Upload config.yaml

rftrans/configs/config.yaml

@@ -0,0 +1,109 @@
+# RFTrans Table II reproduction - end-to-end fine-tune with paper-faithful loss
+# Bug fix: training previously used loss = L_flow + alpha*L_norm with alpha=0.02
+# (flow dominates ~100x, normal supervision essentially ignored).
+# Paper Eq.: L = alpha * L_flow + L_norm with alpha=0.01 (best per Table IV).
+train:
+  datasetsTrain:
+    - images: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/train_cg/RGB'
+      flows: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/train_cg/flow'
+      labels: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/train_cg/normal'
+      masks: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/train_cg/mask'
+
+  datasetsVal:
+    - images: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/valid/RGB'
+      flows: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/valid/flow'
+      labels: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/valid/normal'
+      masks: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/valid/mask'
+
+  datasetsTestReal:
+  datasetsTestSynthetic:
+
+  rgb2flow:
+    model: "drn"
+    numClasses: 2
+    numInputChannels: 3
+    pathWeightsFile: "/data1/liulingfeng/cooperation/ghy/RFTrans-repro/RFTrans/pytorch_networks/refractive_flow/logs-deeplab/exp-001/checkpoints/checkpoint-epoch-0500.pth"
+
+  flow2normal:
+    model: "simple_unet"
+    numClasses: 3
+    numInputChannels: 3
+    pathWeightsFile: "/data1/liulingfeng/cooperation/ghy/RFTrans-repro/RFTrans/pytorch_networks/flow2normal/logs-deeplab/exp-000/checkpoints/checkpoint-epoch-0500.pth"
+
+  model: "drn"
+  batchSize: 8
+  validationBatchSize: 8
+  testBatchSize: 8
+  numEpochs: 100
+  # Train at data's native 512x512 (matches the pretrained R2F/F2N which learned
+  # flow magnitudes in the 512-frame). Paper says 256x256 but our train_cg data
+  # is rendered at 512x512 and the pretrained checkpoints were trained at 512.
+  imgHeight: 512
+  imgWidth: 512
+  numWorkers: 8
+  logsDir: "logs-deeplab"
+  lossFunc: "cosine"
+  percentageDataForTraining: 1.0
+  percentageDataForValidation: 0.5
+
+  outputStride: 8
+  epochSize: 1
+
+  # initialize from previously trained separate checkpoints (RFNet / F2Net)
+  continueTraining: True
+  pathPrevCheckpoint: ""
+  initOptimizerFromCheckpoint: False
+  loadEpochNumberFromCheckpoint: False
+
+  saveImageInterval: 1
+  saveImageIntervalIter: 1000
+  testInterval: 1
+  saveModelInterval: 1
+
+  # NOTE: paper specifies SGD lr=1e-4 momentum 0.9 weight_decay 5e-4 for 100 ep,
+  # but fine-tuning from already-Adam-trained pretrained R2F/F2N with SGD@1e-4
+  # barely moves the weights. Use Adam @ 1e-4 to converge faster from the
+  # pretrained init (loss bug previously suppressed the normal supervision).
+  optimizer: "SGD"
+  optimSgd:
+    learningRate: 1e-4
+    momentum: 0.9
+    weight_decay: 5e-4
+  optimAdam:
+    learningRate: 1e-4
+    weightDecay: 0.0001
+  lrScheduler: "StepLR"
+  lrSchedulerStep:
+    step_size: 30
+    gamma: 0.5
+
+  # alpha for joint loss; paper best = 0.01 with L = alpha * L_flow + L_norm
+  loss_alpha: 0.01
+
+eval:
+  datasetsSynthetic:
+    - images: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/valid/RGB'
+      flows: ''
+      labels: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/valid/normal'
+      masks: '/data1/liulingfeng/cooperation/ghy/RFTrans-repro/data/unity/valid/mask'
+
+  datasetsReal:
+
+  rgb2flow:
+    model: "drn"
+    numClasses: 2
+    numInputChannels: 3
+    pathWeightsFile: ""
+
+  flow2normal:
+    model: "simple_unet"
+    numClasses: 3
+    numInputChannels: 3
+    pathWeightsFile: ""
+
+  model: "drn"
+  batchSize: 4
+  imgHeight: 256
+  imgWidth: 256
+  numWorkers: 4
+  resultsDir: "data/results"