fix

.gitignore

@@ -1 +1,2 @@
-.vscode
+.vscode
+__pycache__

Dockerfile

@@ -0,0 +1,8 @@
+FROM python:3.8.6
+
+WORKDIR /app
+COPY ./requirements.txt /app/requirements.txt
+RUN pip3 install --upgrade pip && pip3 install -r ./requirements.txt
+
+ADD . /workspace
+WORKDIR /workspace

README.md

@@ -2,4 +2,38 @@
 
 ## Overview
 PyTorch implementation of the paper  
-[Xiao Ma, et al., Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate, SIGIR-2018](https://dl.acm.org/doi/abs/10.1145/3209978.3210104)
+[Xiao Ma, et al., Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate, SIGIR-2018](https://dl.acm.org/doi/abs/10.1145/3209978.3210104)
+
+## Dataset
+columns
+- `feature1` ... `feature_n`: categorical feautre column. (Assuming all variables are categorical.)
+- `click`: click label 
+- `conversion`: conversion label (Always 0 if there are no clicks)
+
+sample data
+```
+   feature1  feature2 feature3  click  conversion
+0         1        10       cc      1           1
+1         4        11        a      0           0
+2         6        30        5      0           0
+3        10         3       bb      1           0
+4         3        33       cc      1           1
+5         1         2        d      0           0
+6         4         5       cd      1           0
+```
+
+## Usage
+Environment construction with docker.
+
+Start container
+```
+$ docker-compose up -d --build
+```
+Attach container
+```
+$ docker exec -it esmm /bin/bash
+```
+Run training esmm
+```
+$ python train.py
+```

docker-compose.yml

@@ -0,0 +1,13 @@
+# 開発用
+version: '3'
+services:
+  app:
+    build:
+      context: ./
+      dockerfile: Dockerfile
+    image: esmm
+    volumes:
+      - './:/workspace'
+    container_name: esmm
+    tty: true
+    working_dir: '/workspace'

model/dataset.py

@@ -8,11 +8,12 @@ class EsmmDataset(torch.utils.data.Dataset):
     """
 
     def __init__(self, df: pd.DataFrame):
+        # Drop supervised columns
         df_feature = df.drop(columns=['click', 'conversion'])
 
         self.X = torch.from_numpy(df_feature.values).long()
-        self.click = torch.from_numpy(df['click'].values).float()  # label of click
-        self.conversion = torch.from_numpy(df['conversion'].values).float()  # label of conversion
+        self.click = torch.from_numpy(df['click'].values).float()  # click label
+        self.conversion = torch.from_numpy(df['conversion'].values).float()  # conversion label
 
         self.data_num = len(self.X)
 

requirements.txt

@@ -0,0 +1,8 @@
+pandas==1.4.0
+numpy==1.22.2
+category-encoders==1.3.0
+tqdm==4.62.3
+-f https://download.pytorch.org/whl/cpu/torch_stable.html
+torch==1.10.2+cpu
+torchvision==0.11.3+cpu
+torchaudio==0.10.2+cpu

main.py → train.py

@@ -12,13 +12,20 @@
 
 
 def load_data() -> pd.DataFrame:
-    # TODO implement data_load
+    """
+    Load sample data from local storage.
+
+    Returns:
+        pd.DataFrame: pandas dataframe
+    """
+    df = pd.read_pickle('/workspace/data/sample.pkl')
     return df
 
 
 def get_embedding_size(df: pd.DataFrame, embedding_dim: int) -> List[Tuple[int, int]]:
     """
     Get embedding size
+
     Args:
         df (pd.DataFrame): Train dataset
         embedding_dim (int): Number of embedded dimensions
@@ -36,27 +43,32 @@ def get_embedding_size(df: pd.DataFrame, embedding_dim: int) -> List[Tuple[int,
     return embedding_sizes
 
 
-def train(df: pd.DataFrame):
+def train(df: pd.DataFrame) -> None:
+    """
+    Train ESMM.
+
+    Args:
+        df (pd.DataFrame): Encoded dataset
+    """
     if torch.cuda.is_available():
         device = 'cuda'
     else:
         device = 'cpu'
 
+    # Build dataset
+    dataset = EsmmDataset(df)
+    train_loader = torch.utils.data.DataLoader(dataset, batch_size=64, shuffle=True)
+
     # Build model
     embedding_sizes = get_embedding_size(df, 5)
     model = ESMM(embedding_sizes)
     model = model.to(device)
 
     # Settings
-    batch_size = 64
     loss_fn = loss_fn = nn.BCELoss()
     optimizer = optim.SGD(model.parameters(), lr=0.01)
     epochs = 30
 
-    # Build dataloader
-    dataset = EsmmDataset(df)
-    train_loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True, num_workers=2)
-
     # Start fitting
     model.train()
     for epoch in range(epochs):
@@ -87,16 +99,14 @@ def train(df: pd.DataFrame):
         running_total_loss = running_total_loss / len(train_loader)
         running_ctr_loss = running_ctr_loss / len(train_loader)
         running_ctcvr_loss = running_ctcvr_loss / len(train_loader)
-        print(
-            f'epoch: {epoch+1}, total_loss: {running_total_loss}, ctr_loss: {running_ctr_loss}, ctcvr_loss: {running_ctcvr_loss}'
-        )
+        print(f'epoch: {epoch+1}, train_loss: {running_total_loss}')
 
 
 def main():
     # Load data
     df = load_data()
 
-    # Encode dataset
+    # Encode categorical columns
     category_columns = ['feature1', 'feature2', 'feature3']
     encoder = OrdinalEncoder(cols=category_columns, handle_unknown='impute').fit(df)
     df = encoder.transform(df)