save

ViviHong200709 · Nov 7, 2022 · e31bd24 · e31bd24
1 parent f9e619b
commit e31bd24
Show file tree

Hide file tree

Showing 15 changed files with 341 additions and 72 deletions.
diff --git a/CAT/distillation/MFI/__init__.py b/CAT/distillation/MFI/__init__.py
diff --git a/CAT/distillation/MFI/__pycache__/__init__.cpython-39.pyc b/CAT/distillation/MFI/__pycache__/__init__.cpython-39.pyc
diff --git a/CAT/distillation/MFI/__pycache__/dMFI.cpython-39.pyc b/CAT/distillation/MFI/__pycache__/dMFI.cpython-39.pyc
diff --git a/CAT/distillation/MFI/__pycache__/model.cpython-39.pyc b/CAT/distillation/MFI/__pycache__/model.cpython-39.pyc
diff --git a/CAT/distillation/MFI/__pycache__/tool.cpython-39.pyc b/CAT/distillation/MFI/__pycache__/tool.cpython-39.pyc
diff --git a/CAT/distillation/__pycache__/tool.cpython-39.pyc b/CAT/distillation/__pycache__/tool.cpython-39.pyc
diff --git a/CAT/distillation/prepare_trait.py b/CAT/distillation/prepare_trait.py
@@ -15,8 +15,8 @@
 seed = 0
 random.seed(seed)
 cdm='irt'
-dataset = 'junyi'
-stg='KLI'
+dataset = 'assistment'
+stg='MFI'
 train_triplets = pd.read_csv(
         f'/data/yutingh/CAT/data/{dataset}/train_triples.csv', encoding='utf-8').to_records(index=False)
 ckpt_path = f'/data/yutingh/CAT/ckpt/{dataset}/{cdm}_with_theta.pt'
@@ -58,7 +58,7 @@
     alpha=model.get_alpha(qid)
     beta=model.get_beta(qid)
     item_dict[item_id]=[np.float(alpha[0]),np.float(beta[0])]
-label,k_info,tested_info = get_label_and_k(user_dict,item_dict,50,stg)
+label,k_info,tested_info = get_label_and_k(user_dict,item_dict,50,stg,model)
 trait_dict = {
     'user':user_dict,
     'item':item_dict,
@@ -69,7 +69,7 @@
 
 path_prefix = f"/data/yutingh/CAT/data/{dataset}/{stg}/"
 
-with open(f"{path_prefix}trait.json", "w", encoding="utf-8") as f:
+with open(f"{path_prefix}trait_with_tested_info.json", "w", encoding="utf-8") as f:
     # indent参数保证json数据的缩进，美观 
     # ensure_ascii=False才能输出中文，否则就是Unicode字符
     f.write(json.dumps(trait_dict, ensure_ascii=False))
diff --git a/CAT/distillation/tool.py b/CAT/distillation/tool.py
@@ -21,27 +21,39 @@ def split_data(utrait,label,k_fisher,rate=1,tested_info=None):
     else:
         return ((u_train,label[:max_len],k_fisher[:max_len]),(u_test,label[max_len:],k_fisher[max_len:]))
 
-def get_label_and_k(user_trait,item_trait,k,stg="MFI"):
+def get_label_and_k(user_trait,item_trait,k,stg="MFI", model=None):
     labels=[]
     k_infos=[]
-    tested_ns=[]
-    for theta in tqdm(user_trait.values(),f'get top {k} items'):
+    tested_infos=[]
+    for sid, theta in tqdm(user_trait.items(),f'get top {k} items'):
         if stg=='MFI':
-            k_info,label = get_k_fisher(k, theta, item_trait)
+            k_info, label, tested_info = get_k_fisher(k, theta, item_trait)
+            tested_infos.append(tested_info)
         elif stg=='KLI':
-            k_info,label,tested_n= get_k_kli(k, theta, item_trait)
-            tested_ns.append(tested_n)
+            k_info,label,tested_info= get_k_kli(k, theta, item_trait)
+            tested_infos.append(tested_info)
+        elif stg=='MAAT':
+            get_k_emc(k, sid,theta, item_trait, model)
         k_infos.append(k_info)
         labels.append(label)
-    return labels,k_infos,tested_ns
+    return labels,k_infos,tested_infos
 
 def transform(item_trait, user_trait,labels,k_fishers,tested_infos=None):
     if tested_infos:
         for theta, label, k_fisher,tested_info in zip(user_trait.values(),labels,k_fishers,tested_infos):
             itrait = list(item_trait.values())
             item_n = len(itrait)
+            user_embs=[]
+            for tmp in tested_info:
+                user_emb = [theta]
+                if type(tmp) == list:
+                    user_emb.extend(tmp)
+                else:
+                    user_emb.append(tmp)
+                user_embs.append(user_emb)
             yield pack_batch([
-                torch.tensor(list(zip([theta]*item_n,tested_info))),
+                torch.tensor(user_embs),
+                # torch.tensor(list(zip([theta]*item_n,tested_info))),
                 itrait,
                 label,
                 k_fisher
@@ -65,16 +77,87 @@ def pack_batch(batch):
 
 def get_k_fisher(k, theta, items):
     fisher_arr = []
-    for qid,(alpha,beta) in items.items():
+    items_n=len(items.keys())
+    ns = [random.randint(0,19) for i in range(items_n)]
+    tested_qids = [random.sample(list(range(0,20)),n) for n in ns]
+    avg_embs = np.array(list(items.values())).mean(axis=0)
+    p=0.001
+    avg_tested_embs=[]
+    for tested_qid, (qid,(alpha,beta)) in zip(tested_qids,items.items()):
+        # tested_qid
+        if len(tested_qid)==0:
+            avg_tested_emb=np.array([0,0])
+        else:
+            avg_tested_emb = np.array([items[qid] for qid in tested_qid]).mean(axis=0)
+        item_emb=[alpha,beta]
         pred = alpha * theta + beta
         pred = torch.sigmoid(torch.tensor(pred))
         # pred = 1 / (1 + np.exp(-pred))
         q = 1 - pred
-        fisher_info = float((q*pred*(alpha ** 2)).numpy())
+        diff = ((item_emb-avg_tested_emb)**2).sum()
+        sim = ((item_emb-avg_embs)**2).sum()
+        fisher_info = float((q*pred*(alpha ** 2)).numpy()) + p*diff/sim
         fisher_arr.append((fisher_info,qid))
+        avg_tested_embs.append(avg_tested_emb.tolist())
     fisher_arr_sorted = sorted(fisher_arr, reverse=True)
-    return [i[1] for i in fisher_arr_sorted[:k]],[i[0]for i in fisher_arr]
+    tested_info=[]
+    for avg_tested_emb,n in zip(avg_tested_embs,ns):
+        avg_tested_emb.extend([n])
+        tested_info.append(avg_tested_emb)
+    return [i[1] for i in fisher_arr_sorted[:k]],[i[0]for i in fisher_arr],tested_info
+
+def get_k_emc(k,sid,theta,items,model):
+    epochs = model.config['num_epochs']
+    lr = model.config['learning_rate']
+    device = model.config['device']
+    optimizer = torch.optim.Adam(model.model.parameters(), lr=lr)
+    res_arr = []
+    for qid,(alpha,beta) in items.items():
+        for name, param in model.model.named_parameters():
+            if 'theta' not in name:
+                param.requires_grad = False
+
+        original_weights = model.model.theta.weight.data.clone()
+
+        student_id = torch.LongTensor([sid]).to(device)
+        question_id = torch.LongTensor([qid]).to(device)
+        correct = torch.LongTensor([1]).to(device).float()
+        wrong = torch.LongTensor([0]).to(device).float()
+
+        for ep in range(epochs):
+            optimizer.zero_grad()
+            pred = model.model(student_id, question_id)
+            loss = model._loss_function(pred, correct)
+            loss.backward()
+            optimizer.step()
 
+        pos_weights = model.model.theta.weight.data.clone()
+        model.model.theta.weight.data.copy_(original_weights)
+
+        for ep in range(epochs):
+            optimizer.zero_grad()
+            pred = model.model(student_id, question_id)
+            loss = model._loss_function(pred, wrong)
+            loss.backward()
+            optimizer.step()
+
+        neg_weights = model.model.theta.weight.data.clone()
+        # model.model.theta.weight.data.copy_(original_weights)
+
+        for param in model.model.parameters():
+            param.requires_grad = True
+
+        if type(alpha) == float:
+            alpha = np.array([alpha])
+        if type(theta) == float:
+            theta = np.array([theta])
+        pred = np.matmul(alpha.T, theta) + beta
+        pred = 1 / (1 + np.exp(-pred))
+        result = pred * torch.norm(pos_weights - original_weights).item() + \
+            (1 - pred) * torch.norm(neg_weights - original_weights).item()
+        res_arr.append((result,qid))
+    res_arr_sorted = sorted(res_arr, reverse=True)
+    return [i[1] for i in res_arr_sorted[:k]],[i[0]for i in res_arr]
 
 
 def get_k_kli(k, theta, items):

diff --git a/CAT/distillation/train.py b/CAT/distillation/train.py
@@ -8,33 +8,37 @@
 from CAT.distillation.tool import transform,split_data
 import torch
 import json
+import numpy as np
 dataset='assistment'
 cdm='irt'
-stg='KLI'
-trait = json.load(open(f'/data/yutingh/CAT/data/{dataset}/{stg}/trait.json', 'r'))
+stg='MFI'
+trait = json.load(open(f'/data/yutingh/CAT/data/{dataset}/{stg}/trait_with_tested_info.json', 'r'))
+# stg='KLI'
+# trait = json.load(open(f'/data/yutingh/CAT/data/{dataset}/{stg}/trait.json', 'r'))
 utrait = trait['user']
 itrait = trait['item']
 # print(utrait,itrait)
 label = trait['label']
 k_info = trait['k_info']
-if stg=='KLI':
+if 'tested_info' in trait:
     tested_info= trait['tested_info']
     train_data, test_data = split_data(utrait,label,k_info,0.8,tested_info)
+    user_dim=np.array(tested_info).shape[-1]+1
 else:
+    user_dim=1
     train_data, test_data = split_data(utrait,label,k_info,0.8)
 
 torch.manual_seed(0)
 train_set = transform(itrait,*train_data)
 test_set = transform(itrait,*test_data)
 k=50
 embedding_dim=15  
-epoch=30
+epoch=38
 lr=0.005 if dataset=='assistment' else 0.01
 print(f'lr: {lr}')
-user_dim=2 if stg =='KLI'else 1
 dMFI = dMFIModel(k,embedding_dim,user_dim,device='cuda:4')
 dMFI.train(train_set,test_set,itrait,epoch=epoch,lr=lr)
-dMFI.save(f'/data/yutingh/CAT/ckpt/{dataset}/{cdm}_{stg}_ip.pt')
+dMFI.save(f'/data/yutingh/CAT/ckpt/{dataset}/{cdm}_{stg}_ip_with_tested_info.pt')
 
 
 

diff --git a/CAT/mips/draw_pic.ipynb b/CAT/mips/draw_pic.ipynb
diff --git a/CAT/mips/prepare_ip_trait.py b/CAT/mips/prepare_ip_trait.py
@@ -12,22 +12,27 @@
 
 dataset='assistment'
 cdm='irt'
-stg='KLI'
-trait = json.load(open(f'/data/yutingh/CAT/data/{dataset}/{stg}/trait.json', 'r'))
+stg='MFI'
+trait = json.load(open(f'/data/yutingh/CAT/data/{dataset}/{stg}/trait_with_tested_info.json', 'r'))
 utrait = trait['user']
 itrait = trait['item']
 label = trait['label']
 k_info = trait['k_info']
+if 'tested_info' in trait:
+    tested_info= trait['tested_info']
+    user_dim=np.array(tested_info).shape[-1]+1
+else:
+    user_dim=1
+
 train_data, test_data = split_data(utrait,label,k_info,0.8)
 
 torch.manual_seed(0)
 train_set = transform(itrait,*train_data)
 test_set = transform(itrait,*test_data)
 k=50
 embedding_dim=15
-user_dim=2 if stg =='KLI'else 1
 dMFI = dMFIModel(k,embedding_dim,user_dim,device='cuda:0')
-dMFI.load(f'/data/yutingh/CAT/ckpt/{dataset}/{cdm}_{stg}_ip.pt')
+dMFI.load(f'/data/yutingh/CAT/ckpt/{dataset}/{cdm}_{stg}_ip_with_tested_info.pt')
 # dMFI.eval(test_set,itrait)
 ball_embs=[]
 max_embs_len=torch.tensor(0.)
@@ -47,7 +52,7 @@
 
 path_prefix = f"/data/yutingh/CAT/data/{dataset}/{stg}/"
 
-with open(f"{path_prefix}ball_trait.json", "w", encoding="utf-8") as f:
+with open(f"{path_prefix}ball_trait_with_tested_info.json", "w", encoding="utf-8") as f:
     # indent参数保证json数据的缩进，美观 
     # ensure_ascii=False才能输出中文，否则就是Unicode字符
     f.write(json.dumps(ball_embs, ensure_ascii=False))
diff --git a/CAT/mips/test.py b/CAT/mips/test.py
@@ -23,25 +23,27 @@ def setuplogger():
 
 import os
 os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
-def main(dataset="assistment", cdm="irt", stg = ['Random'], test_length = 20, ctx="cuda:4", lr=0.2, num_epoch=1, efficient=False):
-    lr=0.15 if dataset=='assistment' else 0.2
+def main(dataset="assistment", cdm="irt", stg = ['MFI'], test_length = 20, ctx="cuda:4", lr=0.2, num_epoch=1, efficient=True):
+    # lr=0.05 if dataset=='assistment' else 0.2
     setuplogger()
     seed = 0
     np.random.seed(seed)
     torch.manual_seed(seed)
-
-    config = {
-        'learning_rate': lr,
-        'batch_size': 2048,
-        'num_epochs': num_epoch,
-        'num_dim': 1, # for IRT or MIRT
-        'device': ctx,
-        # for NeuralCD
-        'prednet_len1': 128,
-        'prednet_len2': 64,
-        # 'prednet_len1': 64,
-        # 'prednet_len2': 32,
+    lr_config={
+        "assistment":{
+            "MFI":0.15,
+            "KLI":0.15,
+            "Random":0.05,
+            'MAAT':0.15
+        },
+       "junyi":{
+            "MFI":0.2,
+            "KLI":0.2,
+            "Random":0.2,
+            'MAAT':0.15
+       } 
     }
+
     metadata = json.load(open(f'/data/yutingh/CAT/data/{dataset}/metadata.json', 'r'))
     ckpt_path = f'/data/yutingh/CAT/ckpt/{dataset}/{cdm}.pt'
     # read datasets
@@ -64,6 +66,18 @@ def main(dataset="assistment", cdm="irt", stg = ['Random'], test_length = 20, ct
     df = pd.DataFrame() 
     df1 = pd.DataFrame()
     for i, strategy in enumerate(strategies):
+        config = {
+            'learning_rate': lr_config[dataset][stg[i]],
+            'batch_size': 2048,
+            'num_epochs': num_epoch,
+            'num_dim': 1, # for IRT or MIRT
+            'device': ctx,
+            # for NeuralCD
+            'prednet_len1': 128,
+            'prednet_len2': 64,
+            # 'prednet_len1': 64,
+            # 'prednet_len2': 32,
+        }
         if cdm == 'irt':
             model = CAT.model.IRTModel(**config)
         elif cdm =='ncd':
@@ -72,15 +86,20 @@ def main(dataset="assistment", cdm="irt", stg = ['Random'], test_length = 20, ct
         model.adaptest_load(ckpt_path)
         test_data.reset()
         if efficient:
-            ball_trait = json.load(open(f'/data/yutingh/CAT/data/{dataset}/{stg[i]}/ball_trait.json', 'r'))
+            ball_trait = json.load(open(f'/data/yutingh/CAT/data/{dataset}/{stg[i]}/ball_trait_with_tested_info.json', 'r'))
+            trait = json.load(open(f'/data/yutingh/CAT/data/{dataset}/{stg[i]}/trait_with_tested_info.json', 'r'))
             distill_k=50
             embedding_dim=15
-            user_dim=2 if stg[0]=='KLI' else 1
+            if 'tested_info' in trait:
+                tested_info= trait['tested_info']
+                user_dim=np.array(tested_info).shape[-1]+1
+            else:
+                user_dim=1
             dMFI = dMFIModel(distill_k,embedding_dim,user_dim,device=ctx)
-            dMFI.load(f'/data/yutingh/CAT/ckpt/{dataset}/{cdm}_{stg[i]}_ip.pt')
+            dMFI.load(f'/data/yutingh/CAT/ckpt/{dataset}/{cdm}_{stg[i]}_ip_with_tested_info.pt')
         logging.info('-----------')
         logging.info(f'start adaptive testing with {strategy.name} strategy')
-
+        logging.info('lr: ' + str(config['learning_rate']))
         logging.info(f'Iteration 0')
         res=[]
         time=0
@@ -99,27 +118,36 @@ def main(dataset="assistment", cdm="irt", stg = ['Random'], test_length = 20, ct
             results = tmp_model.evaluate(sid, test_data)
             tmp =[list(results.values())]
             time = datetime.timedelta(microseconds=0)
+            tested_info=[]
             for it in range(1, test_length + 1):
                 starttime = datetime.datetime.now()
                 if efficient:
-                    if stg[i]=='KLI':
-                        theta = tmp_model.model.theta(torch.tensor(sid).to(ctx))
-                        u_emb = dMFI.model.utn(torch.cat((theta,torch.Tensor([it]).to(ctx)),0)).tolist()
+                    theta = tmp_model.model.theta(torch.tensor(sid).to(ctx))
+                    if user_dim==1:
+                        u_emb = dMFI.model.utn(theta).tolist()
                     else:
-                        u_emb = dMFI.model.utn(tmp_model.model.theta(torch.tensor(sid).to(ctx))).tolist()
+                        if stg[i]=='KLI':
+                            u_emb = dMFI.model.utn(torch.cat((theta,torch.Tensor([it]).to(ctx)),0)).tolist()
+                        elif stg[i]=='MFI':
+                            if len(test_data.tested[sid])==0:
+                                avg_tested_emb=np.array([0,0]).tolist()
+                            else:
+                                avg_tested_emb = np.array([trait['item'][str(qid)] for qid in test_data.tested[sid]]).mean(axis=0).tolist()
+                            avg_tested_emb.extend([it])
+                            u_emb = dMFI.model.utn(torch.cat((theta,torch.Tensor(avg_tested_emb).to(ctx)),0)).tolist()
                     candidates=dict(zip(list(range(metadata['num_questions'],metadata['num_questions']+it)),[0]*it))
                     search_metric_tree(candidates,np.array(u_emb),T)
                     untested_qids = set(candidates.keys())-set(test_data.tested[sid])
                     # print(it, untested_qids)
-                    if len(untested_qids) == 1:
-                        max_score = 0 
-                        for k,v in candidates.items():
-                            if k in untested_qids:
-                                if v>max_score:
-                                    qid=k
-                                    max_score=v
-                    else:
-                       qid = strategy.adaptest_select(tmp_model, sid, test_data,item_candidates=untested_qids) 
+                    # if len(untested_qids) == 1:
+                    max_score = 0 
+                    for k,v in candidates.items():
+                        if k in untested_qids:
+                            if v>max_score:
+                                qid=k
+                                max_score=v
+                    # else:
+                    #    qid = strategy.adaptest_select(tmp_model, sid, test_data,item_candidates=untested_qids) 
                 else:
                     qid = strategy.adaptest_select(tmp_model, sid, test_data)
                 test_data.apply_selection(sid, qid)

diff --git a/ckpt/assistment/irt_MFI_ip.pt b/ckpt/assistment/irt_MFI_ip.pt
diff --git a/ckpt/assistment/irt_MFI_ip_with_tested_info.pt b/ckpt/assistment/irt_MFI_ip_with_tested_info.pt
diff --git a/ckpt/junyi/irt_KLI_ip.pt b/ckpt/junyi/irt_KLI_ip.pt