update ZJU-MoCap finetuning

README.md

@@ -83,6 +83,12 @@ python train_net.py --cfg_file configs/enerf/dtu/scan114.yaml
 
 Fine-tuning for 3000 and 11000 iterations takes about 11 minutes and 40 minutes, respectively, on our test machine ( i9-12900K CPU, RTX 3090 GPU).
 
+### Fine-tuning on the ZJU-MoCap dataset
+
+```
+python train_net.py --cfg_file configs/enerf/zjumocap/zjumocap_train.yaml
+```
+
 
 ## Evaluation
 

configs/enerf/llff/flower.yaml

@@ -14,5 +14,8 @@ test_dataset:
     scene: flower
 train:
     epoch: 147 # pretrained epoch + 8
+    sampler_meta:
+        input_views_num: [3, 4] 
+        input_views_prob: [0.4, 0.6]
 save_ep: 1
 eval_ep: 1

configs/enerf/llff/fortress.yaml

@@ -14,5 +14,8 @@ test_dataset:
     scene: fortress
 train:
     epoch: 147 # pretrained epoch + 8
+    sampler_meta:
+        input_views_num: [3, 4] 
+        input_views_prob: [0.4, 0.6]
 save_ep: 1
 eval_ep: 1

configs/enerf/zjumocap/zjumocap_train.yaml

@@ -0,0 +1,56 @@
+parent_cfg: 'configs/enerf/dtu_pretrain.yaml'
+exp_name: zjumocap
+pretrain: dtu_pretrain
+
+train_dataset_module: lib.datasets.zjumocap.enerf
+test_dataset_module: lib.datasets.zjumocap.enerf
+network_module: lib.networks.enerf.network_human
+evaluator_module: lib.evaluators.enerf_human
+
+enerf:
+    sample_on_mask: True
+    train_input_views: [2, 3]
+    train_input_views_prob: [0.9, 0.1]
+    test_input_views: 2
+    cas_config:
+        train_img: [False, False]
+        patch_size: [-1, 64] #
+        num_rays: [4096, 16384] # 
+        num_patchs: [0, 4] # 
+        volume_planes: [32, 8]
+        render_if: [True, True]
+
+train_dataset:
+    data_root: 'zju_mocap'
+    scene: 'CoreView_313'
+    split: train
+    frames: [0, 600, 1]
+    input_views: [0, -1, 2]
+    render_views: [0, -1, 2]
+    input_ratio: 0.5
+
+test_dataset:
+    data_root: 'zju_mocap'
+    scene: 'CoreView_313'
+    split: test
+    frames: [0, 600, 100]
+    input_views: [0, -1, 2]
+    render_views: [1, -1, 2]
+    input_ratio: 0.5
+
+train:
+    batch_size: 1
+    lr: 5e-4
+    epoch: 100
+    scheduler:
+        type: 'exponential'
+        gamma: 0.5
+        decay_epochs: 10
+    sampler_meta:
+        input_views_num: [2, 3] 
+        input_views_prob: [0.9, 0.1]
+
+eval_ep: 1
+
+
+

lib/utils/net_utils.py

@@ -493,7 +493,7 @@ def save_trained_config(cfg):
     yaml.dump(cfg, open(train_config_path, 'w'))
 
 def load_pretrain(net, model_dir):
-    model_dir = os.path.join('data/trained_model', cfg.task, model_dir)
+    model_dir = os.path.join(cfg.workspace, 'trained_model', cfg.task, model_dir)
     if not os.path.exists(model_dir):
         return 1