python - GPU 上的训练速度会随着时间变慢

Question

我的模型训练速度becomes slower over time .每个时代都需要更长的时间来训练。
这是full source code带着我的预处理情绪 tree bank data (将 glove.840B.300d.txt 放入 data/glove 中)。
安装一些python包:

pip install meowlogtool
pip install tqdm

运行命令:

python sentiment.py --emblr 0 --rel_dim 0 --tag_dim 0 --optim adagrad --name basic --lr 0.05 --wd 1e-4 --at_hid_dim 0

模型源码供大家阅读

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable as Var
import utils
import Constants
from model import SentimentModule
from embedding_model import EmbeddingModel

class SimpleGRU(nn.Module):
    """
    w[i] : (300, 1)
    h[i] : (150, 1)
    p[i] : (20, 1)
    r[i] : (20, 1)
    k[i] : (150, 1)
    x[i] : (20 + 150 + 300 + 20 = 490, 1) (490, 1)
    Uz, Ur, Uh : (150, 150) => 67500 => (450, 450)
    Wz, Wr, Wh : (150, 20 + 150 + 300 + 20) (150, 490)
    """
    def __init__(self, cuda, in_dim, hid_dim):
        super(SimpleGRU, self).__init__()
        self.cudaFlag = cuda

        self.Uz = nn.Linear(hid_dim, hid_dim)
        self.Ur = nn.Linear(hid_dim, hid_dim)
        self.Uh = nn.Linear(hid_dim, hid_dim)

        self.Wz = nn.Linear(in_dim, hid_dim)
        self.Wr = nn.Linear(in_dim, hid_dim)
        self.Wh = nn.Linear(in_dim, hid_dim)

        if self.cudaFlag:
            self.Uz = self.Uz.cuda()
            self.Ur = self.Uz.cuda()
            self.Uh = self.Uz.cuda()

            self.Wz = self.Wz.cuda()
            self.Wr = self.Wr.cuda()
            self.Wh = self.Wh.cuda()

    def forward(self, x, h_prev):
        """
    Simple-GRU(compress_x[v], h[t-1]) :
    z[t]         := s(Wz *compress_x[t]+ Uz * h[t-1] + bz)
    r[t]         := s(Wr * compress_x[t] + Ur * h[t-1] + br)
    h_temp[t]     := g(Wh * compress_x[t] + Uh * h[t-1] + bh)
    h[t]         := r[t] .* h[t-1] + (1 - z[t]) .* h_temp[t]
    return h[t]
        :param x: compress_x[t]
        :param h_prev: h[t-1]
        :return:
        """
        z = F.sigmoid(self.Wz(x) + self.Uz(h_prev))
        r = F.sigmoid(self.Wr(x) + self.Ur(h_prev))
        h_temp = F.tanh(self.Wh(x) + self.Uh(h_prev))
        h = r*h_prev + (1-z)*h_temp
        return h



class TreeSimpleGRU(nn.Module):
    def __init__(self, cuda, word_dim, tag_dim, rel_dim, mem_dim, at_hid_dim, criterion, leaf_h = None):
        super(TreeSimpleGRU, self).__init__()
        self.cudaFlag = cuda
        # self.gru_cell = nn.GRUCell(word_dim + tag_dim, mem_dim)
        self.gru_cell = SimpleGRU(self.cudaFlag, word_dim+tag_dim, mem_dim)
        self.gru_at = GRU_AT(self.cudaFlag, word_dim + tag_dim + rel_dim + mem_dim, at_hid_dim ,mem_dim)
        self.mem_dim = mem_dim
        self.in_dim = word_dim
        self.tag_dim = tag_dim
        self.rel_dim = rel_dim
        self.leaf_h = leaf_h # init h for leaf node
        if self.leaf_h == None:
            self.leaf_h = Var(torch.rand(1, self.mem_dim))
            torch.save(self.leaf_h, 'leaf_h.pth')

        if self.cudaFlag:
            self.leaf_h = self.leaf_h.cuda()

        self.criterion = criterion
        self.output_module = None


    def getParameters(self):
        """
        Get flatParameters
        note that getParameters and parameters is not equal in this case
        getParameters do not get parameters of output module
        :return: 1d tensor
        """
        params = []
        for m in [self.gru_cell, self.gru_at]:
            # we do not get param of output module
            l = list(m.parameters())
            params.extend(l)

        one_dim = [p.view(p.numel()) for p in params]
        params = F.torch.cat(one_dim)
        return params


    def set_output_module(self, output_module):
        self.output_module = output_module

    def forward(self, tree, w_emb, tag_emb, rel_emb, training = False):
        loss = Var(torch.zeros(1))  # init zero loss
        if self.cudaFlag:
            loss = loss.cuda()

        for idx in xrange(tree.num_children):
            _, child_loss = self.forward(tree.children[idx], w_emb, tag_emb, rel_emb, training)
            loss = loss + child_loss

        if tree.num_children > 0:
            child_rels, child_k  = self.get_child_state(tree, rel_emb)
            if self.tag_dim > 0:
                tree.state = self.node_forward(w_emb[tree.idx - 1], tag_emb[tree.idx -1], child_rels, child_k)
            else:
                tree.state = self.node_forward(w_emb[tree.idx - 1], None, child_rels, child_k)
        elif tree.num_children == 0:
            if self.tag_dim > 0:
                tree.state = self.leaf_forward(w_emb[tree.idx - 1], tag_emb[tree.idx -1])
            else:
                tree.state = self.leaf_forward(w_emb[tree.idx - 1], None)

        if self.output_module != None:
            output = self.output_module.forward(tree.state, training)
            tree.output = output
            if training and tree.gold_label != None:
                target = Var(utils.map_label_to_target_sentiment(tree.gold_label))
                if self.cudaFlag:
                    target = target.cuda()
                loss = loss + self.criterion(output, target)
        return tree.state, loss


    def leaf_forward(self, word_emb, tag_emb):
        """
        Forward function for leaf node
        :param word_emb:  word embedding of current node u
        :param tag_emb: tag embedding of current node u
        :return: k of current node u
        """
        h = self.leaf_h
        if self.cudaFlag:
            h = h.cuda()
        if self.tag_dim > 0:
            x = F.torch.cat([word_emb, tag_emb], 1)
        else:
            x = word_emb
        k = self.gru_cell(x, h)
        return k


    def node_forward(self, word_emb, tag_emb, child_rels, child_k):
        """
        Foward function for inner node
        :param word_emb: word embedding of current node u
        :param tag_emb: tag embedding of current node u
        :param child_rels (tensor): rels embedding of child node v
        :param child_k (tensor): k of child node v
        :return:
        """
        n_child = child_k.size(0)
        h = Var(torch.zeros(1, self.mem_dim))
        if self.cudaFlag:
            h = h.cuda()

        for i in range(0, n_child):
            k = child_k[i]
            x_list = [word_emb, k]
            if self.rel_dim >0:
                rel = child_rels[i]
                x_list.append(rel)
            if self.tag_dim > 0:
                x_list.append(tag_emb)
            x = F.torch.cat(x_list, 1)
            h = self.gru_at(x, h)
        k = h
        return k

    def get_child_state(self, tree, rels_emb):
        """
        Get child rels, get child k
        :param tree: tree we need to get child
        :param rels_emb (tensor):
        :return:
        """
        if tree.num_children == 0:
            assert False #  never get here
        else:
            child_k = Var(torch.Tensor(tree.num_children, 1, self.mem_dim))
            if self.rel_dim>0:
                child_rels = Var(torch.Tensor(tree.num_children, 1, self.rel_dim))
            else:
                child_rels = None
            if self.cudaFlag:
                child_k = child_k.cuda()
                if self.rel_dim > 0:
                    child_rels = child_rels.cuda()
            for idx in xrange(tree.num_children):
                child_k[idx] = tree.children[idx].state
                if self.rel_dim > 0:
                    child_rels[idx] = rels_emb[tree.children[idx].idx - 1]
        return child_rels, child_k

class AT(nn.Module):
    """
    AT(compress_x[v]) := sigmoid(Wa * tanh(Wb * compress_x[v] + bb) + ba)
    """
    def __init__(self, cuda, in_dim, hid_dim):
        super(AT, self).__init__()
        self.cudaFlag = cuda
        self.in_dim = in_dim
        self.hid_dim = hid_dim

        self.Wa = nn.Linear(hid_dim, 1)
        self.Wb = nn.Linear(in_dim, hid_dim)

        if self.cudaFlag:
            self.Wa = self.Wa.cuda()
            self.Wb = self.Wb.cuda()

    def forward(self, x):
        out = F.sigmoid(self.Wa(F.tanh(self.Wb(x))))
        return out


class GRU_AT(nn.Module):
    def __init__(self, cuda, in_dim, at_hid_dim ,mem_dim):
        super(GRU_AT, self).__init__()
        self.cudaFlag = cuda
        self.in_dim = in_dim
        self.mem_dim = mem_dim
        self.at_hid_dim = at_hid_dim
        if at_hid_dim > 0:
            self.at = AT(cuda, in_dim, at_hid_dim)
        self.gru_cell = SimpleGRU(self.cudaFlag, in_dim, mem_dim)

        if self.cudaFlag:
            if at_hid_dim > 0:
                self.at = self.at.cuda()
            self.gru_cell = self.gru_cell.cuda()

    def forward(self, x, h_prev):
        """

        :param x:
        :param h_prev:
        :return: a * m + (1 - a) * h[t-1]
        """
        m = self.gru_cell(x, h_prev)
        if self.at_hid_dim > 0:
            a = self.at.forward(x)
            h = torch.mm(a, m) + torch.mm((1-a), h_prev)
        else:
            h = m
        return h

class TreeGRUSentiment(nn.Module):
    def __init__(self, cuda, in_dim, tag_dim, rel_dim, mem_dim, at_hid_dim, num_classes, criterion):
        super(TreeGRUSentiment, self).__init__()
        self.cudaFlag = cuda
        self.tree_module = TreeSimpleGRU(cuda, in_dim, tag_dim, rel_dim, mem_dim, at_hid_dim, criterion)
        self.output_module = SentimentModule(cuda, mem_dim, num_classes, dropout=True)
        self.tree_module.set_output_module(self.output_module)

    def get_tree_parameters(self):
        return self.tree_module.getParameters()

    def forward(self, tree, sent_emb, tag_emb, rel_emb, training = False):
        # sent_emb = F.torch.unsqueeze(self.word_embedding.forward(sent_inputs), 1)
        # tag_emb = F.torch.unsqueeze(self.tag_emb.forward(tag_inputs), 1)
        # rel_emb = F.torch.unsqueeze(self.rel_emb.forward(rel_inputs), 1)
        # sent_emb, tag_emb, rel_emb = self.embedding_model(sent_inputs, tag_inputs, rel_inputs)

        tree_state, loss = self.tree_module(tree, sent_emb, tag_emb, rel_emb, training)
        output = tree.output
        return output, loss

Answer 1

Why does neural network learning slow down as the error gets lower?

The reasons for the slowdown are not fully understood, but we have some basic ideas.

For classifiers, most training examples start out as incorrectly classified. Over time, more of them become correctly classified. Early in learning, you might have a nearly 100% error rate, so every example in the minibatch contributes to learning. Late in learning, you might have nearly a 0% error rate, so almost none of the examples in the minibatch contribute to learning. This problem can be resolved to some extent by using hard example mining or importance sampling. Both of these are just techniques for training on more difficult examples more often.

There are other more complicated reasons. One of them is that the condition number of the Hessian tends to worsen a lot as learning progresses, so that the optimal step size becomes smaller and smaller.