From 00d790a66d806f9fc48ee53f87f2e61ac0d7bb86 Mon Sep 17 00:00:00 2001
From: "Dr. Kashif Rasul" <kashif.rasul@gmail.com>
Date: Thu, 16 Jan 2020 10:32:30 +0100
Subject: [PATCH] use pytorch Dataloader and TransformedIterableDataset

---
 pts/dataset/__init__.py |  1 +
 pts/model/estimator.py  | 20 ++++++++++++--------
 pts/trainer.py          | 19 +++++++++++--------
 3 files changed, 24 insertions(+), 16 deletions(-)

diff --git a/pts/dataset/__init__.py b/pts/dataset/__init__.py
index 208479c..0f37e34 100644
--- a/pts/dataset/__init__.py
+++ b/pts/dataset/__init__.py
@@ -29,3 +29,4 @@ from .artificial import (
     generate_sf2,
 )
 from .multivariate_grouper import MultivariateGrouper
+from .transformed_iterable_dataset import TransformedIterableDataset
diff --git a/pts/model/estimator.py b/pts/model/estimator.py
index 0f76f84..ee8b4b9 100644
--- a/pts/model/estimator.py
+++ b/pts/model/estimator.py
@@ -4,8 +4,9 @@ from typing import NamedTuple
 import numpy as np
 import torch
 import torch.nn as nn
+from torch.utils.data import DataLoader
 
-from pts.dataset import Dataset, TrainDataLoader
+from pts.dataset import Dataset, TransformedIterableDataset
 from pts.transform import Transformation
 from pts import Trainer
 
@@ -98,16 +99,19 @@ class PTSEstimator(Estimator):
 
     def train_model(self, training_data: Dataset) -> TrainOutput:
         transformation = self.create_transformation()
-
         transformation.estimate(iter(training_data))
 
-        training_data_loader = TrainDataLoader(
+        training_iter_dataset = TransformedIterableDataset(
             dataset=training_data,
-            transform=transformation,
+            is_train=True,
+            transform=transformation
+        )
+
+        training_data_loader = DataLoader(
+            traning_iter_dataset,
             batch_size=self.trainer.batch_size,
-            num_batches_per_epoch=self.trainer.num_batches_per_epoch,
-            device=self.trainer.device,
-            dtype=self.dtype,
+            num_workers=self.train.num_workers,
+            pip_memory=self.trainer.pip_memory
         )
 
         # ensure that the training network is created on the same device
@@ -116,7 +120,7 @@ class PTSEstimator(Estimator):
         self.trainer(
             net=trained_net,
             input_names=get_module_forward_input_names(trained_net),
-            train_iter=training_data_loader,
+            data_loader=training_data_loader,
         )
 
         return TrainOutput(
diff --git a/pts/trainer.py b/pts/trainer.py
index c61b74d..44a1065 100644
--- a/pts/trainer.py
+++ b/pts/trainer.py
@@ -5,8 +5,7 @@ import torch
 import torch.nn as nn
 from tqdm import tqdm
 
-from pts.dataset import TrainDataLoader
-
+from torch.utils.data import DataLoader
 
 class Trainer:
     def __init__(
@@ -14,6 +13,8 @@ class Trainer:
         epochs: int = 100,
         batch_size: int = 32,
         num_batches_per_epoch: int = 50,
+        num_workers: int = 4,
+        pin_memory: bool = True,
         learning_rate: float = 1e-3,
         weight_decay: float = 1e-6,
         device: Optional[torch.device] = None,
@@ -24,13 +25,12 @@ class Trainer:
         self.learning_rate = learning_rate
         self.weight_decay = weight_decay
         self.device = device
+        self.num_workers = num_workers
+        self.pin_memory = pin_memory
 
     def __call__(
-        self, net: nn.Module, input_names: List[str], train_iter: TrainDataLoader
+        self, net: nn.Module, input_names: List[str], data_loader: DataLoader
     ) -> None:
-
-        net.to(self.device)
-
         optimizer = torch.optim.Adam(
             net.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay
         )
@@ -40,10 +40,10 @@ class Trainer:
             tic = time.time()
             avg_epoch_loss = 0.0
 
-            with tqdm(train_iter) as it:
+            with tqdm(data_loader) as it:
                 for batch_no, data_entry in enumerate(it, start=1):
                     optimizer.zero_grad()
-                    inputs = [data_entry[k] for k in input_names]
+                    inputs = [data_entry[k].to(self.device) for k in input_names]
 
                     output = net(*inputs)
                     if isinstance(output, (list, tuple)):
@@ -63,5 +63,8 @@ class Trainer:
                     loss.backward()
                     optimizer.step()
 
+                    if self.num_batches_per_epoch == batch_no:
+                        break()
+
             # mark epoch end time and log time cost of current epoch
             toc = time.time()