Merge branch 'main' into lmcafee/distrib-opt-nodupe

merge sequence parallelism's layernorm all-reduce into distributed optimizer.

megatron/p2p_communication.py

@@ -61,7 +61,8 @@ def _communicate(tensor_send_next, tensor_send_prev, recv_prev, recv_next,
         tensor_shape = (args.seq_length, args.micro_batch_size, args.hidden_size)
 
     override_scatter_gather_tensors_in_pipeline = False
-    if args.scatter_gather_tensors_in_pipeline:
+    if args.scatter_gather_tensors_in_pipeline and \
+            not args.sequence_parallel:
         tensor_chunk_shape = reduce(operator.mul, tensor_shape, 1)
         if tensor_chunk_shape % mpu.get_tensor_model_parallel_world_size() == 0:
             tensor_chunk_shape = tensor_chunk_shape // \
@@ -93,7 +94,8 @@ def _communicate(tensor_send_next, tensor_send_prev, recv_prev, recv_next,
 
     # Split tensor into smaller chunks if using scatter-gather optimization.
     if not override_scatter_gather_tensors_in_pipeline and \
-            args.scatter_gather_tensors_in_pipeline:
+            args.scatter_gather_tensors_in_pipeline and \
+            not args.sequence_parallel:
         if tensor_send_next is not None:
             tensor_send_next = mpu.split_tensor_into_1d_equal_chunks(tensor_send_next)
 
@@ -138,7 +140,8 @@ def _communicate(tensor_send_next, tensor_send_prev, recv_prev, recv_next,
 
     # If using scatter-gather optimization, gather smaller chunks.
     if not override_scatter_gather_tensors_in_pipeline and \
-            args.scatter_gather_tensors_in_pipeline:
+            args.scatter_gather_tensors_in_pipeline and \
+            not args.sequence_parallel:
         if recv_prev:
             tensor_recv_prev = mpu.gather_split_1d_tensor(
                 tensor_recv_prev).view(tensor_shape).requires_grad_()

megatron/schedules.py

@@ -279,8 +279,12 @@ def forward_backward_pipelining_with_interleaving(forward_step_func,
     pipeline_parallel_rank = mpu.get_pipeline_model_parallel_rank()
 
     args = get_args()
-    tensor_shape = (args.seq_length, args.micro_batch_size, args.hidden_size)
-
+    if args.sequence_parallel:
+        seq_length = args.seq_length // mpu.get_tensor_model_parallel_world_size()
+    else:
+        seq_length = args.seq_length
+    tensor_shape = (seq_length, args.micro_batch_size, args.hidden_size)
+    
     # Compute number of warmup and remaining microbatches.
     num_model_chunks = len(model)
     num_microbatches = get_num_microbatches() * num_model_chunks
@@ -514,18 +518,25 @@ def get_tensor_shapes(rank, model_type):
     # Otherwise, send one tensor (pre-transpose).
     args = get_args()
     tensor_shapes = []
+
+    if args.sequence_parallel:
+        seq_length = args.seq_length // mpu.get_tensor_model_parallel_world_size()
+    else:
+        seq_length = args.seq_length
+
     if model_type == ModelType.encoder_and_decoder:
+        if args.sequence_parallel:
+            decoder_seq_length = args.decoder_seq_length // mpu.get_tensor_model_parallel_world_size()
+        else:
+            decoder_seq_length = args.decoder_seq_length
+
         if mpu.is_pipeline_stage_before_split(rank):
-            # If next rank is after split, then need transpose for encoder_hidden_state.
-            if mpu.is_pipeline_stage_before_split(rank+1):
-                tensor_shapes.append((args.seq_length, args.micro_batch_size, args.hidden_size))
-            else:
-                tensor_shapes.append((args.micro_batch_size, args.seq_length, args.hidden_size))
+            tensor_shapes.append((seq_length, args.micro_batch_size, args.hidden_size))
         else:
-            tensor_shapes.append((args.decoder_seq_length, args.micro_batch_size, args.hidden_size))
-            tensor_shapes.append((args.micro_batch_size, args.seq_length, args.hidden_size))
+            tensor_shapes.append((decoder_seq_length, args.micro_batch_size, args.hidden_size))
+            tensor_shapes.append((seq_length, args.micro_batch_size, args.hidden_size))
     else:
-        tensor_shapes.append((args.seq_length, args.micro_batch_size, args.hidden_size))
+        tensor_shapes.append((seq_length, args.micro_batch_size, args.hidden_size))
     return tensor_shapes
 
 

megatron/training.py

@@ -23,6 +23,7 @@ import time
 _TRAIN_START_TIME = time.time()
 import torch
 from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP
+from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
 
 from megatron import get_args
 from megatron import get_signal_handler
@@ -42,6 +43,7 @@ from megatron.model import ModelType
 from megatron.optimizer import get_megatron_optimizer
 from megatron.initialize import initialize_megatron
 from megatron.initialize import write_args_to_tensorboard
+from megatron.initialize import set_jit_fusion_options
 from megatron.optimizer_param_scheduler import OptimizerParamScheduler
 from megatron.model import DistributedDataParallel as LocalDDP
 from megatron.utils import check_adlr_autoresume_termination
@@ -99,6 +101,8 @@ def pretrain(train_valid_test_dataset_provider,
     # Initalize and get arguments, timers, and Tensorboard writer.
     initialize_megatron(extra_args_provider=extra_args_provider,
                         args_defaults=args_defaults)
+    # Set pytorch JIT layer fusion options and warmup JIT functions.
+    set_jit_fusion_options()
 
     # Adjust the startup time so it reflects the largest value.
     # This will be closer to what scheduler will see (outside of

tasks/finetune_utils.py

@@ -229,7 +229,7 @@ def _train(model, optimizer, opt_param_scheduler, forward_step,
                 prefix = 'iteration {}'.format(iteration)
                 evaluate_and_print_results(prefix, forward_step,
                                            valid_dataloader, model,
-                                           iteration, False)
+                                           iteration, None, False)
 
             # Exiting based on iterations
             if args.exit_interval and iteration % args.exit_interval == 0:

tasks/vision/classification/classification.py

@@ -15,12 +15,15 @@
 
 """Vision-classification finetuning/evaluation."""
 
-from megatron import get_args
+import torch.nn.functional as F
+from functools import partial
+from megatron import get_args, get_timers
 from megatron import print_rank_0
-from megatron.model.vit_model import VitModel
+from megatron.model.vision.classification import VitClassificationModel
 from megatron.data.vit_dataset import build_train_valid_datasets
-from tasks.vision.eval_utils import accuracy_func_provider
+from tasks.vision.classification.eval_utils import accuracy_func_provider
 from tasks.vision.finetune_utils import finetune
+from megatron.utils import average_losses_across_data_parallel_group
 
 
 def classification():
@@ -30,7 +33,7 @@ def classification():
 
         train_ds, valid_ds = build_train_valid_datasets(
             data_path=args.data_path,
-            crop_size=args.img_dim,
+            image_size=(args.img_h, args.img_w),
         )
         return train_ds, valid_ds
 
@@ -40,16 +43,52 @@ def classification():
 
         print_rank_0("building classification model for ImageNet ...")
 
-        return VitModel(num_classes=args.num_classes, finetune=True,
-                        pre_process=pre_process, post_process=post_process)
+        return VitClassificationModel(num_classes=args.num_classes, finetune=True,
+                                      pre_process=pre_process, post_process=post_process)
+
+    def process_batch(batch):
+        """Process batch and produce inputs for the model."""
+        images = batch[0].cuda().contiguous()
+        labels = batch[1].cuda().contiguous()
+        return images, labels
+
+    def cross_entropy_loss_func(labels, output_tensor):
+        logits = output_tensor
+
+        # Cross-entropy loss.
+        loss = F.cross_entropy(logits.contiguous().float(), labels)
+
+        # Reduce loss for logging.
+        averaged_loss = average_losses_across_data_parallel_group([loss])
+
+        return loss, {'lm loss': averaged_loss[0]}
+
+    def _cross_entropy_forward_step(batch, model):
+        """Simple forward step with cross-entropy loss."""
+        timers = get_timers()
+
+        # Get the batch.
+        timers("batch generator").start()
+        try:
+            batch_ = next(batch)
+        except BaseException:
+            batch_ = batch
+        images, labels = process_batch(batch_)
+        timers("batch generator").stop()
+
+        # Forward model.
+        output_tensor = model(images)
+      
+        return output_tensor, partial(cross_entropy_loss_func, labels)
 
     """Finetune/evaluate."""
     finetune(
         train_valid_datasets_provider,
         model_provider,
+        forward_step=_cross_entropy_forward_step,
         end_of_epoch_callback_provider=accuracy_func_provider,
     )
 
-
 def main():
     classification()
+

tasks/vision/classification/eval_utils.py

@@ -33,11 +33,10 @@ def accuracy_func_provider():
     """Provide function that calculates accuracies."""
     args = get_args()
     data_path = args.data_path
-    crop_size = args.img_dim
+    crop_size = (args.img_h, args.img_w)
 
-    # mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
     # Build dataloaders.
-    val_data_path = os.path.join(data_path[0], "val")
+    val_data_path = data_path[1]
     normalize = transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
     transform_val = transforms.Compose(
         [
@@ -54,6 +53,7 @@ def accuracy_func_provider():
         args.micro_batch_size,
         num_workers=args.num_workers,
         drop_last=(mpu.get_data_parallel_world_size() > 1),
+        shuffle=False
     )
 
     def metrics_func(model, epoch):
@@ -71,7 +71,6 @@ def accuracy_func_provider():
 def calculate_correct_answers(model, dataloader, epoch):
     """Calculate correct over total answers"""
 
-    args = get_args()
     forward_backward_func = get_forward_backward_func()
     for m in model:
         m.eval()
@@ -98,7 +97,6 @@ def calculate_correct_answers(model, dataloader, epoch):
         images, labels = process_batch(batch_)
 
         # Forward model.
-        args = get_args()
         output_tensor = model(images)
 
         return output_tensor, partial(loss_func, labels)

tasks/vision/finetune_utils.py

@@ -17,11 +17,10 @@
 
 import torch
 import torch.nn.functional as F
-from functools import partial
 from megatron import get_args
 from megatron import print_rank_0
 from megatron import get_timers
-from megatron import mpu
+from megatron import mpu, utils
 from megatron.checkpointing import load_checkpoint
 from megatron.checkpointing import save_checkpoint
 from megatron.training import evaluate_and_print_results
@@ -29,7 +28,10 @@ from megatron.training import setup_model_and_optimizer
 from megatron.training import train_step
 from megatron.training import training_log
 from megatron.utils import check_adlr_autoresume_termination
-from megatron.utils import average_losses_across_data_parallel_group
+from megatron.utils import average_losses_across_data_parallel_group, print_params_min_max_norm
+from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP
+from megatron.model import DistributedDataParallel as LocalDDP
+from megatron.model import Float16Module, ModelType
 
 
 def process_batch(batch):
@@ -39,45 +41,16 @@ def process_batch(batch):
     return images, labels
 
 
-def cross_entropy_loss_func(labels, output_tensor):
-    logits = output_tensor
-
-    # Cross-entropy loss.
-    loss = F.cross_entropy(logits.contiguous().float(), labels)
-
-    # Reduce loss for logging.
-    averaged_loss = average_losses_across_data_parallel_group([loss])
-
-    return loss, {'lm loss': averaged_loss[0]}
-
-
-def _cross_entropy_forward_step(batch, model):
-    """Simple forward step with cross-entropy loss."""
-    timers = get_timers()
-
-    # Get the batch.
-    timers("batch generator").start()
-    try:
-        batch_ = next(batch)
-    except BaseException:
-        batch_ = batch
-    images, labels = process_batch(batch_)
-    timers("batch generator").stop()
-
-   # Forward model.
-    output_tensor = model(images)
-  
-    return output_tensor, partial(cross_entropy_loss_func, labels)
-
-
-def build_data_loader(dataset, micro_batch_size, num_workers, drop_last):
+def build_data_loader(dataset, micro_batch_size,
+                      num_workers, drop_last, shuffle):
     """Data loader. Note that batch-size is the local (per GPU) batch-size."""
 
     # Sampler.
     world_size = mpu.get_data_parallel_world_size()
     rank = mpu.get_data_parallel_rank()
     sampler = torch.utils.data.distributed.DistributedSampler(
-        dataset, num_replicas=world_size, rank=rank
+        dataset, num_replicas=world_size, rank=rank,
+        drop_last=drop_last, shuffle=shuffle
     )
 
     # Data loader. Note that batch size is the per GPU batch size.
@@ -112,14 +85,14 @@ def _build_train_valid_dataloaders(train_dataset, valid_dataset):
     print_rank_0('building train and validation dataloaders ...')
     # Training dataset.
     train_dataloader = build_data_loader(train_dataset, args.micro_batch_size,
-                                           args.num_workers, not args.keep_last)
+                                         args.num_workers, False, True)
     # Set the training iterations.
     args.train_iters_per_epoch = len(train_dataloader)
     args.train_iters = args.epochs * args.train_iters_per_epoch
     # Validation dataset. For this dataset, we do not need to set up
     # shuffling so we can just use a simple infinite loop.
     valid_dataloader_ = build_data_loader(valid_dataset, args.micro_batch_size,
-                                            args.num_workers, not args.keep_last)
+                                          args.num_workers, True,  False)
     valid_dataloader = _build_infinite_size_dataloader(valid_dataloader_)
 
     # Now that we've built the data loaders, set batch_size arguments
@@ -132,6 +105,7 @@ def _build_train_valid_dataloaders(train_dataset, valid_dataset):
 
     return train_dataloader, valid_dataloader
 
+
 def _train(
     model,
     optimizer,
@@ -140,6 +114,7 @@ def _train(
     train_dataloader,
     valid_dataloader,
     end_of_epoch_callback,
+    process_non_loss_data_func=None
 ):
     """Train the model."""
     args = get_args()
@@ -167,6 +142,7 @@ def _train(
 
         # Set the data loader epoch to shuffle the index iterator.
         train_dataloader.sampler.set_epoch(args.seed + epoch)
+        train_dataloader.dataset.set_epoch(epoch)
 
         # For all the batches in the dataset.
         for iteration_, batch in enumerate(train_dataloader):
@@ -185,8 +161,6 @@ def _train(
 
             # Logging.
             params_norm = None
-            if args.log_params_norm:
-                params_norm = calc_params_l2_norm(model)
 
             report_memory_flag = training_log(
                 losses_dict,
@@ -202,20 +176,16 @@ def _train(
             )
 
             # Autoresume
-            if args.adlr_autoresume and (
-                iteration % args.adlr_autoresume_interval == 0
-            ):
-                check_adlr_autoresume_termination(
-                    iteration, model, optimizer, opt_param_scheduler
-                )
+            if args.adlr_autoresume and \
+                    iteration % args.adlr_autoresume_interval == 0:
+                check_adlr_autoresume_termination(iteration, model, optimizer,
+                                                  opt_param_scheduler)
 
             # Checkpointing
-            if (
-                args.save
-                and args.save_interval
-                and iteration % args.save_interval == 0
-            ):
-                save_checkpoint(iteration, model, optimizer, opt_param_scheduler)
+            if args.save and args.save_interval and \
+                    iteration % args.save_interval == 0:
+                save_checkpoint(iteration, model, optimizer,
+                                opt_param_scheduler)
 
             # Evaluation
             if args.eval_interval and iteration % args.eval_interval == 0:
@@ -226,13 +196,10 @@ def _train(
                     valid_dataloader,
                     model,
                     iteration,
+                    process_non_loss_data_func,
                     False,
                 )
 
-        # Checkpointing at the end of each epoch.
-        if args.save:
-            save_checkpoint(iteration, model, optimizer, opt_param_scheduler)
-
         # Callback at the end of each epoch.
         if end_of_epoch_callback is not None:
             end_of_epoch_callback(model, epoch)
@@ -241,7 +208,9 @@ def _train(
 def finetune(
     train_valid_datasets_provider,
     model_provider,
-    forward_step=_cross_entropy_forward_step,
+    forward_step,
+    model_type=ModelType.encoder_or_decoder,
+    process_non_loss_data_func=None,
     end_of_epoch_callback_provider=None,
 ):
     """Main finetune function used across all tasks."""
@@ -266,7 +235,12 @@ def finetune(
 
     # Build model, optimizer and learning rate scheduler.
     timers("model and optimizer").start()
-    model, optimizer, opt_param_scheduler = setup_model_and_optimizer(model_provider)
+    model, optimizer, opt_param_scheduler = \
+        setup_model_and_optimizer(
+            model_provider,
+            model_type,
+            scale_lr_cond=lambda name, param: ".head." in name,
+            lr_mult=args.head_lr_mult)
     timers("model and optimizer").stop()
 
     # If pretrained checkpoint is provided and we have not trained for
@@ -274,13 +248,34 @@ def finetune(
     # checkpoint.
     timers("pretrained checkpoint").start()
     if args.iteration == 0 and args.pretrained_checkpoint is not None:
-        original_load = args.load
-        args.load = args.pretrained_checkpoint
-        _ = load_checkpoint(model, None, None, strict=False)
-        args.load = original_load
+        if args.pretrained_checkpoint_type == 'default':
+            original_load = args.load
+            args.load = args.pretrained_checkpoint
+            _ = load_checkpoint(model, None, None, strict=False)
+            args.load = original_load
+        elif args.pretrained_checkpoint_type == 'external':
+            unwrap_model = utils.unwrap_model(model)
+            state_dict = torch.load(args.pretrained_checkpoint,
+                                    map_location="cpu")
+            unwrap_model[0].module.backbone.load_state_dict(state_dict,
+                                                            strict=False)
+        elif args.pretrained_checkpoint_type == 'constrastive':
+            unwrap_model = utils.unwrap_model(model)
+            state_dict = torch.load(args.pretrained_checkpoint,
+                                    map_location="cpu")
+            state_dict = state_dict["model"]
+            state_dict = {k.replace("teacher.backbone.", ""): v
+                          for k, v in state_dict.items()
+                          if k.startswith("teacher.backbone.")}
+            unwrap_model[0].module.backbone.load_state_dict(state_dict,
+                                                            strict=False)
+        else:
+            raise Exception("pretrained checkpoint type {} not supported".format(args.pretrained_checkpoint_type))
+
         # This is critical when only model is loaded. We should make sure
         # master parameters are also updated.
         optimizer.reload_model_params()
+
     timers("pretrained checkpoint").stop()
 
     # Print setup timing.
@@ -305,11 +300,13 @@ def finetune(
             train_dataloader,
             valid_dataloader,
             end_of_epoch_callback,
+            process_non_loss_data_func,
         )
     # Or just evaluate.
     else:
         if end_of_epoch_callback is not None:
             print_rank_0("evaluation only mode, setting epoch to -1")
-            end_of_epoch_callback(model, epoch=-1, output_predictions=True)
+            end_of_epoch_callback(model, epoch=-1)
 
     print_rank_0("done :-)")
+

tasks/vision/main.py

@@ -28,32 +28,24 @@ sys.path.append(
 )
 from megatron import get_args
 from megatron.initialize import initialize_megatron
-from classification import main
-
 
 def get_tasks_args(parser):
     """Provide extra arguments required for tasks."""
     group = parser.add_argument_group(title="tasks")
 
-    group.add_argument(
-        "--epochs",
-        type=int,
-        default=None,
-        help="Number of finetunning epochs. Zero results in "
-        "evaluation only.",
-    )
-    group.add_argument(
-        "--pretrained-checkpoint",
-        type=str,
-        default=None,
-        help="Pretrained checkpoint used for finetunning.",
-    )
-    group.add_argument(
-        "--keep-last",
-        action="store_true",
-        help="Keep the last batch (maybe incomplete) in" "the data loader",
-    )
-
+    group.add_argument('--task', type=str, default='segment',
+                       choices=['classify', 'segment_setr', 'segment_segformer'],
+                       help='task name.')
+    group.add_argument("--epochs", type=int, default=None,
+                       help="Number of finetunning epochs. Zero results in "
+                       "evaluation only.")
+    group.add_argument('--pretrained-checkpoint-type', type=str, default='default',
+                       choices=['default', 'external', 'constrastive'],
+                       help='Type of pretrained checkpoint')
+    group.add_argument("--pretrained-checkpoint", type=str, default=None,
+                       help="Pretrained checkpoint used for finetunning.")
+    group.add_argument('--seg-stride', type=int, default=None,
+                       help='sliding window stride during evaluation')
     return parser
 
 
@@ -61,4 +53,14 @@ if __name__ == "__main__":
 
     initialize_megatron(extra_args_provider=get_tasks_args)
     args = get_args()
-    main()
+
+    if args.task == 'classify':
+        from tasks.vision.classification.classification import main
+        main()
+    elif args.task == 'segment_setr':
+        from tasks.vision.segmentation.finetune_setr import main
+        main()
+    elif args.task == 'segment_segformer':
+        from tasks.vision.segmentation.finetune_segformer import main
+        main()
+

tasks/vision/segmentation/cityscapes.py

+# BSD 3-Clause License
+#
+# Copyright (c) Soumith Chintala 2016, 
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+#   list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+#   this list of conditions and the following disclaimer in the documentation
+#   and/or other materials provided with the distribution.
+#
+# * Neither the name of the copyright holder nor the names of its
+#   contributors may be used to endorse or promote products derived from
+#   this software without specific prior written permission.
+
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+# code taken from 
+# https://github.com/pytorch/vision/blob/main/torchvision/datasets/cityscapes.py
+# modified it to change max label index from 255 to 19 (num_classes)
+
+import torch
+import json
+import os
+from collections import namedtuple
+from typing import Any, Callable, Dict, List, Optional, Union, Tuple
+import numpy as np
+from torchvision.datasets.utils import extract_archive, verify_str_arg, iterable_to_str
+from torchvision.datasets import VisionDataset
+from PIL import Image
+from megatron import print_rank_0
+
+
+class Cityscapes(VisionDataset):
+    """`Cityscapes <http://www.cityscapes-dataset.com/>`_ Dataset.
+    Args:
+        root (string): Root directory of dataset where directory ``leftImg8bit``
+            and ``gtFine`` or ``gtCoarse`` are located.
+        split (string, optional): The image split to use, ``train``, ``test`` or ``val`` if mode="fine"
+            otherwise ``train``, ``train_extra`` or ``val``
+        mode (string, optional): The quality mode to use, ``fine`` or ``coarse``
+        target_type (string or list, optional): Type of target to use, ``instance``, ``semantic``, ``polygon``
+            or ``color``. Can also be a list to output a tuple with all specified target types.
+        transform (callable, optional): A function/transform that takes in a PIL image
+            and returns a transformed version. E.g, ``transforms.RandomCrop``
+        target_transform (callable, optional): A function/transform that takes in the
+            target and transforms it.
+        transforms (callable, optional): A function/transform that takes input sample and its target as entry
+            and returns a transformed version.
+    Examples:
+        Get semantic segmentation target
+        .. code-block:: python
+            dataset = Cityscapes('./data/cityscapes', split='train', mode='fine',
+                                 target_type='semantic')
+            img, smnt = dataset[0]
+        Get multiple targets
+        .. code-block:: python
+            dataset = Cityscapes('./data/cityscapes', split='train', mode='fine',
+                                 target_type=['instance', 'color', 'polygon'])
+            img, (inst, col, poly) = dataset[0]
+        Validate on the "coarse" set
+        .. code-block:: python
+            dataset = Cityscapes('./data/cityscapes', split='val', mode='coarse',
+                                 target_type='semantic')
+            img, smnt = dataset[0]
+    """
+    num_classes = 19
+    ignore_index = 19
+    color_table = torch.tensor(
+        [[128, 64, 128],
+         [244, 35, 232],
+         [70, 70, 70],
+         [102, 102, 156],
+         [190, 153, 153],
+         [153, 153, 153],
+         [250, 170, 30],
+         [220, 220, 0],
+         [107, 142, 35],
+         [152, 251, 152],
+         [70, 130, 180],
+         [220, 20, 60],
+         [255, 0, 0],
+         [0, 0, 142],
+         [0, 0, 70],
+         [0, 60, 100],
+         [0, 80, 100],
+         [0, 0, 230],
+         [119, 11, 32],
+         [0, 0, 0]], dtype=torch.float, device='cuda')
+
+
+    # Based on https://github.com/mcordts/cityscapesScripts
+    CityscapesClass = namedtuple('CityscapesClass', ['name', 'id', 'train_id', 
+        'category', 'category_id', 'has_instances', 'ignore_in_eval', 'color'])
+
+    classes = [
+        CityscapesClass('unlabeled', 0, 19, 'void', 0, False, True, (0, 0, 0)),
+        CityscapesClass('ego vehicle', 1, 19, 'void', 0, False, True, (0, 0, 0)),
+        CityscapesClass('rectification border', 2, 19, 'void', 0, False, True, (0, 0, 0)),
+        CityscapesClass('out of roi', 3, 19, 'void', 0, False, True, (0, 0, 0)),
+        CityscapesClass('static', 4, 19, 'void', 0, False, True, (0, 0, 0)),
+        CityscapesClass('dynamic', 5, 19, 'void', 0, False, True, (111, 74, 0)),
+        CityscapesClass('ground', 6, 19, 'void', 0, False, True, (81, 0, 81)),
+        CityscapesClass('road', 7, 0, 'flat', 1, False, False, (128, 64, 128)),
+        CityscapesClass('sidewalk', 8, 1, 'flat', 1, False, False, (244, 35, 232)),
+        CityscapesClass('parking', 9, 19, 'flat', 1, False, True, (250, 170, 160)),
+        CityscapesClass('rail track', 10, 19, 'flat', 1, False, True, (230, 150, 140)),
+        CityscapesClass('building', 11, 2, 'construction', 2, False, False, (70, 70, 70)),
+        CityscapesClass('wall', 12, 3, 'construction', 2, False, False, (102, 102, 156)),
+        CityscapesClass('fence', 13, 4, 'construction', 2, False, False, (190, 153, 153)),
+        CityscapesClass('guard rail', 14, 19, 'construction', 2, False, True, (180, 165, 180)),
+        CityscapesClass('bridge', 15, 19, 'construction', 2, False, True, (150, 100, 100)),
+        CityscapesClass('tunnel', 16, 19, 'construction', 2, False, True, (150, 120, 90)),
+        CityscapesClass('pole', 17, 5, 'object', 3, False, False, (153, 153, 153)),
+        CityscapesClass('polegroup', 18, 19, 'object', 3, False, True, (153, 153, 153)),
+        CityscapesClass('traffic light', 19, 6, 'object', 3, False, False, (250, 170, 30)),
+        CityscapesClass('traffic sign', 20, 7, 'object', 3, False, False, (220, 220, 0)),
+        CityscapesClass('vegetation', 21, 8, 'nature', 4, False, False, (107, 142, 35)),
+        CityscapesClass('terrain', 22, 9, 'nature', 4, False, False, (152, 251, 152)),
+        CityscapesClass('sky', 23, 10, 'sky', 5, False, False, (70, 130, 180)),
+        CityscapesClass('person', 24, 11, 'human', 6, True, False, (220, 20, 60)),
+        CityscapesClass('rider', 25, 12, 'human', 6, True, False, (255, 0, 0)),
+        CityscapesClass('car', 26, 13, 'vehicle', 7, True, False, (0, 0, 142)),
+        CityscapesClass('truck', 27, 14, 'vehicle', 7, True, False, (0, 0, 70)),
+        CityscapesClass('bus', 28, 15, 'vehicle', 7, True, False, (0, 60, 100)),
+        CityscapesClass('caravan', 29, 19, 'vehicle', 7, True, True, (0, 0, 90)),
+        CityscapesClass('trailer', 30, 19, 'vehicle', 7, True, True, (0, 0, 110)),
+        CityscapesClass('train', 31, 16, 'vehicle', 7, True, False, (0, 80, 100)),
+        CityscapesClass('motorcycle', 32, 17, 'vehicle', 7, True, False, (0, 0, 230)),
+        CityscapesClass('bicycle', 33, 18, 'vehicle', 7, True, False, (119, 11, 32)),
+        CityscapesClass('license plate', -1, -1, 'vehicle', 7, False, True, (0, 0, 142)),
+    ]
+
+    # label2trainid
+    label2trainid   = { label.id  : label.train_id for label in classes}
+
+    def __init__(
+            self,
+            root: str,
+            split: str = "train",
+            mode: str = "fine",
+            resolution: int = 1024,
+            transform: Optional[Callable] = None,
+            target_transform: Optional[Callable] = None,
+            transforms: Optional[Callable] = None,
+    ) -> None:
+        super(Cityscapes, self).__init__(root, transforms, transform, target_transform)
+        self.mode = 'gtFine' if mode == 'fine' else 'gtCoarse'
+        self.images_dir = os.path.join(self.root, 'leftImg8bit_trainvaltest/leftImg8bit', split)
+        self.targets_dir = os.path.join(self.root, 'gtFine_trainvaltest/gtFine', split)
+        self.split = split
+        self.resolution = resolution
+        self.images = []
+        self.targets = []
+
+        for city in sorted(os.listdir(self.images_dir)):
+            img_dir = os.path.join(self.images_dir, city)
+            target_dir = os.path.join(self.targets_dir, city)
+            for file_name in os.listdir(img_dir):
+                target_name = '{}_{}_labelIds.png'.format(file_name.split('_leftImg8bit')[0], self.mode)
+                self.images.append(os.path.join(img_dir, file_name))
+                self.targets.append(os.path.join(target_dir, target_name))
+
+
+    def __getitem__(self, index: int) -> Tuple[Any, Any]:
+        """
+        Args:
+            index (int): Index
+        Returns:
+            tuple: (image, target) where target is a tuple of all target types if target_type is a list with more
+            than one item. Otherwise target is a json object if target_type="polygon", else the image segmentation.
+        """
+        image = Image.open(self.images[index]).convert('RGB')
+        
+        target = Image.open(self.targets[index]) 
+        target = np.array(target)
+
+        target_copy = target.copy()
+        for k, v in Cityscapes.label2trainid.items():
+            binary_target = (target == k)
+            target_copy[binary_target] = v
+        target = target_copy
+
+        target = Image.fromarray(target.astype(np.uint8))
+
+        if self.transforms is not None:
+            image, target = self.transforms(image, target)
+
+        return image, target
+
+    def __len__(self) -> int:
+        # len(self.images)
+        return len(self.images)
+

tasks/vision/segmentation/data.py

+import random
+import os
+import math
+import mmcv
+import torch
+import numpy as np
+import torchvision.transforms as T
+from torchvision import datasets
+from torch.utils.data import Dataset
+from megatron.data.autoaugment import ImageNetPolicy
+from tasks.vision.segmentation.cityscapes import Cityscapes
+import tasks.vision.segmentation.transforms as ET
+from megatron.data.autoaugment import ImageNetPolicy
+from megatron import get_args
+from PIL import Image, ImageOps
+
+
+class VitSegmentationJointTransform():
+    def __init__(self, train=True, resolution=None):
+        self.train = train
+        if self.train:
+            self.transform0 = ET.RandomSizeAndCrop(resolution)
+            self.transform1 = ET.RandomHorizontallyFlip()
+
+    def __call__(self, img, mask):
+        if self.train:
+            img, mask = self.transform0(img, mask)
+            img, mask = self.transform1(img, mask)
+        return img, mask
+
+
+class VitSegmentationImageTransform():
+    def __init__(self, train=True, resolution=None):
+        args = get_args()
+        self.train = train
+        assert args.fp16 or args.bf16
+        self.data_type = torch.half if args.fp16 else torch.bfloat16
+        self.mean_std = args.mean_std
+        if self.train:
+            assert resolution is not None
+            self.transform = T.Compose([
+                ET.PhotoMetricDistortion(),
+                T.ToTensor(),
+                T.Normalize(*self.mean_std),
+                T.ConvertImageDtype(self.data_type)
+            ])
+        else:
+            self.transform = T.Compose([
+                T.ToTensor(),
+                T.Normalize(*self.mean_std),
+                T.ConvertImageDtype(self.data_type)
+            ])
+
+    def __call__(self, input):
+        output = self.transform(input)
+        return output
+
+
+class VitSegmentationTargetTransform():
+    def __init__(self, train=True, resolution=None):
+        self.train = train
+
+    def __call__(self, input):
+        output = torch.from_numpy(np.array(input, dtype=np.int32)).long()
+        return output
+
+
+class RandomSeedSegmentationDataset(Dataset):
+    def __init__(self,
+                 dataset,
+                 joint_transform,
+                 image_transform,
+                 target_transform):
+
+        args = get_args()
+        self.base_seed = args.seed
+        self.curr_seed = self.base_seed
+        self.dataset = dataset
+        self.joint_transform = joint_transform
+        self.image_transform = image_transform
+        self.target_transform = target_transform
+
+    def __len__(self):
+        return len(self.dataset)
+
+    def set_epoch(self, epoch):
+        self.curr_seed = self.base_seed + 100 * epoch
+
+    def __getitem__(self, idx):
+        seed = idx + self.curr_seed
+        img, mask = self.dataset[idx]
+
+        torch.manual_seed(seed)
+        random.seed(seed)
+        np.random.seed(seed)
+        img, mask = self.joint_transform(img, mask)
+        img = self.image_transform(img)
+        mask = self.target_transform(mask)
+
+        return img, mask
+
+
+def build_cityscapes_train_valid_datasets(data_path, image_size):
+    args = get_args()
+    args.num_classes = Cityscapes.num_classes
+    args.ignore_index = Cityscapes.ignore_index
+    args.color_table = Cityscapes.color_table
+    args.mean_std = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
+
+    train_joint_transform = \
+        VitSegmentationJointTransform(train=True, resolution=image_size)
+    val_joint_transform = \
+        VitSegmentationJointTransform(train=False, resolution=image_size)
+    train_image_transform = \
+        VitSegmentationImageTransform(train=True, resolution=image_size)
+    val_image_transform = \
+        VitSegmentationImageTransform(train=False, resolution=image_size)
+    train_target_transform = \
+        VitSegmentationTargetTransform(train=True, resolution=image_size)
+    val_target_transform = \
+        VitSegmentationTargetTransform(train=False, resolution=image_size)
+
+    # training dataset
+    train_data = Cityscapes(
+        root=data_path[0],
+        split='train',
+        mode='fine',
+        resolution=image_size
+    )
+    train_data = RandomSeedSegmentationDataset(
+        train_data,
+        joint_transform=train_joint_transform,
+        image_transform=train_image_transform,
+        target_transform=train_target_transform)
+
+    # validation dataset
+    val_data = Cityscapes(
+        root=data_path[0],
+        split='val',
+        mode='fine',
+        resolution=image_size
+    )
+
+    val_data = RandomSeedSegmentationDataset(
+        val_data,
+        joint_transform=val_joint_transform,
+        image_transform=val_image_transform,
+        target_transform=val_target_transform)
+
+    return train_data, val_data
+
+
+def build_train_valid_datasets(data_path, image_size):
+    return build_cityscapes_train_valid_datasets(data_path, image_size)

tasks/vision/segmentation/finetune_segformer.py

+# coding=utf-8
+# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Vision-classification finetuning/evaluation."""
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from functools import partial
+from megatron import get_args, get_timers
+from megatron import mpu, print_rank_0, print_rank_last
+from tasks.vision.finetune_utils import finetune
+from tasks.vision.finetune_utils import build_data_loader
+from megatron.utils import average_losses_across_data_parallel_group
+from megatron.schedules import get_forward_backward_func
+from tasks.vision.segmentation.data import build_train_valid_datasets
+from tasks.vision.segmentation.seg_models import SegformerSegmentationModel
+from megatron.model.vision.utils import resize
+
+
+def calculate_iou(hist_data):
+    acc = np.diag(hist_data).sum() / hist_data.sum()
+    acc_cls = np.diag(hist_data) / hist_data.sum(axis=1)
+    acc_cls = np.nanmean(acc_cls)
+    divisor = hist_data.sum(axis=1) + hist_data.sum(axis=0) - \
+        np.diag(hist_data)
+    iu = np.diag(hist_data) / divisor
+    return iu, acc, acc_cls
+
+
+def fast_hist(pred, gtruth, num_classes):
+    # mask indicates pixels we care about
+    mask = (gtruth >= 0) & (gtruth < num_classes)
+
+    # stretch ground truth labels by num_classes
+    #   class 0  -> 0
+    #   class 1  -> 19
+    #   class 18 -> 342
+    #
+    # TP at 0 + 0, 1 + 1, 2 + 2 ...
+    #
+    # TP exist where value == num_classes*class_id + class_id
+    # FP = row[class].sum() - TP
+    # FN = col[class].sum() - TP
+    hist = np.bincount(num_classes * gtruth[mask].astype(int) + pred[mask],
+                       minlength=num_classes ** 2)
+    hist = hist.reshape(num_classes, num_classes)
+    return hist
+
+
+def segmentation():
+
+    def train_valid_datasets_provider():
+        """Build train and validation dataset."""
+        args = get_args()
+
+        train_ds, valid_ds = build_train_valid_datasets(
+            data_path=args.data_path,
+            image_size=(args.img_h, args.img_w)
+
+        )
+        return train_ds, valid_ds
+
+    def model_provider(pre_process=True, post_process=True):
+        """Build the model."""
+        args = get_args()
+
+        model = SegformerSegmentationModel(num_classes=args.num_classes,
+                                           pre_process=pre_process,
+                                           post_process=post_process)
+        print_rank_0("model = {}".format(model))
+        return model
+
+    def process_batch(batch):
+        """Process batch and produce inputs for the model."""
+        images = batch[0].cuda().contiguous()
+        masks = batch[1].cuda().contiguous()
+        return images, masks
+
+    def calculate_weight(masks, num_classes):
+        bins = torch.histc(masks, bins=num_classes, min=0.0, max=num_classes)
+        hist_norm = bins.float()/bins.sum()
+        hist = ((bins != 0).float() * (1. - hist_norm)) + 1.0
+        return hist
+
+    def cross_entropy_loss_func(images, masks, output_tensor,
+                                non_loss_data=False):
+        args = get_args()
+        ignore_index = args.ignore_index
+        color_table = args.color_table
+        logits = output_tensor.contiguous().float()
+        logits = resize(logits, size=masks.shape[1:],
+                        mode='bilinear', align_corners=False)
+      
+        # Cross-entropy loss.
+        # weight = calculate_weight(masks, num_classes)
+        loss = F.cross_entropy(logits, masks, ignore_index=ignore_index)
+
+        if not non_loss_data:
+            # Reduce loss for logging.
+            averaged_loss = average_losses_across_data_parallel_group([loss])
+            return loss, {'lm loss': averaged_loss[0]}
+        else:
+            seg_mask = logits.argmax(dim=1)
+            output_mask = F.embedding(seg_mask, color_table).permute(0, 3, 1, 2)
+            gt_mask = F.embedding(masks, color_table).permute(0, 3, 1, 2)
+            return torch.cat((images, output_mask, gt_mask), dim=2), loss
+
+    def _cross_entropy_forward_step(batch, model):
+        """Simple forward step with cross-entropy loss."""
+        timers = get_timers()
+
+        # Get the batch.
+        timers("batch generator").start()
+        import types
+        if isinstance(batch, types.GeneratorType):
+            batch_ = next(batch)
+        else:
+            batch_ = batch
+        images, masks = process_batch(batch_)
+        timers("batch generator").stop()
+
+        # Forward model.
+        output_tensor = model(images)
+
+        return output_tensor, partial(cross_entropy_loss_func, images, masks)
+
+    def calculate_correct_answers(model, dataloader, epoch):
+        """Calculate correct over total answers"""
+
+        forward_backward_func = get_forward_backward_func()
+        for m in model:
+            m.eval()
+
+        def loss_func(labels, output_tensor):
+            args = get_args()
+            logits = output_tensor
+            logits = resize(logits, size=labels.shape[1:],
+                            mode='bilinear', align_corners=False)
+
+            loss_dict = {}
+            # Compute the correct answers.
+            probs = logits.contiguous().float().softmax(dim=1)
+            max_probs, preds = torch.max(probs, 1)
+
+            preds = preds.cpu().numpy()
+            performs = fast_hist(preds.flatten(),
+                                 labels.cpu().numpy().flatten(),
+                                 args.ignore_index)
+            loss_dict['performs'] = performs
+            return 0, loss_dict
+
+        # defined inside to capture output_predictions
+        def correct_answers_forward_step(batch, model):
+            try:
+                batch_ = next(batch)
+            except BaseException:
+                batch_ = batch
+            images, labels = process_batch(batch_)
+
+            # Forward model.
+            output_tensor = model(images)
+
+            return output_tensor, partial(loss_func, labels)
+
+        with torch.no_grad():
+            # For all the batches in the dataset.
+            performs = None
+            for _, batch in enumerate(dataloader):
+                loss_dicts = forward_backward_func(correct_answers_forward_step,
+                                                   batch, model,
+                                                   optimizer=None,
+                                                   timers=None,
+                                                   forward_only=True)
+                for loss_dict in loss_dicts:
+                    if performs is None:
+                        performs = loss_dict['performs']
+                    else:
+                        performs += loss_dict['performs']
+
+        for m in model:
+            m.train()
+        # Reduce.
+        if mpu.is_pipeline_last_stage():
+            performs_tensor = torch.cuda.FloatTensor(performs)
+            torch.distributed.all_reduce(performs_tensor,
+                                         group=mpu.get_data_parallel_group())
+            hist = performs_tensor.cpu().numpy()
+            iu, acc, acc_cls = calculate_iou(hist)
+            miou = np.nanmean(iu)
+
+            return iu, miou
+
+    def accuracy_func_provider():
+        """Provide function that calculates accuracies."""
+        args = get_args()
+
+        train_ds, valid_ds = build_train_valid_datasets(
+            data_path=args.data_path,
+            image_size=(args.img_h, args.img_w)
+        )
+        dataloader = build_data_loader(
+            valid_ds,
+            args.micro_batch_size,
+            num_workers=args.num_workers,
+            drop_last=(mpu.get_data_parallel_world_size() > 1),
+            shuffle=False
+        )
+
+        def metrics_func(model, epoch):
+            print_rank_0("calculating metrics ...")
+            iou, miou = calculate_correct_answers(model, dataloader, epoch)
+            print_rank_last(
+                " >> |epoch: {}| overall: iou = {},"
+                "miou = {:.4f} %".format(epoch, iou, miou*100.0)
+            )
+        return metrics_func
+
+    def dump_output_data(data, iteration, writer):
+        for (output_tb, loss) in data:
+            # output_tb[output_tb < 0] = 0
+            # output_tb[output_tb > 1] = 1
+            writer.add_images("image-outputseg-realseg", output_tb,
+                              global_step=None, walltime=None,
+                              dataformats='NCHW')
+
+    """Finetune/evaluate."""
+    finetune(
+        train_valid_datasets_provider,
+        model_provider,
+        forward_step=_cross_entropy_forward_step,
+        process_non_loss_data_func=dump_output_data,
+        end_of_epoch_callback_provider=accuracy_func_provider,
+    )
+
+
+def main():
+    segmentation()
+

tasks/vision/segmentation/finetune_setr.py

+# coding=utf-8
+# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Vision-classification finetuning/evaluation."""
+
+import torch
+import torch.nn.functional as F
+from functools import partial
+from megatron import get_args, get_timers
+from megatron import mpu, print_rank_0, print_rank_last
+from tasks.vision.finetune_utils import finetune
+from tasks.vision.finetune_utils import build_data_loader
+from megatron.utils import average_losses_across_data_parallel_group
+from megatron.schedules import get_forward_backward_func
+from tasks.vision.segmentation.metrics import CFMatrix
+from tasks.vision.segmentation.data import build_train_valid_datasets
+from tasks.vision.segmentation.seg_models import SetrSegmentationModel
+from tasks.vision.segmentation.utils import slidingcrops, slidingjoins
+
+def segmentation():
+    def train_valid_datasets_provider():
+        """Build train and validation dataset."""
+        args = get_args()
+
+        train_ds, valid_ds = build_train_valid_datasets(
+            data_path=args.data_path,
+            image_size=(args.img_h, args.img_w)
+
+        )
+        return train_ds, valid_ds
+
+    def model_provider(pre_process=True, post_process=True):
+        """Build the model."""
+        args = get_args()
+
+        return SetrSegmentationModel(num_classes=args.num_classes,
+                                     pre_process=pre_process,
+                                     post_process=post_process)
+
+    def process_batch(batch):
+        """Process batch and produce inputs for the model."""
+        images = batch[0].cuda().contiguous()
+        masks = batch[1].cuda().contiguous()
+        return images, masks
+
+    def calculate_weight(masks, num_classes):
+        bins = torch.histc(masks, bins=num_classes, min=0.0, max=num_classes)
+        hist_norm = bins.float()/bins.sum()
+        hist = ((bins != 0).float() * (1. - hist_norm)) + 1.0
+        return hist
+
+    def cross_entropy_loss_func(images, masks, output_tensor, non_loss_data=False):
+        args = get_args()
+        ignore_index = args.ignore_index
+        color_table = args.color_table
+        weight = calculate_weight(masks, args.num_classes)
+        logits = output_tensor.contiguous().float()
+        loss = F.cross_entropy(logits, masks, weight=weight, ignore_index=ignore_index)
+
+        if not non_loss_data:
+            # Reduce loss for logging.
+            averaged_loss = average_losses_across_data_parallel_group([loss])
+
+            return loss, {'lm loss': averaged_loss[0]}
+        else:
+            seg_mask = logits.argmax(dim=1)
+            output_mask = F.embedding(seg_mask, color_table).permute(0, 3, 1, 2)
+            gt_mask = F.embedding(masks, color_table).permute(0, 3, 1, 2)
+            return torch.cat((images, output_mask, gt_mask), dim=2), loss
+
+    def _cross_entropy_forward_step(batch, model):
+        """Simple forward step with cross-entropy loss."""
+        args = get_args()
+        timers = get_timers()
+
+        # Get the batch.
+        timers("batch generator").start()
+        import types
+        if isinstance(batch, types.GeneratorType):
+            batch_ = next(batch)
+        else:
+            batch_ = batch
+        images, masks = process_batch(batch_)
+        timers("batch generator").stop()
+
+        # Forward model.
+        if not model.training:
+            images, masks, _, _ = slidingcrops(images, masks)
+        #print_rank_0("images size = {}".format(images.size()))
+       
+        if not model.training:
+            output_tensor = torch.cat([model(image) for image in torch.split(images, args.micro_batch_size)])
+        else:
+            output_tensor = model(images)
+
+        return output_tensor, partial(cross_entropy_loss_func, images, masks)
+
+    def calculate_correct_answers(model, dataloader, epoch):
+        """Calculate correct over total answers"""
+
+        forward_backward_func = get_forward_backward_func()
+        for m in model:
+            m.eval()
+
+        def loss_func(labels, slices_info, img_size, output_tensor):
+            args = get_args()
+            logits = output_tensor
+
+            loss_dict = {}
+            # Compute the correct answers.
+            probs = logits.contiguous().float().softmax(dim=1)
+            max_probs, preds = torch.max(probs, 1)
+            preds = preds.int()
+            preds, labels = slidingjoins(preds, max_probs, labels, slices_info, img_size)
+            _, performs = CFMatrix()(preds, labels, args.ignore_index)
+
+            loss_dict['performs'] = performs
+            return 0, loss_dict
+
+        # defined inside to capture output_predictions
+        def correct_answers_forward_step(batch, model):
+            args = get_args()
+            try:
+                batch_ = next(batch)
+            except BaseException:
+                batch_ = batch
+            images, labels = process_batch(batch_)
+
+            assert not model.training
+            images, labels, slices_info, img_size = slidingcrops(images, labels)
+            # Forward model.
+            output_tensor = torch.cat([model(image) for image in torch.split(images, args.micro_batch_size)])
+
+            return output_tensor, partial(loss_func, labels, slices_info, img_size)
+
+        with torch.no_grad():
+            # For all the batches in the dataset.
+            performs = None
+            for _, batch in enumerate(dataloader):
+                loss_dicts = forward_backward_func(correct_answers_forward_step,
+                                                   batch, model,
+                                                   optimizer=None,
+                                                   timers=None,
+                                                   forward_only=True)
+                for loss_dict in loss_dicts:
+                    if performs is None:
+                        performs = loss_dict['performs']
+                    else:
+                        performs += loss_dict['performs']
+
+        for m in model:
+            m.train()
+        # Reduce.
+        if mpu.is_pipeline_last_stage():
+            torch.distributed.all_reduce(performs,
+                                         group=mpu.get_data_parallel_group())
+            # Print on screen.
+            # performs[int(ch), :] = [nb_tp, nb_fp, nb_tn, nb_fn]
+            true_positive = performs[:, 0]
+            false_positive = performs[:, 1]
+            false_negative = performs[:, 3]
+
+            iou = true_positive / (true_positive + false_positive + false_negative)
+            miou = iou[~torch.isnan(iou)].mean()
+
+            return iou.tolist(), miou.item()
+
+    def accuracy_func_provider():
+        """Provide function that calculates accuracies."""
+        args = get_args()
+
+        train_ds, valid_ds = build_train_valid_datasets(
+            data_path=args.data_path,
+            image_size=(args.img_h, args.img_w)
+        )
+        dataloader = build_data_loader(
+            valid_ds,
+            args.micro_batch_size,
+            num_workers=args.num_workers,
+            drop_last=(mpu.get_data_parallel_world_size() > 1),
+            shuffle=False
+        )
+
+        def metrics_func(model, epoch):
+            print_rank_0("calculating metrics ...")
+            iou, miou = calculate_correct_answers(model, dataloader, epoch)
+            print_rank_last(
+                " >> |epoch: {}| overall: iou = {},"
+                "miou = {:.4f} %".format(epoch, iou, miou*100.0)
+            )
+        return metrics_func
+
+    def dump_output_data(data, iteration, writer):
+        for (output_tb, loss) in data:
+            # output_tb[output_tb < 0] = 0
+            # output_tb[output_tb > 1] = 1
+            writer.add_images("image-outputseg-realseg", output_tb,
+                              global_step=None, walltime=None,
+                              dataformats='NCHW')
+
+    """Finetune/evaluate."""
+    finetune(
+        train_valid_datasets_provider,
+        model_provider,
+        forward_step=_cross_entropy_forward_step,
+        process_non_loss_data_func=dump_output_data,
+        end_of_epoch_callback_provider=accuracy_func_provider,
+    )
+
+
+def main():
+    segmentation()
+

tasks/vision/segmentation/seg_heads.py

+# coding=utf-8
+# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+import einops
+import torch
+import apex
+import torch.nn.functional as F
+from megatron import get_args
+from megatron.model import LayerNorm
+from megatron.model.module import MegatronModule
+from megatron.model.vision.utils import resize
+
+
+class SetrSegmentationHead(MegatronModule):
+    def __init__(self, hidden_size, num_classes):
+        super(SetrSegmentationHead, self).__init__()
+        args = get_args()
+        self.hidden_size = hidden_size
+        self.num_classes = num_classes
+        self.img_h = args.img_h
+        self.img_w = args.img_w
+        self.patch_dim = args.patch_dim
+
+        self.layernorm = LayerNorm(hidden_size, eps=args.layernorm_epsilon)
+        self.conv_0 = torch.nn.Conv2d(hidden_size, hidden_size,
+                                      1, 1, bias=False)
+        self.norm_0 = apex.parallel.SyncBatchNorm(hidden_size)
+        self.conv_1 = torch.nn.Conv2d(hidden_size, num_classes, 1, 1)
+
+    def to_2D(self, x):
+        n, hw, c = x.shape
+        h = self.img_h // self.patch_dim
+        w = self.img_w // self.patch_dim
+        assert(hw == h * w)
+        x = x.transpose(1, 2).reshape(n, c, h, w)
+        return x
+
+    def forward(self, hidden_states):
+        # [b c h w]
+        hidden_states = self.layernorm(hidden_states)
+        hidden_states = self.to_2D(hidden_states)
+
+        hidden_states = self.conv_0(hidden_states)
+        hidden_states = self.norm_0(hidden_states)
+        hidden_states = torch.tanh(hidden_states)
+        hidden_states = self.conv_1(hidden_states)
+
+        # [b c h w]
+        result = F.interpolate(hidden_states,
+                               size=(self.img_h, self.img_w),
+                               mode='bilinear')
+
+        return result
+
+
+class MLP(torch.nn.Module):
+    """
+    Linear Embedding
+    """
+    def __init__(self, input_dim=2048, embed_dim=768):
+        super().__init__()
+        self.proj = torch.nn.Linear(input_dim, embed_dim)
+
+    def forward(self, x):
+        x = x.flatten(2).transpose(1, 2)
+        x = self.proj(x)
+        return x
+
+
+class SegformerSegmentationHead(MegatronModule):
+    def __init__(self, feature_strides, in_channels,
+                 embedding_dim, dropout_ratio):
+        super(SegformerSegmentationHead, self).__init__()
+        assert len(feature_strides) == len(in_channels)
+        assert min(feature_strides) == feature_strides[0]
+        args = get_args()
+        self.feature_strides = feature_strides
+        self.in_channels = in_channels
+        self.embedding_dim = embedding_dim
+        self.num_classes = args.num_classes
+        self.dropout_ratio = dropout_ratio
+
+        c1_in_channels, c2_in_channels, c3_in_channels, c4_in_channels = \
+            self.in_channels
+
+        self.linear_c4 = MLP(input_dim=c4_in_channels,
+                             embed_dim=self.embedding_dim)
+        self.linear_c3 = MLP(input_dim=c3_in_channels,
+                             embed_dim=self.embedding_dim)
+        self.linear_c2 = MLP(input_dim=c2_in_channels,
+                             embed_dim=self.embedding_dim)
+        self.linear_c1 = MLP(input_dim=c1_in_channels,
+                             embed_dim=self.embedding_dim)
+
+        self.conv_fuse = torch.nn.Conv2d(self.embedding_dim*4,
+                                         self.embedding_dim, 1, 1)
+        self.norm = apex.parallel.SyncBatchNorm(self.embedding_dim)
+
+        self.dropout = torch.nn.Dropout2d(self.dropout_ratio)
+        self.linear_pred = torch.nn.Conv2d(self.embedding_dim,
+                                           self.num_classes,
+                                           kernel_size=1)
+
+    def forward(self, inputs):
+        c1, c2, c3, c4 = inputs
+
+        ############## MLP decoder on C1-C4 ###########
+        n, _, h, w = c4.shape
+
+        _c4 = self.linear_c4(c4).permute(0, 2, 1).reshape(n, -1, c4.shape[2], c4.shape[3])
+        _c4 = resize(_c4, size=c1.size()[2:], mode='bilinear', align_corners=False)
+
+        _c3 = self.linear_c3(c3).permute(0, 2, 1).reshape(n, -1, c3.shape[2], c3.shape[3])
+        _c3 = resize(_c3, size=c1.size()[2:], mode='bilinear', align_corners=False)
+
+        _c2 = self.linear_c2(c2).permute(0, 2, 1).reshape(n, -1, c2.shape[2], c2.shape[3])
+        _c2 = resize(_c2, size=c1.size()[2:], mode='bilinear', align_corners=False)
+
+        _c1 = self.linear_c1(c1).permute(0, 2, 1).reshape(n, -1, c1.shape[2], c1.shape[3])
+
+        _c = self.conv_fuse(torch.cat([_c4, _c3, _c2, _c1], dim=1))
+        x = self.norm(_c)
+        x = F.relu(x, inplace=True)
+        x = self.dropout(x)
+        x = self.linear_pred(x)
+
+        return x
+

tasks/vision/segmentation/seg_models.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+import einops
+import torch
+import apex
+import torch.nn.functional as F
+from megatron import get_args
+from megatron.model.module import MegatronModule
+from megatron.model.vision.vit_backbone import VitBackbone, VitMlpHead
+from megatron.model.vision.mit_backbone import mit_b3, mit_b5
+from tasks.vision.segmentation.seg_heads import SetrSegmentationHead, SegformerSegmentationHead
+
+
+class SetrSegmentationModel(MegatronModule):
+
+    def __init__(self,
+                 num_classes,
+                 pre_process=True,
+                 post_process=True):
+        super(SetrSegmentationModel, self).__init__()
+        args = get_args()
+        assert post_process & pre_process
+        self.hidden_size = args.hidden_size
+        self.num_classes = num_classes
+        self.backbone = VitBackbone(
+            pre_process=pre_process,
+            post_process=post_process,
+            class_token=False,
+            post_layer_norm=False,
+            drop_path_rate=0.1
+        )
+
+        self.head = SetrSegmentationHead(
+            self.hidden_size,
+            self.num_classes
+        )
+
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        pass
+
+    def forward(self, input):
+        # [b hw c]
+        hidden_states = self.backbone(input)
+        result_final = self.head(hidden_states)
+        return result_final
+
+
+class SegformerSegmentationModel(MegatronModule):
+
+    def __init__(self,
+                 num_classes,
+                 pre_process=True,
+                 post_process=True):
+        super(SegformerSegmentationModel, self).__init__()
+        args = get_args()
+        self.hidden_size = args.hidden_size
+        self.num_classes = num_classes
+        self.pre_process = pre_process
+        self.post_process = post_process
+
+        self.backbone = mit_b5()
+        self.head = SegformerSegmentationHead(
+            feature_strides=[4, 8, 16, 32],
+            in_channels=[64, 128, 320, 512],
+            embedding_dim=768,
+            dropout_ratio=0.1
+        )
+
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        pass
+
+    def forward(self, input):
+        # [b hw c]
+        hidden_states = self.backbone(input)
+        hidden_states = self.head(hidden_states)
+        return hidden_states
+

tasks/vision/segmentation/transforms.py

+# Copyright (c) 2020 The MMSegmenation Authors.
+#
+# This source code is licensed under the Apache license found in the
+# LICENSE file in the root directory of this source tree.
+
+import random
+import os
+import math
+import mmcv
+import torch
+import numpy as np
+import torchvision.transforms as T
+from torchvision import datasets
+from torch.utils.data import Dataset
+from megatron import print_rank_0
+from megatron import get_args
+from PIL import Image, ImageOps, ImageEnhance
+import torchvision.transforms as torch_tr
+
+def _is_pil_image(img):
+    return isinstance(img, Image.Image)
+
+class PhotoMetricDistortion(object):
+    """Apply photometric distortion to image sequentially, every transformation
+    is applied with a probability of 0.5. The position of random contrast is in
+    second or second to last.
+    1. random brightness
+    2. random contrast (mode 0)
+    3. convert color from BGR to HSV
+    4. random saturation
+    5. random hue
+    6. convert color from HSV to BGR
+    7. random contrast (mode 1)
+    8. randomly swap channels
+    Args:
+        brightness_delta (int): delta of brightness.
+        contrast_range (tuple): range of contrast.
+        saturation_range (tuple): range of saturation.
+        hue_delta (int): delta of hue.
+    """
+
+    def __init__(self,
+                 brightness_delta=32,
+                 contrast_range=(0.5, 1.5),
+                 saturation_range=(0.5, 1.5),
+                 hue_delta=18):
+        self.brightness_delta = brightness_delta
+        self.contrast_lower, self.contrast_upper = contrast_range
+        self.saturation_lower, self.saturation_upper = saturation_range
+        self.hue_delta = hue_delta
+
+    def convert(self, img, alpha=1, beta=0):
+        """Multiple with alpha and add beat with clip."""
+        img = img.astype(np.float32) * alpha + beta
+        img = np.clip(img, 0, 255)
+        return img.astype(np.uint8)
+
+    def brightness(self, img):
+        """Brightness distortion."""
+        if random.randint(0, 1):
+            return self.convert(
+                img,
+                beta=random.uniform(-self.brightness_delta,
+                                    self.brightness_delta))
+        return img
+
+    def contrast(self, img):
+        """Contrast distortion."""
+        if random.randint(0, 1):
+            return self.convert(
+                img,
+                alpha=random.uniform(self.contrast_lower, self.contrast_upper))
+        return img
+
+    def saturation(self, img):
+        """Saturation distortion."""
+        if random.randint(0, 1):
+            img = mmcv.bgr2hsv(img)
+            img[:, :, 1] = self.convert(
+                img[:, :, 1],
+                alpha=random.uniform(self.saturation_lower,
+                                     self.saturation_upper))
+            img = mmcv.hsv2bgr(img)
+        return img
+
+    def hue(self, img):
+        """Hue distortion."""
+        if random.randint(0, 1):
+            img = mmcv.bgr2hsv(img)
+            img[:, :,
+                0] = (img[:, :, 0].astype(int) +
+                      random.randint(-self.hue_delta, self.hue_delta)) % 180
+            img = mmcv.hsv2bgr(img)
+        return img
+
+    def __call__(self, img):
+        """Call function to perform photometric distortion on images.
+        Args:
+            results (dict): Result dict from loading pipeline.
+        Returns:
+            dict: Result dict with images distorted.
+        """
+        img = np.array(img)
+
+        # random brightness
+        img = self.brightness(img)
+
+        # mode == 0 --> do random contrast first
+        # mode == 1 --> do random contrast last
+        mode = random.randint(0, 1)
+        if mode == 1:
+            img = self.contrast(img)
+
+        # random saturation
+        img = self.saturation(img)
+
+        # random hue
+        img = self.hue(img)
+
+        # random contrast
+        if mode == 0:
+            img = self.contrast(img)
+
+        img = Image.fromarray(img.astype(np.uint8)).convert('RGB')
+        return img
+
+
+class RandomCrop(object):
+    """
+    Take a random crop from the image.
+
+    First the image or crop size may need to be adjusted if the incoming image
+    is too small...
+
+    If the image is smaller than the crop, then:
+         the image is padded up to the size of the crop
+         unless 'nopad', in which case the crop size is shrunk to fit the image
+
+    A random crop is taken such that the crop fits within the image.
+
+
+    if cfg.DATASET.TRANSLATION_AUG_FIX is set, we insure that there's always
+    translation randomness of at least that value around the image.
+
+    if image < crop_size:
+        # slide crop within image, random offset
+    else:
+        # slide image within crop
+    """
+    def __init__(self, crop_size):
+        args = get_args()
+        self.size = crop_size
+        self.cat_max_ratio = 0.75
+        self.ignore_index = args.ignore_index
+        self.pad_color = (0, 0, 0)
+
+    def get_crop_bbox(self, img):
+        """Randomly get a crop bounding box."""
+        img_w, img_h = img.size
+        target_h, target_w = self.size  #[H W]
+        margin_h = max(img_h - target_h, 0)
+        margin_w = max(img_w - target_w, 0)
+        offset_h = random.randint(0, margin_h)
+        offset_w = random.randint(0, margin_w)
+        crop_y1, crop_y2 = offset_h, offset_h + target_h
+        crop_x1, crop_x2 = offset_w, offset_w + target_w
+
+        return crop_y1, crop_y2, crop_x1, crop_x2
+
+    def crop(self, img, crop_bbox):
+        """Crop from ``img``"""
+        crop_y1, crop_y2, crop_x1, crop_x2 = crop_bbox
+        img = img.crop((crop_x1, crop_y1, crop_x2, crop_y2))
+        return img
+
+    @staticmethod
+    def crop_in_image(target_w, target_h, w, h, img, mask):
+        if w == target_w:
+            x1 = 0
+        else:
+            x1 = random.randint(0, w - target_w)
+        if h == target_h:
+            y1 = 0
+        else:
+            y1 = random.randint(0, h - target_h)
+
+        return [img.crop((x1, y1, x1 + target_w, y1 + target_h)),
+                mask.crop((x1, y1, x1 + target_w, y1 + target_h))]
+
+
+    def __call__(self, img, mask):
+        w, h = img.size
+        target_h, target_w = self.size   # ASSUME H, W
+
+        if w == target_w and h == target_h:
+            return img, mask
+
+        # Pad image if image < crop
+        if target_h > h:
+            pad_h = (target_h - h) // 2 + 1
+        else:
+            pad_h = 0
+        if target_w > w:
+            pad_w = (target_w - w) // 2 + 1
+        else:
+            pad_w = 0
+        border = (pad_w, pad_h, pad_w, pad_h)
+        if pad_h or pad_w:
+            img = ImageOps.expand(img, border=border, fill=(0, 0, 0))
+            mask = ImageOps.expand(mask, border=border, fill=self.ignore_index)
+            w, h = img.size
+
+        crop_bbox = self.get_crop_bbox(img)
+        if self.cat_max_ratio < 1.:
+            # Repeat 10 times
+            for _ in range(10):
+                seg_temp = self.crop(mask, crop_bbox)
+                labels, cnt = np.unique(seg_temp, return_counts=True)
+                cnt = cnt[labels != self.ignore_index]
+                if len(cnt) > 1 and np.max(cnt) / np.sum(
+                        cnt) < self.cat_max_ratio:
+                    break
+                crop_bbox = self.get_crop_bbox(img)
+
+        # crop the image
+        img = self.crop(img, crop_bbox)
+
+        # crop semantic seg
+        mask = self.crop(mask, crop_bbox)
+        assert(img.size[0] == self.size[1] and img.size[1] == self.size[0])
+          
+        return img, mask
+
+
+class RandomSizeAndCrop(object):
+    def __init__(self,
+                 crop_size,
+                 scale_min=0.5,
+                 scale_max=2.0):
+        self.crop = RandomCrop(crop_size)
+        self.scale_min = scale_min
+        self.scale_max = scale_max
+
+    def __call__(self, img, mask):
+
+        scale_amt = random.uniform(self.scale_min, self.scale_max)
+        w, h = [int(i * scale_amt) for i in img.size]
+
+        resized_img = img.resize((w, h), Image.BICUBIC)
+        resized_mask = mask.resize((w, h), Image.NEAREST)
+        img, mask = self.crop(resized_img, resized_mask)
+        return img, mask
+
+class RandomHorizontallyFlip(object):
+    def __call__(self, img, mask):
+        if random.random() < 0.5:
+            return img.transpose(Image.FLIP_LEFT_RIGHT), mask.transpose(
+                Image.FLIP_LEFT_RIGHT)
+        return img, mask
+
+
+def adjust_brightness(img, brightness_factor):
+    """Adjust brightness of an Image.
+
+    Args:
+        img (PIL Image): PIL Image to be adjusted.
+        brightness_factor (float):  How much to adjust the brightness. Can be
+            any non negative number. 0 gives a black image, 1 gives the
+            original image while 2 increases the brightness by a factor of 2.
+
+    Returns:
+        PIL Image: Brightness adjusted image.
+    """
+    if not _is_pil_image(img):
+        raise TypeError('img should be PIL Image. Got {}'.format(type(img)))
+
+    enhancer = ImageEnhance.Brightness(img)
+    img = enhancer.enhance(brightness_factor)
+    return img
+
+
+def adjust_contrast(img, contrast_factor):
+    """Adjust contrast of an Image.
+
+    Args:
+        img (PIL Image): PIL Image to be adjusted.
+        contrast_factor (float): How much to adjust the contrast. Can be any
+            non negative number. 0 gives a solid gray image, 1 gives the
+            original image while 2 increases the contrast by a factor of 2.
+
+    Returns:
+        PIL Image: Contrast adjusted image.
+    """
+    if not _is_pil_image(img):
+        raise TypeError('img should be PIL Image. Got {}'.format(type(img)))
+
+    enhancer = ImageEnhance.Contrast(img)
+    img = enhancer.enhance(contrast_factor)
+    return img
+
+
+def adjust_saturation(img, saturation_factor):
+    """Adjust color saturation of an image.
+
+    Args:
+        img (PIL Image): PIL Image to be adjusted.
+        saturation_factor (float):  How much to adjust the saturation. 0 will
+            give a black and white image, 1 will give the original image while
+            2 will enhance the saturation by a factor of 2.
+
+    Returns:
+        PIL Image: Saturation adjusted image.
+    """
+    if not _is_pil_image(img):
+        raise TypeError('img should be PIL Image. Got {}'.format(type(img)))
+
+    enhancer = ImageEnhance.Color(img)
+    img = enhancer.enhance(saturation_factor)
+    return img
+
+
+def adjust_hue(img, hue_factor):
+    """Adjust hue of an image.
+
+    The image hue is adjusted by converting the image to HSV and
+    cyclically shifting the intensities in the hue channel (H).
+    The image is then converted back to original image mode.
+
+    `hue_factor` is the amount of shift in H channel and must be in the
+    interval `[-0.5, 0.5]`.
+
+    See https://en.wikipedia.org/wiki/Hue for more details on Hue.
+
+    Args:
+        img (PIL Image): PIL Image to be adjusted.
+        hue_factor (float):  How much to shift the hue channel. Should be in
+            [-0.5, 0.5]. 0.5 and -0.5 give complete reversal of hue channel in
+            HSV space in positive and negative direction respectively.
+            0 means no shift. Therefore, both -0.5 and 0.5 will give an image
+            with complementary colors while 0 gives the original image.
+
+    Returns:
+        PIL Image: Hue adjusted image.
+    """
+    if not(-0.5 <= hue_factor <= 0.5):
+        raise ValueError('hue_factor is not in [-0.5, 0.5].'.format(hue_factor))
+
+    if not _is_pil_image(img):
+        raise TypeError('img should be PIL Image. Got {}'.format(type(img)))
+
+    input_mode = img.mode
+    if input_mode in {'L', '1', 'I', 'F'}:
+        return img
+
+    h, s, v = img.convert('HSV').split()
+
+    np_h = np.array(h, dtype=np.uint8)
+    # uint8 addition take cares of rotation across boundaries
+    with np.errstate(over='ignore'):
+        np_h += np.uint8(hue_factor * 255)
+    h = Image.fromarray(np_h, 'L')
+
+    img = Image.merge('HSV', (h, s, v)).convert(input_mode)
+    return img
+
+
+class ColorJitter(object):
+    """Randomly change the brightness, contrast and saturation of an image.
+
+    Args:
+        brightness (float): How much to jitter brightness. brightness_factor
+            is chosen uniformly from [max(0, 1 - brightness), 1 + brightness].
+        contrast (float): How much to jitter contrast. contrast_factor
+            is chosen uniformly from [max(0, 1 - contrast), 1 + contrast].
+        saturation (float): How much to jitter saturation. saturation_factor
+            is chosen uniformly from [max(0, 1 - saturation), 1 + saturation].
+        hue(float): How much to jitter hue. hue_factor is chosen uniformly from
+            [-hue, hue]. Should be >=0 and <= 0.5.
+    """
+    def __init__(self, brightness=0, contrast=0, saturation=0, hue=0):
+        self.brightness = brightness
+        self.contrast = contrast
+        self.saturation = saturation
+        self.hue = hue
+
+    @staticmethod
+    def get_params(brightness, contrast, saturation, hue):
+        """Get a randomized transform to be applied on image.
+
+        Arguments are same as that of __init__.
+
+        Returns:
+            Transform which randomly adjusts brightness, contrast and
+            saturation in a random order.
+        """
+        transforms = []
+        if brightness > 0:
+            brightness_factor = np.random.uniform(max(0, 1 - brightness), 1 + brightness)
+            transforms.append(
+                torch_tr.Lambda(lambda img: adjust_brightness(img, brightness_factor)))
+
+        if contrast > 0:
+            contrast_factor = np.random.uniform(max(0, 1 - contrast), 1 + contrast)
+            transforms.append(
+                torch_tr.Lambda(lambda img: adjust_contrast(img, contrast_factor)))
+
+        if saturation > 0:
+            saturation_factor = np.random.uniform(max(0, 1 - saturation), 1 + saturation)
+            transforms.append(
+                torch_tr.Lambda(lambda img: adjust_saturation(img, saturation_factor)))
+
+        if hue > 0:
+            hue_factor = np.random.uniform(-hue, hue)
+            transforms.append(
+                torch_tr.Lambda(lambda img: adjust_hue(img, hue_factor)))
+
+        np.random.shuffle(transforms)
+        transform = torch_tr.Compose(transforms)
+
+        return transform
+
+    def __call__(self, img):
+        """
+        Args:
+            img (PIL Image): Input image.
+
+        Returns:
+            PIL Image: Color jittered image.
+        """
+        transform = self.get_params(self.brightness, self.contrast,
+                                    self.saturation, self.hue)
+        return transform(img)
+

tasks/vision/segmentation/utils.py

+import math
+import torch
+import numpy as np
+from megatron import get_args
+
+def slidingcrops(img, mask):
+    # img: [b c h w]
+    # mask: [b h w]
+    args = get_args()
+    assert args.img_h == args.img_w
+    crop_size = args.img_h
+    stride = args.seg_stride
+    ignore_index = args.ignore_index
+    n, c, h, w = img.shape
+    assert h >= crop_size
+    assert w >= crop_size
+    long_size = max(h, w)
+
+    img_slices, mask_slices, slices_info = [], [], []
+    if long_size > crop_size:
+        assert stride <= crop_size
+        h_step_num = int(math.ceil((h - crop_size) / float(stride))) + 1
+        w_step_num = int(math.ceil((w - crop_size) / float(stride))) + 1
+        for yy in range(h_step_num):
+            for xx in range(w_step_num):
+                sy, sx = yy * stride, xx * stride
+                ey, ex = sy + crop_size, sx + crop_size
+                img_sub = img[:, :, sy: ey, sx: ex]
+                mask_sub = mask[:, sy: ey, sx: ex]
+
+                # padding
+                sub_h, sub_w = img_sub.shape[2:]
+                pad_h = max(crop_size - sub_h, 0)
+                pad_w = max(crop_size - sub_w, 0)
+                img_sub = torch.nn.functional.pad(img_sub, pad=(0, pad_w, 0, pad_h), value=ignore_index)
+                mask_sub = torch.nn.functional.pad(mask_sub, pad=(0, pad_w, 0, pad_h))
+
+                img_slices.append(img_sub)
+                mask_slices.append(mask_sub)
+                slices_info.append([sy, ey, sx, ex, sub_h, sub_w])
+
+        return torch.cat(img_slices), torch.cat(mask_slices), slices_info, (h, w)
+    else:
+        return img, mask, [[0, h, 0, w, h, w]], (h, w)
+
+
+def slidingjoins(preds, probs, labels, slices_info, img_size):
+    args = get_args()
+    num_slices = len(slices_info)
+
+    if num_slices == 1:
+        return preds, labels
+
+    h, w = img_size
+    split_size = args.micro_batch_size
+
+    preds_split = torch.split(preds, split_size)
+    probs_split = torch.split(probs, split_size)
+    labels_split = torch.split(labels, split_size)
+
+    assert(len(preds_split) == num_slices)
+
+    total_max_probs = torch.zeros((split_size, h, w), dtype=torch.float, device='cuda')
+    total_preds = torch.zeros((split_size, h, w), dtype=torch.int, device='cuda')
+    total_labels = torch.zeros((split_size, h, w), dtype=torch.int, device='cuda')
+
+    for i in range(num_slices):
+        sy, ey, sx, ex, sub_h, sub_w = slices_info[i]
+        assert sy + sub_h <= h
+        assert sx + sub_w <= w
+        curr_max_probs = total_max_probs[:, sy:sy + sub_h, sx:sx + sub_w]
+        curr_preds = total_preds[:, sy:sy + sub_h, sx:sx + sub_w]
+
+        local_max_probs = probs_split[i][:, :sub_h, : sub_w]
+        local_preds = preds_split[i][:, :sub_h, :sub_w]
+
+        result_max_probs = torch.maximum(curr_max_probs, local_max_probs)
+        result_preds = torch.where(curr_max_probs >= local_max_probs, curr_preds, local_preds)
+
+        total_max_probs[:, sy:sy + sub_h, sx:sx + sub_w] = result_max_probs
+        total_preds[:, sy:sy + sub_h, sx:sx + sub_w] = result_preds
+        total_labels[:, sy:sy + sub_h, sx:sx + sub_w] = labels_split[i][0, :sub_h, :sub_w]
+
+    return total_preds, total_labels
+

tools/merge_datasets.py

+import os
+import sys
+import json
+import argparse
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
+                                             os.path.pardir)))
+
+from megatron.data import indexed_dataset
+
+
+def main(args):
+
+    prefixes = set()
+    for basename in os.listdir(args.input):
+        prefix, ext = os.path.splitext(basename)
+
+        if prefix in prefixes:
+            continue
+
+        if not os.path.isfile(os.path.join(args.input, basename)):
+            continue
+
+        ext_pair = '.bin' if ext == '.idx' else '.idx'
+        assert os.path.isfile(os.path.join(args.input, prefix) + ext_pair), \
+               f'ERROR: {ext_pair} file not provided for {os.path.join(args.input, prefix)}'
+
+        prefixes.add(prefix)
+
+    builder = None
+    for prefix in sorted(prefixes):
+        if builder is None:
+            dataset = indexed_dataset.make_dataset(os.path.join(args.input, prefix), 'infer')
+
+            if isinstance(dataset, indexed_dataset.MMapIndexedDataset):
+                builder = indexed_dataset.MMapIndexedDatasetBuilder(args.output_prefix + '.bin', dtype=dataset._index.dtype)
+            else:
+                builder = indexed_dataset.IndexedDatasetBuilder(args.output_prefix + '.bin')
+
+            del dataset
+
+        builder.merge_file_(os.path.join(args.input, prefix))
+
+    builder.finalize(args.output_prefix + '.idx')
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+
+    group = parser.add_argument_group(title='input data')
+    group.add_argument('--input', type=str, required=True,
+                       help='Path to directory containing all document files to merge')
+
+    group = parser.add_argument_group(title='output data')
+    group.add_argument('--output-prefix', type=str, required=True,
+                       help='Path to binary output file without suffix')
+
+    args = parser.parse_args()
+
+    assert os.path.isdir(args.input), \
+           f'ERROR: {args.input} is not a directory or does not exist'
+
+    assert os.path.isdir(os.path.dirname(args.output_prefix)), \
+           f'ERROR: {os.path.dirname(args.output_prefix)} is not a directory or does not exist'
+
+    main(args)
+

tools/preprocess_data.py

@@ -122,8 +122,10 @@ def get_args():
                        choices=['lazy', 'cached', 'mmap'])
 
     group = parser.add_argument_group(title='runtime')
-    group.add_argument('--workers', type=int, default=1,
+    group.add_argument('--workers', type=int, required=True,
                        help='Number of worker processes to launch')
+    group.add_argument('--chunk-size', type=int, required=True,
+                       help='Chunk size assigned to each worker process')
     group.add_argument('--log-interval', type=int, default=100,
                        help='Interval between progress updates')
     args = parser.parse_args()
@@ -154,7 +156,7 @@ def main():
     encoder = Encoder(args)
     tokenizer = build_tokenizer(args)
     pool = multiprocessing.Pool(args.workers, initializer=encoder.initializer)
-    encoded_docs = pool.imap(encoder.encode, fin, 25)
+    encoded_docs = pool.imap(encoder.encode, fin, args.chunk_size)
     #encoded_docs = map(encoder.encode, fin)
 
     level = "document"