changed for movie dataset

TaNP.py

@@ -0,0 +1,245 @@
+import torch
+import numpy as np
+from random import randint
+from copy import deepcopy
+from torch.autograd import Variable
+from torch.nn import functional as F
+from collections import OrderedDict
+from embeddings_TaNP import Item, User, Encoder, MuSigmaEncoder, Decoder, Gating_Decoder, TaskEncoder, MemoryUnit,Movie_item,Movie_user
+import torch.nn as nn
+
+class NP(nn.Module):
+    def __init__(self, config):
+        super(NP, self).__init__()
+
+        # change for Movie lens
+        # self.x_dim = config['second_embedding_dim'] * 2
+        self.x_dim = 32 * 8
+
+
+        # use one-hot or not?
+        self.y_dim = 1
+        self.z1_dim = config['z1_dim']
+        self.z2_dim = config['z2_dim']
+        # z is the dimension size of mu and sigma.
+        self.z_dim = config['z_dim']
+        # the dimension size of rc.
+        self.enc_h1_dim = config['enc_h1_dim']
+        self.enc_h2_dim = config['enc_h2_dim']
+
+        self.dec_h1_dim = config['dec_h1_dim']
+        self.dec_h2_dim = config['dec_h2_dim']
+        self.dec_h3_dim = config['dec_h3_dim']
+
+        self.taskenc_h1_dim = config['taskenc_h1_dim']
+        self.taskenc_h2_dim = config['taskenc_h2_dim']
+        self.taskenc_final_dim = config['taskenc_final_dim']
+
+        self.clusters_k = config['clusters_k']
+        self.temperture = config['temperature']
+        self.dropout_rate = config['dropout_rate']
+
+        # Initialize networks
+
+        # change for Movie Lens
+        # self.item_emb = Item(config)
+        # self.user_emb = User(config)
+        self.item_emb = Movie_item(config)
+        self.user_emb = Movie_user(config)
+
+        # This encoder is used to generated z actually, it is a latent encoder in ANP.
+        self.xy_to_z = Encoder(self.x_dim, self.y_dim, self.enc_h1_dim, self.enc_h2_dim, self.z1_dim, self.dropout_rate)
+        self.z_to_mu_sigma = MuSigmaEncoder(self.z1_dim, self.z2_dim, self.z_dim)
+        # This encoder is used to generated r actually, it is a deterministic encoder in ANP.
+        self.xy_to_task = TaskEncoder(self.x_dim, self.y_dim, self.taskenc_h1_dim, self.taskenc_h2_dim, self.taskenc_final_dim,
+                                      self.dropout_rate)
+        self.memoryunit = MemoryUnit(self.clusters_k, self.taskenc_final_dim, self.temperture)
+        #self.xz_to_y = Gating_Decoder(self.x_dim, self.z_dim, self.taskenc_final_dim, self.dec_h1_dim, self.dec_h2_dim, self.dec_h3_dim, self.y_dim, self.dropout_rate)
+        self.xz_to_y = Decoder(self.x_dim, self.z_dim, self.taskenc_final_dim, self.dec_h1_dim, self.dec_h2_dim, self.dec_h3_dim, self.y_dim, self.dropout_rate)
+
+    def aggregate(self, z_i):
+        return torch.mean(z_i, dim=0)
+
+    def xy_to_mu_sigma(self, x, y):
+        # Encode each point into a representation r_i
+        z_i = self.xy_to_z(x, y)
+        # Aggregate representations r_i into a single representation r
+        z = self.aggregate(z_i)
+        # Return parameters of distribution
+        return self.z_to_mu_sigma(z)
+
+    # embedding each (item, user) as the x for np
+    def embedding(self, x):
+        # change for Movie lens
+        rate_idx = Variable(x[:, 0], requires_grad=False)
+        genre_idx = Variable(x[:, 1:26], requires_grad=False)
+        director_idx = Variable(x[:, 26:2212], requires_grad=False)
+        actor_idx = Variable(x[:, 2212:10242], requires_grad=False)
+        gender_idx = Variable(x[:, 10242], requires_grad=False)
+        age_idx = Variable(x[:, 10243], requires_grad=False)
+        occupation_idx = Variable(x[:, 10244], requires_grad=False)
+        area_idx = Variable(x[:, 10245], requires_grad=False)
+        item_emb = self.item_emb(rate_idx, genre_idx, director_idx, actor_idx)
+        user_emb = self.user_emb(gender_idx, age_idx, occupation_idx, area_idx)
+        x = torch.cat((item_emb, user_emb), 1)
+        return x
+
+    def forward(self, x_context, y_context, x_target, y_target):
+        # change for Movie lens
+        x_context_embed = self.embedding(x_context)
+        x_target_embed = self.embedding(x_target)
+
+        if self.training:
+            # sigma is log_sigma actually
+            mu_target, sigma_target, z_target = self.xy_to_mu_sigma(x_target_embed, y_target)
+            mu_context, sigma_context, z_context = self.xy_to_mu_sigma(x_context_embed, y_context)
+            task = self.xy_to_task(x_context_embed, y_context)
+            mean_task = self.aggregate(task)
+            C_distribution, new_task_embed = self.memoryunit(mean_task)
+            p_y_pred = self.xz_to_y(x_target_embed, z_target, new_task_embed)
+            return p_y_pred, mu_target, sigma_target, mu_context, sigma_context, C_distribution
+        else:
+            mu_context, sigma_context, z_context = self.xy_to_mu_sigma(x_context_embed, y_context)
+            task = self.xy_to_task(x_context_embed, y_context)
+            mean_task = self.aggregate(task)
+            C_distribution, new_task_embed = self.memoryunit(mean_task)
+            p_y_pred = self.xz_to_y(x_target_embed, z_context, new_task_embed)
+            return p_y_pred
+
+
+class Trainer(torch.nn.Module):
+    def __init__(self, config):
+        self.opt = config
+        super(Trainer, self).__init__()
+        self.use_cuda = config['use_cuda']
+        self.np = NP(self.opt)
+        self._lambda = config['lambda']
+        self.optimizer = torch.optim.Adam(self.np.parameters(), lr=config['lr'])
+
+    # our kl divergence
+    def kl_div(self, mu_target, logsigma_target, mu_context, logsigma_context):
+        target_sigma = torch.exp(logsigma_target)
+        context_sigma = torch.exp(logsigma_context)
+        kl_div = (logsigma_context - logsigma_target) - 0.5 + (((target_sigma ** 2) + (mu_target - mu_context) ** 2) / 2 * context_sigma ** 2)
+        #kl_div = (t.exp(posterior_var) + (posterior_mu-prior_mu) ** 2) / t.exp(prior_var) - 1. + (prior_var - posterior_var)
+        #kl_div = 0.5 * kl_div.sum()
+        kl_div = kl_div.sum()
+        return kl_div
+
+    # new kl divergence -- kl(st|sc)
+    def new_kl_div(self, prior_mu, prior_var, posterior_mu, posterior_var):
+        kl_div = (torch.exp(posterior_var) + (posterior_mu-prior_mu) ** 2) / torch.exp(prior_var) - 1. + (prior_var - posterior_var)
+        kl_div = 0.5 * kl_div.sum()
+        return kl_div
+
+    def loss(self, p_y_pred, y_target, mu_target, sigma_target, mu_context, sigma_context):
+        #print('p_y_pred size is ', p_y_pred.size())
+        regression_loss = F.mse_loss(p_y_pred, y_target.view(-1, 1))
+        #print('regession loss size is ', regression_loss.size())
+        # kl divergence between target and context
+        #print('regession_loss is ', regression_loss.item())
+        kl = self.new_kl_div(mu_context, sigma_context, mu_target, sigma_target)
+        #print('KL_loss is ', kl.item())
+        return regression_loss+kl
+
+    def context_target_split(self, support_set_x, support_set_y, query_set_x, query_set_y):
+        total_x = torch.cat((support_set_x, query_set_x), 0)
+        total_y = torch.cat((support_set_y, query_set_y), 0)
+        total_size = total_x.size(0)
+        context_min = self.opt['context_min']
+        context_max = self.opt['context_max']
+        extra_tar_min = self.opt['target_extra_min']
+        #here we simply use the total_size as the maximum of target size.
+        num_context = randint(context_min, context_max)
+        num_target = randint(extra_tar_min, total_size - num_context)
+        sampled = np.random.choice(total_size, num_context+num_target, replace=False)
+        x_context = total_x[sampled[:num_context], :]
+        y_context = total_y[sampled[:num_context]]
+        x_target = total_x[sampled, :]
+        y_target = total_y[sampled]
+        return x_context, y_context, x_target, y_target
+
+    def new_context_target_split(self, support_set_x, support_set_y, query_set_x, query_set_y):
+        total_x = torch.cat((support_set_x, query_set_x), 0)
+        total_y = torch.cat((support_set_y, query_set_y), 0)
+        total_size = total_x.size(0)
+        context_min = self.opt['context_min']
+
+        # change for Movie lens
+        context_min = min(context_min, total_size - 1)
+
+        num_context = np.random.randint(context_min, total_size)
+        num_target = np.random.randint(0, total_size - num_context)
+        sampled = np.random.choice(total_size, num_context+num_target, replace=False)
+        x_context = total_x[sampled[:num_context], :]
+        y_context = total_y[sampled[:num_context]]
+        x_target = total_x[sampled, :]
+        y_target = total_y[sampled]
+        return x_context, y_context, x_target, y_target
+
+    def global_update(self, support_set_xs, support_set_ys, query_set_xs, query_set_ys):
+        batch_sz = len(support_set_xs)
+        losses = []
+        C_distribs = []
+        if self.use_cuda:
+            for i in range(batch_sz):
+                support_set_xs[i] = support_set_xs[i].cuda()
+                support_set_ys[i] = support_set_ys[i].cuda()
+                query_set_xs[i] = query_set_xs[i].cuda()
+                query_set_ys[i] = query_set_ys[i].cuda()
+        for i in range(batch_sz):
+            x_context, y_context, x_target, y_target = self.new_context_target_split(support_set_xs[i], support_set_ys[i],
+                                                                                 query_set_xs[i], query_set_ys[i])
+            # print("inja1: x_context_size:",x_context.size())
+            p_y_pred, mu_target, sigma_target, mu_context, sigma_context, C_distribution = self.np(x_context, y_context, x_target,
+                                                                                  y_target)
+            C_distribs.append(C_distribution)
+            loss = self.loss(p_y_pred, y_target, mu_target, sigma_target, mu_context, sigma_context)
+            #print('Each task has loss: ', loss)
+            losses.append(loss)
+        # calculate target distribution for clustering in batch manner.
+        # batchsize * k
+        C_distribs = torch.stack(C_distribs)
+        # batchsize * k
+        C_distribs_sq = torch.pow(C_distribs, 2)
+        # 1*k
+        C_distribs_sum = torch.sum(C_distribs, dim=0, keepdim=True)
+        # batchsize * k
+        temp = C_distribs_sq / C_distribs_sum
+        # batchsize * 1
+        temp_sum = torch.sum(temp, dim=1, keepdim=True)
+        target_distribs = temp / temp_sum
+        # calculate the kl loss
+        clustering_loss = self._lambda * F.kl_div(C_distribs.log(), target_distribs, reduction='batchmean')
+        #print('The clustering loss is %.6f' % (clustering_loss.item()))
+        np_losses_mean = torch.stack(losses).mean(0)
+        total_loss = np_losses_mean + clustering_loss
+        self.optimizer.zero_grad()
+        total_loss.backward()
+        self.optimizer.step()
+        return total_loss.item(), C_distribs.cpu().detach().numpy()
+
+    def query_rec(self, support_set_xs, support_set_ys, query_set_xs, query_set_ys):
+        batch_sz = 1
+        # used for calculating the rmse.
+        losses_q = []
+        if self.use_cuda:
+            for i in range(batch_sz):
+                support_set_xs[i] = support_set_xs[i].cuda()
+                support_set_ys[i] = support_set_ys[i].cuda()
+                query_set_xs[i] = query_set_xs[i].cuda()
+                query_set_ys[i] = query_set_ys[i].cuda()
+        for i in range(batch_sz):
+            #query_set_y_pred = self.forward(support_set_xs[i], support_set_ys[i], query_set_xs[i], num_local_update)
+            query_set_y_pred = self.np(support_set_xs[i], support_set_ys[i], query_set_xs[i], query_set_ys[i])
+            # obtain the mean of gaussian distribution
+            #(interation_size, y_dim)
+            #query_set_y_pred = query_set_y_pred.loc.detach()
+            #print('test_y_pred size is ', query_set_y_pred.size())
+            loss_q = F.mse_loss(query_set_y_pred, query_set_ys[i].view(-1, 1))
+            losses_q.append(loss_q)
+        losses_q = torch.stack(losses_q).mean(0)
+        output_list, recommendation_list = query_set_y_pred.view(-1).sort(descending=True)
+        return losses_q.item(), recommendation_list
+
+

TaNP_training.py

@@ -0,0 +1,36 @@
+import os
+import torch
+import pickle
+import random
+from eval import testing
+
+def training(trainer, opt, train_dataset, test_dataset, batch_size, num_epoch, model_save=True, model_filename=None, logger=None):
+    training_set_size = len(train_dataset)
+    for epoch in range(num_epoch):
+        random.shuffle(train_dataset)
+        num_batch = int(training_set_size / batch_size)
+        a, b, c, d = zip(*train_dataset)
+        trainer.train()
+        all_C_distribs = []
+        for i in range(num_batch):
+            try:
+                supp_xs = list(a[batch_size*i:batch_size*(i+1)])
+                supp_ys = list(b[batch_size*i:batch_size*(i+1)])
+                query_xs = list(c[batch_size*i:batch_size*(i+1)])
+                query_ys = list(d[batch_size*i:batch_size*(i+1)])
+            except IndexError:
+                continue
+            train_loss, batch_C_distribs = trainer.global_update(supp_xs, supp_ys, query_xs, query_ys)
+            all_C_distribs.append(batch_C_distribs)
+
+        P5, NDCG5, MAP5, P7, NDCG7, MAP7, P10, NDCG10, MAP10 = testing(trainer, opt, test_dataset)
+        logger.log(
+            "{}\t{:.6f}\t TOP-5 {:.4f}\t{:.4f}\t{:.4f}\t TOP-7: {:.4f}\t{:.4f}\t{:.4f}"
+            "\t TOP-10: {:.4f}\t{:.4f}\t{:.4f}".
+                format(epoch, train_loss, P5, NDCG5, MAP5, P7, NDCG7, MAP7, P10, NDCG10, MAP10))
+        if epoch == (num_epoch-1):
+            with open('output_att', 'wb') as fp:
+                pickle.dump(all_C_distribs, fp)
+
+    if model_save:
+        torch.save(trainer.state_dict(), model_filename)

embeddings_TaNP.py

@@ -0,0 +1,372 @@
+import torch
+import torch.nn as nn
+import torch.nn.init as init
+import torch.nn.functional as F
+
+class Item(torch.nn.Module):
+    def __init__(self, config):
+        super(Item, self).__init__()
+        self.feature_dim = config['if_dim']
+        self.first_embedding_dim = config['first_embedding_dim']
+        self.second_embedding_dim = config['second_embedding_dim']
+
+        self.first_embedding_layer = torch.nn.Linear(
+            in_features=self.feature_dim,
+            out_features=self.first_embedding_dim,
+            bias=True
+        )
+
+        self.second_embedding_layer = torch.nn.Linear(
+            in_features=self.first_embedding_dim,
+            out_features=self.second_embedding_dim,
+            bias=True
+        )
+
+    def forward(self, x, vars=None):
+        first_hidden = self.first_embedding_layer(x)
+        first_hidden = F.relu(first_hidden)
+        sec_hidden = self.second_embedding_layer(first_hidden)
+        return F.relu(sec_hidden)
+
+class Movie_item(torch.nn.Module):
+    def __init__(self, config):
+        super(Movie_item, self).__init__()
+        self.num_rate = config['num_rate']
+        self.num_genre = config['num_genre']
+        self.num_director = config['num_director']
+        self.num_actor = config['num_actor']
+        self.embedding_dim = config['embedding_dim']
+
+        # change for Movie
+        self.feature_dim = 4 * self.embedding_dim
+
+        self.embedding_rate = torch.nn.Embedding(
+            num_embeddings=self.num_rate, 
+            embedding_dim=self.embedding_dim
+        )
+        
+        self.embedding_genre = torch.nn.Linear(
+            in_features=self.num_genre,
+            out_features=self.embedding_dim,
+            bias=False
+        )
+        
+        self.embedding_director = torch.nn.Linear(
+            in_features=self.num_director,
+            out_features=self.embedding_dim,
+            bias=False
+        )
+        
+        self.embedding_actor = torch.nn.Linear(
+            in_features=self.num_actor,
+            out_features=self.embedding_dim,
+            bias=False
+        )
+
+    def forward(self, rate_idx, genre_idx, director_idx, actors_idx, vars=None):
+        rate_emb = self.embedding_rate(rate_idx)
+        genre_emb = self.embedding_genre(genre_idx.float()) / torch.sum(genre_idx.float(), 1).view(-1, 1)
+        director_emb = self.embedding_director(director_idx.float()) / torch.sum(director_idx.float(), 1).view(-1, 1)
+        actors_emb = self.embedding_actor(actors_idx.float()) / torch.sum(actors_idx.float(), 1).view(-1, 1)
+        return torch.cat((rate_emb, genre_emb, director_emb, actors_emb), 1)
+
+class User(torch.nn.Module):
+    def __init__(self, config):
+        super(User, self).__init__()
+        self.feature_dim = config['uf_dim']
+        self.first_embedding_dim = config['first_embedding_dim']
+        self.second_embedding_dim = config['second_embedding_dim']
+
+        self.first_embedding_layer = torch.nn.Linear(
+            in_features=self.feature_dim,
+            out_features=self.first_embedding_dim,
+            bias=True
+        )
+
+        self.second_embedding_layer = torch.nn.Linear(
+            in_features=self.first_embedding_dim,
+            out_features=self.second_embedding_dim,
+            bias=True
+        )
+
+    def forward(self, x, vars=None):
+        first_hidden = self.first_embedding_layer(x)
+        first_hidden = F.relu(first_hidden)
+        sec_hidden = self.second_embedding_layer(first_hidden)
+        return F.relu(sec_hidden)
+
+    
+class Movie_user(torch.nn.Module):
+    def __init__(self, config):
+        super(Movie_user, self).__init__()
+        self.num_gender = config['num_gender']
+        self.num_age = config['num_age']
+        self.num_occupation = config['num_occupation']
+        self.num_zipcode = config['num_zipcode']
+        self.embedding_dim = config['embedding_dim']
+
+        self.embedding_gender = torch.nn.Embedding(
+            num_embeddings=self.num_gender,
+            embedding_dim=self.embedding_dim
+        )
+
+        self.embedding_age = torch.nn.Embedding(
+            num_embeddings=self.num_age,
+            embedding_dim=self.embedding_dim
+        )
+
+        self.embedding_occupation = torch.nn.Embedding(
+            num_embeddings=self.num_occupation,
+            embedding_dim=self.embedding_dim
+        )
+
+        self.embedding_area = torch.nn.Embedding(
+            num_embeddings=self.num_zipcode,
+            embedding_dim=self.embedding_dim
+        )
+
+    def forward(self, gender_idx, age_idx, occupation_idx, area_idx):
+        gender_emb = self.embedding_gender(gender_idx)
+        age_emb = self.embedding_age(age_idx)
+        occupation_emb = self.embedding_occupation(occupation_idx)
+        area_emb = self.embedding_area(area_idx)
+        return torch.cat((gender_emb, age_emb, occupation_emb, area_emb), 1)
+    
+class Encoder(nn.Module):
+    #Maps an (x_i, y_i) pair to a representation r_i.
+    # Add the dropout into encoder ---03.31
+    def __init__(self, x_dim, y_dim, h1_dim, h2_dim, z1_dim, dropout_rate):
+        super(Encoder, self).__init__()
+
+        self.x_dim = x_dim
+        self.y_dim = y_dim
+        self.h1_dim = h1_dim
+        self.h2_dim = h2_dim
+        self.z1_dim = z1_dim
+        self.dropout_rate = dropout_rate
+
+        layers = [nn.Linear(self.x_dim + self.y_dim, self.h1_dim),
+                  torch.nn.Dropout(self.dropout_rate),
+                  nn.ReLU(inplace=True),
+                  nn.Linear(self.h1_dim, self.h2_dim),
+                  torch.nn.Dropout(self.dropout_rate),
+                  nn.ReLU(inplace=True),
+                  nn.Linear(self.h2_dim, self.z1_dim)]
+
+        self.input_to_hidden = nn.Sequential(*layers)
+
+    def forward(self, x, y):
+        y = y.view(-1, 1)
+        input_pairs = torch.cat((x, y), dim=1)
+        return self.input_to_hidden(input_pairs)
+
+
+class MuSigmaEncoder(nn.Module):
+    def __init__(self, z1_dim, z2_dim, z_dim):
+        super(MuSigmaEncoder, self).__init__()
+
+        self.z1_dim = z1_dim
+        self.z2_dim = z2_dim
+        self.z_dim = z_dim
+        self.z_to_hidden = nn.Linear(self.z1_dim, self.z2_dim)
+        self.hidden_to_mu = nn.Linear(self.z2_dim, z_dim)
+        self.hidden_to_logsigma = nn.Linear(self.z2_dim, z_dim)
+
+    def forward(self, z_input):
+        hidden = torch.relu(self.z_to_hidden(z_input))
+        mu = self.hidden_to_mu(hidden)
+        log_sigma = self.hidden_to_logsigma(hidden)
+        std = torch.exp(0.5 * log_sigma)
+        eps = torch.randn_like(std)
+        z = eps.mul(std).add_(mu)
+        return mu, log_sigma, z
+
+class TaskEncoder(nn.Module):
+    def __init__(self, x_dim, y_dim, h1_dim, h2_dim, final_dim, dropout_rate):
+        super(TaskEncoder, self).__init__()
+        self.x_dim = x_dim
+        self.y_dim = y_dim
+        self.h1_dim = h1_dim
+        self.h2_dim = h2_dim
+        self.final_dim = final_dim
+        self.dropout_rate = dropout_rate
+        layers = [nn.Linear(self.x_dim + self.y_dim, self.h1_dim),
+                  torch.nn.Dropout(self.dropout_rate),
+                  nn.ReLU(inplace=True),
+                  nn.Linear(self.h1_dim, self.h2_dim),
+                  torch.nn.Dropout(self.dropout_rate),
+                  nn.ReLU(inplace=True),
+                  nn.Linear(self.h2_dim, self.final_dim)]
+
+        self.input_to_hidden = nn.Sequential(*layers)
+
+    def forward(self, x, y):
+        y = y.view(-1, 1)
+        input_pairs = torch.cat((x, y), dim=1)
+        return self.input_to_hidden(input_pairs)
+
+class MemoryUnit(nn.Module):
+    # clusters_k is k keys
+    def __init__(self, clusters_k, emb_size, temperature):
+        super(MemoryUnit, self).__init__()
+        self.clusters_k = clusters_k
+        self.embed_size = emb_size
+        self.temperature = temperature
+        self.array = nn.Parameter(init.xavier_uniform_(torch.FloatTensor(self.clusters_k, self.embed_size)))
+
+    def forward(self, task_embed):
+        res = torch.norm(task_embed-self.array, p=2, dim=1, keepdim=True)
+        res = torch.pow((res / self.temperature) + 1, (self.temperature + 1) / -2)
+        # 1*k
+        C = torch.transpose(res / res.sum(), 0, 1)
+        # 1*k, k*d, 1*d
+        value = torch.mm(C, self.array)
+        # simple add operation
+        new_task_embed = value + task_embed
+        # calculate target distribution
+        return C, new_task_embed
+
+
+class Decoder(nn.Module):
+    """
+    Maps target input x_target and z, r to predictions y_target.
+    """
+    def __init__(self, x_dim, z_dim, task_dim, h1_dim, h2_dim, h3_dim, y_dim, dropout_rate):
+        super(Decoder, self).__init__()
+        self.x_dim = x_dim
+        self.z_dim = z_dim
+        self.task_dim = task_dim
+        self.h1_dim = h1_dim
+        self.h2_dim = h2_dim
+        self.h3_dim = h3_dim
+        self.y_dim = y_dim
+        self.dropout_rate = dropout_rate
+        self.dropout = nn.Dropout(self.dropout_rate)
+
+        self.hidden_layer_1 = nn.Linear(self.x_dim + self.z_dim, self.h1_dim)
+        self.hidden_layer_2 = nn.Linear(self.h1_dim, self.h2_dim)
+        self.hidden_layer_3 = nn.Linear(self.h2_dim, self.h3_dim)
+
+        self.film_layer_1_beta = nn.Linear(self.task_dim, self.h1_dim, bias=False)
+        self.film_layer_1_gamma = nn.Linear(self.task_dim, self.h1_dim, bias=False)
+        self.film_layer_2_beta = nn.Linear(self.task_dim, self.h2_dim, bias=False)
+        self.film_layer_2_gamma = nn.Linear(self.task_dim, self.h2_dim, bias=False)
+        self.film_layer_3_beta = nn.Linear(self.task_dim, self.h3_dim, bias=False)
+        self.film_layer_3_gamma = nn.Linear(self.task_dim, self.h3_dim, bias=False)
+
+        self.final_projection = nn.Linear(self.h3_dim, self.y_dim)
+
+    def forward(self, x, z, task):
+        interaction_size, _ = x.size()
+        z = z.unsqueeze(0).repeat(interaction_size, 1)
+        # Input is concatenation of z with every row of x
+        inputs = torch.cat((x, z), dim=1)
+        hidden_1 = self.hidden_layer_1(inputs)
+        beta_1 = torch.tanh(self.film_layer_1_beta(task))
+        gamma_1 = torch.tanh(self.film_layer_1_gamma(task))
+        hidden_1 = torch.mul(hidden_1, gamma_1) + beta_1
+        hidden_1 = self.dropout(hidden_1)
+        hidden_2 = F.relu(hidden_1)
+
+        hidden_2 = self.hidden_layer_2(hidden_2)
+        beta_2 = torch.tanh(self.film_layer_2_beta(task))
+        gamma_2 = torch.tanh(self.film_layer_2_gamma(task))
+        hidden_2 = torch.mul(hidden_2, gamma_2) + beta_2
+        hidden_2 = self.dropout(hidden_2)
+        hidden_3 = F.relu(hidden_2)
+
+        hidden_3 = self.hidden_layer_3(hidden_3)
+        beta_3 = torch.tanh(self.film_layer_3_beta(task))
+        gamma_3 = torch.tanh(self.film_layer_3_gamma(task))
+        hidden_final = torch.mul(hidden_3, gamma_3) + beta_3
+        hidden_final = self.dropout(hidden_final)
+        hidden_final = F.relu(hidden_final)
+
+        y_pred = self.final_projection(hidden_final)
+        return y_pred
+
+
+
+class Gating_Decoder(nn.Module):
+
+    def __init__(self, x_dim, z_dim, task_dim, h1_dim, h2_dim, h3_dim, y_dim, dropout_rate):
+        super(Gating_Decoder, self).__init__()
+        self.x_dim = x_dim
+        self.z_dim = z_dim
+        self.task_dim = task_dim
+        self.h1_dim = h1_dim
+        self.h2_dim = h2_dim
+        self.h3_dim = h3_dim
+        self.y_dim = y_dim
+        self.dropout_rate = dropout_rate
+        self.dropout = nn.Dropout(self.dropout_rate)
+
+        self.hidden_layer_1 = nn.Linear(self.x_dim + self.z_dim, self.h1_dim)
+        self.hidden_layer_2 = nn.Linear(self.h1_dim, self.h2_dim)
+        self.hidden_layer_3 = nn.Linear(self.h2_dim, self.h3_dim)
+
+        self.film_layer_1_beta = nn.Linear(self.task_dim, self.h1_dim, bias=False)
+        self.film_layer_1_gamma = nn.Linear(self.task_dim, self.h1_dim, bias=False)
+        self.film_layer_1_eta = nn.Linear(self.task_dim, self.h1_dim, bias=False)
+        self.film_layer_1_delta = nn.Linear(self.task_dim, self.h1_dim, bias=False)
+
+        self.film_layer_2_beta = nn.Linear(self.task_dim, self.h2_dim, bias=False)
+        self.film_layer_2_gamma = nn.Linear(self.task_dim, self.h2_dim, bias=False)
+        self.film_layer_2_eta = nn.Linear(self.task_dim, self.h2_dim, bias=False)
+        self.film_layer_2_delta = nn.Linear(self.task_dim, self.h2_dim, bias=False)
+
+
+        self.film_layer_3_beta = nn.Linear(self.task_dim, self.h3_dim, bias=False)
+        self.film_layer_3_gamma = nn.Linear(self.task_dim, self.h3_dim, bias=False)
+        self.film_layer_3_eta = nn.Linear(self.task_dim, self.h3_dim, bias=False)
+        self.film_layer_3_delta = nn.Linear(self.task_dim, self.h3_dim, bias=False)
+
+
+        self.final_projection = nn.Linear(self.h3_dim, self.y_dim)
+
+    def forward(self, x, z, task):
+        interaction_size, _ = x.size()
+        z = z.unsqueeze(0).repeat(interaction_size, 1)
+        # Input is concatenation of z with every row of x
+        inputs = torch.cat((x, z), dim=1)
+        hidden_1 = self.hidden_layer_1(inputs)
+        beta_1 = torch.tanh(self.film_layer_1_beta(task))
+        gamma_1 = torch.tanh(self.film_layer_1_gamma(task))
+        eta_1 = torch.tanh(self.film_layer_1_eta(task))
+        delta_1 = torch.sigmoid(self.film_layer_1_delta(task))
+
+        gamma_1 = gamma_1 * delta_1 + eta_1 * (1-delta_1)
+        beta_1 = beta_1 * delta_1 + eta_1 * (1-delta_1)
+
+        hidden_1 = torch.mul(hidden_1, gamma_1) + beta_1
+        hidden_1 = self.dropout(hidden_1)
+        hidden_2 = F.relu(hidden_1)
+
+        hidden_2 = self.hidden_layer_2(hidden_2)
+        beta_2 = torch.tanh(self.film_layer_2_beta(task))
+        gamma_2 = torch.tanh(self.film_layer_2_gamma(task))
+        eta_2 = torch.tanh(self.film_layer_2_eta(task))
+        delta_2 = torch.sigmoid(self.film_layer_2_delta(task))
+
+        gamma_2 = gamma_2 * delta_2 + eta_2 * (1 - delta_2)
+        beta_2 = beta_2 * delta_2 + eta_2 * (1 - delta_2)
+
+
+        hidden_2 = torch.mul(hidden_2, gamma_2) + beta_2
+        hidden_2 = self.dropout(hidden_2)
+        hidden_3 = F.relu(hidden_2)
+        hidden_3 = self.hidden_layer_3(hidden_3)
+        beta_3 = torch.tanh(self.film_layer_3_beta(task))
+        gamma_3 = torch.tanh(self.film_layer_3_gamma(task))
+        eta_3 = torch.tanh(self.film_layer_3_eta(task))
+        delta_3 = torch.sigmoid(self.film_layer_3_delta(task))
+
+        gamma_3 = gamma_3 * delta_3 + eta_3 * (1 - delta_3)
+        beta_3 = beta_3 * delta_3 + eta_3 * (1 - delta_3)
+
+        hidden_final = torch.mul(hidden_3, gamma_3) + beta_3
+        hidden_final = self.dropout(hidden_final)
+        hidden_final = F.relu(hidden_final)
+
+        y_pred = self.final_projection(hidden_final)
+        return y_pred

eval.py

@@ -0,0 +1,46 @@
+"""
+Run evaluation with saved models.
+"""
+import random
+import argparse
+from tqdm import tqdm
+import torch
+from utils.scorer import *
+
+
+def testing(trainer, opt, test_dataset):
+    test_dataset_len = len(test_dataset)
+    #batch_size = opt["batch_size"]
+    minibatch_size = 1
+    a, b, c, d = zip(*test_dataset)
+    trainer.eval()
+    all_loss = 0
+    pre5 = []
+    ap5 = []
+    ndcg5 = []
+    pre7 = []
+    ap7 = []
+    ndcg7 = []
+    pre10 = []
+    ap10 = []
+    ndcg10 = []
+    for i in range(test_dataset_len):
+        try:
+            supp_xs = list(a[minibatch_size * i:minibatch_size * (i + 1)])
+            supp_ys = list(b[minibatch_size * i:minibatch_size * (i + 1)])
+            query_xs = list(c[minibatch_size * i:minibatch_size * (i + 1)])
+            query_ys = list(d[minibatch_size * i:minibatch_size * (i + 1)])
+        except IndexError:
+            continue
+        test_loss, recommendation_list = trainer.query_rec(supp_xs, supp_ys, query_xs, query_ys)
+        all_loss += test_loss
+
+        add_metric(recommendation_list, query_ys[0].cpu().detach().numpy(), pre5, ap5, ndcg5, 5)
+        add_metric(recommendation_list, query_ys[0].cpu().detach().numpy(), pre7, ap7, ndcg7, 7)
+        add_metric(recommendation_list, query_ys[0].cpu().detach().numpy(), pre10, ap10, ndcg10, 10)
+
+    mpre5, mndcg5, map5 = cal_metric(pre5, ap5, ndcg5)
+    mpre7, mndcg7, map7 = cal_metric(pre7, ap7, ndcg7)
+    mpre10, mndcg10, map10 = cal_metric(pre10, ap10, ndcg10)
+
+    return mpre5, mndcg5, map5, mpre7, mndcg7, map7, mpre10, mndcg10, map10

train.sh

@@ -0,0 +1,55 @@
+first_embedding_dim=32
+second_embedding_dim=16
+z1_dim=32
+z2_dim=32
+z_dim=32
+enc_h1_dim=32
+enc_h2_dim=16
+taskenc_h1_dim=32
+taskenc_h2_dim=32
+taskenc_final_dim=16
+clusters_k=10
+temperature=1.0
+lambda=1.0
+dec_h1_dim=32
+dec_h2_dim=32
+dec_h3_dim=16
+dropout_rate=0
+lr=0.0001
+optim='adam'
+num_epoch=100
+batch_size=32
+train_ratio=0.7
+valid_ratio=0.1
+support_size=20
+query_size=10
+max_len=200
+context_min=20
+CUDA_VISIBLE_DEVICES=0 python train_TaNP.py \
+ --first_embedding_dim $first_embedding_dim \
+ --second_embedding_dim $second_embedding_dim \
+ --z1_dim $z1_dim \
+ --z2_dim $z2_dim \
+ --z_dim $z_dim \
+ --enc_h1_dim $enc_h1_dim \
+ --enc_h2_dim $enc_h2_dim \
+ --taskenc_h1_dim $taskenc_h1_dim \
+ --taskenc_h2_dim $taskenc_h2_dim \
+ --taskenc_final_dim $taskenc_final_dim \
+ --clusters_k $clusters_k \
+ --lambda $lambda \
+ --temperature $temperature \
+ --dec_h1_dim $dec_h1_dim \
+ --dec_h2_dim $dec_h2_dim \
+ --dec_h3_dim $dec_h3_dim \
+ --lr $lr \
+ --dropout_rate $dropout_rate \
+ --optim $optim \
+ --num_epoch $num_epoch \
+ --batch_size $batch_size \
+ --train_ratio $train_ratio \
+ --valid_ratio $valid_ratio \
+ --support_size $support_size \
+ --query_size $query_size \
+ --max_len $max_len \
+ --context_min $context_min

train_TaNP.py

@@ -0,0 +1,193 @@
+import os
+from datetime import datetime
+import time
+import numpy as np
+import random
+import argparse
+import pickle
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from torch.autograd import Variable
+import json
+from utils.loader import Preprocess
+from TaNP import Trainer
+from TaNP_training import training
+from utils import helper
+from eval import testing
+
+parser = argparse.ArgumentParser()
+# parser.add_argument('--data_dir', type=str, default='data/lastfm_20')#1
+# parser.add_argument('--model_save_dir', type=str, default='save_model_dir')#1
+parser.add_argument('--data_dir', type=str, default='/media/external_10TB/10TB/maheri/melu_data')#1
+parser.add_argument('--model_save_dir', type=str, default='/media/external_10TB/10TB/maheri/tanp_data/tanp_models')#1
+
+parser.add_argument('--id', type=str, default='1', help='used for save hyper-parameters.')#1
+
+parser.add_argument('--first_embedding_dim', type=int, default=32, help='Embedding dimension for item and user.')#1
+parser.add_argument('--second_embedding_dim', type=int, default=16, help='Embedding dimension for item and user.')#1
+
+parser.add_argument('--z1_dim', type=int, default=32, help='The dimension of z1 in latent path.')
+parser.add_argument('--z2_dim', type=int, default=32, help='The dimension of z2 in latent path.')
+parser.add_argument('--z_dim', type=int, default=32, help='The dimension of z in latent path.')
+
+parser.add_argument('--enc_h1_dim', type=int, default=64, help='The hidden first dimension of encoder.')
+parser.add_argument('--enc_h2_dim', type=int, default=64, help='The hidden second dimension of encoder.')
+
+parser.add_argument('--taskenc_h1_dim', type=int, default=128, help='The hidden first dimension of task encoder.')
+parser.add_argument('--taskenc_h2_dim', type=int, default=64, help='The hidden second dimension of task encoder.')
+parser.add_argument('--taskenc_final_dim', type=int, default=64, help='The hidden second dimension of task encoder.')
+
+parser.add_argument('--clusters_k', type=int, default=7, help='Cluster numbers of tasks.')
+parser.add_argument('--temperature', type=float, default=1.0, help='used for student-t distribution.')
+parser.add_argument('--lambda', type=float, default=0.1, help='used to balance the clustering loss and NP loss.')
+
+parser.add_argument('--dec_h1_dim', type=int, default=128, help='The hidden first dimension of encoder.')
+parser.add_argument('--dec_h2_dim', type=int, default=128, help='The hidden second dimension of encoder.')
+parser.add_argument('--dec_h3_dim', type=int, default=128, help='The hidden third dimension of encoder.')
+
+# used for movie datasets
+parser.add_argument('--num_gender', type=int, default=2, help='User information.')#1
+parser.add_argument('--num_age', type=int, default=7, help='User information.')#1
+parser.add_argument('--num_occupation', type=int, default=21, help='User information.')#1
+parser.add_argument('--num_zipcode', type=int, default=3402, help='User information.')#1
+parser.add_argument('--num_rate', type=int, default=6, help='Item information.')#1
+parser.add_argument('--num_genre', type=int, default=25, help='Item information.')#1
+parser.add_argument('--num_director', type=int, default=2186, help='Item information.')#1
+parser.add_argument('--num_actor', type=int, default=8030, help='Item information.')#1
+
+parser.add_argument('--dropout_rate', type=float, default=0, help='used in encoder and decoder.')
+parser.add_argument('--lr', type=float, default=1e-4, help='Applies to SGD and Adagrad.')#1
+parser.add_argument('--optim', type=str, default='adam', help='sgd, adagrad, adam or adamax.')
+parser.add_argument('--num_epoch', type=int, default=150)#1
+parser.add_argument('--batch_size', type=int, default=32)#1
+parser.add_argument('--train_ratio', type=float, default=0.7, help='Warm user ratio for training.')#1
+parser.add_argument('--valid_ratio', type=float, default=0.1, help='Cold user ratio for validation.')#1
+parser.add_argument('--seed', type=int, default=2020)#1
+parser.add_argument('--save', type=int, default=0)#1
+parser.add_argument('--use_cuda', type=bool, default=torch.cuda.is_available())#1
+parser.add_argument('--cpu', action='store_true', help='Ignore CUDA.')#1
+parser.add_argument('--support_size', type=int, default=20)#1
+parser.add_argument('--query_size', type=int, default=10)#1
+parser.add_argument('--max_len', type=int, default=200, help='The max length of interactions for each user.')
+parser.add_argument('--context_min', type=int, default=20, help='Minimum size of context range.')
+
+
+# change for Movie lens
+parser.add_argument('--embedding_dim', type=int, default=32, help='embedding dimension for each item/user feature of Movie lens')
+parser.add_argument('--first_fc_hidden_dim', type=int, default=64, help='embedding dimension for each item/user feature of Movie lens')
+parser.add_argument('--second_fc_hidden_dim', type=int, default=64, help='embedding dimension for each item/user feature of Movie lens')
+
+args = parser.parse_args()
+
+def seed_everything(seed=1023):
+    random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    np.random.seed(seed)
+    os.environ['PYTHONHASHSEED'] = str(seed)
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+
+seed = args.seed
+seed_everything(seed)
+
+if args.cpu:
+    args.use_cuda = False
+elif args.use_cuda:
+    torch.cuda.manual_seed(args.seed)
+
+opt = vars(args)
+
+# print model info
+helper.print_config(opt)
+helper.ensure_dir(opt["model_save_dir"], verbose=True)
+# save model config
+helper.save_config(opt, opt["model_save_dir"] + "/" +opt["id"] + '.config', verbose=True)
+# record training log
+file_logger = helper.FileLogger(opt["model_save_dir"] + '/' + opt['id'] + ".log",
+                                header="# epoch\ttrain_loss\tprecision5\tNDCG5\tMAP5\tprecision7"
+                                       "\tNDCG7\tMAP7\tprecision10\tNDCG10\tMAP10")
+
+# change for Movie Lens
+# preprocess = Preprocess(opt)
+print("Preprocess is done.")
+print("Create model TaNP...")
+
+# opt['uf_dim'] = preprocess.uf_dim
+# opt['if_dim'] = preprocess.if_dim
+
+trainer = Trainer(opt)
+if opt['use_cuda']:
+    trainer.cuda()
+
+model_filename = "{}/{}.pt".format(opt['model_save_dir'], opt["id"])
+
+# /4 since sup_x, sup_y, query_x, query_y
+
+# change for Movie lens
+# training_set_size = int(len(os.listdir("{}/{}/{}".format(opt["data_dir"], "training", "log"))) / 4)
+training_set_size = int(len(os.listdir("{}/{}".format(opt["data_dir"], "warm_state"))) / 4)
+
+supp_xs_s = []
+supp_ys_s = []
+query_xs_s = []
+query_ys_s = []
+for idx in range(training_set_size):
+    # supp_xs_s.append(pickle.load(open("{}/{}/{}/supp_x_{}.pkl".format(opt["data_dir"], "training", "log", idx), "rb")))
+    # supp_ys_s.append(pickle.load(open("{}/{}/{}/supp_y_{}.pkl".format(opt["data_dir"], "training", "log", idx), "rb")))
+    # query_xs_s.append(pickle.load(open("{}/{}/{}/query_x_{}.pkl".format(opt["data_dir"], "training", "log", idx), "rb")))
+    # query_ys_s.append(pickle.load(open("{}/{}/{}/query_y_{}.pkl".format(opt["data_dir"], "training", "log", idx), "rb")))
+    supp_xs_s.append(pickle.load(open("{}/{}/supp_x_{}.pkl".format(opt["data_dir"], "warm_state", idx), "rb")))
+    supp_ys_s.append(pickle.load(open("{}/{}/supp_y_{}.pkl".format(opt["data_dir"], "warm_state", idx), "rb")))
+    query_xs_s.append(pickle.load(open("{}/{}/query_x_{}.pkl".format(opt["data_dir"], "warm_state", idx), "rb")))
+    query_ys_s.append(pickle.load(open("{}/{}/query_y_{}.pkl".format(opt["data_dir"], "warm_state", idx), "rb")))
+
+train_dataset = list(zip(supp_xs_s, supp_ys_s, query_xs_s, query_ys_s))
+
+del (supp_xs_s, supp_ys_s, query_xs_s, query_ys_s)
+
+# change for Movie lens
+# testing_set_size = int(len(os.listdir("{}/{}/{}".format(opt["data_dir"], "testing", "log"))) / 4)
+testing_set_size = int(len(os.listdir("{}/{}".format(opt["data_dir"], "user_cold_state"))) / 4)
+
+supp_xs_s = []
+supp_ys_s = []
+query_xs_s = []
+query_ys_s = []
+for idx in range(testing_set_size):
+    # change for Movie lens
+    # supp_xs_s.append(
+    #     pickle.load(open("{}/{}/{}/supp_x_{}.pkl".format(opt["data_dir"], "testing", "log", idx), "rb")))
+    # supp_ys_s.append(
+    #     pickle.load(open("{}/{}/{}/supp_y_{}.pkl".format(opt["data_dir"], "testing", "log", idx), "rb")))
+    # query_xs_s.append(
+    #     pickle.load(open("{}/{}/{}/query_x_{}.pkl".format(opt["data_dir"], "testing", "log", idx), "rb")))
+    # query_ys_s.append(
+    #     pickle.load(open("{}/{}/{}/query_y_{}.pkl".format(opt["data_dir"], "testing", "log", idx), "rb")))
+    supp_xs_s.append(
+        pickle.load(open("{}/{}/supp_x_{}.pkl".format(opt["data_dir"], "user_cold_state", idx), "rb")))
+    supp_ys_s.append(
+        pickle.load(open("{}/{}/supp_y_{}.pkl".format(opt["data_dir"], "user_cold_state", idx), "rb")))
+    query_xs_s.append(
+        pickle.load(open("{}/{}/query_x_{}.pkl".format(opt["data_dir"], "user_cold_state", idx), "rb")))
+    query_ys_s.append(
+        pickle.load(open("{}/{}/query_y_{}.pkl".format(opt["data_dir"], "user_cold_state", idx), "rb")))
+test_dataset = list(zip(supp_xs_s, supp_ys_s, query_xs_s, query_ys_s))
+
+del (supp_xs_s, supp_ys_s, query_xs_s, query_ys_s)
+
+print("# epoch\ttrain_loss\tprecision5\tNDCG5\tMAP5\tprecision7\tNDCG7\tMAP7\tprecision10\tNDCG10\tMAP10")
+
+if not os.path.exists(model_filename):
+    print("Start training...")
+    training(trainer, opt, train_dataset, test_dataset, batch_size=opt['batch_size'], num_epoch=opt['num_epoch'],
+            model_save=opt["save"], model_filename=model_filename, logger=file_logger)
+
+else:
+    print("Load pre-trained model...")
+    opt = helper.load_config(model_filename[:-2]+"config")
+    helper.print_config(opt)
+    trained_state_dict = torch.load(model_filename)
+    trainer.load_state_dict(trained_state_dict)
+

utils/helper.py

@@ -0,0 +1,74 @@
+"""
+Helper functions.
+"""
+
+import os
+import json
+import argparse
+
+
+### IO
+def check_dir(d):
+    if not os.path.exists(d):
+        print("Directory {} does not exist. Exit.".format(d))
+        exit(1)
+
+
+def check_files(files):
+    for f in files:
+        if f is not None and not os.path.exists(f):
+            print("File {} does not exist. Exit.".format(f))
+            exit(1)
+
+
+def ensure_dir(d, verbose=True):
+    if not os.path.exists(d):
+        if verbose:
+            print("Directory {} do not exist; creating...".format(d))
+        os.makedirs(d)
+
+
+def save_config(config, path, verbose=True):
+    with open(path, 'w') as outfile:
+        json.dump(config, outfile, indent=2)
+    if verbose:
+        print("Config saved to file {}".format(path))
+    return config
+
+
+def load_config(path, verbose=True):
+    with open(path) as f:
+        config = json.load(f)
+    if verbose:
+        print("Config loaded from file {}".format(path))
+    return config
+
+
+def print_config(config):
+    info = "Running with the following configs:\n"
+    for k, v in config.items():
+        info += "\t{} : {}\n".format(k, str(v))
+    print("\n" + info + "\n")
+    return
+
+
+class FileLogger(object):
+    """
+    A file logger that opens the file periodically and write to it.
+    """
+
+    def __init__(self, filename, header=None):
+        self.filename = filename
+        if os.path.exists(filename):
+            # remove the old file
+            os.remove(filename)
+        if header is not None:
+            with open(filename, 'w') as out:
+                print(header, file=out)
+
+    def log(self, message):
+        with open(self.filename, 'a') as out:
+            print(message)
+            print(message, file=out)
+
+

utils/loader.py

@@ -0,0 +1,218 @@
+import json
+import random
+import torch
+import numpy as np
+import pickle
+import codecs
+import re
+import os
+import datetime
+import tqdm
+import pandas as pd
+
+#convert userids to userdict key-id(int), val:onehot_vector(tensor)
+#element in list is str type.
+def to_onehot_dict(list):
+    dict={}
+    length = len(list)
+    for index, element in enumerate(list):
+        vector = torch.zeros(1, length).long()
+        element = int(element)
+        vector[:, element] = 1.0
+        dict[element] = vector
+    return dict
+
+def load_list(fname):
+    list_ = []
+    with open(fname, encoding="utf-8") as f:
+        for line in f.readlines():
+            list_.append(line.strip())
+    return list_
+
+# used for merge dictionaries.
+def merge_key(dict1, dict2):
+    res = {**dict1, **dict2}
+    return res
+
+def merge_value(dict1, dict2): # merge and item_cold
+    for key, value in dict2.items():
+        if key in dict1.keys():
+            # if list(set(dict1[key]+value)) the final number of movies-1m is 1000205
+            new_value = dict1[key]+value
+            dict1[key] = new_value
+        else:
+            print('Unexpected key.')
+
+def count_values(dict):
+    count_val = 0
+    for key, value in dict.items():
+        count_val += len(value)
+    return count_val
+
+def construct_dictionary(user_list, total_dict):
+    dict = {}
+    for i in range(len(user_list)):
+        dict[str(user_list[i])] = total_dict[str(user_list[i])]
+    return dict
+
+class Preprocess(object):
+    """
+    Preprocess the training, validation and test data.
+    Generate the episode-style data.
+    """
+
+    def __init__(self, opt):
+        self.batch_size = opt["batch_size"]
+        self.opt = opt
+        # warm data ratio
+        self.train_ratio = opt['train_ratio']
+        self.valid_ratio = opt['valid_ratio']
+        self.test_ratio = 1 - self.train_ratio - self.valid_ratio
+        self.dataset_path = opt["data_dir"]
+        self.support_size = opt['support_size']
+        self.query_size = opt['query_size']
+        self.max_len = opt['max_len']
+        # save one-hot dimension length
+        uf_dim, if_dim = self.preprocess(self.dataset_path)
+        self.uf_dim = uf_dim
+        self.if_dim = if_dim
+
+    def preprocess(self, dataset_path):
+        """ Preprocess the data and convert to ids. """
+        #Create training-validation-test datasets
+        print('Create training, validation and test data from scratch!')
+        with open('./{}/interaction_dict_x.json'.format(dataset_path), 'r', encoding='utf-8') as f:
+            inter_dict_x = json.loads(f.read())
+        with open('./{}/interaction_dict_y.json'.format(dataset_path), 'r', encoding='utf-8') as f:
+            inter_dict_y = json.loads(f.read())
+        print('The size of total interactions is %d.' % (count_values(inter_dict_x)))  # 42346
+        assert count_values(inter_dict_x) == count_values(inter_dict_y)
+
+        with open('./{}/user_list.json'.format(dataset_path), 'r', encoding='utf-8') as f:
+            userids = json.loads(f.read())
+
+        with open('./{}/item_list.json'.format(dataset_path), 'r', encoding='utf-8') as f:
+            itemids = json.loads(f.read())
+
+        #userids = list(inter_dict_x.keys())
+        random.shuffle(userids)
+        warm_user_size = int(len(userids) * self.train_ratio)
+        valid_user_size = int(len(userids) * self.valid_ratio)
+        warm_users = userids[:warm_user_size]
+        valid_users = userids[warm_user_size:warm_user_size+valid_user_size]
+        cold_users = userids[warm_user_size+valid_user_size:]
+        assert len(userids) == len(warm_users)+len(valid_users)+len(cold_users)
+
+
+            # Construct the training data dict
+        training_dict_x = construct_dictionary(warm_users, inter_dict_x)
+        training_dict_y = construct_dictionary(warm_users, inter_dict_y)
+
+            #Avoid the new items shown in test data in the case of cold user.
+        item_set = set()
+        for i in training_dict_x.values():
+            i = set(i)
+            item_set = item_set.union(i)
+
+        # Construct one-hot dictionary
+        user_dict = to_onehot_dict(userids)
+        # only items contained in all data are encoded.
+        item_dict = to_onehot_dict(itemids)
+
+        # This part of data is not used, so we do not process it temporally.
+        valid_dict_x = construct_dictionary(valid_users, inter_dict_x)
+        valid_dict_y = construct_dictionary(valid_users, inter_dict_y)
+        assert count_values(valid_dict_x) == count_values(valid_dict_y)
+
+        test_dict_x = construct_dictionary(cold_users, inter_dict_x)
+        test_dict_y = construct_dictionary(cold_users, inter_dict_y)
+        assert count_values(test_dict_x) == count_values(test_dict_y)
+
+        print('Before delete new items in test data, test data has %d interactions.' % (count_values(test_dict_x)))
+
+        #Delete the new items in test data.
+        unseen_count = 0
+        for key, value in test_dict_x.items():
+            assert len(value) == len(test_dict_y[key])
+            unseen_item_index = [index for index, i in enumerate(value) if i not in item_set]
+            unseen_count+=len(unseen_item_index)
+            if len(unseen_item_index) == 0:
+                continue
+            else:
+                new_value_x = [element for index, element in enumerate(value) if index not in unseen_item_index]
+                new_value_y = [test_dict_y[key][index] for index, element in enumerate(value) if index not in unseen_item_index]
+                test_dict_x[key] = new_value_x
+                test_dict_y[key] = new_value_y
+        print('After delete new items in test data, test data has %d interactions.' % (count_values(test_dict_x)))
+        assert count_values(test_dict_x) == count_values(test_dict_y)
+        print('The number of total unseen interactions is %d.' % (unseen_count))
+
+        pickle.dump(training_dict_x, open("{}/training_dict_x_{:2f}.pkl".format(dataset_path, self.train_ratio), "wb"))
+        pickle.dump(training_dict_y, open("{}/training_dict_y_{:2f}.pkl".format(dataset_path, self.train_ratio), "wb"))
+        pickle.dump(valid_dict_x, open("{}/valid_dict_x_{:2f}.pkl".format(dataset_path, self.valid_ratio), "wb"))
+        pickle.dump(valid_dict_y, open("{}/valid_dict_y_{:2f}.pkl".format(dataset_path, self.valid_ratio), "wb"))
+        pickle.dump(test_dict_x, open("{}/test_dict_x_{:2f}.pkl".format(dataset_path, self.test_ratio), "wb"))
+        pickle.dump(test_dict_y, open("{}/test_dict_y_{:2f}.pkl".format(dataset_path, self.test_ratio), "wb"))
+
+        def generate_episodes(dict_x, dict_y, category, support_size, query_size, max_len, dir="log"):
+            idx = 0
+            if not os.path.exists("{}/{}/{}".format(dataset_path, category, dir)):
+                os.makedirs("{}/{}/{}".format(dataset_path, category, dir))
+                os.makedirs("{}/{}/{}".format(dataset_path, category, "evidence"))
+                for _, user_id in enumerate(dict_x.keys()):
+                    u_id = int(user_id)
+                    seen_music_len = len(dict_x[str(u_id)])
+                    indices = list(range(seen_music_len))
+                    # filter some users with their interactions, i.e., tasks
+                    if seen_music_len < (support_size + query_size) or seen_music_len > max_len:
+                        continue
+                    random.shuffle(indices)
+                    tmp_x = np.array(dict_x[str(u_id)])
+                    tmp_y = np.array(dict_y[str(u_id)])
+
+                    support_x_app = None
+                    for m_id in tmp_x[indices[:support_size]]:
+                        m_id = int(m_id)
+                        tmp_x_converted = torch.cat((item_dict[m_id], user_dict[u_id]), 1)
+                        try:
+                            support_x_app = torch.cat((support_x_app, tmp_x_converted), 0)
+                        except:
+                            support_x_app = tmp_x_converted
+
+                    query_x_app = None
+                    for m_id in tmp_x[indices[support_size:]]:
+                        m_id = int(m_id)
+                        u_id = int(user_id)
+                        tmp_x_converted = torch.cat((item_dict[m_id], user_dict[u_id]), 1)
+                        try:
+                            query_x_app = torch.cat((query_x_app, tmp_x_converted), 0)
+                        except:
+                            query_x_app = tmp_x_converted
+
+                    support_y_app = torch.FloatTensor(tmp_y[indices[:support_size]])
+                    query_y_app = torch.FloatTensor(tmp_y[indices[support_size:]])
+
+                    pickle.dump(support_x_app, open("{}/{}/{}/supp_x_{}.pkl".format(dataset_path, category, dir, idx), "wb"))
+                    pickle.dump(support_y_app, open("{}/{}/{}/supp_y_{}.pkl".format(dataset_path, category, dir, idx), "wb"))
+                    pickle.dump(query_x_app, open("{}/{}/{}/query_x_{}.pkl".format(dataset_path, category, dir, idx), "wb"))
+                    pickle.dump(query_y_app, open("{}/{}/{}/query_y_{}.pkl".format(dataset_path, category, dir, idx), "wb"))
+                    # used for evidence candidate selection
+                    with open("{}/{}/{}/supp_x_{}_u_m_ids.txt".format(dataset_path, category, "evidence", idx), "w") as f:
+                        for m_id in tmp_x[indices[:support_size]]:
+                            f.write("{}\t{}\n".format(u_id, m_id))
+                    with open("{}/{}/{}/query_x_{}_u_m_ids.txt".format(dataset_path, category, "evidence", idx), "w") as f:
+                        for m_id in tmp_x[indices[support_size:]]:
+                            f.write("{}\t{}\n".format(u_id, m_id))
+                    idx+=1
+
+        print("Generate eposide data for training.")
+        generate_episodes(training_dict_x, training_dict_y, "training", self.support_size, self.query_size, self.max_len)
+        print("Generate eposide data for validation.")
+        generate_episodes(valid_dict_x, valid_dict_y, "validation", self.support_size, self.query_size, self.max_len)
+        print("Generate eposide data for testing.")
+        generate_episodes(test_dict_x, test_dict_y, "testing", self.support_size, self.query_size, self.max_len)
+
+        return len(userids), len(itemids)
+
+
+

utils/scorer.py

@@ -0,0 +1,77 @@
+import math
+
+def AP(ranked_list, ground_truth, topn):
+    hits, sum_precs = 0, 0.0
+    t = [a for a in ground_truth]
+    t.sort(reverse=True)
+    t=t[:topn]
+    for i in range(topn):
+        id = ranked_list[i]
+        if ground_truth[id] in t:
+            hits += 1
+            sum_precs += hits / (i+1.0)
+            t.remove(ground_truth[id])
+    if hits > 0:
+        return sum_precs / topn
+    else:
+        return 0.0
+
+def RR(ranked_list, ground_truth,topn):
+    t = [a for a in ground_truth]
+    t.sort(reverse=True)
+    t = t[:topn]
+    for i in range(topn):
+        id = ranked_list[i]
+        if ground_truth[id] in t:
+            return 1 / (i + 1.0)
+    return 0
+
+def precision(ranked_list,ground_truth,topn):
+    t = [a for a in ground_truth]
+    t.sort(reverse=True)
+    t = t[:topn]
+    hits = 0
+    for i in range(topn):
+        id = ranked_list[i]
+        if ground_truth[id] in t:
+            t.remove(ground_truth[id])
+            hits += 1
+    pre = hits/topn
+    return pre
+
+
+def nDCG(ranked_list, ground_truth, topn):
+    dcg = 0
+    idcg = IDCG(ground_truth, topn)
+    # print(ranked_list)
+    # input()
+    for i in range(topn):
+        id = ranked_list[i]
+        dcg += ((2 ** ground_truth[id]) -1)/ math.log(i+2, 2)
+    # print('dcg is ', dcg, " n is ", topn)
+    # print('idcg is ', idcg, " n is ", topn)
+    return dcg / idcg
+
+def IDCG(ground_truth,topn):
+    t = [a for a in ground_truth]
+    t.sort(reverse=True)
+    idcg = 0
+    for i in range(topn):
+        idcg += ((2**t[i]) - 1) / math.log(i+2, 2)
+    return idcg
+
+def add_metric(recommend_list, ALL_group_list, precision_list, ap_list, ndcg_list, topn):
+    ndcg = nDCG(recommend_list, ALL_group_list, topn)
+    ap = AP(recommend_list, ALL_group_list, topn)
+    pre = precision(recommend_list, ALL_group_list, topn)
+    precision_list.append(pre)
+    ap_list.append(ap)
+    ndcg_list.append(ndcg)
+
+
+
+def cal_metric(precision_list,ap_list,ndcg_list):
+    mpre = sum(precision_list) / len(precision_list)
+    map = sum(ap_list) / len(ap_list)
+    mndcg = sum(ndcg_list) / len(ndcg_list)
+    return mpre, mndcg, map

utils/torch_utils.py

@@ -0,0 +1,168 @@
+"""
+Utility functions for torch.
+"""
+
+import torch
+from torch import nn, optim
+from torch.optim.optimizer import Optimizer
+
+### class
+class MyAdagrad(Optimizer):
+    """My modification of the Adagrad optimizer that allows to specify an initial
+    accumulater value. This mimics the behavior of the default Adagrad implementation
+    in Tensorflow. The default PyTorch Adagrad uses 0 for initial acculmulator value.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-2)
+        lr_decay (float, optional): learning rate decay (default: 0)
+        init_accu_value (float, optional): initial accumulater value.
+        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
+    """
+
+    def __init__(self, params, lr=1e-2, lr_decay=0, init_accu_value=0.1, weight_decay=0):
+        defaults = dict(lr=lr, lr_decay=lr_decay, init_accu_value=init_accu_value, \
+                weight_decay=weight_decay)
+        super(MyAdagrad, self).__init__(params, defaults)
+
+        for group in self.param_groups:
+            for p in group['params']:
+                state = self.state[p]
+                state['step'] = 0
+                state['sum'] = torch.ones(p.data.size()).type_as(p.data) *\
+                        init_accu_value
+
+    def share_memory(self):
+        for group in self.param_groups:
+            for p in group['params']:
+                state = self.state[p]
+                state['sum'].share_memory_()
+
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            loss = closure()
+
+        for group in self.param_groups:
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+
+                grad = p.grad.data
+                state = self.state[p]
+
+                state['step'] += 1
+
+                if group['weight_decay'] != 0:
+                    if p.grad.data.is_sparse:
+                        raise RuntimeError("weight_decay option is not compatible with sparse gradients ")
+                    grad = grad.add(group['weight_decay'], p.data)
+
+                clr = group['lr'] / (1 + (state['step'] - 1) * group['lr_decay'])
+
+                if p.grad.data.is_sparse:
+                    grad = grad.coalesce()  # the update is non-linear so indices must be unique
+                    grad_indices = grad._indices()
+                    grad_values = grad._values()
+                    size = torch.Size([x for x in grad.size()])
+
+                    def make_sparse(values):
+                        constructor = type(p.grad.data)
+                        if grad_indices.dim() == 0 or values.dim() == 0:
+                            return constructor()
+                        return constructor(grad_indices, values, size)
+                    state['sum'].add_(make_sparse(grad_values.pow(2)))
+                    std = state['sum']._sparse_mask(grad)
+                    std_values = std._values().sqrt_().add_(1e-10)
+                    p.data.add_(-clr, make_sparse(grad_values / std_values))
+                else:
+                    state['sum'].addcmul_(1, grad, grad)
+                    std = state['sum'].sqrt().add_(1e-10)
+                    p.data.addcdiv_(-clr, grad, std)
+
+        return loss
+
+### torch specific functions
+def get_optimizer(name, parameters, lr, l2=0):
+    if name == 'sgd':
+        return torch.optim.SGD(parameters, lr=lr, weight_decay=l2)
+    elif name in ['adagrad', 'myadagrad']:
+        # use my own adagrad to allow for init accumulator value
+        return MyAdagrad(parameters, lr=lr, init_accu_value=0.1, weight_decay=l2)
+    elif name == 'adam':
+        return torch.optim.Adam(parameters, weight_decay=l2) # use default lr
+    elif name == 'adamax':
+        return torch.optim.Adamax(parameters, weight_decay=l2) # use default lr
+    elif name == 'adadelta':
+        return torch.optim.Adadelta(parameters, lr=lr, weight_decay=l2)
+    else:
+        raise Exception("Unsupported optimizer: {}".format(name))
+
+
+def change_lr(optimizer, new_lr):
+    for param_group in optimizer.param_groups:
+        param_group['lr'] = new_lr
+
+
+def flatten_indices(seq_lens, width):
+    flat = []
+    for i, l in enumerate(seq_lens):
+        for j in range(l):
+            flat.append(i * width + j)
+    return flat
+
+
+def set_cuda(var, cuda):
+    if cuda:
+        return var.cuda()
+    return var
+
+
+def keep_partial_grad(grad, topk):
+    """
+    Keep only the topk rows of grads.
+    """
+    assert topk < grad.size(0)
+    grad.data[topk:].zero_()
+    return grad
+
+### model IO
+def save(model, optimizer, opt, filename):
+    params = {
+        'model': model.state_dict(),
+        'optimizer': optimizer.state_dict(),
+        'config': opt
+    }
+    try:
+        torch.save(params, filename)
+    except BaseException:
+        print("[ Warning: model saving failed. ]")
+
+
+def load(model, optimizer, filename):
+    try:
+        dump = torch.load(filename)
+    except BaseException:
+        print("[ Fail: model loading failed. ]")
+    if model is not None:
+        model.load_state_dict(dump['model'])
+    if optimizer is not None:
+        optimizer.load_state_dict(dump['optimizer'])
+    opt = dump['config']
+    return model, optimizer, opt
+
+
+def load_config(filename):
+    try:
+        dump = torch.load(filename)
+    except BaseException:
+        print("[ Fail: model loading failed. ]")
+    return dump['config']
+