y_pred = model.predict(X)
lake_df['rmse_pred'] = y_pred
y_act = lake_df['rmse']
lake_df.sort_values(by=['rmse_pred'], inplace=True)
lowest_rmse = lake_df.iloc[0]['rmse_pred']
top_ids = [str(j) for j in lake_df.iloc[:int(k)]['site_id']]
best_site = top_ids[0]
#define target test data to use
data_dir_target = "../../data/processed/lake_data/"+target_id+"/"
(_, _, tst_data_target, tst_dates_target, unique_tst_dates_target, all_data_target, all_phys_data_target, all_dates_target,
hypsography_target) = buildLakeDataForRNN_manylakes_finetune2(target_id, data_dir_target, seq_length, n_features,
win_shift = win_shift, begin_loss_ind = begin_loss_ind,
outputFullTestMatrix=True, allTestSeq=False, oldFeat=False)
#useful values, LSTM params
batch_size = all_data_target.size()[0]
u_depths_target = np.unique(tst_data_target[:,0,0])
n_depths = torch.unique(all_data_target[:,:,0]).size()[0]
n_test_dates_target = unique_tst_dates_target.shape[0]
#define LSTM model
class LSTM(nn.Module):
def __init__(self, input_size, hidden_size, batch_size):
super(LSTM, self).__init__()
self.input_size = input_size
###########################33
n_ep = 400 #number of epochs
win_shift = 176 #how much to slide the window on training set each time
pretrain_path = "../../../models/WRR_69Lake/"+args.lake_name+"/pretrain_experiment_trial"+str(trial)
save_path = "../../../models/WRR_69Lake/"+args.lake_name+"/pgrnn_finetune_trial"+str(trial)
###############################
# data preprocess
##################################
#create train and test sets
(trn_data, trn_dates, tst_data, tst_dates, unique_tst_dates, all_data, all_phys_data,
all_dates, hypsography) = buildLakeDataForRNN_manylakes_finetune2(args.lake_name, \
args.data_path, \
seq_length, n_features, \
win_shift=win_shift, \
begin_loss_ind=begin_loss_ind, \
latter_third_test=True, \
outputFullTestMatrix=True, \
sparseTen=False, \
realization='none', \
allTestSeq=True)
batch_size = trn_data.size()[0]
n_test_dates = unique_tst_dates.shape[0]
n_depths = torch.unique(all_data[:, :, 0]).size()[0]
u_depths = np.unique(tst_data[:, 0, 0])
#format training data for loading
train_data = TemperatureTrainDataset(trn_data)
load_path = "../../models/" + lakename + "/PGRNN_sparse_" + str(
n_prof) + "_" + str(seed)
###############################
# data preprocess
##################################
#create train and test sets
(trn_data, trn_dates, tst_data, tst_dates, unique_tst_dates,
all_data, all_phys_data, all_dates,
hypsography) = buildLakeDataForRNN_manylakes_finetune2(
lakename,
data_dir,
seq_length,
n_features,
win_shift=win_shift,
begin_loss_ind=begin_loss_ind,
outputFullTestMatrix=False,
allTestSeq=False,
sparseCustom=n_prof,
randomSeed=seed)
#if error code is returned (trn data as int), skip and record id
if isinstance(trn_data, int):
target_id = lakename
if trn_data == 25:
no_25.append(target_id)
continue
elif trn_data == 30:
no_30.append(target_id)
continue
elif trn_data == 35:
#paths to save
pretrain_path = "../../models/" + lakename + "/pretrain_source_model"
save_path = "../../models/" + lakename + "/PGRNN_source_model_0.7"
###############################
# data preprocess
##################################
#create train and test sets
(trn_data, trn_dates, tst_data, tst_dates, unique_tst_dates, all_data,
all_phys_data, all_dates,
hypsography) = buildLakeDataForRNN_manylakes_finetune2(
lakename,
data_dir,
seq_length,
n_features,
win_shift=win_shift,
begin_loss_ind=begin_loss_ind,
outputFullTestMatrix=True,
allTestSeq=True)
n_depths = torch.unique(all_data[:, :, 0]).size()[0]
u_depths = np.unique(tst_data[:, 0, 0])
trn_data = tst_data
batch_size = trn_data.size()[0]
#Dataset classes
class TemperatureTrainDataset(Dataset):
#training dataset class, allows Dataloader to load both input/target
def __init__(self, trn_data):
