Merge pull request #405 from duyiqi17/mind

update reademe and fix some configs

Author: seemingwang

models/recall/mind/README.md

@@ -76,6 +76,12 @@ os : windows/linux/macos
 
 在mind模型目录的快速执行命令如下：
 ```
+# 安装faiss
+# CPU
+pip install faiss-cpu
+# GPU
+# pip install faiss-gpu
+
 # 进入模型目录
 # cd models/recall/mind # 在任意目录均可运行
 # 动态图训练
@@ -99,7 +105,7 @@ python -u static_infer.py -m config.yaml -top_n 50  #对测试数据进行预测
 在全量数据下模型的指标如下：
 | 模型 |  batch_size | epoch_num| Recall@50 | NDCG@50 | HitRate@50 |Time of each epoch |
 | :------| :------ | :------ | :------| :------ | :------|  :------ | 
-| mind | 128 | 20 | 8.43% | 13.28% | 17.22% | -- |
+| mind | 128 | 20 | 8.43% | 13.28% | 17.22% | 398.64s(CPU) |
 
 
 1. 确认您当前所在目录为PaddleRec/models/recall/mind
@@ -108,12 +114,26 @@ python -u static_infer.py -m config.yaml -top_n 50  #对测试数据进行预测
 cd ../../../datasets/AmazonBook
 sh run.sh
 ``` 
-3. 切回模型目录,执行命令运行全量数据
+3. 安装依赖，我们使用[faiss](https://github.com/facebookresearch/faiss)来进行向量召回
+```bash
+# CPU-only version(pip)
+pip install faiss-cpu
+
+# GPU(+CPU) version(pip)
+#pip install faiss-gpu
+
+# CPU-only version(conda)
+#conda install -c pytorch faiss-cpu
+
+# GPU(+CPU) version(conda)
+#conda install -c pytorch faiss-gpu
+```
+4. 切回模型目录,执行命令运行全量数据
 ```bash
 cd - # 切回模型目录
 # 动态图训练
 python -u ../../../tools/trainer.py -m config_bigdata.yaml # 全量数据运行config_bigdata
-python -u infer.py -m config_bigdata.yaml # 全量数据运行config_bigdata
+python -u infer.py -m config_bigdata.yaml -top_n 50 # 全量数据运行config_bigdata
 ```
 
 ## 进阶使用

models/recall/mind/config.yaml

@@ -15,18 +15,19 @@
 runner:
   train_data_dir: "data/train"
   train_reader_path: "mind_reader" # importlib format
-  use_gpu: True
+  use_gpu: False
   use_auc: False
   train_batch_size: 128
-  epochs: 2
-  print_interval: 500
+  epochs: 1
+  print_interval: 10
   model_save_path: "output_model_mind"
   infer_batch_size: 128
   infer_reader_path: "mind_infer_reader" # importlib format
   test_data_dir: "data/valid"
   infer_load_path: "output_model_mind"
   infer_start_epoch: 0
   infer_end_epoch: 1
+  batches_per_epoch: 100
 
   # distribute_config
   # sync_mode: "async"
@@ -45,6 +46,6 @@ hyper_parameters:
   item_count: 367983
   embedding_dim: 64
   hidden_size: 64
-  neg_samples: 1280
+  neg_samples: 128
   maxlen: 20
   pow_p: 1.0

models/recall/mind/config_bigdata.yaml

@@ -15,10 +15,10 @@
 runner:
   train_data_dir: "../../../datasets/AmazonBook/train"
   train_reader_path: "mind_reader" # importlib format
-  use_gpu: True
+  use_gpu: False
   use_auc: False
   train_batch_size: 128
-  epochs: 6
+  epochs: 20
   print_interval: 500
   model_save_path: "output_model_mind"
   infer_batch_size: 128

models/recall/mind/mind_reader.py

@@ -24,6 +24,7 @@ def __init__(self, file_list, config):
         self.file_list = file_list
         self.maxlen = config.get("hyper_parameters.maxlen", 30)
         self.batch_size = config.get("runner.train_batch_size", 128)
+        self.batches_per_epoch = config.get("runner.batches_per_epoch", 1000)
         self.init()
         self.count = 0
 
@@ -52,7 +53,7 @@ def init(self):
     def __iter__(self):
         while True:
             user_id_list = random.sample(self.users, self.batch_size)
-            if self.count >= 1000 * self.batch_size:
+            if self.count >= self.batches_per_epoch * self.batch_size:
                 self.count = 0
                 break
             for user_id in user_id_list: