torch-rechub 0.0.1__py3-none-any.whl → 0.0.4__py3-none-any.whl

torch_rechub/trainers/match_trainer.py

@@ -0,0 +1,239 @@
+import os
+
+import torch
+import tqdm
+from sklearn.metrics import roc_auc_score
+
+from ..basic.callback import EarlyStopper
+from ..basic.loss_func import BPRLoss, RegularizationLoss
+
+
+class MatchTrainer(object):
+    """A general trainer for Matching/Retrieval
+
+    Args:
+        model (nn.Module): any matching model.
+        mode (int, optional): the training mode, `{0:point-wise, 1:pair-wise, 2:list-wise}`. Defaults to 0.
+        optimizer_fn (torch.optim): optimizer function of pytorch (default = `torch.optim.Adam`).
+        optimizer_params (dict): parameters of optimizer_fn.
+        scheduler_fn (torch.optim.lr_scheduler) : torch scheduling class, eg. `torch.optim.lr_scheduler.StepLR`.
+        scheduler_params (dict): parameters of optimizer scheduler_fn.
+        n_epoch (int): epoch number of training.
+        earlystop_patience (int): how long to wait after last time validation auc improved (default=10).
+        device (str): `"cpu"` or `"cuda:0"`
+        gpus (list): id of multi gpu (default=[]). If the length >=1, then the model will wrapped by nn.DataParallel.
+        model_path (str): the path you want to save the model (default="./"). Note only save the best weight in the validation data.
+    """
+
+    def __init__(
+        self,
+        model,
+        mode=0,
+        optimizer_fn=torch.optim.Adam,
+        optimizer_params=None,
+        regularization_params=None,
+        scheduler_fn=None,
+        scheduler_params=None,
+        n_epoch=10,
+        earlystop_patience=10,
+        device="cpu",
+        gpus=None,
+        model_path="./",
+    ):
+        self.model = model  # for uniform weights save method in one gpu or multi gpu
+        if gpus is None:
+            gpus = []
+        self.gpus = gpus
+        if len(gpus) > 1:
+            print('parallel running on these gpus:', gpus)
+            self.model = torch.nn.DataParallel(self.model, device_ids=gpus)
+        # torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+        self.device = torch.device(device)
+        self.model.to(self.device)
+        if optimizer_params is None:
+            optimizer_params = {"lr": 1e-3, "weight_decay": 1e-5}
+        if regularization_params is None:
+            regularization_params = {"embedding_l1": 0.0, "embedding_l2": 0.0, "dense_l1": 0.0, "dense_l2": 0.0}
+        self.mode = mode
+        if mode == 0:  # point-wise loss, binary cross_entropy
+            self.criterion = torch.nn.BCELoss()  # default loss binary cross_entropy
+        elif mode == 1:  # pair-wise loss
+            self.criterion = BPRLoss()
+        elif mode == 2:  # list-wise loss, softmax
+            self.criterion = torch.nn.CrossEntropyLoss()
+        else:
+            raise ValueError("mode only contain value in %s, but got %s" % ([0, 1, 2], mode))
+        self.optimizer = optimizer_fn(self.model.parameters(), **optimizer_params)  # default optimizer
+        self.scheduler = None
+        if scheduler_fn is not None:
+            self.scheduler = scheduler_fn(self.optimizer, **scheduler_params)
+        self.evaluate_fn = roc_auc_score  # default evaluate function
+        self.n_epoch = n_epoch
+        self.early_stopper = EarlyStopper(patience=earlystop_patience)
+        self.model_path = model_path
+        # Initialize regularization loss
+        self.reg_loss_fn = RegularizationLoss(**regularization_params)
+
+    def train_one_epoch(self, data_loader, log_interval=10):
+        self.model.train()
+        total_loss = 0
+        tk0 = tqdm.tqdm(data_loader, desc="train", smoothing=0, mininterval=1.0)
+        for i, (x_dict, y) in enumerate(tk0):
+            x_dict = {k: v.to(self.device) for k, v in x_dict.items()}  # tensor to GPU
+            y = y.to(self.device)
+            if self.mode == 0:
+                y = y.float()  # torch._C._nn.binary_cross_entropy expected Float
+            else:
+                y = y.long()  #
+            if self.mode == 1:  # pair_wise
+                pos_score, neg_score = self.model(x_dict)
+                loss = self.criterion(pos_score, neg_score)
+            else:
+                y_pred = self.model(x_dict)
+                loss = self.criterion(y_pred, y)
+
+            # Add regularization loss
+            reg_loss = self.reg_loss_fn(self.model)
+            loss = loss + reg_loss
+
+            # used for debug
+            # if i == 0:
+            #     print()
+            #     if self.mode == 0:
+            #         print('pred: ', [f'{float(each):5.2g}' for each in y_pred.detach().cpu().tolist()])
+            #         print('truth:', [f'{float(each):5.2g}' for each in y.detach().cpu().tolist()])
+            #     elif self.mode == 2:
+            #         pred = y_pred.detach().cpu().mean(0)
+            #         pred = torch.softmax(pred, dim=0).tolist()
+            #         print('pred: ', [f'{float(each):4.2g}' for each in pred])
+            #     elif self.mode == 1:
+            #         print('pos:', [f'{float(each):5.2g}' for each in pos_score.detach().cpu().tolist()])
+            #         print('neg: ', [f'{float(each):5.2g}' for each in neg_score.detach().cpu().tolist()])
+
+            self.model.zero_grad()
+            loss.backward()
+            self.optimizer.step()
+            total_loss += loss.item()
+            if (i + 1) % log_interval == 0:
+                tk0.set_postfix(loss=total_loss / log_interval)
+                total_loss = 0
+
+    def fit(self, train_dataloader, val_dataloader=None):
+        for epoch_i in range(self.n_epoch):
+            print('epoch:', epoch_i)
+            self.train_one_epoch(train_dataloader)
+            if self.scheduler is not None:
+                if epoch_i % self.scheduler.step_size == 0:
+                    print("Current lr : {}".format(self.optimizer.state_dict()['param_groups'][0]['lr']))
+                self.scheduler.step()  # update lr in epoch level by scheduler
+
+            if val_dataloader:
+                auc = self.evaluate(self.model, val_dataloader)
+                print('epoch:', epoch_i, 'validation: auc:', auc)
+                if self.early_stopper.stop_training(auc, self.model.state_dict()):
+                    print(f'validation: best auc: {self.early_stopper.best_auc}')
+                    self.model.load_state_dict(self.early_stopper.best_weights)
+                    break
+        torch.save(self.model.state_dict(), os.path.join(self.model_path, "model.pth"))  # save best auc model
+
+    def evaluate(self, model, data_loader):
+        model.eval()
+        targets, predicts = list(), list()
+        with torch.no_grad():
+            tk0 = tqdm.tqdm(data_loader, desc="validation", smoothing=0, mininterval=1.0)
+            for i, (x_dict, y) in enumerate(tk0):
+                x_dict = {k: v.to(self.device) for k, v in x_dict.items()}
+                y = y.to(self.device)
+                y_pred = model(x_dict)
+                targets.extend(y.tolist())
+                predicts.extend(y_pred.tolist())
+        return self.evaluate_fn(targets, predicts)
+
+    def predict(self, model, data_loader):
+        model.eval()
+        predicts = list()
+        with torch.no_grad():
+            tk0 = tqdm.tqdm(data_loader, desc="predict", smoothing=0, mininterval=1.0)
+            for i, (x_dict, y) in enumerate(tk0):
+                x_dict = {k: v.to(self.device) for k, v in x_dict.items()}
+                y = y.to(self.device)
+                y_pred = model(x_dict)
+                predicts.extend(y_pred.tolist())
+        return predicts
+
+    def inference_embedding(self, model, mode, data_loader, model_path):
+        # inference
+        assert mode in ["user", "item"], "Invalid mode={}.".format(mode)
+        model.mode = mode
+        model.load_state_dict(torch.load(os.path.join(model_path, "model.pth"), map_location=self.device, weights_only=True))
+        model = model.to(self.device)
+        model.eval()
+        predicts = []
+        with torch.no_grad():
+            tk0 = tqdm.tqdm(data_loader, desc="%s inference" % (mode), smoothing=0, mininterval=1.0)
+            for i, x_dict in enumerate(tk0):
+                x_dict = {k: v.to(self.device) for k, v in x_dict.items()}
+                y_pred = model(x_dict)
+                predicts.append(y_pred.data)
+        return torch.cat(predicts, dim=0)
+
+    def export_onnx(self, output_path, mode=None, dummy_input=None, batch_size=2, seq_length=10, opset_version=14, dynamic_batch=True, device=None, verbose=False):
+        """Export the trained matching model to ONNX format.
+
+        This method exports matching/retrieval models (e.g., DSSM, YoutubeDNN, MIND)
+        to ONNX format. For dual-tower models, you can export user tower and item
+        tower separately for efficient online serving.
+
+        Args:
+            output_path (str): Path to save the ONNX model file.
+            mode (str, optional): Export mode for dual-tower models:
+                - "user": Export only the user tower (for user embedding inference)
+                - "item": Export only the item tower (for item embedding inference)
+                - None: Export the full model (default)
+            dummy_input (dict, optional): Example input dict {feature_name: tensor}.
+                If not provided, dummy inputs will be generated automatically.
+            batch_size (int): Batch size for auto-generated dummy input (default: 2).
+            seq_length (int): Sequence length for SequenceFeature (default: 10).
+            opset_version (int): ONNX opset version (default: 14).
+            dynamic_batch (bool): Enable dynamic batch size (default: True).
+            device (str, optional): Device for export ('cpu', 'cuda', etc.).
+                If None, defaults to 'cpu' for maximum compatibility.
+            verbose (bool): Print export details (default: False).
+
+        Returns:
+            bool: True if export succeeded, False otherwise.
+
+        Example:
+            >>> trainer = MatchTrainer(dssm_model, mode=0, ...)
+            >>> trainer.fit(train_dl)
+
+            >>> # Export user tower for user embedding inference
+            >>> trainer.export_onnx("user_tower.onnx", mode="user")
+
+            >>> # Export item tower for item embedding inference
+            >>> trainer.export_onnx("item_tower.onnx", mode="item")
+
+            >>> # Export full model (for online similarity computation)
+            >>> trainer.export_onnx("full_model.onnx")
+
+            >>> # Export on specific device
+            >>> trainer.export_onnx("user_tower.onnx", mode="user", device="cpu")
+        """
+        from ..utils.onnx_export import ONNXExporter
+
+        # Handle DataParallel wrapped model
+        model = self.model.module if hasattr(self.model, 'module') else self.model
+
+        # Store original mode
+        original_mode = getattr(model, 'mode', None)
+
+        # Use provided device or default to 'cpu'
+        export_device = device if device is not None else 'cpu'
+
+        try:
+            exporter = ONNXExporter(model, device=export_device)
+            return exporter.export(output_path=output_path, mode=mode, dummy_input=dummy_input, batch_size=batch_size, seq_length=seq_length, opset_version=opset_version, dynamic_batch=dynamic_batch, verbose=verbose)
+        finally:
+            # Restore original mode
+            if hasattr(model, 'mode'):
+                model.mode = original_mode

torch_rechub/trainers/matching.md

@@ -0,0 +1,3 @@
+# Matching
+
+召回使用文档

torch_rechub/trainers/mtl_trainer.py

@@ -1,11 +1,15 @@
 import os
-import tqdm
+
 import numpy as np
 import torch
 import torch.nn as nn
+import tqdm
+
 from ..basic.callback import EarlyStopper
-from ..basic.utils import get_loss_func, get_metric_func
+from ..basic.loss_func import RegularizationLoss
 from ..models.multi_task import ESMM
+from ..utils.data import get_loss_func, get_metric_func
+from ..utils.mtl import MetaBalance, gradnorm, shared_task_layers
 
 
 class MTLTrainer(object):
@@ -18,7 +22,7 @@ class MTLTrainer(object):
         optimizer_params (dict): parameters of optimizer_fn.
         scheduler_fn (torch.optim.lr_scheduler) : torch scheduling class, eg. `torch.optim.lr_scheduler.StepLR`.
         scheduler_params (dict): parameters of optimizer scheduler_fn.
-        adaptive_params (dict): parameters of adaptive loss weight method. Now only support `{"method" : "uwl"}`. 
+        adaptive_params (dict): parameters of adaptive loss weight method. Now only support `{"method" : "uwl"}`.
         n_epoch (int): epoch number of training.
         earlystop_taskid (int): task id of earlystop metrics relies between multi task (default = 0).
         earlystop_patience (int): how long to wait after last time validation auc improved (default = 10).
@@ -32,10 +36,8 @@ class MTLTrainer(object):
         model,
         task_types,
         optimizer_fn=torch.optim.Adam,
-        optimizer_params={
-            "lr": 1e-3,
-            "weight_decay": 1e-5
-        },
+        optimizer_params=None,
+        regularization_params=None,
         scheduler_fn=None,
         scheduler_params=None,
         adaptive_params=None,
@@ -43,10 +45,16 @@ class MTLTrainer(object):
         earlystop_taskid=0,
         earlystop_patience=10,
         device="cpu",
-        gpus=[],
+        gpus=None,
         model_path="./",
     ):
         self.model = model
+        if gpus is None:
+            gpus = []
+        if optimizer_params is None:
+            optimizer_params = {"lr": 1e-3, "weight_decay": 1e-5}
+        if regularization_params is None:
+            regularization_params = {"embedding_l1": 0.0, "embedding_l2": 0.0, "dense_l1": 0.0, "dense_l2": 0.0}
         self.task_types = task_types
         self.n_task = len(task_types)
         self.loss_weight = None
@@ -56,7 +64,27 @@ class MTLTrainer(object):
                 self.adaptive_method = "uwl"
                 self.loss_weight = nn.ParameterList(nn.Parameter(torch.zeros(1)) for _ in range(self.n_task))
                 self.model.add_module("loss weight", self.loss_weight)
-        self.optimizer = optimizer_fn(self.model.parameters(), **optimizer_params)  #default Adam optimizer
+            elif adaptive_params["method"] == "metabalance":
+                self.adaptive_method = "metabalance"
+                share_layers, task_layers = shared_task_layers(self.model)
+                self.meta_optimizer = MetaBalance(share_layers)
+                self.share_optimizer = optimizer_fn(share_layers, **optimizer_params)
+                self.task_optimizer = optimizer_fn(task_layers, **optimizer_params)
+            elif adaptive_params["method"] == "gradnorm":
+                self.adaptive_method = "gradnorm"
+                self.alpha = adaptive_params.get("alpha", 0.16)
+                share_layers = shared_task_layers(self.model)[0]
+                # gradnorm calculate the gradients of each loss on the last
+                # fully connected shared layer weight(dimension is 2)
+                for i in range(len(share_layers)):
+                    if share_layers[-i].ndim == 2:
+                        self.last_share_layer = share_layers[-i]
+                        break
+                self.initial_task_loss = None
+                self.loss_weight = nn.ParameterList(nn.Parameter(torch.ones(1)) for _ in range(self.n_task))
+                self.model.add_module("loss weight", self.loss_weight)
+        if self.adaptive_method != "metabalance":
+            self.optimizer = optimizer_fn(self.model.parameters(), **optimizer_params)  # default Adam optimizer
         self.scheduler = None
         if scheduler_fn is not None:
             self.scheduler = scheduler_fn(self.optimizer, **scheduler_params)
@@ -65,27 +93,32 @@ class MTLTrainer(object):
         self.n_epoch = n_epoch
         self.earlystop_taskid = earlystop_taskid
         self.early_stopper = EarlyStopper(patience=earlystop_patience)
-        self.device = torch.device(device)
 
         self.gpus = gpus
         if len(gpus) > 1:
             print('parallel running on these gpus:', gpus)
             self.model = torch.nn.DataParallel(self.model, device_ids=gpus)
+        # torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+        self.device = torch.device(device)
+        self.model.to(self.device)
         self.model_path = model_path
+        # Initialize regularization loss
+        self.reg_loss_fn = RegularizationLoss(**regularization_params)
 
     def train_one_epoch(self, data_loader):
         self.model.train()
         total_loss = np.zeros(self.n_task)
         tk0 = tqdm.tqdm(data_loader, desc="train", smoothing=0, mininterval=1.0)
         for iter_i, (x_dict, ys) in enumerate(tk0):
-            x_dict = {k: v.to(self.device) for k, v in x_dict.items()}  #tensor to GPU
+            x_dict = {k: v.to(self.device) for k, v in x_dict.items()}  # tensor to GPU
             ys = ys.to(self.device)
             y_preds = self.model(x_dict)
             loss_list = [self.loss_fns[i](y_preds[:, i], ys[:, i].float()) for i in range(self.n_task)]
             if isinstance(self.model, ESMM):
-                loss = sum(loss_list[1:])  #ESSM only compute loss for ctr and ctcvr task
+                # ESSM only compute loss for ctr and ctcvr task
+                loss = sum(loss_list[1:])
             else:
-                if self.adaptive_method != None:
+                if self.adaptive_method is not None:
                     if self.adaptive_method == "uwl":
                         loss = 0
                         for loss_i, w_i in zip(loss_list, self.loss_weight):
@@ -93,38 +126,74 @@ class MTLTrainer(object):
                             loss += 2 * loss_i * torch.exp(-w_i) + w_i
                 else:
                     loss = sum(loss_list) / self.n_task
-            self.model.zero_grad()
-            loss.backward()
-            self.optimizer.step()
+
+            # Add regularization loss
+            reg_loss = self.reg_loss_fn(self.model)
+            loss = loss + reg_loss
+            if self.adaptive_method == 'metabalance':
+                self.share_optimizer.zero_grad()
+                self.task_optimizer.zero_grad()
+                self.meta_optimizer.step(loss_list)
+                self.share_optimizer.step()
+                self.task_optimizer.step()
+            elif self.adaptive_method == "gradnorm":
+                self.optimizer.zero_grad()
+                if self.initial_task_loss is None:
+                    self.initial_task_loss = [l.item() for l in loss_list]
+                gradnorm(loss_list, self.loss_weight, self.last_share_layer, self.initial_task_loss, self.alpha)
+                self.optimizer.step()
+                # renormalize
+                loss_weight_sum = sum([w.item() for w in self.loss_weight])
+                normalize_coeff = len(self.loss_weight) / loss_weight_sum
+                for w in self.loss_weight:
+                    w.data = w.data * normalize_coeff
+            else:
+                self.model.zero_grad()
+                loss.backward()
+                self.optimizer.step()
             total_loss += np.array([l.item() for l in loss_list])
         log_dict = {"task_%d:" % (i): total_loss[i] / (iter_i + 1) for i in range(self.n_task)}
+        loss_list = [total_loss[i] / (iter_i + 1) for i in range(self.n_task)]
         print("train loss: ", log_dict)
         if self.loss_weight:
             print("loss weight: ", [w.item() for w in self.loss_weight])
 
-    def fit(self, train_dataloader, val_dataloader):
-        self.model.to(self.device)
+        return loss_list
+
+    def fit(self, train_dataloader, val_dataloader, mode='base', seed=0):
+        total_log = []
+
         for epoch_i in range(self.n_epoch):
-            self.train_one_epoch(train_dataloader)
+            _log_per_epoch = self.train_one_epoch(train_dataloader)
+
             if self.scheduler is not None:
                 if epoch_i % self.scheduler.step_size == 0:
                     print("Current lr : {}".format(self.optimizer.state_dict()['param_groups'][0]['lr']))
-                self.scheduler.step()  #update lr in epoch level by scheduler
+                self.scheduler.step()  # update lr in epoch level by scheduler
             scores = self.evaluate(self.model, val_dataloader)
             print('epoch:', epoch_i, 'validation scores: ', scores)
+
+            for score in scores:
+                _log_per_epoch.append(score)
+
+            total_log.append(_log_per_epoch)
+
             if self.early_stopper.stop_training(scores[self.earlystop_taskid], self.model.state_dict()):
                 print('validation best auc of main task %d: %.6f' % (self.earlystop_taskid, self.early_stopper.best_auc))
                 self.model.load_state_dict(self.early_stopper.best_weights)
-                torch.save(self.early_stopper.best_weights, os.path.join(self.model_path, "model.pth"))  #save best auc model
                 break
 
+        torch.save(self.model.state_dict(), os.path.join(self.model_path, "model_{}_{}.pth".format(mode, seed)))  # save best auc model
+
+        return total_log
+
     def evaluate(self, model, data_loader):
         model.eval()
         targets, predicts = list(), list()
         with torch.no_grad():
             tk0 = tqdm.tqdm(data_loader, desc="validation", smoothing=0, mininterval=1.0)
             for i, (x_dict, ys) in enumerate(tk0):
-                x_dict = {k: v.to(self.device) for k, v in x_dict.items()}  #tensor to GPU
+                x_dict = {k: v.to(self.device) for k, v in x_dict.items()}  # tensor to GPU
                 ys = ys.to(self.device)
                 y_preds = self.model(x_dict)
                 targets.extend(ys.tolist())
@@ -142,4 +211,49 @@ class MTLTrainer(object):
                 x_dict = {k: v.to(self.device) for k, v in x_dict.items()}
                 y_preds = model(x_dict)
                 predicts.extend(y_preds.tolist())
-        return predicts
+        return predicts
+
+    def export_onnx(self, output_path, dummy_input=None, batch_size=2, seq_length=10, opset_version=14, dynamic_batch=True, device=None, verbose=False):
+        """Export the trained multi-task model to ONNX format.
+
+        This method exports multi-task learning models (e.g., MMOE, PLE, ESMM, SharedBottom)
+        to ONNX format for deployment. The exported model will have multiple outputs
+        corresponding to each task.
+
+        Note:
+            The ONNX model will output a tensor of shape [batch_size, n_task] where
+            n_task is the number of tasks in the multi-task model.
+
+        Args:
+            output_path (str): Path to save the ONNX model file.
+            dummy_input (dict, optional): Example input dict {feature_name: tensor}.
+                If not provided, dummy inputs will be generated automatically.
+            batch_size (int): Batch size for auto-generated dummy input (default: 2).
+            seq_length (int): Sequence length for SequenceFeature (default: 10).
+            opset_version (int): ONNX opset version (default: 14).
+            dynamic_batch (bool): Enable dynamic batch size (default: True).
+            device (str, optional): Device for export ('cpu', 'cuda', etc.).
+                If None, defaults to 'cpu' for maximum compatibility.
+            verbose (bool): Print export details (default: False).
+
+        Returns:
+            bool: True if export succeeded, False otherwise.
+
+        Example:
+            >>> trainer = MTLTrainer(mmoe_model, task_types=["classification", "classification"], ...)
+            >>> trainer.fit(train_dl, val_dl)
+            >>> trainer.export_onnx("mmoe.onnx")
+
+            >>> # Export on specific device
+            >>> trainer.export_onnx("mmoe.onnx", device="cpu")
+        """
+        from ..utils.onnx_export import ONNXExporter
+
+        # Handle DataParallel wrapped model
+        model = self.model.module if hasattr(self.model, 'module') else self.model
+
+        # Use provided device or default to 'cpu'
+        export_device = device if device is not None else 'cpu'
+
+        exporter = ONNXExporter(model, device=export_device)
+        return exporter.export(output_path=output_path, dummy_input=dummy_input, batch_size=batch_size, seq_length=seq_length, opset_version=opset_version, dynamic_batch=dynamic_batch, verbose=verbose)