|
- # ConvLSTM Cell
- import random
-
- import numpy as np
- import torch
- from tqdm import tqdm
- from sklearn.metrics import mean_squared_error
- from torch import nn
- from torch.utils.data import Dataset, DataLoader
-
-
- class ConvLSTMCell(nn.Module):
- def __init__(self, input_dim, hidden_dim, kernel_size):
- super().__init__()
- self.input_dim = input_dim
- self.hidden_dim = hidden_dim
- pad = kernel_size[0] // 2, kernel_size[1] // 2
- # 卷积操作Wx*Xt+Wh*Ht-1
- self.conv = nn.Conv2d(in_channels=input_dim + hidden_dim, out_channels=4 * hidden_dim, kernel_size=kernel_size,
- padding=pad)
-
- def initialize(self, inputs):
- # device = inputs.device
- N, _, H, W = inputs.size()
- # 初始化隐藏层状态Ht
- self.hidden_state = torch.zeros(N, self.hidden_dim, H, W).cuda()
- # 初始化记忆细胞状态ct
- self.cell_state = torch.zeros(N, self.hidden_dim, H, W).cuda()
- # 初始化记忆单元状态Mt
- self.memory_state = torch.zeros(N, self.hidden_dim, H, W).cuda()
-
- def forward(self, inputs, first_step=False):
- # 如果当前是第一个时间步,初始化Ht、ct、Mt
- if first_step:
- self.initialize(inputs)
-
- # ConvLSTM部分
- # 拼接Xt和Ht
- combined = torch.cat([inputs, self.hidden_state], dim=1) # (N, C, H, W), C=input_dim+hidden_dim
- # 进行卷积操作
- combined_conv = self.conv(combined)
- # 得到四个门控单元it、ft、ot、gt
- cc_i, cc_f, cc_o, cc_g = torch.split(combined_conv, self.hidden_dim, dim=1)
- i = torch.sigmoid(cc_i)
- f = torch.sigmoid(cc_f)
- o = torch.sigmoid(cc_o)
- g = torch.tanh(cc_g)
- # 得到当前时间步的记忆细胞状态ct=ft·ct-1+it·gt
- self.cell_state = f * self.cell_state + i * g
- # 得到当前时间步的隐藏层状态Ht=ot·tanh(ct)
- self.hidden_state = o * torch.tanh(self.cell_state)
-
-
- return self.hidden_state
-
-
- # 构建3dcnn-ConvLSTM模型 #输入为(1184,5,7,12,15)
- class ConvLSTM(nn.Module):
- def __init__(self, input_dim, hidden_dim, kernel_size):
- super().__init__()
- self.input_dim = input_dim
- self.hidden_dim = hidden_dim
- self.num_layers = len(hidden_dim)
- # self.num_layers = 1
- self.conv1 = torch.nn.Sequential(
- torch.nn.Conv3d(6, 1, 3, stride=1, padding=1),
- # 4,4 3d输入格式(batch_size, channel, depth, height, width) (in_channel,out_channel,kernel_size,strid, pad)
- # torch.nn.BatchNorm3d(1),
- )
-
- layers = []
- for i in range(self.num_layers):
- cur_input_dim = self.input_dim if i == 0 else self.hidden_dim[i - 1]
- layers.append(ConvLSTMCell(input_dim=cur_input_dim, hidden_dim=self.hidden_dim[i],
- kernel_size=kernel_size))
- self.layers = nn.ModuleList(layers)
-
- self.conv_output = nn.Conv2d(self.hidden_dim[-1], 1, kernel_size=1)
-
- def forward(self, input_x, input_frames=10, future_frames=1, output_frames=10,
- teacher_forcing=False, scheduled_sampling_ratio=0, train=True):
- # input_x = input_x.permute(0, 1, 4, 2, 3).contiguous()
- x = self.conv1(input_x) # 把 channel变成1 输入(N,C,B,H,W)
- x1 = x.permute(0, 2, 3, 4, 1) #(N,C,B,H,W) --->(N,B,H,W,C) 把channel放到最后
- input_x = x1.permute(0, 1, 4, 2, 3).contiguous()
- # print(x1.shape) # torch.Size([1, 10, 40, 200, 1])
-
- total_steps = input_frames + future_frames - 1
- outputs = [None] * total_steps
-
- # 对于每一个时间步
- for t in range(total_steps):
- # 在前12个月,使用每个月的输入样本Xt
- if t < input_frames:
- input_ = input_x[:, t]
- # 若不使用teacher forcing,则以上一个时间步的预测标签作为当前时间步的输入
- elif not teacher_forcing:
- input_ = outputs[t - 1]
- # 若使用teacher forcing,则以ratio的概率使用上一个时间步的实际标签作为当前时间步的输入
- # else:
- # mask = teacher_forcing_mask[:, t - input_frames].float().to(device)
- # input_ = input_x[:, t].to(device) * mask + outputs[t - 1] * (1 - mask)
- first_step = (t == 0)
- input_ = input_.float()
-
- # 将当前时间步的输入通过隐藏层
- for layer_idx in range(self.num_layers):
- input_ = self.layers[layer_idx](input_, first_step=first_step)
-
- # 记录每个时间步的输出
- if train or (t >= (input_frames - 1)):
- outputs[t] = self.conv_output(input_)
-
- outputs = [x for x in outputs if x is not None]
-
- # 确认输出序列的长度
- if train:
- assert len(outputs) == output_frames
- else:
- assert len(outputs) == future_frames
-
- # 得到sst的预测序列
- outputs = torch.stack(outputs, dim=1)[:, :, 0] # (N, 37, H, W)
- outputs1 = outputs[:,-future_frames:,:,:]
- # 对sst的预测序列在nino3.4区域取三个月的平均值就得到nino3.4指数的预测序列
- # nino_pred = outputs[:, -future_frames:, 10:13, 19:30].mean(dim=[2, 3]) # (N, 26)
- # nino_pred = nino_pred.unfold(dimension=1, size=3, step=1).mean(dim=2) # (N, 24)
-
- return outputs1
-
-
- for seed in range(2022, 2023):
- for date_append in range(5, 14):
- def setup_seed(seed):
- torch.manual_seed(seed)
- torch.cuda.manual_seed_all(seed)
- np.random.seed(seed)
- random.seed(seed)
- torch.backends.cudnn.deterministic = True
-
-
- # 设置随机数种子
- setup_seed(seed)
-
- # 需要 mld u v sss temp 降水 蒸发 混合层下的盐度
-
- data = np.load(r'/dataset/10day_for_14day_all_variables_deep5_pacific_10_19_SSS_learn.npz')
- print(data.files)
- # 2000-01-01 ---> 2019-12-31
- # surface_latent_heat_flux = data['surface_latent_heat_flux'][:] # (7305, 10, 40, 200)
- # surface_sensible_heat_flux = data['surface_sensible_heat_flux'][:] # (9851, 10, 40, 200)
- # surface_net_radiation = data['surface_net_radiation'][:] # (9851, 10, 40, 200)
- evaporation = data['evaporation'][:]
- total_precipitation = data['total_precipitation'][:]
- mld = data['mld'][:]
- # sst_surface = data['sst_surface'][:]
- sss_surface = data['sst_deep'][:]
- u_surface = data['u_deep'][:]
- v_surface = data['v_deep'][:]
- sss_surface_label = data['sst_deep_label'][:]
- # print(sst_surface)
-
- # print(sst_surface.shape)
- print(sss_surface.shape) #(3642, 10, 40, 200)
-
- print(sss_surface_label.shape) #(3638, 40, 200)
-
- dS_dt = data['dT_dt'][:]
- dS_dx = data['dT_dx'][:]
- dS_dy = data['dT_dy'][:]
-
-
-
- train_size = 2208
- valid_size = 2912 # 前20% 作为验证 剩下的20%的作为测试
- # surface_latent_heat_flux = surface_latent_heat_flux + surface_sensible_heat_flux + surface_net_radiation
-
- evaporation = evaporation.reshape(-1, 1, 10, 40, 200)
- evaporation = torch.Tensor(evaporation)
- evaporation_train = evaporation[0:train_size, :, :, :, :]
- evaporation_valid = evaporation[train_size:valid_size, :, :, :, :]
- # evaporation_test = evaporation[valid_size:,:,:,:,:]
-
- total_precipitation = total_precipitation.reshape(-1, 1, 10, 40, 200)
- total_precipitation = torch.Tensor(total_precipitation)
- total_precipitation_train = total_precipitation[0:train_size, :, :, :, :]
- total_precipitation_valid = total_precipitation[train_size:valid_size, :, :, :, :]
- # total_precipitation_test = total_precipitation[valid_size:,:,:,:,:]
-
- mld = mld.reshape(-1, 1, 10, 40, 200)
- mld = torch.Tensor(mld)
- mld_train = mld[0:train_size, :, :, :, :]
- mld_valid = mld[train_size:valid_size, :, :, :, :]
- # mld_test = mld[valid_size:,:,:,:,:]
-
- # sst_surface = sst_surface.reshape(-1, 1, 10, 40, 200)
- # sst_surface = torch.Tensor(sst_surface)
- # sst_surface_train = sst_surface[0:train_size, :, :, :, :]
- # sst_surface_valid = sst_surface[train_size:valid_size, :, :, :, :]
- # # sst_surface_test = sst_surface[valid_size:,:,:,:,:]
-
- sss_surface = sss_surface.reshape(-1, 1, 10, 40, 200)
- sss_surface = torch.Tensor(sss_surface)
- sss_surface_train = sss_surface[0:train_size, :, :, :, :]
- sss_surface_valid = sss_surface[train_size:valid_size, :, :, :, :]
- # sss_surface_test = sss_surface[valid_size:,:,:,:,:]
-
- u_surface = u_surface.reshape(-1, 1, 10, 40, 200)
- u_surface = torch.Tensor(u_surface)
- u_surface_train = u_surface[0:train_size, :, :, :, :]
- u_surface_valid = u_surface[train_size:valid_size, :, :, :, :]
- # u_surface_test = u_surface[valid_size:,:,:,:,:]
-
- v_surface = v_surface.reshape(-1, 1, 10, 40, 200)
- v_surface = torch.Tensor(v_surface)
- v_surface_train = v_surface[0:train_size, :, :, :, :]
- v_surface_valid = v_surface[train_size:valid_size, :, :, :, :]
- # # v_surface_test = v_surface[valid_size:,:,:,:,:]
-
- dS_dt = dS_dt.transpose(1,0,2,3)
- dS_dt = dS_dt.reshape(-1, 1, 10, 40, 200)
- dS_dt = torch.Tensor(dS_dt)
- dS_dt_train = dS_dt[0:train_size, :, :, :]
- dS_dt_valid = dS_dt[train_size:valid_size, :, :, :]
-
- dS_dx = dS_dx.transpose(1,2,0,3)
- dS_dx = dS_dx.reshape(-1, 1, 10, 40, 200)
- dS_dx = torch.Tensor(dS_dx)
- dS_dx_train = dS_dx[0:train_size, :, :, :]
- dS_dx_valid = dS_dx[train_size:valid_size, :, :, :]
-
- dS_dy = dS_dy.transpose(1,2,3,0)
- dS_dy = dS_dy.reshape(-1, 1, 10, 40, 200)
- dS_dy = torch.Tensor(dS_dy)
- dS_dy_train = dS_dy[0:train_size, :, :, :]
- dS_dy_valid = dS_dy[train_size:valid_size, :, :, :]
-
- # train_data = torch.cat((surface_latent_heat_flux_train, surface_sensible_heat_flux_train, surface_net_radiation_train,
- # evaporation_train, total_precipitation_train, mld_train, sst_surface_train, sss_surface_train,
- # u_surface_train, v_surface_train, T_d_train, S_d_train, u_d_train, v_d_train,
- # xx_train, yy_train), dim=2) # train_data.shape:torch.Size([5920, 10, 16, 40, 200])
-
- # valid_data = torch.cat((surface_latent_heat_flux_valid, surface_sensible_heat_flux_valid, surface_net_radiation_valid,
- # evaporation_valid, total_precipitation_valid, mld_valid, sst_surface_valid, sss_surface_valid,
- # u_surface_valid, v_surface_valid, T_d_valid, S_d_valid, u_d_valid, v_d_valid,
- # xx_valid, yy_valid), dim=2)
-
- # test_data = torch.cat((surface_latent_heat_flux_test, surface_sensible_heat_flux_test, surface_net_radiation_test,
- # evaporation_test, total_precipitation_test, mld_test, sst_surface_test, sss_surface_test,
- # u_surface_test, v_surface_test, T_d_test, S_d_test, u_d_test, v_d_test,
- # xx_test, yy_test), dim=2)
-
- train_data = torch.cat((evaporation_train, total_precipitation_train, sss_surface_train, mld_train, u_surface_train, v_surface_train,
- dS_dt_train, dS_dx_train, dS_dy_train
- ), dim=1) # train_data.shape:torch.Size([5920, 10, 16, 40, 200])
-
- valid_data = torch.cat((evaporation_valid, total_precipitation_valid, sss_surface_valid, mld_valid, u_surface_valid, v_surface_valid,
- dS_dt_valid, dS_dx_valid, dS_dy_valid
- ), dim=1)
-
- # train_data = sst_surface_train
- # # train_data.shape:torch.Size([5920, 10, 16, 40, 200])
-
- # valid_data = sst_surface_valid
-
-
-
-
- # train_data = torch.cat((sst_surface_train, sss_surface_train,), dim=1) # train_data.shape:torch.Size([5920, 10, 16, 40, 200])
-
- # valid_data = torch.cat((sst_surface_valid, sss_surface_valid,), dim=1)
-
- print(train_data.shape)
- print(valid_data.shape)
-
- # sst_train_label = sst_surface_label[10 + date_append: train_size + 10 + date_append, :, :]
- # sst_valid_label = sst_surface_label[train_size + 10 + date_append: valid_size + 10 + date_append, :, :]
- # sst_test_label = sst_surface_label[valid_size + 14:,:,:]
- # sss_surface (3642, 10, 40, 200)
- sss_train_label = sss_surface_label[10 :train_size + 10 , date_append, :, :]
- sss_valid_label = sss_surface_label[train_size + 10 : valid_size + 10 ,date_append, :, :]
- # sss_test_label = sss_surface_label[valid_size + 10 + date_append:,:,:]
-
- # sst_train_label = sst_train_label.reshape(-1, 1, 40, 200)
- # sst_valid_label = sst_valid_label.reshape(-1, 1, 40, 200)
-
- sss_train_label = sss_train_label.reshape(-1, 1, 40, 200)
- sss_valid_label = sss_valid_label.reshape(-1, 1, 40, 200)
-
- # train_label = np.concatenate((sst_train_label, sss_train_label), axis=1)
- # valid_label = np.concatenate((sst_valid_label, sss_valid_label), axis=1)
- train_label = sss_train_label
- valid_label = sss_valid_label
- print(train_label.shape)
- # train_label = sst_train_label
- # valid_label = sst_valid_label
-
- # print('train_label.shape:{}'.format(train_label.shape))
- # print('valid_label.shape:{}'.format(valid_label.shape))
- # print('train_data.shape:{}'.format(train_data.shape))
- # print('valid_data.shape:{}'.format(valid_data.shape))
-
- # print('train_data.shape:{}'.format(train_data.shape)) # train_data.shape:torch.Size([5920, 10, 16, 40, 200])
- # print('valid_data.shape:{}'.format(valid_data.shape)) # valid_data.shape:torch.Size([1952, 10, 16, 40, 200])
- # print('test_data.shape:{}'.format(test_data.shape)) # test_data.shape:torch.Size([1979, 10, 16, 40, 200])
- # print('sst_train_label.shape:{}'.format(sst_train_label.shape)) # sst_train_label.shape:(5920, 14, 40, 200)
- # print('sst_valid_label.shape:{}'.format(sst_valid_label.shape)) # sst_valid_label.shape:(1952, 14, 40, 200)
- # print('sst_test_label.shape:{}'.format(sst_test_label.shape)) # sst_test_label.shape:(1961, 14, 40, 200)
- # print('sss_train_label.shape:{}'.format(sss_train_label.shape)) # sss_train_label.shape:(5920, 14, 40, 200)
- # print('sss_valid_label.shape:{}'.format(sss_valid_label.shape)) # sss_valid_label.shape:(1952, 14, 40, 200)
- # print('sss_test_label.shape:{}'.format(sss_test_label.shape)) # sss_test_label.shape:(1961, 14, 40, 200)
-
- # 构建数据管道
- class MyDataset(Dataset):
- def __init__(self, data, label):
- self.data = torch.Tensor(data)
- self.label = torch.Tensor(label)
-
- def __len__(self):
- return len(self.label)
-
- def __getitem__(self, idx):
- return self.data[idx], self.label[idx]
-
-
- batch_size1 = 32
- batch_size2 = 32
- batch_size3 = 3000
-
- trainset = MyDataset(train_data, train_label)
- trainloader = DataLoader(trainset, batch_size=batch_size1, shuffle=True, drop_last=False, pin_memory=True,
- num_workers=4)
-
- validset = MyDataset(valid_data, valid_label)
- validloader = DataLoader(validset, batch_size=batch_size2, shuffle=True, drop_last=False, pin_memory=True,
- num_workers=4)
-
- # testset = MyDataset(test_data, sss_test_label)
- # testloader = DataLoader(testset, batch_size=batch_size3, shuffle=False, drop_last=False,pin_memory=True, num_workers=0)
-
- # print('Qnet_train.shape:{}'.format(sst1_train.shape))
-
- model_weights1 = '/model/epo200_lay3_lr0.001_e{}_ConvLSTM_3D_{}day_model_weights_SST_deep5.pth'.format(seed, date_append)
-
- input_dim = 1
- # 隐藏层节点数
- hidden_dim = (32, 32, 32)
- # 卷积核大小
- kernel_size = (3, 3)
-
- model = ConvLSTM(input_dim, hidden_dim, kernel_size).cuda()
-
- criterion = nn.MSELoss()
- # 定义优化器
- optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
-
- epochs = 200
- train_losses, valid_losses = [], []
- # best_loss = 2
- best_score = float('inf')
- best_score1 = float('inf')
-
- pred_val = np.zeros((704, 1, 40, 200))
-
- sores = []
-
-
- def rmse(y_true, y_preds):
- return np.sqrt(mean_squared_error(y_pred=y_preds, y_true=y_true))
-
-
- for epoch in range(epochs):
- print('Epoch: {}/{}'.format(epoch + 1, epochs))
- # print(var_y)
- # 模型训练
- model.train()
- losses = 0
- loss1 = 0
- for data, label in tqdm(trainloader):
- data = data.cuda()
- label = label.cuda()
- optimizer.zero_grad()
- # print('data1.shape:{}'.format(data1.shape)) # data1.shape:torch.Size([32, 10, 6, 40, 200]) surface_latent_heat_flux_train surface_sensible_heat_flux_train surface_net_radiation_train evaporation_train total_precipitation_train mld_train
- # print('data2.shape:{}'.format(data2.shape)) # data2.shape:torch.Size([32, 10, 3, 40, 200]) sss_surface_train u_surface_train v_surface_train
- # print('data3.shape:{}'.format(data3.shape)) # data3.shape:torch.Size([32, 10, 40, 200]) S_d_train
- # print('data4.shape:{}'.format(data4.shape)) # data4.shape:torch.Size([32, 10, 2, 40, 200]) u_d_train v_d_train
- # print('label.shape:{}'.format(label.shape)) #label.shape:torch.Size([32, 14, 40, 200])
- # print('data_train.shape:{}'.format(data.shape)) # data_train.shape:torch.Size([32, 12, 10, 40, 200])
- data_train = data[:,:6,:,:,:]
- out = model(data_train)
- # print('out.shape:{}'.format(out.shape))
- # print('label.shape:{}'.format(label.shape))
- # print(out)
- # 偏S/偏t - (E - P) * (S / h) - [u * 偏S/偏x + v * 偏S/偏y ] + H * (w_h + dh/dt * ((S - S_h) / h)) = 0 loss1
- # 偏T/偏t - Q / (p * C_p * h_m) - u * (偏T/偏x) - v * (偏T/偏y) + w_e * ((T - T_d) / h) = 0
-
- # sst_label = label[:,0,:,:]
- # sss_label = label[:,1,:,:]
-
- # sst_out = out[:,0,:,:]
- # sss_out = out[:,1,:,:]
- out = out.reshape(-1, 1, 40, 200)
- loss = criterion(out, label)
- # loss2 = criterion(sss_out, sss_label)
-
- losses += loss
-
- loss.backward()
- optimizer.step()
- train_loss = losses / len(trainloader)
- train_losses.append(train_loss)
-
- print('Training Loss: {:.10f}'.format((train_loss)))
-
- # model.eval()
- losses = 0
- with torch.no_grad():
- for i, data in tqdm(enumerate(validloader)):
- data, label = data
- data = data.cuda()
- label = label.cuda()
- optimizer.zero_grad()
-
- data_train = data[:,:6,:,:,:]
- out = model(data_train)
- out = out.reshape(-1, 1, 40, 200)
- loss = criterion(out, label)
-
- losses += float(loss)
-
- out1 = out.detach().cpu().numpy()
- pred_val[i * batch_size2:(i + 1) * batch_size2] = np.array(out1)
-
- valid_loss = losses / len(validloader)
- valid_losses.append(valid_loss)
-
- valid_label1 = valid_label.reshape(-1, 1)
- preds1 = pred_val.reshape(-1, 1)
-
- s = rmse(valid_label1, preds1)
- sores.append(s)
- print('Score: {:.3f}'.format(s))
-
- if valid_loss < best_score1: # 求s的最小值 ---》最大值反过来 inf符号也要反过来
- best_score1 = valid_loss
- checkpoint = {'best_score': valid_loss,
- 'state_dict': model.state_dict()}
- torch.save(checkpoint, model_weights1) # if valid_loss < best_loss:
- best_loss = valid_loss
- torch.save(model.state_dict(),
- '/model/_ConvLSTM_3D_lr0.001_model_200_layer3_{}day_e{}_SSS_deep5.pt'.format(date_append, seed))
-
- print(sores)
- print(best_score)
- print(s)
|