Support DistributedDataParallel

Closes #116 (closed)

With this fix, we can

• Train on multiple GPUs
• Continue a training started with 1 GPU on multiple GPUs
• Continue a training started with multiple GPUs on 1 GPU
• Load any pre-trained model (trained with 0, 1 or more GPUs)

It also closes #143 (closed) and #118 (closed)

dan/manager/ocr.py

@@ -127,7 +127,6 @@ class OCRDatasetManager:
             batch_size=self.params["batch_size"],
             shuffle=True if self.train_sampler is None else False,
             drop_last=False,
-            batch_sampler=self.train_sampler,
             sampler=self.train_sampler,
             num_workers=self.params["num_gpu"] * self.params["worker_per_gpu"],
             pin_memory=self.pin_memory,
@@ -141,7 +140,6 @@ class OCRDatasetManager:
                 self.valid_datasets[key],
                 batch_size=1,
                 sampler=self.valid_samplers[key],
-                batch_sampler=self.valid_samplers[key],
                 shuffle=False,
                 num_workers=self.params["num_gpu"] * self.params["worker_per_gpu"],
                 pin_memory=self.pin_memory,