[dependency] removed torchvision (#833)

* [dependency] removed torchvision

* fixed transforms

colossalai/registry/__init__.py

 import torch.distributed.optim as dist_optim
 import torch.nn as nn
 import torch.optim as optim
-import torchvision.models as tv_models
-import torchvision.datasets as tv_datasets
-from torchvision import transforms
 
 from .registry import Registry
 
 LAYERS = Registry("layers", third_party_library=[nn])
 LOSSES = Registry("losses")
-MODELS = Registry("models", third_party_library=[tv_models])
+MODELS = Registry("models")
 OPTIMIZERS = Registry("optimizers", third_party_library=[optim, dist_optim])
-DATASETS = Registry("datasets", third_party_library=[tv_datasets])
+DATASETS = Registry("datasets")
 DIST_GROUP_INITIALIZER = Registry("dist_group_initializer")
 GRADIENT_HANDLER = Registry("gradient_handler")
 LOSSES = Registry("losses", third_party_library=[nn])
 HOOKS = Registry("hooks")
-TRANSFORMS = Registry("transforms", third_party_library=[transforms])
+TRANSFORMS = Registry("transforms")
 DATA_SAMPLERS = Registry("data_samplers")
 LR_SCHEDULERS = Registry("lr_schedulers")
 SCHEDULE = Registry("schedules")

requirements/requirements-test.txt

 pytest
-rpyc
-matplotlib
-tensorboard
+torchvision
 transformers

requirements/requirements.txt

 torch>=1.8
-torchvision>=0.9
 numpy
 tqdm
 psutil
-tensorboard
 packaging
 pre-commit
 rich

tests/test_data/test_cifar10_dataset.py

@@ -10,23 +10,10 @@ from torch.utils.data import DataLoader
 
 from colossalai.builder import build_dataset, build_transform
 from colossalai.context import Config
+from torchvision.transforms import ToTensor
 
-TRAIN_DATA = dict(
-    dataset=dict(
-        type='CIFAR10',
-        root=Path(os.environ['DATA']),
-        train=True,
-        download=True
-    ),
-    dataloader=dict(batch_size=4, shuffle=True, num_workers=2),
-    transform_pipeline=[
-        dict(type='ToTensor'),
-        dict(type='Normalize',
-             mean=(0.5, 0.5, 0.5),
-             std=(0.5, 0.5, 0.5)
-             )
-    ]
-)
+TRAIN_DATA = dict(dataset=dict(type='CIFAR10', root=Path(os.environ['DATA']), train=True, download=True),
+                  dataloader=dict(batch_size=4, shuffle=True, num_workers=2))
 
 
 @pytest.mark.cpu
@@ -37,7 +24,7 @@ def test_cifar10_dataset():
     transform_cfg = config.transform_pipeline
 
     # build transform
-    transform_pipeline = [build_transform(cfg) for cfg in transform_cfg]
+    transform_pipeline = [ToTensor()]
     transform_pipeline = transforms.Compose(transform_pipeline)
     dataset_cfg['transform'] = transform_pipeline
 

tests/test_data/test_data_parallel_sampler.py

@@ -12,26 +12,25 @@ import torch.multiprocessing as mp
 from torch.utils.data import DataLoader
 
 import colossalai
-from colossalai.builder import build_dataset, build_transform
+from colossalai.builder import build_dataset
 from torchvision import transforms
 from colossalai.context import ParallelMode, Config
 from colossalai.core import global_context as gpc
 from colossalai.utils import get_dataloader, free_port
 from colossalai.testing import rerun_if_address_is_in_use
+from torchvision.transforms import ToTensor
 
 CONFIG = Config(
     dict(
-        train_data=dict(dataset=dict(
-            type='CIFAR10',
-            root=Path(os.environ['DATA']),
-            train=True,
-            download=True,
+        train_data=dict(
+            dataset=dict(
+                type='CIFAR10',
+                root=Path(os.environ['DATA']),
+                train=True,
+                download=True,
+            ),
+            dataloader=dict(batch_size=8,),
         ),
-                        dataloader=dict(batch_size=8,),
-                        transform_pipeline=[
-                            dict(type='ToTensor'),
-                            dict(type='Normalize', mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
-                        ]),
         parallel=dict(
             pipeline=dict(size=1),
             tensor=dict(size=1, mode=None),
@@ -45,7 +44,7 @@ def run_data_sampler(rank, world_size, port):
     colossalai.launch(**dist_args)
     print('finished initialization')
 
-    transform_pipeline = [build_transform(cfg) for cfg in gpc.config.train_data.transform_pipeline]
+    transform_pipeline = [ToTensor()]
     transform_pipeline = transforms.Compose(transform_pipeline)
     gpc.config.train_data.dataset['transform'] = transform_pipeline
     dataset = build_dataset(gpc.config.train_data.dataset)

tests/test_data/test_deterministic_dataloader.py

@@ -13,26 +13,24 @@ from torchvision import transforms
 from torch.utils.data import DataLoader
 
 import colossalai
-from colossalai.builder import build_dataset, build_transform
+from colossalai.builder import build_dataset
 from colossalai.context import ParallelMode, Config
 from colossalai.core import global_context as gpc
 from colossalai.utils import free_port
 from colossalai.testing import rerun_if_address_is_in_use
+from torchvision import transforms
 
 CONFIG = Config(
     dict(
-        train_data=dict(dataset=dict(
-            type='CIFAR10',
-            root=Path(os.environ['DATA']),
-            train=True,
-            download=True,
+        train_data=dict(
+            dataset=dict(
+                type='CIFAR10',
+                root=Path(os.environ['DATA']),
+                train=True,
+                download=True,
+            ),
+            dataloader=dict(num_workers=2, batch_size=2, shuffle=True),
         ),
-                        dataloader=dict(num_workers=2, batch_size=2, shuffle=True),
-                        transform_pipeline=[
-                            dict(type='ToTensor'),
-                            dict(type='RandomCrop', size=32),
-                            dict(type='Normalize', mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
-                        ]),
         parallel=dict(
             pipeline=dict(size=1),
             tensor=dict(size=1, mode=None),
@@ -50,7 +48,7 @@ def run_data_sampler(rank, world_size, port):
     transform_cfg = gpc.config.train_data.transform_pipeline
 
     # build transform
-    transform_pipeline = [build_transform(cfg) for cfg in transform_cfg]
+    transform_pipeline = [transforms.ToTensor(), transforms.RandomCrop(size=32)]
     transform_pipeline = transforms.Compose(transform_pipeline)
     dataset_cfg['transform'] = transform_pipeline