090944e816fd3ff8e861fba4452851e0a901491d,scripts/language_model/large_word_language_model.py,,train,#,193

Before Change

                for j, (X, y, m, s, h) in enumerate(zip(data, target, mask, sample, hiddens)):
                    output, h, new_target = model(X, y, h, s)
                    output = output.reshape((-3, -1))
                    new_target = new_target.reshape((-1,))
                    l = loss(output, new_target) * m.reshape((-1,))
                    Ls.append(l/args.batch_size)
                    hiddens[j] = h

            autograd.backward(Ls)

After Change

            for _, batch in enumerate(zip(data, target, mask, sample, hiddens)):
                parallel.put(batch)

            for _ in range(len(data)):
                hidden, ls = parallel.get()
                // hidden states are ordered by context id
                index = context.index(hidden[0].context)

Instances