[SGD] Imagenet example (basic) (#8020)

* Checkpoint the image-models example

* Update cluster definition

* Fix copyright info

* Use original args

* Checkpoint fixes

* Add README

* Add some missing features

* Format

* Get rid of the unused Namespace class

* Address comments

* Link the imagenet example in docs

* Cleanup

* Fix lint

python/ray/util/sgd/torch/examples/image_models/.gitignore

@@ -0,0 +1,2 @@
+imagenette2.tgz
+/data

python/ray/util/sgd/torch/examples/image_models/LICENSE_THIRDPARTY

@@ -0,0 +1,202 @@
+Components of timm (args.py):
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "{}"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2019 Ross Wightman
+
+   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.

python/ray/util/sgd/torch/examples/image_models/README.md

@@ -0,0 +1,48 @@
+# Imagenet Models RaySGD Example
+
+Based on [the timm package](https://github.com/rwightman/pytorch-image-models).
+
+# Usage
+
+## Ray autoscaler
+
+- `ray up cluster.yaml`
+- `ray rsync-up -A cluster.yaml`
+- `ray submit train.py -- data -n=4`
+
+## Manual
+
+- Make `train.py` and `args.py` available on the remote host.
+- `pip install timm`
+- Download and unpack an ImageNet-compatible dataset (has to be full size). Internally we use [Imagenette](https://github.com/fastai/imagenette) for development purposes.
+- Optional: setup a ray cluster (`ray start --head` on the head node and `ray start --redis-address HEAD_ADDRESS` on each of the worker nodes).
+- Run `python train.py DATA_DIRECTORY` on the head node.
+
+## Manual (single node)
+
+- `pip install timm`
+- `ray start --head`
+- `python train.py DATA_DIRECTORY`
+- Use the `-n` argument to control the number of processes + GPUs used.
+
+# Advantages
+
+Compared to the original `timm` package, the RaySGD train script has a few advantages:
+
+- Compatibility with Ray autoscaler (automatic simple cluster provisioning).
+- Built-in fault tolerance (epochs will checkpoint and restart if a worker fails). This means you can, for example, make all your worker nodes preemtible (e.g. AWS spot requests). **Note:** the head node *must* be non-preemtible.
+- Since a Ray cluster is already setup, you can run other distributed tasks.
+
+# Limitations
+
+Support for some command line flags from the original `timm` package has been intentionally dropped:
+- `-j/--workers` - Ray can start multiple training processes with `-n/--num-workers` instead.
+- `--num-gpu` - Ray can use multiple GPUs by launching multiple processes with `-n/--num-workers` instead. (`DistributedDataParallel` is faster than simple `DataParallel` in practice anyway)
+
+Other features are still in the works:
+- Logging
+- Compatibility with timm checkpoints
+- EMA
+- Sync batch norm
+- Learning rate scheduling
+- Some testing

python/ray/util/sgd/torch/examples/image_models/args.py

@@ -0,0 +1,477 @@
+# Modified by Maksim Smolin in 2020
+#
+# Original work by Ross Wightman as part of the timm package
+# (see LICENSE_THIRDPARTY)
+
+# Note: other authors MUST include themselves in the above copyright notice
+#       in order to abide by the terms of the Apache license
+
+import logging
+import argparse
+import yaml
+
+config_parser = parser = argparse.ArgumentParser(
+    description="Training Config", add_help=False)
+parser.add_argument(
+    "-c",
+    "--config",
+    default="",
+    type=str,
+    metavar="FILE",
+    help="YAML config file specifying default arguments")
+
+parser = argparse.ArgumentParser(description="PyTorch ImageNet Training")
+# Dataset / Model parameters
+parser.add_argument("data", metavar="DIR", help="path to dataset")
+parser.add_argument(
+    "--model",
+    default="resnet101",
+    type=str,
+    metavar="MODEL",
+    help="Name of model to train (default: 'countception'")
+parser.add_argument(
+    "--pretrained",
+    action="store_true",
+    default=False,
+    help="Start with pretrained version of specified network (if avail)")
+parser.add_argument(
+    "--initial-checkpoint",
+    default="",
+    type=str,
+    metavar="PATH",
+    help="Initialize model from this checkpoint (default: none)")
+parser.add_argument(
+    "--resume",
+    default="",
+    type=str,
+    metavar="PATH",
+    help="Resume full model and optimizer state from checkpoint "
+    "(default: none)")
+parser.add_argument(
+    "--no-resume-opt",
+    action="store_true",
+    default=False,
+    help="prevent resume of optimizer state when resuming model")
+parser.add_argument(
+    "--num-classes",
+    type=int,
+    default=1000,
+    metavar="N",
+    help="number of label classes (default: 1000)")
+parser.add_argument(
+    "--gp",
+    default="avg",
+    type=str,
+    metavar="POOL",
+    help=("Type of global pool, 'avg', 'max', 'avgmax', 'avgmaxc' "
+          "(default: 'avg')"))
+parser.add_argument(
+    "--img-size",
+    type=int,
+    default=None,
+    metavar="N",
+    help="Image patch size (default: None => model default)")
+parser.add_argument(
+    "--crop-pct",
+    default=None,
+    type=float,
+    metavar="N",
+    help="Input image center crop percent (for validation only)")
+parser.add_argument(
+    "--mean",
+    type=float,
+    nargs="+",
+    default=None,
+    metavar="MEAN",
+    help="Override mean pixel value of dataset")
+parser.add_argument(
+    "--std",
+    type=float,
+    nargs="+",
+    default=None,
+    metavar="STD",
+    help="Override std deviation of of dataset")
+parser.add_argument(
+    "--interpolation",
+    default="",
+    type=str,
+    metavar="NAME",
+    help="Image resize interpolation type (overrides model)")
+parser.add_argument(
+    "-b",
+    "--batch-size",
+    type=int,
+    default=32,
+    metavar="N",
+    help="input batch size for training (default: 32)")
+parser.add_argument(
+    "-vb",
+    "--validation-batch-size-multiplier",
+    type=int,
+    default=1,
+    metavar="N",
+    help="ratio of validation batch size to training batch size (default: 1)")
+parser.add_argument(
+    "--drop",
+    type=float,
+    default=0.0,
+    metavar="PCT",
+    help="Dropout rate (default: 0.)")
+parser.add_argument(
+    "--drop-connect",
+    type=float,
+    default=None,
+    metavar="PCT",
+    help="Drop connect rate, DEPRECATED, use drop-path (default: None)")
+parser.add_argument(
+    "--drop-path",
+    type=float,
+    default=None,
+    metavar="PCT",
+    help="Drop path rate (default: None)")
+parser.add_argument(
+    "--drop-block",
+    type=float,
+    default=None,
+    metavar="PCT",
+    help="Drop block rate (default: None)")
+parser.add_argument(
+    "--jsd",
+    action="store_true",
+    default=False,
+    help="Enable Jensen-Shannon Divergence + CE loss. Use with `--aug-splits`."
+)
+# Optimizer parameters
+parser.add_argument(
+    "--opt",
+    default="sgd",
+    type=str,
+    metavar="OPTIMIZER",
+    help="Optimizer (default: 'sgd'")
+parser.add_argument(
+    "--opt-eps",
+    default=1e-8,
+    type=float,
+    metavar="EPSILON",
+    help="Optimizer Epsilon (default: 1e-8)")
+parser.add_argument(
+    "--momentum",
+    type=float,
+    default=0.9,
+    metavar="M",
+    help="SGD momentum (default: 0.9)")
+parser.add_argument(
+    "--weight-decay",
+    type=float,
+    default=0.0001,
+    help="weight decay (default: 0.0001)")
+# Learning rate schedule parameters
+parser.add_argument(
+    "--sched",
+    default="step",
+    type=str,
+    metavar="SCHEDULER",
+    help="LR scheduler (default: 'step'")
+parser.add_argument(
+    "--lr",
+    type=float,
+    default=0.01,
+    metavar="LR",
+    help="learning rate (default: 0.01)")
+parser.add_argument(
+    "--lr-noise",
+    type=float,
+    nargs="+",
+    default=None,
+    metavar="pct, pct",
+    help="learning rate noise on/off epoch percentages")
+parser.add_argument(
+    "--lr-noise-pct",
+    type=float,
+    default=0.67,
+    metavar="PERCENT",
+    help="learning rate noise limit percent (default: 0.67)")
+parser.add_argument(
+    "--lr-noise-std",
+    type=float,
+    default=1.0,
+    metavar="STDDEV",
+    help="learning rate noise std-dev (default: 1.0)")
+parser.add_argument(
+    "--warmup-lr",
+    type=float,
+    default=0.0001,
+    metavar="LR",
+    help="warmup learning rate (default: 0.0001)")
+parser.add_argument(
+    "--min-lr",
+    type=float,
+    default=1e-5,
+    metavar="LR",
+    help="lower lr bound for cyclic schedulers that hit 0 (1e-5)")
+parser.add_argument(
+    "--epochs",
+    type=int,
+    default=200,
+    metavar="N",
+    help="number of epochs to train (default: 2)")
+parser.add_argument(
+    "--start-epoch",
+    default=None,
+    type=int,
+    metavar="N",
+    help="manual epoch number (useful on restarts)")
+parser.add_argument(
+    "--decay-epochs",
+    type=float,
+    default=30,
+    metavar="N",
+    help="epoch interval to decay LR")
+parser.add_argument(
+    "--warmup-epochs",
+    type=int,
+    default=3,
+    metavar="N",
+    help="epochs to warmup LR, if scheduler supports")
+parser.add_argument(
+    "--cooldown-epochs",
+    type=int,
+    default=10,
+    metavar="N",
+    help="epochs to cooldown LR at min_lr, after cyclic schedule ends")
+parser.add_argument(
+    "--patience-epochs",
+    type=int,
+    default=10,
+    metavar="N",
+    help="patience epochs for Plateau LR scheduler (default: 10")
+parser.add_argument(
+    "--decay-rate",
+    "--dr",
+    type=float,
+    default=0.1,
+    metavar="RATE",
+    help="LR decay rate (default: 0.1)")
+# Augmentation parameters
+parser.add_argument(
+    "--color-jitter",
+    type=float,
+    default=0.4,
+    metavar="PCT",
+    help="Color jitter factor (default: 0.4)")
+parser.add_argument(
+    "--aa",
+    type=str,
+    default=None,
+    metavar="NAME",
+    help="Use AutoAugment policy. 'v0' or 'original'. (default: None)"),
+# parser.add_argument(
+#     "--aug-splits",
+#     type=int,
+#     default=0,
+#     help="Number of augmentation splits (default: 0, valid: 0 or >=2)")
+parser.add_argument(
+    "--reprob",
+    type=float,
+    default=0.,
+    metavar="PCT",
+    help="Random erase prob (default: 0.)")
+parser.add_argument(
+    "--remode",
+    type=str,
+    default="const",
+    help="Random erase mode (default: 'const')")
+parser.add_argument(
+    "--recount", type=int, default=1, help="Random erase count (default: 1)")
+parser.add_argument(
+    "--resplit",
+    action="store_true",
+    default=False,
+    help="Do not random erase first (clean) augmentation split")
+parser.add_argument(
+    "--mixup",
+    type=float,
+    default=0.0,
+    help="mixup alpha, mixup enabled if > 0. (default: 0.)")
+parser.add_argument(
+    "--mixup-off-epoch",
+    default=0,
+    type=int,
+    metavar="N",
+    help="turn off mixup after this epoch, disabled if 0 (default: 0)")
+parser.add_argument(
+    "--smoothing",
+    type=float,
+    default=0.1,
+    help="label smoothing (default: 0.1)")
+parser.add_argument(
+    "--train-interpolation",
+    type=str,
+    default="random",
+    help="Training interpolation (random, bilinear, bicubic default: 'random')"
+)
+# Batch norm parameters
+# (only works with gen_efficientnet based models currently)
+parser.add_argument(
+    "--bn-tf",
+    action="store_true",
+    default=False,
+    help="Use Tensorflow BatchNorm defaults for models that support it "
+    "(default: False)")
+parser.add_argument(
+    "--bn-momentum",
+    type=float,
+    default=None,
+    help="BatchNorm momentum override (if not None)")
+parser.add_argument(
+    "--bn-eps",
+    type=float,
+    default=None,
+    help="BatchNorm epsilon override (if not None)")
+parser.add_argument(
+    "--sync-bn",
+    action="store_true",
+    help="Enable NVIDIA Apex or Torch synchronized BatchNorm.")
+parser.add_argument(
+    "--dist-bn",
+    type=str,
+    default="",
+    help=("Distribute BatchNorm stats between nodes after each epoch "
+          "('broadcast', 'reduce', or '')"))
+# parser.add_argument(
+#     "--split-bn",
+#     action="store_true",
+#     help="Enable separate BN layers per augmentation split.")
+# Model Exponential Moving Average
+parser.add_argument(
+    "--model-ema",
+    action="store_true",
+    default=False,
+    help="Enable tracking moving average of model weights")
+parser.add_argument(
+    "--model-ema-force-cpu",
+    action="store_true",
+    default=False,
+    help="Force ema to be tracked on CPU, rank=0 node only. "
+    "Disables EMA validation.")
+parser.add_argument(
+    "--model-ema-decay",
+    type=float,
+    default=0.9998,
+    help="decay factor for model weights moving average (default: 0.9998)")
+# Misc
+parser.add_argument(
+    "--seed",
+    type=int,
+    default=42,
+    metavar="S",
+    help="random seed (default: 42)")
+parser.add_argument(
+    "--log-interval",
+    type=int,
+    default=50,
+    metavar="N",
+    help="how many batches to wait before logging training status")
+parser.add_argument(
+    "--recovery-interval",
+    type=int,
+    default=0,
+    metavar="N",
+    help="how many batches to wait before writing recovery checkpoint")
+parser.add_argument(
+    "--no-gpu",
+    action="store_true",
+    default=False,
+    help="do not use a GPU even if available")
+parser.add_argument(
+    "--save-images",
+    action="store_true",
+    default=False,
+    help="save images of input bathes every log interval for debugging")
+parser.add_argument(
+    "--amp",
+    action="store_true",
+    default=False,
+    help="use NVIDIA amp for mixed precision training")
+parser.add_argument(
+    "--pin-mem",
+    action="store_true",
+    default=False,
+    help="Pin CPU memory in DataLoader for more efficient (sometimes) "
+    "transfer to GPU.")
+parser.add_argument(
+    "--no-prefetcher",
+    action="store_true",
+    default=False,
+    help="disable fast prefetcher")
+parser.add_argument(
+    "--output",
+    default="",
+    type=str,
+    metavar="PATH",
+    help="path to output folder (default: none, current dir)")
+parser.add_argument(
+    "--eval-metric",
+    default="prec1",
+    type=str,
+    metavar="EVAL_METRIC",
+    help="Best metric (default: 'prec1'")
+parser.add_argument(
+    "--tta",
+    type=int,
+    default=0,
+    metavar="N",
+    help="Test/inference time augmentation (oversampling) factor. 0=None "
+    "(default: 0)")
+parser.add_argument("--local_rank", default=0, type=int)
+
+# ray
+parser.add_argument(
+    "--ray-address",
+    default="auto",
+    metavar="ADDR",
+    help="Ray cluster address. [default=auto]")
+parser.add_argument(
+    "-n",
+    "--ray-num-workers",
+    type=int,
+    default=1,
+    metavar="N",
+    help="Number of Ray replicas to use. [default=1]")
+
+
+def parse_args():
+    # Do we have a config file to parse?
+    args_config, remaining = config_parser.parse_known_args()
+    if args_config.config:
+        with open(args_config.config, "r") as f:
+            cfg = yaml.safe_load(f)
+            parser.set_defaults(**cfg)
+
+    # The main arg parser parses the rest of the args, the usual
+    # defaults will have been overridden if config file specified.
+    args = parser.parse_args(remaining)
+
+    # Cache the args as a text string to save them in the output dir later
+    args_text = yaml.safe_dump(args.__dict__, default_flow_style=False)
+
+    # Arguments pre-processing from the original train.py
+    args.prefetcher = not args.no_prefetcher
+    args.distributed = False  # ray SGD handles this (DistributedSampler)
+    args.device = "cuda"  # ray should handle this
+
+    if args.no_gpu == 0 and args.prefetcher:
+        logging.warning("Prefetcher needs CUDA currently "
+                        "(might be a bug in timm). "
+                        "Disabling it.")
+        args.prefetcher = False
+
+    # assert args.aug_splits == 0 or args.aug_splits > 1, (
+    #     "Split must be 0 or 2+")
+
+    # args.num_aug_splits = args.aug_splits
+    args.num_aug_splits = 0  # todo:
+
+    args.split_bn = False  # todo:
+
+    return args, args_text

python/ray/util/sgd/torch/examples/image_models/cluster.yaml

@@ -0,0 +1,102 @@
+# An unique identifier for the head node and workers of this cluster.
+cluster_name: sgd-pytorch-imagenet
+
+# The maximum number of workers nodes to launch in addition to the head
+# node. This takes precedence over min_workers. min_workers default to 0.
+min_workers: 1
+initial_workers: 1
+max_workers: 1
+
+target_utilization_fraction: 0.9
+
+# If a node is idle for this many minutes, it will be removed.
+idle_timeout_minutes: 10
+# docker:
+#     image: tensorflow/tensorflow:1.5.0-py3
+#     container_name: ray_docker
+
+# Cloud-provider specific configuration.
+provider:
+    type: aws
+    region: us-east-1
+    availability_zone: us-east-1c
+
+# How Ray will authenticate with newly launched nodes.
+auth:
+    ssh_user: ubuntu
+    # ssh_private_key: ...
+
+head_node:
+    InstanceType: p3.2xlarge
+    ImageId: ami-0698bcaf8bd9ef56d
+    # KeyName: ...
+    InstanceMarketOptions:
+        MarketType: spot
+        SpotOptions:
+            BlockDurationMinutes: 360
+    BlockDeviceMappings:
+        - DeviceName: /dev/sda1
+          Ebs:
+              VolumeSize: 300
+           # SpotOptions:
+           #     MaxPrice: "9.0"
+
+
+worker_nodes:
+    InstanceType: p3.8xlarge
+    ImageId: ami-0698bcaf8bd9ef56d
+    # KeyName: ...
+    InstanceMarketOptions:
+        MarketType: spot
+        SpotOptions:
+            BlockDurationMinutes: 360
+    BlockDeviceMappings:
+        - DeviceName: /dev/sda1
+          Ebs:
+              VolumeSize: 300
+        # SpotOptions:
+        #     MaxPrice: "9.0"
+    #     # Run workers on spot by default. Comment this out to use on-demand.
+    #     InstanceMarketOptions:
+    #         MarketType: spot
+
+setup_commands:
+    # This replaces the standard anaconda Ray installation
+    - ray || pip install -U https://s3-us-west-2.amazonaws.com/ray-wheels/latest/ray-0.9.0.dev0-cp36-cp36m-manylinux1_x86_64.whl
+    # Uncomment this and the filemount to update the Ray installation with your local Ray code
+    # - rm -rf ./anaconda3/lib/python3.6/site-packages/ray/util/sgd/
+    # - cp -rf ~/sgd ./anaconda3/lib/python3.6/site-packages/ray/util/
+
+    # Installing this without -U to make sure we don't replace the existing Ray installation
+    - pip install ray[rllib]
+    - pip install -U ipdb torch torchvision tqdm
+    # Install Apex
+    - rm -rf apex || true
+    - git clone https://github.com/NVIDIA/apex && cd apex && pip install -v --no-cache-dir  ./ || true
+
+    # Install timm and get data
+    - pip install timm
+    - ls data || (curl https://s3.amazonaws.com/fast-ai-imageclas/imagenette2.tgz -O && tar zxvf imagenette2.tgz && mv imagenette2 data)
+
+
+file_mounts: {
+    # This should point to ray/python/ray/util/sgd.
+    # ~/anaconda3/lib/python3.6/site-packages/ray/util/sgd/: ../../../../sgd,
+}
+
+# Custom commands that will be run on the head node after common setup.
+head_setup_commands: []
+
+# Custom commands that will be run on worker nodes after common setup.
+worker_setup_commands: []
+
+# # Command to start ray on the head node. You don't need to change this.
+head_start_ray_commands:
+    - ray stop
+    - ray start --head --redis-port=6379 --object-manager-port=8076 --autoscaling-config=~/ray_bootstrap_config.yaml --object-store-memory=1000000000
+
+# Command to start ray on worker nodes. You don't need to change this.
+worker_start_ray_commands:
+    - ray stop
+    - ray start --address=$RAY_HEAD_IP:6379 --object-manager-port=8076 --object-store-memory=1000000000
+

python/ray/util/sgd/torch/examples/image_models/get_data.sh

@@ -0,0 +1,4 @@
+#!/usr/bin/env bash
+curl https://s3.amazonaws.com/fast-ai-imageclas/imagenette2.tgz -O
+tar zxvf imagenette2.tgz
+mv imagenette2 data

python/ray/util/sgd/torch/examples/image_models/train.py

@@ -0,0 +1,151 @@
+# Based on work by Ross Wightman as part of the timm package
+# (see LICENSE_THIRDPARTY)
+#
+# As modified by
+# - Maksim Smolin in 2020
+
+# Note: other authors MUST include themselves in the above copyright notice
+#       in order to abide by the terms of the Apache license
+
+from os.path import join
+
+from tqdm import trange
+
+import torch.nn as nn
+
+from timm.data import Dataset, create_loader
+from timm.data import resolve_data_config, FastCollateMixup
+from timm.models import create_model, convert_splitbn_model
+from timm.optim import create_optimizer
+from timm.utils import setup_default_logging
+
+import ray
+from ray.util.sgd.utils import BATCH_SIZE
+
+from ray.util.sgd import TorchTrainer
+# from ray.util.sgd.torch import TrainingOperator
+
+from ray.util.sgd.torch.examples.image_models.args import parse_args
+
+
+def model_creator(config):
+    args = config["args"]
+
+    model = create_model(
+        "resnet101",  # args.model,
+        pretrained=args.pretrained,
+        num_classes=args.num_classes,
+        drop_rate=args.drop,
+        drop_connect_rate=args.drop_connect,  # DEPRECATED, use drop_path
+        drop_path_rate=args.drop_path,
+        drop_block_rate=args.drop_block,
+        global_pool=args.gp,
+        bn_tf=args.bn_tf,
+        bn_momentum=args.bn_momentum,
+        bn_eps=args.bn_eps,
+        checkpoint_path=args.initial_checkpoint)
+
+    # always false right now
+    if args.split_bn:
+        assert args.num_aug_splits > 1 or args.resplit
+        model = convert_splitbn_model(model, max(args.num_aug_splits, 2))
+
+    return model
+
+
+def data_creator(config):
+    # torch.manual_seed(args.seed + torch.distributed.get_rank())
+
+    args = config["args"]
+
+    # todo: verbose should depend on rank
+    data_config = resolve_data_config(vars(args), verbose=True)
+
+    dataset_train = Dataset(join(args.data, "train"))
+    dataset_eval = Dataset(join(args.data, "val"))
+
+    collate_fn = None
+    if args.prefetcher and args.mixup > 0:
+        # collate conflict (need to support deinterleaving in collate mixup)
+        assert args.num_aug_splits == 0
+        collate_fn = FastCollateMixup(args.mixup, args.smoothing,
+                                      args.num_classes)
+
+    common_params = dict(
+        input_size=data_config["input_size"],
+        use_prefetcher=args.prefetcher,
+        mean=data_config["mean"],
+        std=data_config["std"],
+        num_workers=1,
+        distributed=args.distributed,
+        pin_memory=args.pin_mem)
+
+    train_loader = create_loader(
+        dataset_train,
+        is_training=True,
+        batch_size=config[BATCH_SIZE],
+        re_prob=args.reprob,
+        re_mode=args.remode,
+        re_count=args.recount,
+        re_split=args.resplit,
+        collate_fn=collate_fn,
+        color_jitter=args.color_jitter,
+        auto_augment=args.aa,
+        interpolation=args.train_interpolation,
+        num_aug_splits=args.num_aug_splits,  # always 0 right now
+        **common_params)
+    eval_loader = create_loader(
+        dataset_eval,
+        is_training=False,
+        batch_size=args.validation_batch_size_multiplier * config[BATCH_SIZE],
+        interpolation=data_config["interpolation"],
+        crop_pct=data_config["crop_pct"],
+        **common_params)
+
+    return train_loader, eval_loader
+
+
+def optimizer_creator(model, config):
+    args = config["args"]
+    return create_optimizer(args, model)
+
+
+def loss_creator(config):
+    # there should be more complicated logic here, but we don't support
+    # separate train and eval losses yet
+    return nn.CrossEntropyLoss()
+
+
+def main():
+    setup_default_logging()
+
+    args, args_text = parse_args()
+
+    ray.init(address=args.ray_address)
+
+    trainer = TorchTrainer(
+        model_creator=model_creator,
+        data_creator=data_creator,
+        optimizer_creator=optimizer_creator,
+        loss_creator=loss_creator,
+        use_tqdm=True,
+        use_fp16=args.amp,
+        apex_args={"opt_level": "O1"},
+        config={
+            "args": args,
+            BATCH_SIZE: args.batch_size
+        },
+        num_workers=args.ray_num_workers)
+
+    pbar = trange(args.epochs, unit="epoch")
+    for i in pbar:
+        trainer.train()
+
+        val_stats = trainer.validate()
+        pbar.set_postfix(dict(acc=val_stats["val_accuracy"]))
+
+    trainer.shutdown()
+
+
+if __name__ == "__main__":
+    main()