Merge pull request #239 from pykt-team/youh_dev

add new

Author: lycyhrc

README.md

@@ -76,8 +76,12 @@ The hyper parameter tunning results of our experiments about all the DLKT models
 22. DTransformer: Tracing Knowledge Instead of Patterns: Stable Knowledge Tracing with Diagnostic Transformer
 23. stableKT: Enhancing Length Generalization for Attention Based Knowledge Tracing Models with Linear Biases
 24. extraKT: Extending Context Window of Attention Based Knowledge Tracing Models via Length Extrapolation
-25. ReKT: Revisiting Knowledge Tracing: A Simple and Powerful Model
-
+25. csKT: Addressing Cold-start Problem in Knowledge Tracing via Kernel Bias and Cone Attention
+26. LefoKT: Rethinking and Improving Student Learning and Forgetting Processes for Attention Based Knowledge Tracing Models
+27. FlucKT: Cognitive Fluctuations Enhanced Attention Network for Knowledge Tracing
+28. UKT: Uncertainty-aware Knowledge Tracing
+29. HCGKT: Hierarchical Contrastive Graph Knowledge Tracing with Multi-level Feature Learning
+30. RobustKT: Enhancing Knowledge Tracing through Decoupling Cognitive Pattern from Error-Prone Data
 
 ## Citation
 

docs/pics/robustkt.png

@@ -0,0 +1,32 @@
+program: wandb_robustkt_train.py
+method: bayes
+metric:
+    goal: maximize
+    name: validauc
+parameters:
+    model_name:
+        values: ["robustkt"]
+    dataset_name:
+        values: ["xes"]
+    emb_type:
+        values: ["qid"]
+    save_dir:
+        values: ["models/akt_tiaocan"]
+    d_model:
+        values: [64, 256]
+    d_ff:
+        values: [64, 256]
+    dropout:
+        values: [0.05,0.1,0.3,0.5]
+    learning_rate:
+        values: [1e-3, 1e-4, 1e-5]
+    num_attn_heads:
+        values: [4, 8]
+    n_blocks:
+        values: [1, 2, 4]
+    kernel_size:
+        values: [4,5,8,16,32]
+    seed:
+        values: [42, 3407]
+    fold:
+        values: [0, 1, 2, 3, 4]

examples/seedwandb/robustkt.yaml

@@ -0,0 +1,27 @@
+import argparse
+from wandb_train import main
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--dataset_name", type=str, default="assist2009")
+    parser.add_argument("--model_name", type=str, default="robustkt")
+    parser.add_argument("--emb_type", type=str, default="qid")
+    parser.add_argument("--save_dir", type=str, default="saved_model")
+    parser.add_argument("--seed", type=int, default=3407)
+    parser.add_argument("--fold", type=int, default=0)
+    parser.add_argument("--dropout", type=float, default=0.2)
+    
+    parser.add_argument("--d_model", type=int, default=256)
+    parser.add_argument("--d_ff", type=int, default=512)
+    parser.add_argument("--num_attn_heads", type=int, default=8)
+    parser.add_argument("--n_blocks", type=int, default=4)
+    parser.add_argument("--learning_rate", type=float, default=1e-4)
+    parser.add_argument("--ks", type=float, default=5)
+
+    parser.add_argument("--use_wandb", type=int, default=0)
+    parser.add_argument("--add_uuid", type=int, default=0)
+   
+    args = parser.parse_args()
+
+    params = vars(args)
+    main(params)

examples/wandb_robustkt_train.py

@@ -39,7 +39,7 @@ def main(params):
     with open("../configs/kt_config.json") as f:
         config = json.load(f)
         train_config = config["train_config"]
-        if model_name in ["dkvmn","deep_irt", "sakt", "saint","saint++", "akt","folibikt", "atkt", "lpkt", "skvmn", "dimkt"]:
+        if model_name in ["dkvmn","deep_irt", "sakt", "saint","saint++", "akt", "robustkt", "folibikt", "atkt", "lpkt", "skvmn", "dimkt"]:
             train_config["batch_size"] = 64 ## because of OOM
         if model_name in ["simplekt","stablekt", "bakt_time", "sparsekt"]:
             train_config["batch_size"] = 64 ## because of OOM

examples/wandb_train.py

@@ -114,7 +114,7 @@ def evaluate(model, test_loader, model_name, rel=None, save_path=""):
             elif model_name == "saint":
                 y = model(cq.long(), cc.long(), r.long())
                 y = y[:, 1:]
-            elif model_name in ["akt","extrakt","folibikt", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx", "lefokt_akt", "fluckt"]:                                
+            elif model_name in ["akt","extrakt","folibikt", "robustkt", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx", "lefokt_akt", "fluckt"]:                                
                 y, reg_loss = model(cc.long(), cr.long(), cq.long())
                 y = y[:,1:]
             elif model_name in ["dtransformer"]:
@@ -183,7 +183,7 @@ def early_fusion(curhs, model, model_name):
         que_diff = model.diff_layer(curhs[1])#equ 13
         p = torch.sigmoid(3.0*stu_ability-que_diff)#equ 14
         p = p.squeeze(-1)
-    elif model_name in ["akt","extrakt", "folibikt","dtransformer","simplekt","stablekt","cskt", "fluckt", "bakt_time", "sparsekt", "lefokt_akt", "ukt", "hcgkt", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx"]:
+    elif model_name in ["akt","extrakt", "folibikt","robustkt", "dtransformer","simplekt","stablekt","cskt", "fluckt", "bakt_time", "sparsekt", "lefokt_akt", "ukt", "hcgkt", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx"]:
         output = model.out(curhs[0]).squeeze(-1)
         m = nn.Sigmoid()
         p = m(output)
@@ -229,7 +229,7 @@ def effective_fusion(df, model, model_name, fusion_type):
 
     curhs, curr = [[], []], []
     dcur = {"late_trues": [], "qidxs": [], "questions": [], "concepts": [], "row": [], "concept_preds": []}
-    hasearly = ["dkvmn","deep_irt", "skvmn", "kqn", "akt","extrakt", "folibikt", "dtransformer", "simplekt","stablekt","cskt","fluckt", "ukt", "hcgkt", "bakt_time", "sparsekt","lefokt_akt",  "saint", "sakt", "hawkes", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx", "lpkt"]
+    hasearly = ["dkvmn","deep_irt", "skvmn", "kqn", "akt","extrakt", "folibikt", "robustkt", "dtransformer", "simplekt","stablekt","cskt","fluckt", "ukt", "hcgkt", "bakt_time", "sparsekt","lefokt_akt",  "saint", "sakt", "hawkes", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx", "lpkt"]
     for ui in df:
         # 一题一题处理
         curdf = ui[1]
@@ -277,7 +277,7 @@ def group_fusion(dmerge, model, model_name, fusion_type, fout):
     if cq.shape[1] == 0:
         cq = cc
 
-    hasearly = ["dkvmn","deep_irt", "skvmn", "kqn", "dtransformer", "akt","extrakt", "folibikt","simplekt","stablekt","cskt", "fluckt", "ukt",  "hcgkt", "bakt_time", "sparsekt","lefokt_akt",  "saint", "sakt", "hawkes", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx", "lpkt"]
+    hasearly = ["dkvmn","deep_irt", "skvmn", "kqn", "dtransformer", "akt","robustkt", "extrakt", "folibikt","simplekt","stablekt","cskt", "fluckt", "ukt",  "hcgkt", "bakt_time", "sparsekt","lefokt_akt",  "saint", "sakt", "hawkes", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx", "lpkt"]
 
     alldfs, drest = [], dict() # not predict infos!
     # print(f"real bz in group fusion: {rs.shape[0]}")
@@ -374,7 +374,7 @@ def evaluate_question(model, test_loader, model_name, fusion_type=["early_fusion
     # dkvmn / akt / saint: give cur -> predict cur
     # sakt: give past+cur -> predict cur
     # kqn: give past+cur -> predict cur
-    hasearly = ["dkvmn","deep_irt", "skvmn", "kqn", "dtransformer", "akt","extrakt","folibikt", "simplekt","cskt","fluckt", "stablekt", "ukt", "hcgkt", "bakt_time", "sparsekt", "lefokt_akt", "saint", "sakt", "hawkes", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx", "lpkt"]
+    hasearly = ["dkvmn","deep_irt", "skvmn", "kqn", "dtransformer", "akt","extrakt","folibikt", "robustkt", "simplekt","cskt","fluckt", "stablekt", "ukt", "hcgkt", "bakt_time", "sparsekt", "lefokt_akt", "saint", "sakt", "hawkes", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx", "lpkt"]
     if save_path != "":
         fout = open(save_path, "w", encoding="utf8")
         if model_name in hasearly:
@@ -433,7 +433,7 @@ def evaluate_question(model, test_loader, model_name, fusion_type=["early_fusion
                 y = y[:,1:]
             elif model_name in ["rekt"]:
                 y, h = model(dcurori, qtest=True, train=False)
-            elif model_name in ["akt","extrakt", "folibikt","fluckt","lefokt_akt", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx"]:
+            elif model_name in ["akt","extrakt", "folibikt","fluckt","robustkt", "lefokt_akt", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx"]:
                 y, reg_loss, h = model(cc.long(), cr.long(), cq.long(), True)
                 y = y[:,1:]
             elif model_name in ["dtransformer"]:
@@ -934,7 +934,7 @@ def predict_each_group(dtotal, dcur, dforget, curdforget, is_repeat, qidx, uid,
             # 应该用预测的r更新memory value，但是这里一个知识点一个知识点预测，所以curr不起作用！
             y = model(cin.long(), rin.long())
             pred = y[0][-1]
-        elif model_name in ["akt","extrakt","folibikt","fluckt", "lefokt_akt", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx"]:  
+        elif model_name in ["akt","extrakt","folibikt","fluckt", "robustkt","lefokt_akt", "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx"]:  
             #### 输入有question！     
             if qout != None:
                 curq = torch.tensor([[qout.item()]]).to(device)
@@ -1318,7 +1318,7 @@ def predict_each_group2(dtotal, dcur, dforget, curdforget, is_repeat, qidx, uid,
         elif model_name == "saint":
             y = model(ccq.long(), ccc.long(), curr.long())
             y = y[:, 1:]
-        elif model_name in ["akt","extrakt","folibikt", "cakt","fluckt","lefokt_akt",  "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx"]:                                
+        elif model_name in ["akt","extrakt","folibikt", "robustkt", "cakt","fluckt","lefokt_akt",  "akt_vector", "akt_norasch", "akt_mono", "akt_attn", "aktattn_pos", "aktmono_pos", "akt_raschx", "akt_raschy", "aktvec_raschx"]:                                
             y, reg_loss = model(ccc.long(), ccr.long(), ccq.long())
             y = y[:,1:]
         elif model_name in ["dtransformer"]:

pykt/models/evaluate_model.py

@@ -37,6 +37,7 @@
 from .lefokt_akt import LEFOKT_AKT
 from .ukt import UKT
 from .hcgkt import HCGKT
+from .robustkt import Robustkt
 
 device = "cpu" if not torch.cuda.is_available() else "cuda"
 
@@ -135,6 +136,8 @@ def init_model(model_name, model_config, data_config, emb_type):
         model = UKT(data_config["num_c"], data_config["num_q"], **model_config, emb_type=emb_type, emb_path=data_config["emb_path"]).to(device)
     elif model_name == "hcgkt":
         model = HCGKT(data_config["num_c"], data_config["num_q"], **model_config, emb_type=emb_type, emb_path=data_config["emb_path"]).to(device)
+    elif model_name == "robustkt":
+        model = Robustkt(data_config["num_c"], data_config["num_q"], **model_config, emb_type=emb_type, emb_path=data_config["emb_path"]).to(device)
     elif model_name == "dtransformer":
         model = DTransformer(data_config["num_c"], data_config["num_q"], **model_config, emb_type=emb_type,
                      emb_path=data_config["emb_path"]).to(device)