upgrade: replace CURL with GAPIC

community-content/pytorch_efficient_training/README.md

@@ -15,19 +15,15 @@ pip install -r requirements.txt
 * resnet_dp.py - Train ResNet-50 on single node multiple GPUs with `DataParallel` strategy.
 * resnet_ddp.py - Train ResNet-50 on single node multiple GPUs with `DistributedDataParallel` strategy.
 * resnet_ddp_wds.py - Train ResNet-50 on single node multiple GPUs with `DistributedDataParallel` strategy and `Webdataset`.
-* resnet_fsdp.py - Train ResNet-50 on single node multiple GPUs with `FullyShardedDataParallel` strategy.
-* resnet_fsdp_wds.py - Train ResNet-50 on single node multiple GPUs with `FullyShardedDataParallel` strategy and `Webdataset`.
 * shard_imagenet.py - Shard ImagNet individual files into `tar` files.
 
 ## Benchmark
 
 When run the benchmark on Nvidia T4 GPUs using ImageNet validation dataset, you can get the result like:
-Strategy               | Seconds/Epoch - Local Data | Seconds/Epoch - Cloud Data
----------------------- | -------------------------- | --------------------------
-On 1 GPU               | 489                        | 804 (2x slower)
-On 4 GPUs (DP)         | 157                        | 738 (5x slower)
-On 4 GPUs (DDP)        | 134                        | 432 (3x slower)
-On 4 GPUs (DDP + WDS)  | 131                        | 133 (same performance)
-On 4 GPUs (FSDP)       | 139                        | 353 (3x slower)
-On 4 GPUs (FSDP + WDS) | 138                        | 135 (same performance)
+Strategy              | Seconds/Epoch - Local Data | Seconds/Epoch - Cloud Data
+--------------------- | -------------------------- | --------------------------
+On 1 GPU              | 489                        | 804 (2x slower)
+On 4 GPUs (DP)        | 157                        | 738 (5x slower)
+On 4 GPUs (DDP)       | 134                        | 432 (3x slower)
+On 4 GPUs (DDP + WDS) | 131                        | 133 (same performance)
 

community-content/pytorch_efficient_training/resnet_fsdp.py

@@ -1,242 +0,0 @@
-# Copyright 2022 Google LLC
-#
-# Licensed under the Apache License, Version 2.0 (the \"License\");
-# you may not use this file except in compliance with the License.\n",
-# 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.
-
-"""Train resnet on multiple GPUs with FSDP."""
-
-import argparse
-import functools
-import os
-import time
-
-from PIL import Image
-import torch
-from torch import nn
-import torch.distributed as dist
-from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
-from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy
-import torch.multiprocessing as mp
-import torchmetrics
-import torchvision
-from torchvision.models import resnet50
-
-
-class ImageFolder(torchvision.datasets.ImageFolder):
-  """Class for loading imagenet."""
-
-  def __init__(self, image_list_file, transform=None, target_transform=None):
-    self.samples = self._make_dataset(image_list_file)
-    self.loader = self._loader
-
-    self.imgs = self.samples
-    self.targets = [s[1] for s in self.samples]
-
-    self.transform = transform
-    self.target_transform = target_transform
-
-  def _make_dataset(self, image_list_file):
-    items = []
-    with open(image_list_file, 'r') as f:
-      for line in f:
-        item = line.strip().split(' ')
-        items.append((item[0], int(item[1])))
-    return items
-
-  def _loader(self, image_path):
-    with open(image_path, 'rb') as f:
-      img = Image.open(f)
-      img = img.convert('RGB')
-      return img
-
-
-def train(model, device, dataloader, optimizer):
-  model.train()
-  for image, target in dataloader:
-    image = image.to(device, non_blocking=True)
-    target = target.to(device, non_blocking=True)
-    pred = model(image)
-    # pred.shape (N, C), target.shape (N)
-    loss = nn.functional.cross_entropy(pred, target)
-    optimizer.zero_grad()
-    loss.backward()
-    optimizer.step()
-  return loss
-
-
-def evaluate(model, device, dataloader, metric):
-  model.eval()
-  with torch.no_grad():
-    for image, target in dataloader:
-      image = image.to(device, non_blocking=True)
-      target = target.to(device, non_blocking=True)
-      pred = model(image)
-      metric.update(pred, target)
-  accuracy = metric.compute()
-  metric.reset()
-  return accuracy
-
-
-def worker(gpu, args):
-  """Run training and evaluation."""
-  # Init process group.
-  print(f'Initiating process {gpu}')
-  dist.init_process_group(
-      backend='nccl',
-      init_method='env://',
-      world_size=args.gpus,
-      rank=gpu)
-
-  # Create train dataloader.
-  train_dataset = ImageFolder(
-      image_list_file=args.train_data_path,
-      transform=torchvision.transforms.Compose([
-          torchvision.transforms.RandomResizedCrop(224),
-          torchvision.transforms.RandomHorizontalFlip(),
-          torchvision.transforms.ToTensor(),
-          torchvision.transforms.Normalize(
-              mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-      ]))
-  train_sampler = torch.utils.data.distributed.DistributedSampler(
-      train_dataset, num_replicas=args.gpus, rank=gpu)
-  train_dataloader = torch.utils.data.DataLoader(
-      dataset=train_dataset,
-      batch_size=args.train_batch_size,
-      shuffle=False,
-      num_workers=args.dataloader_num_workers,
-      pin_memory=True,
-      sampler=train_sampler)
-  if gpu == 0:
-    print(f'Train dataloader | samples: {len(train_dataloader.dataset)}, '
-          f'num workers: {train_dataloader.num_workers}, '
-          f'global batch size: {args.train_batch_size * args.gpus}, '
-          f'batches/epoch: {len(train_dataloader)}')
-
-  # Create eval dataloader.
-  eval_dataset = ImageFolder(
-      image_list_file=args.eval_data_path,
-      transform=torchvision.transforms.Compose([
-          torchvision.transforms.Resize(256),
-          torchvision.transforms.CenterCrop(224),
-          torchvision.transforms.ToTensor(),
-          torchvision.transforms.Normalize(
-              mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-      ]))
-  eval_sampler = torch.utils.data.distributed.DistributedSampler(
-      eval_dataset, num_replicas=args.gpus, rank=gpu)
-  eval_dataloader = torch.utils.data.DataLoader(
-      dataset=eval_dataset,
-      batch_size=args.eval_batch_size,
-      shuffle=False,
-      num_workers=args.dataloader_num_workers,
-      pin_memory=True,
-      drop_last=True,
-      sampler=eval_sampler)
-  if gpu == 0:
-    print(f'Eval dataloader | samples: {len(eval_dataloader.dataset)}, '
-          f'num workers: {eval_dataloader.num_workers}, '
-          f'batch size: {args.eval_batch_size}, '
-          f'batches/epoch: {len(eval_dataloader)}')
-
-  # Wrap policy.
-  my_auto_wrap_policy = functools.partial(
-      size_based_auto_wrap_policy, min_num_params=100)
-  torch.cuda.set_device(gpu)
-
-  # Create model.
-  model = resnet50(weights=None)
-  model.to(args.device)
-  model = FSDP(model, auto_wrap_policy=my_auto_wrap_policy)
-
-  # Optimizer.
-  optimizer = torch.optim.SGD(model.parameters(), 0.1)
-
-  # Main loop.
-  metric = torchmetrics.classification.Accuracy(top_k=1).to(args.device)
-  for epoch in range(1, args.epochs + 1):
-    if gpu == 0:
-      print(f'Running epoch {epoch}')
-    train_sampler.set_epoch(epoch)
-
-    start = time.time()
-    train(model, args.device, train_dataloader, optimizer)
-    end = time.time()
-    if gpu == 0:
-      print(f'Training finished in {(end - start):>0.3f} seconds')
-
-    start = time.time()
-    evaluate(model, args.device, eval_dataloader, metric)
-    end = time.time()
-    if gpu == 0:
-      print(f'Evaluation finished in {(end - start):>0.3f} seconds')
-
-  if gpu == 0:
-    print('Done')
-  dist.destroy_process_group()
-
-
-def create_args():
-  """Create main args."""
-  parser = argparse.ArgumentParser(
-      formatter_class=argparse.ArgumentDefaultsHelpFormatter)
-  parser.add_argument(
-      '--gpus',
-      default=4,
-      type=int,
-      help='number of gpus to use')
-  parser.add_argument(
-      '--epochs',
-      default=2,
-      type=int,
-      help='number of total epochs to run')
-  parser.add_argument(
-      '--dataloader_num_workers',
-      default=2,
-      type=int,
-      help='number of workders for dataloader')
-  parser.add_argument(
-      '--train_data_path',
-      default='',
-      type=str,
-      help='path to training data')
-  parser.add_argument(
-      '--train_batch_size',
-      default=32,
-      type=int,
-      help='batch size for training per gpu')
-  parser.add_argument(
-      '--eval_data_path',
-      default='',
-      type=str,
-      help='path to evaluation data')
-  parser.add_argument(
-      '--eval_batch_size',
-      default=32,
-      type=int,
-      help='batch size for evaluation per gpu')
-  args = parser.parse_args()
-  return args
-
-
-def main():
-  args = create_args()
-
-  os.environ['MASTER_ADDR'] = 'localhost'
-  os.environ['MASTER_PORT'] = '8888'
-
-  args.device = 'cuda' if torch.cuda.is_available() else 'cpu'
-  print(f'Launch job on {args.gpus} GPUs with FSDP')
-  mp.spawn(worker, nprocs=args.gpus, args=(args,))
-
-
-if __name__ == '__main__':
-  main()

community-content/pytorch_efficient_training/resnet_fsdp_wds.py

@@ -1,240 +0,0 @@
-"""Train resnet on multiple GPUs with DDP."""
-
-import argparse
-import functools
-import itertools
-import math
-import os
-import time
-
-import torch
-from torch import nn
-import torch.distributed as dist
-from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
-from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy
-import torch.multiprocessing as mp
-import torchmetrics
-from torchvision.models import resnet50
-from torchvision.transforms import transforms
-import webdataset as wds
-
-
-def wds_split(src, rank, world_size):
-  """Shards split function for webdataset."""
-  # The context of caller of this function is within multiple processes
-  # (by DDP world_size) and multiple workers (by dataloader_num_workers).
-  # So we totally have (world_size * num_workers) workers for processing data.
-  # NOTE: Raw data should be sharded to enough shards to make sure one process
-  # can handle at least one shard, otherwise the process may hang.
-  worker_id = 0
-  num_workers = 1
-  worker_info = torch.utils.data.get_worker_info()
-  if worker_info:
-    worker_id = worker_info.id
-    num_workers = worker_info.num_workers
-  for s in itertools.islice(src, rank * num_workers + worker_id, None,
-                            world_size * num_workers):
-    yield s
-
-
-def identity(x):
-  return x
-
-
-def create_wds_dataloader(rank, args, mode):
-  """Create webdataset dataset and dataloader."""
-  if mode == 'train':
-    transform = transforms.Compose([
-        transforms.RandomResizedCrop(224),
-        transforms.RandomHorizontalFlip(),
-        transforms.ToTensor(),
-        transforms.Normalize(
-            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    ])
-    data_path = args.train_data_path
-    data_size = args.train_data_size
-    batch_size_local = args.train_batch_size
-    batch_size_global = args.train_batch_size * args.gpus
-    # Since webdataset disallows partial batch, we pad the last batch for train.
-    batches = int(math.ceil(data_size / batch_size_global))
-  else:
-    transform = transforms.Compose([
-        transforms.Resize(256),
-        transforms.CenterCrop(224),
-        transforms.ToTensor(),
-        transforms.Normalize(
-            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    ])
-    data_path = args.eval_data_path
-    data_size = args.eval_data_size
-    batch_size_local = args.eval_batch_size
-    batch_size_global = args.eval_batch_size * args.gpus
-    # Since webdataset disallows partial batch, we drop the last batch for eval.
-    batches = int(data_size / batch_size_global)
-
-  dataset = wds.DataPipeline(
-      wds.SimpleShardList(data_path),
-      functools.partial(wds_split, rank=rank, world_size=args.gpus),
-      wds.tarfile_to_samples(),
-      wds.decode('pil'),
-      wds.to_tuple('jpg;png;jpeg cls'),
-      wds.map_tuple(transform, identity),
-      wds.batched(batch_size_local, partial=False),
-  )
-  num_workers = args.dataloader_num_workers
-  dataloader = wds.WebLoader(
-      dataset=dataset,
-      batch_size=None,
-      shuffle=False,
-      num_workers=num_workers,
-      persistent_workers=True if num_workers > 0 else False,
-      pin_memory=True).repeat(nbatches=batches)
-  print(f'{mode} dataloader | samples: {data_size}, '
-        f'num_workers: {num_workers}, '
-        f'local batch size: {batch_size_local}, '
-        f'global batch size: {batch_size_global}, '
-        f'batches: {batches}')
-  return dataloader
-
-
-def train(model, device, dataloader, optimizer):
-  model.train()
-  for image, target in dataloader:
-    image = image.to(device, non_blocking=True)
-    target = target.to(device, non_blocking=True)
-    pred = model(image)
-    # pred.shape (N, C), target.shape (N)
-    loss = nn.functional.cross_entropy(pred, target)
-    optimizer.zero_grad()
-    loss.backward()
-    optimizer.step()
-  return loss
-
-
-def evaluate(model, device, dataloader, metric):
-  model.eval()
-  with torch.no_grad():
-    for image, target in dataloader:
-      image = image.to(device, non_blocking=True)
-      target = target.to(device, non_blocking=True)
-      pred = model(image)
-      metric.update(pred, target)
-  accuracy = metric.compute()
-  metric.reset()
-  return accuracy
-
-
-def worker(gpu, args):
-  """Run training and evaluation."""
-  # Init process group.
-  print(f'Initiating process {gpu}')
-  dist.init_process_group(
-      backend='nccl',
-      init_method='env://',
-      world_size=args.gpus,
-      rank=gpu)
-
-  # Create dataloader.
-  train_dataloader = create_wds_dataloader(gpu, args, 'train')
-  eval_dataloader = create_wds_dataloader(gpu, args, 'eval')
-
-  # Wrap policy.
-  my_auto_wrap_policy = functools.partial(
-      size_based_auto_wrap_policy, min_num_params=100)
-  torch.cuda.set_device(gpu)
-
-  # Create model.
-  model = resnet50(weights=None)
-  model.to(args.device)
-  model = FSDP(model, auto_wrap_policy=my_auto_wrap_policy)
-
-  # Optimizer.
-  optimizer = torch.optim.SGD(model.parameters(), 0.1)
-
-  # Main loop.
-  metric = torchmetrics.classification.Accuracy(top_k=1).to(args.device)
-  for epoch in range(1, args.epochs + 1):
-    if gpu == 0:
-      print(f'Running epoch {epoch}')
-
-    start = time.time()
-    train(model, args.device, train_dataloader, optimizer)
-    end = time.time()
-    if gpu == 0:
-      print(f'Training finished in {(end - start):>0.3f} seconds')
-
-    start = time.time()
-    evaluate(model, args.device, eval_dataloader, metric)
-    end = time.time()
-    if gpu == 0:
-      print(f'Evaluation finished in {(end - start):>0.3f} seconds')
-
-  if gpu == 0:
-    print('Done')
-
-
-def create_args():
-  """Create main args."""
-  parser = argparse.ArgumentParser(
-      formatter_class=argparse.ArgumentDefaultsHelpFormatter)
-  parser.add_argument(
-      '--gpus',
-      default=4,
-      type=int,
-      help='number of gpus to use')
-  parser.add_argument(
-      '--epochs',
-      default=2,
-      type=int,
-      help='number of total epochs to run')
-  parser.add_argument(
-      '--dataloader_num_workers',
-      default=2,
-      type=int,
-      help='number of workders for dataloader')
-  parser.add_argument(
-      '--train_data_path',
-      default='',
-      type=str,
-      help='path to training data')
-  parser.add_argument(
-      '--train_batch_size',
-      default=32,
-      type=int,
-      help='batch size for training per gpu')
-  parser.add_argument(
-      '--train_data_size',
-      default=50000,
-      type=int,
-      help='data size for training')
-  parser.add_argument(
-      '--eval_data_path',
-      default='',
-      type=str,
-      help='path to evaluation data')
-  parser.add_argument(
-      '--eval_batch_size',
-      default=32,
-      type=int,
-      help='batch size for evaluation per gpu')
-  parser.add_argument(
-      '--eval_data_size',
-      default=50000,
-      type=int,
-      help='data size for evaluation')
-  args = parser.parse_args()
-  return args
-
-
-def main():
-  args = create_args()
-  os.environ['MASTER_ADDR'] = 'localhost'
-  os.environ['MASTER_PORT'] = '8888'
-
-  args.device = 'cuda' if torch.cuda.is_available() else 'cpu'
-  print(f'Launch job on {args.gpus} GPUs with FSDP')
-  mp.spawn(worker, nprocs=args.gpus, args=(args,))
-
-
-if __name__ == '__main__':
-  main()

notebooks/official/tabular_workflows/tabnet_on_vertex_pipelines.ipynb

@@ -97,7 +97,7 @@
         "### Dataset\n",
         "\n",
         "The dataset you will be using is [Bank Marketing](https://archive.ics.uci.edu/ml/datasets/bank+marketing).\n",
-        "The data is for direct marketing campaigns (phone calls) of a Portuguese banking institution. The binary classification goal is to predict if a client subscribe a term deposit. For this notebook, you randomly selected 90% of the rows in the original dataset and saved them in a train.csv file hosted on Cloud Storage. To download the file, click [here](https://storage.googleapis.com/cloud-samples-data-us-central1/vertex-ai/tabular-workflows/datasets/bank-marketing/train.csv)."
+        "The data is for direct marketing campaigns (phone calls) of a Portuguese banking institution. The binary classification goal is to predict if a client subscribe a term deposit. For this notebook, you randomly selected 90% of the rows in the original dataset and saved them in a train.csv file hosted on Cloud Storage. To download the file, click [here](https://storage.googleapis.com/cloud-samples-data/vertex-ai/tabular-workflows/datasets/bank-marketing/train.csv)."
       ]
     },
     {
@@ -746,7 +746,7 @@
       },
       "outputs": [],
       "source": [
-        "data_source_csv_filenames = \"gs://cloud-samples-data-us-central1/vertex-ai/tabular-workflows/datasets/bank-marketing/train.csv\"\n",
+        "data_source_csv_filenames = \"gs://cloud-samples-data/vertex-ai/tabular-workflows/datasets/bank-marketing/train.csv\"\n",
         "data_source_bigquery_table_path = (\n",
         "    None  # @param {type:\"string\"}, format: bq://bq_project.bq_dataset.bq_table\n",
         ")"

notebooks/official/tabular_workflows/wide_and_deep_on_vertex_pipelines.ipynb

@@ -97,7 +97,7 @@
         "### Dataset\n",
         "\n",
         "The dataset you will be using is [Bank Marketing](https://archive.ics.uci.edu/ml/datasets/bank+marketing).\n",
-        "The data is for direct marketing campaigns (phone calls) of a Portuguese banking institution. The binary classification goal is to predict if a client will subscribe a term deposit. For this notebook, we randomly selected 90% of the rows in the original dataset and saved them in a train.csv file hosted on Cloud Storage. To download the file, click [here](https://storage.googleapis.com/cloud-samples-data-us-central1/vertex-ai/tabular-workflows/datasets/bank-marketing/train.csv)."
+        "The data is for direct marketing campaigns (phone calls) of a Portuguese banking institution. The binary classification goal is to predict if a client will subscribe a term deposit. For this notebook, we randomly selected 90% of the rows in the original dataset and saved them in a train.csv file hosted on Cloud Storage. To download the file, click [here](https://storage.googleapis.com/cloud-samples-data/vertex-ai/tabular-workflows/datasets/bank-marketing/train.csv)."
       ]
     },
     {
@@ -667,7 +667,7 @@
       },
       "outputs": [],
       "source": [
-        "data_source_csv_filenames = \"gs://cloud-samples-data-us-central1/vertex-ai/tabular-workflows/datasets/bank-marketing/train.csv\"\n",
+        "data_source_csv_filenames = \"gs://cloud-samples-data/vertex-ai/tabular-workflows/datasets/bank-marketing/train.csv\"\n",
         "data_source_bigquery_table_path = (\n",
         "    None  # @param {type:\"string\"}, format: bq://bq_project.bq_dataset.bq_table\n",
         ")"

notebooks/official/tensorboard/tensorboard_profiler_custom_training.ipynb

@@ -29,7 +29,7 @@
         "id": "l2mMvIUG9meX"
       },
       "source": [
-        "# Profile model training performance using Vertex AI TensorBoard Profiler\n",
+        "# Profile model training performance using Profiler\n",
         "\n",
         "<table align=\"left\">\n",
         "\n",
@@ -49,7 +49,7 @@
         "      <img src=\"https://lh3.googleusercontent.com/UiNooY4LUgW_oTvpsNhPpQzsstV5W8F7rYgxgGBD85cWJoLmrOzhVs_ksK_vgx40SHs7jCqkTkCk=e14-rj-sc0xffffff-h130-w32\" alt=\"Vertex AI logo\">\n",
         "      Open in Vertex AI Workbench\n",
         "    </a>\n",
-        "  </td>\n",
+        "  </td>                                                                                               \n",
         "</table>"
       ]
     },
@@ -331,16 +331,22 @@
         "! gcloud projects add-iam-policy-binding $PROJECT_ID \\\n",
         "        --member=\"serviceAccount:$SERVICE_ACCOUNT\" \\\n",
         "        --role=\"roles/storage.admin\" \\\n",
-        "        --quiet\n",
-        "\n",
+        "        --quiet"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "S_8_5jm-Gk6w"
+      },
+      "outputs": [],
+      "source": [
         "# Grant AI Platform permission.\n",
         "! gcloud projects add-iam-policy-binding $PROJECT_ID \\\n",
         "        --member=\"serviceAccount:$SERVICE_ACCOUNT\" \\\n",
         "        --role=\"roles/aiplatform.user\" \\\n",
-        "        --quiet\n",
-        "\n",
-        "! gcloud projects get-iam-policy $PROJECT_ID \\\n",
-        "        --filter=bindings.members:serviceAccount:$SERVICE_ACCOUNT"
+        "        --quiet"
       ]
     },
     {
@@ -762,7 +768,7 @@
         "WORKDIR /root\n",
         "\n",
         "# Installs additional packages as you need.\n",
-        "RUN pip3 install google-cloud-aiplatform[cloud_profiler]>=1.19.1\n",
+        "RUN pip3 install google-cloud-aiplatform[cloud_profiler]\n",
         "\n",
         "# Copies the trainer code to the docker image.\n",
         "RUN mkdir /root/trainer\n",
@@ -792,7 +798,7 @@
         "IMAGE_NAME = \"tensorboard-custom-container\"\n",
         "IMAGE_URI = f\"{REGION}-docker.pkg.dev/{PROJECT_ID}/{DOCKER_REPOSITORY}/{IMAGE_NAME}\"\n",
         "\n",
-        "! gcloud builds submit --project {PROJECT_ID} --region={REGION} --tag {IMAGE_URI} --timeout=3600s --quiet"
+        "! gcloud builds submit --project {PROJECT_ID} --region={REGION} --tag {IMAGE_URI} --timeout=60m --quiet"
       ]
     },
     {