Merge pull request #10 from PRIS-CV/onion-release

Onion release

.gitignore

@@ -5,4 +5,5 @@ __pycache__
 .ipynb_checkpoints
 cam_images/
 .vscode
+ut.py
 

configs/api-net/api-net-resnet101.yml → configs/api_net/api_net_resnet101.yml

@@ -114,8 +114,11 @@ TRANSFORMS:
 
 OPTIMIZER:
   NAME: "SGD"
-  MOMENTUM: 0.9
+  ARGS: 
+    - momentum: 0.9
+    - weight_decay: 0.0005
   LR: 
+    base: 0.002
     backbone: 0.002
     encoder: 0.01
     necks: 0.01
@@ -124,7 +127,7 @@ OPTIMIZER:
 ITERATION_NUM: ~
 EPOCH_NUM: 100
 START_EPOCH: 0
-UPDATE_STRATEGY: "general_updating"
+UPDATE_STRATEGY: "general_strategy"
 
 
 # Validation details

configs/cal/cal_resnet101.yml

@@ -105,9 +105,11 @@ TRANSFORMS:
 
 OPTIMIZER:
   NAME: SGD
-  MOMENTUM: 0.9
-  WEIGHT_DECAY: 1e-5
+  ARGS: 
+    - momentum: 0.9
+    - weight_decay: 0.0005
   LR: 
+    base: 0.001
     backbone: 0.001
     encoder: 0.001
     necks: 0.001
@@ -116,14 +118,14 @@ OPTIMIZER:
 ITERATION_NUM: ~
 EPOCH_NUM: 160
 START_EPOCH: 0
-UPDATE_STRATEGY: "general_updating"
+UPDATE_STRATEGY: "general_strategy"
 LR_SCHEDULE:
   NAME: "adjusting_schedule"
   ARGS: 
     - base_lr: 0.001
     - base_duration: 2.0
     - base_rate: 0.9
-    - update_level: "batch_update"
+    # - update_level: "batch_update"
 
 # Validation details
 PER_ITERATION: ~

configs/mutual_channel_loss/mcl_vgg16.yml

@@ -102,8 +102,11 @@ TRANSFORMS:
 
 OPTIMIZER:
   NAME: SGD
-  MOMENTUM: 0.9
+  ARGS: 
+    - momentum: 0.9
+    - weight_decay: 0.0005
   LR: 
+    base: 0.001
     backbone: 0.001
     encoder: 0.01
     necks: 0.01
@@ -112,7 +115,7 @@ OPTIMIZER:
 ITERATION_NUM: ~
 EPOCH_NUM: 200
 START_EPOCH: 0
-UPDATE_STRATEGY: "general_updating"
+UPDATE_STRATEGY: "general_strategy"
 
 # Validation details
 PER_ITERATION: ~
@@ -134,9 +137,4 @@ METRICS:
     metric: "precision"
     top_k: ~
     threshold: 0.5
-
-INTERPRETER:
-  NAME: "cam"
-  METHOD: "gradcam"
-  TARGET_LAYERS: 
-    - "layer4"
+

configs/progressive_multi_granularity_learning/pmg_resnet50.yml

@@ -106,8 +106,11 @@ TRANSFORMS:
 
 OPTIMIZER:
   NAME: "SGD"
-  MOMENTUM: 0.9
+  ARGS: 
+    - momentum: 0.9
+    - weight_decay: 0.0005
   LR: 
+    base: 0.0002 
     backbone: 0.0002
     encoder: 0.002
     necks: 0.002
@@ -140,8 +143,3 @@ METRICS:
     top_k: ~
     threshold: 0.5
 
-INTERPRETER:
-  NAME: "cam"
-  METHOD: "gradcam"
-  TARGET_LAYERS: 
-    - "layer4"

configs/progressive_multi_granularity_learning/pmg_v2_resnet50.yml

@@ -28,6 +28,16 @@ DATASET:
 MODEL:
   NAME: "PMG_V2"
   CLASS_NUM: 200
+  ARGS: 
+    - outputs_num: 3
+    - BLOCKS: 
+      - [8, 8, 0, 0]
+      - [4, 4, 4, 0]
+      - [2, 2, 2, 2]
+    - alpha: 
+      - 0.01
+      - 0.05
+      - 0.1
   CRITERIONS: 
     - name: "cross_entropy_loss"
       args: []
@@ -108,8 +118,11 @@ TRANSFORMS:
 
 OPTIMIZER:
   NAME: "SGD"
-  MOMENTUM: 0.9
+  ARGS: 
+    - momentum: 0.9
+    - weight_decay: 0.0005
   LR: 
+    base: 0.0005 
     backbone: 0.0005
     encoder: ~
     necks: 0.005
@@ -140,9 +153,3 @@ METRICS:
     metric: "precision"
     top_k: ~
     threshold: 0.5
-
-INTERPRETER:
-  NAME: "cam"
-  METHOD: "gradcam"
-  TARGET_LAYERS: 
-    - "layer4"

configs/resnet/resnet50.yml

@@ -88,17 +88,20 @@ TRANSFORMS:
 
 OPTIMIZER:
   NAME: "SGD"
-  MOMENTUM: 0.9
-  LR: 
+  ARGS: 
+    - momentum: 0.9
+    - weight_decay: 0.0005
+  LR:
+    base: 0.0002 
     backbone: 0.0002
     encoder: 0.002
     necks: 0.002
     heads: 0.002
 
 ITERATION_NUM: ~
-EPOCH_NUM: 1
+EPOCH_NUM: 50
 START_EPOCH: 0
-UPDATE_STRATEGY: "general_updating"
+UPDATE_STRATEGY: "general_strategy"
 
 # Validation details
 PER_ITERATION: ~

configs/resnet/resnet50_cutmix.yml

@@ -88,8 +88,11 @@ TRANSFORMS:
 
 OPTIMIZER:
   NAME: "SGD"
-  MOMENTUM: 0.9
-  LR: 
+  ARGS: 
+    - momentum: 0.9
+    - weight_decay: 0.0005
+  LR:
+    base: 0.0002 
     backbone: 0.0002
     encoder: 0.002
     necks: 0.002

configs/swin_transformer/swin_transformer.yml

@@ -0,0 +1,168 @@
+EXP_NAME: "SwinT"
+
+RESUME_WEIGHT: ~
+
+DISTRIBUTED: False
+
+WEIGHT:
+  NAME: "swinT.pth"
+  SAVE_DIR: "/mnt/sdb/data/wangxinran/weight/fgvclib"
+
+LOGGER:
+  NAME: "txt_logger"
+
+DATASET:
+  NAME: "CUB_200_2011"
+  ROOT: "/mnt/sdb/data/wangxinran/dataset"
+  TRAIN:
+    BATCH_SIZE: 16
+    POSITIVE: 0
+    PIN_MEMORY: True
+    SHUFFLE: True
+    NUM_WORKERS: 4
+  TEST:
+    BATCH_SIZE: 16
+    POSITIVE: 0
+    PIN_MEMORY: False
+    SHUFFLE: True
+    NUM_WORKERS: 4
+
+MODEL:
+  NAME: "SwinTransformer"
+  CLASS_NUM: 200
+  ARGS:
+    - img_size: 384
+    - fpn_size: 1536
+    - lambda_s: 0.0
+    - lambda_n: 5.0
+    - lambda_b: 0.5
+    - lambda_c: 1.0
+    - update_freq: 2
+    - use_selection: True
+    - use_fpn: True
+    - use_combiner: True
+    - num_select:
+      - layer1: 2048
+      - layer2: 512
+      - layer3: 128
+      - layer4: 32
+
+  CRITERIONS:
+    - name: "cross_entropy_loss"
+      args: []
+      w: 1.0
+    - name: "mean_square_error_loss"
+      args: []
+      w: 1.0
+  BACKBONE:
+    NAME: "swin_large_patch4_window12_384_in22k"
+    ARGS:
+      - pretrained: True
+  ENCODER:
+    NAME: ~
+  NECKS:
+    NAME: ~
+  HEADS:
+    NAME: "GCN_combiner"
+    ARGS:
+      - num_selects:
+        - layer1: 2048
+        - layer2: 512
+        - layer3: 128
+        - layer4: 32
+      - total_num_selects: 2720
+      - num_classes: 200
+      - fpn_size: 1536
+
+TRANSFORMS:
+  TRAIN:
+    - name: "resize"
+      size:
+        - 510
+        - 510
+    - name: "random_crop"
+      size: 384
+      padding: 0
+    - name: "random_horizontal_flip"
+      prob: 0.5
+    - name: "randomApply_gaussianBlur"
+      prob: 0.1
+    - name: "randomAdjust_sharpness"
+      sharpness_factor: 1.5
+      prob: 0.1
+    - name: "to_tensor"
+    - name: "normalize"
+      mean:
+         - 0.485
+         - 0.456
+         - 0.406
+      std:
+        - 0.229
+        - 0.224
+        - 0.225
+  TEST:
+    - name: "resize"
+      size:
+        - 510
+        - 510
+    - name: "center_crop"
+      size: 384
+    - name: "to_tensor"
+    - name: "normalize"
+      mean:
+         - 0.485
+         - 0.456
+         - 0.406
+      std:
+        - 0.229
+        - 0.224
+        - 0.225
+
+OPTIMIZER:
+  NAME: "AdamW"
+  ARGS: 
+    - weight_decay: 0.0005
+  LR:
+    base: 0.0001
+    backbone: 0.0001
+    encoder: 0.0001
+    necks: 0.0001
+    heads: 0.0001
+
+LR_SCHEDULE:
+  NAME: "warmup_cosine_decay_schedule"
+  ARGS: 
+    - warmup_steps: 800
+    - max_lr: 0.0001
+    - max_epochs: 50
+    - decay_type: 1
+
+ITERATION_NUM: ~
+EPOCH_NUM: 50
+START_EPOCH: 0
+AMP: True
+UPDATE_STRATEGY: ""
+UPDATE_FUNCTION: "update_swin_transformer"
+EVALUATE_FUNCTION: "swin_transformer_evaluate"
+
+# Validation details
+PER_ITERATION: ~
+PER_EPOCH: ~
+METRICS:
+  - name: "accuracy(topk=1)"
+    metric: "accuracy"
+    top_k: 1
+    threshold: ~
+  - name: "accuracy(topk=5)"
+    metric: "accuracy"
+    top_k: 5
+    threshold: ~
+  - name: "recall(threshold=0.5)"
+    metric: "recall"
+    top_k: ~
+    threshold: 0.5
+  - name: "precision(threshold=0.5)"
+    metric: "precision"
+    top_k: ~
+    threshold: 0.5
+