:
   - name: EvaluateHook

configs/swav/swav_r50_100ep.yaml

@@ -1,7 +1,10 @@
 epochs: 100
+use_simclr_iters: True
+global_batch_size: 256
 output_dir: output_dir
 seed: 0
 device: gpu
+total_iters: 51
 
 model:
   name: SwAV
@@ -28,10 +31,10 @@ model:
 dataloader:
   train:
     loader:
-      num_workers: 16
+      num_workers: 8
       use_shared_memory: True
     sampler:
-      batch_size: 128
+      batch_size: 32
       shuffle: true
       drop_last: true
     dataset:
@@ -43,13 +46,13 @@ dataloader:
       max_scale_crops: [1., 0.14]
 
 lr_scheduler:
-  name: CosineWarmup
-  learning_rate: 4.8
-  T_max: 31200
-  warmup_steps: 3120
-  start_lr: 0.3
-  end_lr: 4.8
-  eta_min: 0.0048
+  name: simclrCosineWarmup
+  learning_rate_scaling: linear
+  total_images: 1281167
+  warmup_epochs: 10
+  start_lr: 0
+  end_lr: 0.3
+  T_max: 300
 
 optimizer:
   name: LarsMomentumOptimizer

passl_v110/datasets/builder.py

@@ -16,6 +16,7 @@
 import numpy as np
 import math
 import paddle
+import random
 from paddle.io import DistributedBatchSampler
 
 from ..utils.registry import Registry, build_from_config
@@ -92,11 +93,17 @@ def build_dataloader(cfg, device):
 
     sampler_name = sampler_cfg.pop('name', 'DistributedBatchSampler')
 
+    def worker_init_fn(worker_id):
+        """ set seed in subproces for dataloader when num_workers > 0"""
+        np.random.seed(cfg.seed + worker_id)
+        random.seed(cfg.seed + worker_id)
+
     sampler = eval("{}".format(sampler_name))(dataset, **sampler_cfg)
 
     dataloader = paddle.io.DataLoader(dataset,
                                       batch_sampler=sampler,
                                       places=device,
+                                      worker_init_fn=worker_init_fn,
                                       **loader_cfg)
 
     #setup mixup / cutmix