Optimizer.step() doesn't work

PyTorch Live
1

I tried to train model.
But, output did not change when validation data was entered into the model after training.
Then, I checked parameters of model using model.parameters(), But weights were same.
What is the problem?

I tried to check whether grad is 0 or not using this code.

loss.backward()

for param in net.parameters():
    print(param.grad.data.sum())

# start debugger
import pdb; pdb.set_trace()

optimizer.step()

But this error is happened.
AttributeError: ‘NoneType’ object has no attribute ‘data’

Also, I changed learning_rate 0.001 to 0.01, But nothing changed.

Batch sizes, etc. are set to small values for testing purposes.

How can I fix it ?

# from positional_encoding import positional_encoding
import torch
import torch.nn as nn
import time
import numpy as np
import random
import copy
import math
# from timm.scheduler import CosineLRScheduler
from transformers import get_cosine_schedule_with_warmup
from encoder_MRL import TransformerEncoder
from loss_MRL import MarginRankingLoss
from make_data import positional_encoding
from make_input import make_input_test

def train(model): 
    model.train()
    batch_size = 10
    train_loss_list = []
    #乱数
    l = random.sample(range(0,1000), k=1000)
    # l = random.sample(range(0,287113), k=287113)
    # batch_loss = []
    count = 0
    #ミニバッチ学習 (minibatch training)
    iterate = math.ceil(1000 / batch_size) #小数点以下繰り上げ
    for batch in range(iterate):
        if batch == iterate: #最後のバッチ (last batch)
            batch_size = 287113 % batch_size
        train_data_, train_answer_, SBse_pos_, ExSum_id_ = [], [], [], []
        # for i in l[count:count+batch_size]:
        for i in range(1):
            train_data = np.load(f'/home1/user/sata/train_data_MRL/{i}.npy')
            train_answer = np.load(f'./cnn_dm/train_answer_MRL/{i}.npy')
            SBse_pos = np.load(f'./cnn_dm/train_SBse_pos/{i}.npy')
            ExSum_id = np.load(f'./cnn_dm/train_ExSum_id_MRL/{i}.npy')
            
        # for i in range(10):
            # train_data_.append(train_data)
            # train_answer_.append(train_answer)
            # SBse_pos_.append(SBse_pos)
            # ExSum_id_.append(ExSum_id)

            train_answer = torch.tensor(train_answer, dtype=torch.float32)  
            # train_data_ = np.array(train_data_)
            # print(train_data_.shape)

            input = torch.tensor(train_data, dtype=torch.float32).to(device)
            cos_sim, output = model(input, SBse_pos) #output: encoderの出力
            output = output.to('cpu')
            # cos_sim = np.array(cos_sim)
            # print(cos_sim.shape, flush=True)
            # output.detach().numpy()
            # print(output.shape, flush=True)

        #cos類似度(cos_sim)を並び替え (sort cos_sim)
        cos_sim_sorted = []
        for j in ExSum_id:
            cos_sim_sorted.append(cos_sim[j])
        cos_sim_sorted = torch.tensor(cos_sim)

        #outputのソート, 抽出 (sort output)
        output_sorted = []
        for i in ExSum_id[:3]:
            output_sorted.append(output[i].tolist())
        output_sorted = torch.tensor(output_sorted)

        criterion = MarginRankingLoss(0.001) #引数はmarginの値
        loss = criterion.get_loss(cos_sim_sorted, train_answer, output_sorted)
        # loss.detach().numpy()/
        # print(loss.shape, flush=True)
        # batch_loss.append(loss) #temp
    
        # loss = sum(batch_loss) / len(batch_loss) #temp
        # loss.clone().detach().requires_grad_(True)
        # loss = torch.tensor(loss, requires_grad=True) ##temp
        loss.requires_grad_(True)

        #重みの更新 (update weights)
        optimizer.zero_grad()
        loss.backward()
        for n, param in model.named_parameters():
            if param.grad is not None:
                print(n, param.grad.data.sum())
        optimizer.step()

        train_loss_list.append(loss)
        batch_loss = []
        
        if (count) % 1000 == 0:
            print(f'Epoch {epoch+1} Deta {count} Loss {loss}', flush=True)

        count += batch_size

    train_loss_list = torch.tensor(train_loss_list).to('cpu').detach().numpy().copy()
    train_loss = np.mean(train_loss_list)

    return train_loss

def evaluate(model):
    model.eval()
    with torch.no_grad():
        valid_loss_list = []
        for i in range(100):
        # for i in range(13368):
            #データの読み込み (load data)
            valid_data = np.load(f'/home1/user/sata/validation_data_MRL/{i}.npy')
            valid_answer = np.load(f'./cnn_dm/validation_answer_MRL/{i}.npy')
            SBse_pos = np.load(f'./cnn_dm/validation_SBse_pos/{i}.npy')
            ExSum_id = np.load(f'./cnn_dm/validation_ExSum_id_MRL/{i}.npy')
            
            #torch.tensorに変換 np.array to torch.tensor
            # valid_data = torch.tensor(valid_data, dtype=torch.float32)
            valid_answer = torch.tensor(valid_answer, dtype=torch.float32)

            input = torch.tensor(valid_data, dtype=torch.float32).to(device)
            cos_sim, output = model(input, SBse_pos)
            output = output.to('cpu')
            #cos類似度(cos_sim)を並び替え (sort cos_sim)
            cos_sim_sorted = []
            for j in ExSum_id:
                cos_sim_sorted.append(cos_sim[j])
            cos_sim_sorted = torch.tensor(cos_sim)
        
            #outputのソート, 抽出 (sort output)
            output_sorted = []
            for i in ExSum_id[:3]:
                output_sorted.append(output[i].tolist())
            output_sorted = torch.tensor(output_sorted)

            criterion = MarginRankingLoss(0.001) #引数はmarginの値 (0.001 is margin)
            loss = criterion.get_loss(cos_sim_sorted, valid_answer, output_sorted)

            valid_loss_list.append(loss)
            if (i+1) % 100 == 0:
            # if (i+1) % 1000 == 0:
                print(f'Deta {i+1} Loss {loss}', flush=True)
        valid_loss_list = torch.tensor(valid_loss_list).to('cpu').detach().numpy().copy()
        valid_loss = np.mean(valid_loss_list)
    return valid_loss

if __name__ == '__main__':
    device = torch.device('cuda:0')

    num_layers = 6
    d_model = 768
    num_heads = 8
    dff = 3072
    # input_limit = 100
    dropout_rate = 0.1
    # batch_size = 10
    batch_size = 100
    lr = 0.001
    # lr = 0.01
    EPOCHS = 30

    print('loading models...', flush=True)
    model = TransformerEncoder(d_model, num_heads, dff, dropout_rate, num_layers).to(device)
    optimizer = torch.optim.Adam(model.parameters(), lr=lr, betas=(0.9, 0.98), eps=1e-9)
    scheduler = get_cosine_schedule_with_warmup(optimizer, num_warmup_steps=3, num_training_steps=EPOCHS)
    # scheduler = CosineLRScheduler(optimizer, t_initial=100, lr_min=1e-6, warmup_t=3, warmup_lr_init=1e-6, warmup_prefix=True)
    print('models are ready !\n', flush=True)

    checkpoint_path = './checkpoints/model.pt'

    for epoch in range(EPOCHS):
        start = time.time()

        print('training now...', flush=True)
        train_loss = train(model)
        print('evaluating now...', flush=True)
        valid_loss = evaluate(model)
        print('processing scheduler.step() ...', flush=True)
        scheduler.step()
        # scheduler.step(epoch)
        
        print(f'Epoch {epoch+1} train_Loss {train_loss}', flush=True)
        print(f'Epoch {epoch+1} valid_Loss {valid_loss}', flush=True)
        print(f'Time taken for 1 epoch: {time.time() - start} secs\n', flush=True)
        
        #チェックポイントの保存 (save checkpoints)
        if (epoch + 1) % 5 == 0:
            print('saving checkpoint...\n', flush=True)
            torch.save({
            'model_state_dict': model.state_dict(),
            'optimizer_state_dict': optimizer.state_dict(),
            "scheduler": scheduler.state_dict(),
            # 'epoch': epoch,
            # 'loss': loss,
            # }, checkpoint_path)
            }, f'./checkpoints/model-{epoch+1}_MRL.pt')

    print('training is done', flush=True)
2

You are detaching the outputs of your model by rewrapping them into tensors. The backward() call will then raise a proper error which you masked by calling .requires_grad_(True) on the loss. Use torch.stack instead of recreating a new tensor and it should work.

1 Like
3

I understand.
Is there a problem with calculating the loss after sorting the output of the model?
Thank you for your help.

Edit
Or should they be sorted within the model?

4

Sorting is not the problem, detaching the computation graph by creating new tensors is.

1 Like
5

Thank you.
I modified the code like this.

"""encoder_MRL.py"""
import torch
import torch.nn as nn
import numpy as np

class TransformerEncoder(nn.Module):
    def __init__(self, d_model, num_heads, dff, dropout_rate, num_layers):
        super().__init__()
        encoder_layer = nn.TransformerEncoderLayer(d_model, num_heads, dff, dropout_rate)
        self.transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers)

    def forward(self, input, SBse_pos, ExSum_id): 
        output = self.transformer_encoder(input) #(len(article),100,768)

        indices = torch.arange(output.size(0))
        MTse = output[indices, indices]
 
        cos_sim = torch.cosine_similarity(MTse, SBse_pos, dim=1)
        cos_sim_sorted = torch.gather(cos_sim, 0, ExSum_id)

        return cos_sim_sorted, output

"""train.py"""
cos_sim, output = model(input, SBse_pos, ExSum_id)
criterion = MarginRankingLoss(0.001) 
loss = criterion.get_loss(cos_sim, train_answer, output)
loss.requires_grad_(True)
optimizer.zero_grad()
loss.backward()
optimizer.step()

But validation loss didn’t cahange. When w/o loss.requires_grad_(True), This error happened.

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

I think It seem to detaching the computation graph. Which code make this problem?
I don’t think I rewrapped them into tensors.
If you don’t mind, could you give me the code to fix it?

editted:
I was using “TotalLoss = torch.tensor(0, dtype=torch.float32).to(device)” to initialize TotalLoss to 0.
After changing code initializing TotalLoss, RuntimeError no longer occurs w/o loss.requires_grad_(True). But valid_loss doesn’t change.

"""loss_MRL.py"""
import torch 
import torch.nn as nn
device = torch.device('cuda:0')

class MarginRankingLoss():
    def __init__(self, margin):
        super(MarginRankingLoss, self).__init__()
        # self._init_paramm_map
        self.margin = margin
        self.loss_func = nn.MarginRankingLoss(margin)

    def get_loss(self, cos_sim, real, pred):
        # TotalLoss = torch.tensor(0, dtype=torch.float32).to(device)
        # initializing TotalLoss to 0
        ones = torch.ones(cos_sim.size()).to(device)
        loss_func = torch.nn.MarginRankingLoss(0.0)
        TotalLoss = loss_func(cos_sim, cos_sim, ones)

        #MRL
        n = cos_sim.size(0)
        for i in range(1,n):
            pos_score = torch.tensor(cos_sim[:-i], dtype=torch.float32).to(device)
            neg_score = torch.tensor(cos_sim[i:], dtype=torch.float32).to(device)
            ones = torch.ones(pos_score.size()).to(device)
            loss_func = torch.nn.MarginRankingLoss(self.margin * i)
            TotalLoss += loss_func(pos_score, neg_score, ones)
        
        return TotalLoss
6

I fixed it modifying code like this. valid_loss now changes as training progresses.

"""loss_MRL.py"""
pos_score = cos_sim[:-i]
neg_score = cos_sim[i:]