Merge pull request #16 from mhpi/highflow-expert

high-flow commit 2

environment.yml

@@ -1,4 +1,4 @@
-name: mhpihydrodl
+name: hydrodl_new
 channels:
   - pytorch
   - nvidia

example/StreamflowExample-DI.py

@@ -46,7 +46,6 @@
 gpuid = 6
 torch.cuda.set_device(gpuid)
 
-
 # Set hyperparameters
 EPOCH = 300
 BATCH_SIZE = 100
@@ -57,6 +56,7 @@
 forType = 'daymet'
 trainBuff = 365
 loadTrain = True
+subset_train = 'All'  #give the list of basins to train on or else fix 'All' to use all
 
 # Fix random seed
 seedid = 111111
@@ -104,7 +104,7 @@
                'geol_2nd_class', 'glim_2nd_class_frac', 'carbonate_rocks_frac', 'geol_porostiy', 'geol_permeability']
 optData = default.optDataCamels
 optData = default.update(
-    optData, varT=varF, varC=attrLst, tRange=Ttrain, forType=forType)  # Update the training period
+    optData, varT=varF, varC=attrLst, tRange=Ttrain, forType=forType, subset=subset_train)  # Update the training period
 
 # if (interfaceOpt == 1) and (2 not in Action):
 if (interfaceOpt == 1) and (loadTrain is True):
@@ -171,7 +171,7 @@
 )
 
 # define output folder for model results
-exp_name = f"CAMELSDemo-torch={torch.__version__}"
+exp_name = f"CAMELSDemo"
 exp_disp = "TestRun"
 save_path = os.path.join(
     exp_name,
@@ -199,7 +199,7 @@
         ny = yTrain.shape[-1]
         # load model for training
         if torch.cuda.is_available():
-            model = CudnnLstmModel(nx=nx, ny=ny, hiddenSize=HIDDENSIZE, warmUpDay=trainBuff)
+            model = CudnnLstmModel(nx=nx, ny=ny, hiddenSize=HIDDENSIZE)
         else:
             model = CpuLstmModel(nx=nx, ny=ny, hiddenSize=HIDDENSIZE)
         optModel = default.update(optModel, nx=nx, ny=ny)

hydroDL/data/camels.py

@@ -646,7 +646,7 @@ def __init__(self, *, subset="All", tRange, forType='nldas'):
         elif type(subset) is list:
             self.usgsId = np.array(subset)
             crd = np.zeros([len(self.usgsId), 2])
-            ind = np.full(len(self.usgsId), np.nan).astype(int)
+            ind = np.full(len(self.usgsId), -1, dtype=int)
             for ii in range(len(self.usgsId)):
                 tempind = np.where(gageDict["id"] == self.usgsId[ii])
                 ind[ii] = tempind[0][0]

hydroDL/model/dropout.py

@@ -2,7 +2,10 @@
 import torch.nn
 
 
-def createMask(x, dr):
+
+def createMask(x, dr, seed):
+    generator = torch.Generator(device="cuda")
+    generator.manual_seed(seed)
     mask = x.new().resize_as_(x).bernoulli_(1 - dr).div_(1 - dr).detach_()
     # print('droprate='+str(dr))
     return mask

hydroDL/model/rnn/CudnnLstm.py

@@ -10,7 +10,7 @@
 
 
 class CudnnLstm(nn.Module):
-    def __init__(self, *, inputSize, hiddenSize, dr=0.5, drMethod="drW", gpu=0):
+    def __init__(self, *, inputSize, hiddenSize, dr=0.5, drMethod="drW", gpu=0, seed=42):
         super(CudnnLstm, self).__init__()
         self.inputSize = inputSize
         self.hiddenSize = hiddenSize
@@ -23,6 +23,7 @@ def __init__(self, *, inputSize, hiddenSize, dr=0.5, drMethod="drW", gpu=0):
         self._all_weights = [["w_ih", "w_hh", "b_ih", "b_hh"]]
         self.cuda()
         self.name = "CudnnLstm"
+        self.seed = seed
         self.is_legacy = True
 
         self.reset_mask()
@@ -42,8 +43,8 @@ def __setstate__(self, d):
         self._all_weights = [["w_ih", "w_hh", "b_ih", "b_hh"]]
 
     def reset_mask(self):
-        self.maskW_ih = createMask(self.w_ih, self.dr)
-        self.maskW_hh = createMask(self.w_hh, self.dr)
+        self.maskW_ih = createMask(self.w_ih, self.dr, self.seed)
+        self.maskW_hh = createMask(self.w_hh, self.dr, self.seed)
 
     def reset_parameters(self):
         stdv = 1.0 / math.sqrt(self.hiddenSize)