add new combined dataset

Author: yonigozlan

.gitignore

@@ -84,6 +84,7 @@ target/
 lib/pycocotools/_mask.c
 lib/nms/cpu_nms.c
 
+output/*
 outputs/*
 models/*
 logs/*

experiments/coco/hrnet/w48_384x288_adam_lr1e-3.yaml

@@ -25,7 +25,8 @@ DATASET:
 MODEL:
   INIT_WEIGHTS: true
   NAME: pose_hrnet
-  NUM_JOINTS: 17
+  NUM_JOINTS: 58
+  # NUM_JOINTS: 17
   PRETRAINED: "models/pytorch/imagenet/hrnet_w48-8ef0771d.pth"
   TARGET_TYPE: gaussian
   IMAGE_SIZE:
@@ -112,7 +113,8 @@ TEST:
   BBOX_THRE: 1.0
   IMAGE_THRE: 0.0
   IN_VIS_THRE: 0.2
-  MODEL_FILE: "models/pytorch/pose_coco/pose_hrnet_w48_384x288.pth"
+  MODEL_FILE: "outputs/sherlock/model_best1e-3_62_epochs.pth"
+  # MODEL_FILE: "models/pytorch/pose_coco/pose_hrnet_w48_384x288.pth"
   NMS_THRE: 1.0
   OKS_THRE: 0.9
   USE_GT_BBOX: true

experiments/infinity_coco/hrnet/w48_384x288_adam_lr1e-3.yaml

@@ -5,9 +5,9 @@ CUDNN:
   ENABLED: true
 DATA_DIR: ""
 GPUS: (0,)
-OUTPUT_DIR: "outputs/output_infinity_coco_unfrozen"
-LOG_DIR: "logs/log_infinity_coco_unfrozen"
-WORKERS: 2
+OUTPUT_DIR: "outputs/output_new_infinity"
+LOG_DIR: "logs/output_new_infinity"
+WORKERS: 12
 PRINT_FREQ: 10
 
 DATASET:
@@ -17,14 +17,14 @@ DATASET:
   FLIP: true
   NUM_JOINTS_HALF_BODY: 27
   PROB_HALF_BODY: 0.3
-  ROOT: "../infinity_dataset_combined"
+  ROOT: "../combined_infinity_dataset"
   ROOT_COCO: "/scratch/users/yonigoz/coco_dataset/"
   ROT_FACTOR: 45
   SCALE_FACTOR: 0.35
   TEST_SET: "test"
   TRAIN_SET: "train"
   TRAIN_SET_COCO: "train2017"
-  COCO_INFINITY_RATIO: 10
+  COCO_INFINITY_RATIO: 3
 MODEL:
   INIT_WEIGHTS: true
   NAME: pose_hrnet

experiments/infinity_coco/hrnet/w48_384x288_adam_lr1e-3_local.yaml

@@ -5,9 +5,9 @@ CUDNN:
   ENABLED: true
 DATA_DIR: ""
 GPUS: (0,)
-OUTPUT_DIR: "outputs/output_infinity_coco_local"
-LOG_DIR: "logs/log_infinity_coco_local"
-WORKERS: 2
+OUTPUT_DIR: "outputs/output_new_infinity"
+LOG_DIR: "logs/output_new_infinity"
+WORKERS: 12
 PRINT_FREQ: 10
 
 DATASET:
@@ -17,7 +17,7 @@ DATASET:
   FLIP: true
   NUM_JOINTS_HALF_BODY: 27
   PROB_HALF_BODY: 0.3
-  ROOT: "../infinity_dataset_combined"
+  ROOT: "../combined_infinity_dataset"
   ROOT_COCO: "data/coco/"
   ROT_FACTOR: 45
   SCALE_FACTOR: 0.35
@@ -63,6 +63,7 @@ MODEL:
       - "stage3"
       - "transition3"
       - "stage4"
+    PRETRAIN_FINAL_LAYER: true
     FINAL_CONV_KERNEL: 1
     STAGE2:
       NUM_MODULES: 1

lib/core/function.py

@@ -131,6 +131,7 @@ def validate(
     model.eval()
 
     num_samples = len(val_dataset)
+    # all_preds = np.zeros((num_samples, 17, 3), dtype=np.float32)
     all_preds = np.zeros((num_samples, config.MODEL.NUM_JOINTS, 3), dtype=np.float32)
     all_boxes = np.zeros((num_samples, 6))
     image_path = []
@@ -142,13 +143,15 @@ def validate(
         for i, (input, target, target_weight, meta) in enumerate(val_loader):
             # compute output
             outputs = model(input)
+            # outputs = model(input)[:, :17, :, :]
             if isinstance(outputs, list):
                 output = outputs[-1]
             else:
                 output = outputs
 
             if config.TEST.FLIP_TEST:
                 input_flipped = input.flip(3)
+                # outputs_flipped = model(input_flipped)[:, :17, :, :]
                 outputs_flipped = model(input_flipped)
 
                 if isinstance(outputs_flipped, list):

lib/models/pose_hrnet.py

@@ -343,7 +343,10 @@ def __init__(self, cfg, **kwargs):
         )
 
         self.pretrained_layers = extra["PRETRAINED_LAYERS"]
-        self.pretrain_final_layer = extra["PRETRAIN_FINAL_LAYER"]
+        if "PRETRAIN_FINAL_LAYER" in extra:
+            self.pretrain_final_layer = extra["PRETRAIN_FINAL_LAYER"]
+        else:
+            self.pretrain_final_layer = False
 
     def _make_transition_layer(self, num_channels_pre_layer, num_channels_cur_layer):
         num_branches_cur = len(num_channels_cur_layer)
@@ -511,6 +514,7 @@ def init_weights(self, pretrained=""):
             self.load_state_dict(need_init_state_dict, strict=False)
 
             if self.pretrain_final_layer:
+                print("load final layer")
                 self.final_layer.weight.data[:17, :, :, :] = pretrained_state_dict[
                     "final_layer.weight"
                 ]