Merge pull request #533 from rwightman/pit_and_vit_update

Addition of PiT models and update/cleanup of ViT, new NFNet weight, TFDS wrapper fix, few misc fixes/updates

README.md

@@ -23,6 +23,22 @@ I'm fortunate to be able to dedicate significant time and money of my own suppor
 
 ## What's New
 
+### April 1, 2021
+* Add snazzy `benchmark.py` script for bulk `timm` model benchmarking of train and/or inference
+* Add Pooling-based Vision Transformer (PiT) models (from https://github.com/naver-ai/pit)
+  * Merged distilled variant into main for torchscript compatibility
+  * Some `timm` cleanup/style tweaks and weights have hub download support
+* Cleanup Vision Transformer (ViT) models
+  * Merge distilled (DeiT) model into main so that torchscript can work
+  * Support updated weight init (defaults to old still) that closer matches original JAX impl (possibly better training from scratch)
+  * Separate hybrid model defs into different file and add several new model defs to fiddle with, support patch_size != 1 for hybrids
+  * Fix fine-tuning num_class changes (PiT and ViT) and pos_embed resizing (Vit) with distilled variants
+  * nn.Sequential for block stack (does not break downstream compat)
+* TnT (Transformer-in-Transformer) models contributed by author (from https://gitee.com/mindspore/mindspore/tree/master/model_zoo/research/cv/TNT)
+* Add RegNetY-160 weights from DeiT teacher model
+* Add new NFNet-L0 w/ SE attn (rename `nfnet_l0b`->`nfnet_l0`) weights 82.75 top-1 @ 288x288
+* Some fixes/improvements for TFDS dataset wrapper
+
 ### March 17, 2021
 * Add new ECA-NFNet-L0 (rename `nfnet_l0c`->`eca_nfnet_l0`) weights trained by myself.
   * 82.6 top-1 @ 288x288, 82.8 @ 320x320, trained at 224x224
@@ -189,6 +205,7 @@ A full version of the list below with source links can be found in the [document
 * NFNet-F - https://arxiv.org/abs/2102.06171
 * NF-RegNet / NF-ResNet - https://arxiv.org/abs/2101.08692
 * PNasNet - https://arxiv.org/abs/1712.00559
+* Pooling-based Vision Transformer (PiT) - https://arxiv.org/abs/2103.16302
 * RegNet - https://arxiv.org/abs/2003.13678
 * RepVGG - https://arxiv.org/abs/2101.03697
 * ResNet/ResNeXt
@@ -204,6 +221,7 @@ A full version of the list below with source links can be found in the [document
 * ReXNet - https://arxiv.org/abs/2007.00992
 * SelecSLS - https://arxiv.org/abs/1907.00837
 * Selective Kernel Networks - https://arxiv.org/abs/1903.06586
+* Transformer-iN-Transformer (TNT) - https://arxiv.org/abs/2103.00112
 * TResNet - https://arxiv.org/abs/2003.13630
 * Vision Transformer - https://arxiv.org/abs/2010.11929
 * VovNet V2 and V1 - https://arxiv.org/abs/1911.06667

benchmark.py

@@ -0,0 +1,481 @@
+#!/usr/bin/env python3
+""" Model Benchmark Script
+
+An inference and train step benchmark script for timm models.
+
+Hacked together by Ross Wightman (https://github.com/rwightman)
+"""
+import argparse
+import os
+import csv
+import json
+import time
+import logging
+import torch
+import torch.nn as nn
+import torch.nn.parallel
+from collections import OrderedDict
+from contextlib import suppress
+from functools import partial
+
+from timm.models import create_model, is_model, list_models
+from timm.optim import create_optimizer_v2
+from timm.data import resolve_data_config
+from timm.utils import AverageMeter, setup_default_logging
+
+
+has_apex = False
+try:
+    from apex import amp
+    has_apex = True
+except ImportError:
+    pass
+
+has_native_amp = False
+try:
+    if getattr(torch.cuda.amp, 'autocast') is not None:
+        has_native_amp = True
+except AttributeError:
+    pass
+
+torch.backends.cudnn.benchmark = True
+_logger = logging.getLogger('validate')
+
+
+parser = argparse.ArgumentParser(description='PyTorch Benchmark')
+
+# benchmark specific args
+parser.add_argument('--model-list', metavar='NAME', default='',
+                    help='txt file based list of model names to benchmark')
+parser.add_argument('--bench', default='both', type=str,
+                    help="Benchmark mode. One of 'inference', 'train', 'both'. Defaults to 'inference'")
+parser.add_argument('--detail', action='store_true', default=False,
+                    help='Provide train fwd/bwd/opt breakdown detail if True. Defaults to False')
+parser.add_argument('--results-file', default='', type=str, metavar='FILENAME',
+                    help='Output csv file for validation results (summary)')
+parser.add_argument('--num-warm-iter', default=10, type=int,
+                    metavar='N', help='Number of warmup iterations (default: 10)')
+parser.add_argument('--num-bench-iter', default=40, type=int,
+                    metavar='N', help='Number of benchmark iterations (default: 40)')
+
+# common inference / train args
+parser.add_argument('--model', '-m', metavar='NAME', default='resnet50',
+                    help='model architecture (default: resnet50)')
+parser.add_argument('-b', '--batch-size', default=256, type=int,
+                    metavar='N', help='mini-batch size (default: 256)')
+parser.add_argument('--img-size', default=None, type=int,
+                    metavar='N', help='Input image dimension, uses model default if empty')
+parser.add_argument('--input-size', default=None, nargs=3, type=int,
+                    metavar='N N N', help='Input all image dimensions (d h w, e.g. --input-size 3 224 224), uses model default if empty')
+parser.add_argument('--num-classes', type=int, default=None,
+                    help='Number classes in dataset')
+parser.add_argument('--gp', default=None, type=str, metavar='POOL',
+                    help='Global pool type, one of (fast, avg, max, avgmax, avgmaxc). Model default if None.')
+parser.add_argument('--channels-last', action='store_true', default=False,
+                    help='Use channels_last memory layout')
+parser.add_argument('--amp', action='store_true', default=False,
+                    help='use PyTorch Native AMP for mixed precision training. Overrides --precision arg.')
+parser.add_argument('--precision', default='float32', type=str,
+                    help='Numeric precision. One of (amp, float32, float16, bfloat16, tf32)')
+parser.add_argument('--torchscript', dest='torchscript', action='store_true',
+                    help='convert model torchscript for inference')
+
+
+# train optimizer parameters
+parser.add_argument('--opt', default='sgd', type=str, metavar='OPTIMIZER',
+                    help='Optimizer (default: "sgd"')
+parser.add_argument('--opt-eps', default=None, type=float, metavar='EPSILON',
+                    help='Optimizer Epsilon (default: None, use opt default)')
+parser.add_argument('--opt-betas', default=None, type=float, nargs='+', metavar='BETA',
+                    help='Optimizer Betas (default: None, use opt default)')
+parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
+                    help='Optimizer momentum (default: 0.9)')
+parser.add_argument('--weight-decay', type=float, default=0.0001,
+                    help='weight decay (default: 0.0001)')
+parser.add_argument('--clip-grad', type=float, default=None, metavar='NORM',
+                    help='Clip gradient norm (default: None, no clipping)')
+parser.add_argument('--clip-mode', type=str, default='norm',
+                    help='Gradient clipping mode. One of ("norm", "value", "agc")')
+
+
+# model regularization / loss params that impact model or loss fn
+parser.add_argument('--smoothing', type=float, default=0.1,
+                    help='Label smoothing (default: 0.1)')
+parser.add_argument('--drop', type=float, default=0.0, metavar='PCT',
+                    help='Dropout rate (default: 0.)')
+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)')
+
+
+def timestamp(sync=False):
+    return time.perf_counter()
+
+
+def cuda_timestamp(sync=False, device=None):
+    if sync:
+        torch.cuda.synchronize(device=device)
+    return time.perf_counter()
+
+
+def count_params(model: nn.Module):
+    return sum([m.numel() for m in model.parameters()])
+
+
+def resolve_precision(precision: str):
+    assert precision in ('amp', 'float16', 'bfloat16', 'float32')
+    use_amp = False
+    model_dtype = torch.float32
+    data_dtype = torch.float32
+    if precision == 'amp':
+        use_amp = True
+    elif precision == 'float16':
+        model_dtype = torch.float16
+        data_dtype = torch.float16
+    elif precision == 'bfloat16':
+        model_dtype = torch.bfloat16
+        data_dtype = torch.bfloat16
+    return use_amp, model_dtype, data_dtype
+
+
+class BenchmarkRunner:
+    def __init__(
+            self, model_name, detail=False, device='cuda', torchscript=False, precision='float32',
+            num_warm_iter=10, num_bench_iter=50, **kwargs):
+        self.model_name = model_name
+        self.detail = detail
+        self.device = device
+        self.use_amp, self.model_dtype, self.data_dtype = resolve_precision(precision)
+        self.channels_last = kwargs.pop('channels_last', False)
+        self.amp_autocast = torch.cuda.amp.autocast if self.use_amp else suppress
+
+        self.model = create_model(
+            model_name,
+            num_classes=kwargs.pop('num_classes', None),
+            in_chans=3,
+            global_pool=kwargs.pop('gp', 'fast'),
+            scriptable=torchscript)
+        self.model.to(
+            device=self.device,
+            dtype=self.model_dtype,
+            memory_format=torch.channels_last if self.channels_last else None)
+        self.num_classes = self.model.num_classes
+        self.param_count = count_params(self.model)
+        _logger.info('Model %s created, param count: %d' % (model_name, self.param_count))
+        if torchscript:
+            self.model = torch.jit.script(self.model)
+
+        data_config = resolve_data_config(kwargs, model=self.model, use_test_size=True)
+        self.input_size = data_config['input_size']
+        self.batch_size = kwargs.pop('batch_size', 256)
+
+        self.example_inputs = None
+        self.num_warm_iter = num_warm_iter
+        self.num_bench_iter = num_bench_iter
+        self.log_freq = num_bench_iter // 5
+        if 'cuda' in self.device:
+            self.time_fn = partial(cuda_timestamp, device=self.device)
+        else:
+            self.time_fn = timestamp
+
+    def _init_input(self):
+        self.example_inputs = torch.randn(
+            (self.batch_size,) + self.input_size, device=self.device, dtype=self.data_dtype)
+        if self.channels_last:
+            self.example_inputs = self.example_inputs.contiguous(memory_format=torch.channels_last)
+
+
+class InferenceBenchmarkRunner(BenchmarkRunner):
+
+    def __init__(self, model_name, device='cuda', torchscript=False, **kwargs):
+        super().__init__(model_name=model_name, device=device, torchscript=torchscript, **kwargs)
+        self.model.eval()
+
+    def run(self):
+        def _step():
+            t_step_start = self.time_fn()
+            with self.amp_autocast():
+                output = self.model(self.example_inputs)
+            t_step_end = self.time_fn(True)
+            return t_step_end - t_step_start
+
+        _logger.info(
+            f'Running inference benchmark on {self.model_name} for {self.num_bench_iter} steps w/ '
+            f'input size {self.input_size} and batch size {self.batch_size}.')
+
+        with torch.no_grad():
+            self._init_input()
+
+            for _ in range(self.num_warm_iter):
+                _step()
+
+            total_step = 0.
+            num_samples = 0
+            t_run_start = self.time_fn()
+            for i in range(self.num_bench_iter):
+                delta_fwd = _step()
+                total_step += delta_fwd
+                num_samples += self.batch_size
+                num_steps = i + 1
+                if num_steps % self.log_freq == 0:
+                    _logger.info(
+                        f"Infer [{num_steps}/{self.num_bench_iter}]."
+                        f" {num_samples / total_step:0.2f} samples/sec."
+                        f" {1000 * total_step / num_steps:0.3f} ms/step.")
+            t_run_end = self.time_fn(True)
+            t_run_elapsed = t_run_end - t_run_start
+
+        results = dict(
+            samples_per_sec=round(num_samples / t_run_elapsed, 2),
+            step_time=round(1000 * total_step / self.num_bench_iter, 3),
+            batch_size=self.batch_size,
+            img_size=self.input_size[-1],
+            param_count=round(self.param_count / 1e6, 2),
+        )
+
+        _logger.info(
+            f"Inference benchmark of {self.model_name} done. "
+            f"{results['samples_per_sec']:.2f} samples/sec, {results['step_time']:.2f} ms/step")
+
+        return results
+
+
+class TrainBenchmarkRunner(BenchmarkRunner):
+
+    def __init__(self, model_name, device='cuda', torchscript=False, **kwargs):
+        super().__init__(model_name=model_name, device=device, torchscript=torchscript, **kwargs)
+        self.model.train()
+
+        if kwargs.pop('smoothing', 0) > 0:
+            self.loss = nn.CrossEntropyLoss().to(self.device)
+        else:
+            self.loss = nn.CrossEntropyLoss().to(self.device)
+        self.target_shape = tuple()
+
+        self.optimizer = create_optimizer_v2(
+            self.model,
+            optimizer_name=kwargs.pop('opt', 'sgd'),
+            learning_rate=kwargs.pop('lr', 1e-4))
+
+    def _gen_target(self, batch_size):
+        return torch.empty(
+            (batch_size,) + self.target_shape, device=self.device, dtype=torch.long).random_(self.num_classes)
+
+    def run(self):
+        def _step(detail=False):
+            self.optimizer.zero_grad()  # can this be ignored?
+            t_start = self.time_fn()
+            t_fwd_end = t_start
+            t_bwd_end = t_start
+            with self.amp_autocast():
+                output = self.model(self.example_inputs)
+                if isinstance(output, tuple):
+                    output = output[0]
+                if detail:
+                    t_fwd_end = self.time_fn(True)
+                target = self._gen_target(output.shape[0])
+                self.loss(output, target).backward()
+                if detail:
+                    t_bwd_end = self.time_fn(True)
+            self.optimizer.step()
+            t_end = self.time_fn(True)
+            if detail:
+                delta_fwd = t_fwd_end - t_start
+                delta_bwd = t_bwd_end - t_fwd_end
+                delta_opt = t_end - t_bwd_end
+                return delta_fwd, delta_bwd, delta_opt
+            else:
+                delta_step = t_end - t_start
+                return delta_step
+
+        _logger.info(
+            f'Running train benchmark on {self.model_name} for {self.num_bench_iter} steps w/ '
+            f'input size {self.input_size} and batch size {self.batch_size}.')
+
+        self._init_input()
+
+        for _ in range(self.num_warm_iter):
+            _step()
+
+        t_run_start = self.time_fn()
+        if self.detail:
+            total_fwd = 0.
+            total_bwd = 0.
+            total_opt = 0.
+            num_samples = 0
+            for i in range(self.num_bench_iter):
+                delta_fwd, delta_bwd, delta_opt = _step(True)
+                num_samples += self.batch_size
+                total_fwd += delta_fwd
+                total_bwd += delta_bwd
+                total_opt += delta_opt
+                num_steps = (i + 1)
+                if num_steps % self.log_freq == 0:
+                    total_step = total_fwd + total_bwd + total_opt
+                    _logger.info(
+                        f"Train [{num_steps}/{self.num_bench_iter}]."
+                        f" {num_samples / total_step:0.2f} samples/sec."
+                        f" {1000 * total_fwd / num_steps:0.3f} ms/step fwd,"
+                        f" {1000 * total_bwd / num_steps:0.3f} ms/step bwd,"
+                        f" {1000 * total_opt / num_steps:0.3f} ms/step opt."
+                    )
+            total_step = total_fwd + total_bwd + total_opt
+            t_run_elapsed = self.time_fn() - t_run_start
+            results = dict(
+                samples_per_sec=round(num_samples / t_run_elapsed, 2),
+                step_time=round(1000 * total_step / self.num_bench_iter, 3),
+                fwd_time=round(1000 * total_fwd / self.num_bench_iter, 3),
+                bwd_time=round(1000 * total_bwd / self.num_bench_iter, 3),
+                opt_time=round(1000 * total_opt / self.num_bench_iter, 3),
+                batch_size=self.batch_size,
+                img_size=self.input_size[-1],
+                param_count=round(self.param_count / 1e6, 2),
+            )
+        else:
+            total_step = 0.
+            num_samples = 0
+            for i in range(self.num_bench_iter):
+                delta_step = _step(False)
+                num_samples += self.batch_size
+                total_step += delta_step
+                num_steps = (i + 1)
+                if num_steps % self.log_freq == 0:
+                    _logger.info(
+                        f"Train [{num_steps}/{self.num_bench_iter}]."
+                        f" {num_samples / total_step:0.2f} samples/sec."
+                        f" {1000 * total_step / num_steps:0.3f} ms/step.")
+            t_run_elapsed = self.time_fn() - t_run_start
+            results = dict(
+                samples_per_sec=round(num_samples / t_run_elapsed, 2),
+                step_time=round(1000 * total_step / self.num_bench_iter, 3),
+                batch_size=self.batch_size,
+                img_size=self.input_size[-1],
+                param_count=round(self.param_count / 1e6, 2),
+            )
+
+        _logger.info(
+            f"Train benchmark of {self.model_name} done. "
+            f"{results['samples_per_sec']:.2f} samples/sec, {results['step_time']:.2f} ms/sample")
+
+        return results
+
+
+def decay_batch_exp(batch_size, factor=0.5, divisor=16):
+    out_batch_size = batch_size * factor
+    if out_batch_size > divisor:
+        out_batch_size = (out_batch_size + 1) // divisor * divisor
+    else:
+        out_batch_size = batch_size - 1
+    return max(0, int(out_batch_size))
+
+
+def _try_run(model_name, bench_fn, initial_batch_size, bench_kwargs):
+    batch_size = initial_batch_size
+    results = dict()
+    while batch_size >= 1:
+        try:
+            bench = bench_fn(model_name=model_name, batch_size=batch_size, **bench_kwargs)
+            results = bench.run()
+            return results
+        except RuntimeError as e:
+            torch.cuda.empty_cache()
+            batch_size = decay_batch_exp(batch_size)
+            print(f'Error: {str(e)} while running benchmark. Reducing batch size to {batch_size} for retry.')
+    return results
+
+
+def benchmark(args):
+    if args.amp:
+        _logger.warning("Overriding precision to 'amp' since --amp flag set.")
+        args.precision = 'amp'
+    _logger.info(f'Benchmarking in {args.precision} precision. '
+                 f'{"NHWC" if args.channels_last else "NCHW"} layout. '
+                 f'torchscript {"enabled" if args.torchscript else "disabled"}')
+
+    bench_kwargs = vars(args).copy()
+    bench_kwargs.pop('amp')
+    model = bench_kwargs.pop('model')
+    batch_size = bench_kwargs.pop('batch_size')
+
+    bench_fns = (InferenceBenchmarkRunner,)
+    prefixes = ('infer',)
+    if args.bench == 'both':
+        bench_fns = (
+            InferenceBenchmarkRunner,
+            TrainBenchmarkRunner
+        )
+        prefixes = ('infer', 'train')
+    elif args.bench == 'train':
+        bench_fns = TrainBenchmarkRunner,
+        prefixes = 'train',
+
+    model_results = OrderedDict(model=model)
+    for prefix, bench_fn in zip(prefixes, bench_fns):
+        run_results = _try_run(model, bench_fn, initial_batch_size=batch_size, bench_kwargs=bench_kwargs)
+        if prefix:
+            run_results = {'_'.join([prefix, k]): v for k, v in run_results.items()}
+        model_results.update(run_results)
+    param_count = model_results.pop('infer_param_count', model_results.pop('train_param_count', 0))
+    model_results.setdefault('param_count', param_count)
+    model_results.pop('train_param_count', 0)
+    return model_results
+
+
+def main():
+    setup_default_logging()
+    args = parser.parse_args()
+    model_cfgs = []
+    model_names = []
+
+    if args.model_list:
+        args.model = ''
+        with open(args.model_list) as f:
+            model_names = [line.rstrip() for line in f]
+        model_cfgs = [(n, None) for n in model_names]
+    elif args.model == 'all':
+        # validate all models in a list of names with pretrained checkpoints
+        args.pretrained = True
+        model_names = list_models(pretrained=True, exclude_filters=['*in21k'])
+        model_cfgs = [(n, None) for n in model_names]
+    elif not is_model(args.model):
+        # model name doesn't exist, try as wildcard filter
+        model_names = list_models(args.model)
+        model_cfgs = [(n, None) for n in model_names]
+
+    if len(model_cfgs):
+        results_file = args.results_file or './benchmark.csv'
+        _logger.info('Running bulk validation on these pretrained models: {}'.format(', '.join(model_names)))
+        results = []
+        try:
+            for m, _ in model_cfgs:
+                if not m:
+                    continue
+                args.model = m
+                r = benchmark(args)
+                results.append(r)
+        except KeyboardInterrupt as e:
+            pass
+        sort_key = 'train_samples_per_sec' if 'train' in args.bench else 'infer_samples_per_sec'
+        results = sorted(results, key=lambda x: x[sort_key], reverse=True)
+        if len(results):
+            write_results(results_file, results)
+
+        import json
+        json_str = json.dumps(results, indent=4)
+        print(json_str)
+    else:
+        benchmark(args)
+
+
+def write_results(results_file, results):
+    with open(results_file, mode='w') as cf:
+        dw = csv.DictWriter(cf, fieldnames=results[0].keys())
+        dw.writeheader()
+        for r in results:
+            dw.writerow(r)
+        cf.flush()
+
+
+if __name__ == '__main__':
+    main()

tests/test_models.py

@@ -14,7 +14,7 @@ if hasattr(torch._C, '_jit_set_profiling_executor'):
     torch._C._jit_set_profiling_mode(False)
 
 # transformer models don't support many of the spatial / feature based model functionalities
-NON_STD_FILTERS = ['vit_*', 'tnt_*']
+NON_STD_FILTERS = ['vit_*', 'tnt_*', 'pit_*']
 NUM_NON_STD = len(NON_STD_FILTERS)
 
 # exclude models that cause specific test failures

timm/data/config.py

@@ -5,7 +5,7 @@ from .constants import *
 _logger = logging.getLogger(__name__)
 
 
-def resolve_data_config(args, default_cfg={}, model=None, use_test_size=False, verbose=True):
+def resolve_data_config(args, default_cfg={}, model=None, use_test_size=False, verbose=False):
     new_config = {}
     default_cfg = default_cfg
     if not default_cfg and model is not None and hasattr(model, 'default_cfg'):

timm/data/dataset.py

@@ -73,12 +73,13 @@ class IterableImageDataset(data.IterableDataset):
             batch_size=None,
             class_map='',
             load_bytes=False,
+            repeats=0,
             transform=None,
     ):
         assert parser is not None
         if isinstance(parser, str):
             self.parser = create_parser(
-                parser, root=root, split=split, is_training=is_training, batch_size=batch_size)
+                parser, root=root, split=split, is_training=is_training, batch_size=batch_size, repeats=repeats)
         else:
             self.parser = parser
         self.transform = transform

timm/data/dataset_factory.py

@@ -23,6 +23,7 @@ def create_dataset(name, root, split='validation', search_split=True, is_trainin
             root, parser=name, split=split, is_training=is_training, batch_size=batch_size, **kwargs)
     else:
         # FIXME support more advance split cfg for ImageFolder/Tar datasets in the future
+        kwargs.pop('repeats', 0)  # FIXME currently only Iterable dataset support the repeat multiplier
         if search_split and os.path.isdir(root):
             root = _search_split(root, split)
         ds = ImageDataset(root, parser=name, **kwargs)

timm/data/parsers/parser_tfds.py

@@ -29,6 +29,11 @@ SHUFFLE_SIZE = 16834  # samples to shuffle in DS queue
 PREFETCH_SIZE = 4096  # samples to prefetch
 
 
+def even_split_indices(split, n, num_samples):
+    partitions = [round(i * num_samples / n) for i in range(n + 1)]
+    return [f"{split}[{partitions[i]}:{partitions[i+1]}]" for i in range(n)]
+
+
 class ParserTfds(Parser):
     """ Wrap Tensorflow Datasets for use in PyTorch
 
@@ -52,7 +57,7 @@ class ParserTfds(Parser):
         components.
 
     """
-    def __init__(self, root, name, split='train', shuffle=False, is_training=False, batch_size=None):
+    def __init__(self, root, name, split='train', shuffle=False, is_training=False, batch_size=None, repeats=0):
         super().__init__()
         self.root = root
         self.split = split
@@ -62,6 +67,8 @@ class ParserTfds(Parser):
             assert batch_size is not None,\
                 "Must specify batch_size in training mode for reasonable behaviour w/ TFDS wrapper"
         self.batch_size = batch_size
+        self.repeats = repeats
+        self.subsplit = None
 
         self.builder = tfds.builder(name, data_dir=root)
         # NOTE: please use tfds command line app to download & prepare datasets, I don't want to call
@@ -95,6 +102,7 @@ class ParserTfds(Parser):
         if worker_info is not None:
             self.worker_info = worker_info
             num_workers = worker_info.num_workers
+            global_num_workers = self.dist_num_replicas * num_workers
             worker_id = worker_info.id
 
             # FIXME I need to spend more time figuring out the best way to distribute/split data across
@@ -114,19 +122,31 @@ class ParserTfds(Parser):
             #     split = split + '[{}:]'.format(start)
             # else:
             #     split = split + '[{}:{}]'.format(start, start + split_size)
-
-        input_context = tf.distribute.InputContext(
-            num_input_pipelines=self.dist_num_replicas * num_workers,
-            input_pipeline_id=self.dist_rank * num_workers + worker_id,
-            num_replicas_in_sync=self.dist_num_replicas  # FIXME does this have any impact?
-        )
-
-        read_config = tfds.ReadConfig(input_context=input_context)
-        ds = self.builder.as_dataset(split=split, shuffle_files=self.shuffle, read_config=read_config)
+            if not self.is_training and '[' not in self.split:
+                # If not training, and split doesn't define a subsplit, manually split the dataset
+                # for more even samples / worker
+                self.subsplit = even_split_indices(self.split, global_num_workers, self.num_samples)[
+                    self.dist_rank * num_workers + worker_id]
+
+        if self.subsplit is None:
+            input_context = tf.distribute.InputContext(
+                num_input_pipelines=self.dist_num_replicas * num_workers,
+                input_pipeline_id=self.dist_rank * num_workers + worker_id,
+                num_replicas_in_sync=self.dist_num_replicas  # FIXME does this arg have any impact?
+            )
+        else:
+            input_context = None
+
+        read_config = tfds.ReadConfig(
+            shuffle_seed=42,
+            shuffle_reshuffle_each_iteration=True,
+            input_context=input_context)
+        ds = self.builder.as_dataset(
+            split=self.subsplit or self.split, shuffle_files=self.shuffle, read_config=read_config)
         # avoid overloading threading w/ combo fo TF ds threads + PyTorch workers
         ds.options().experimental_threading.private_threadpool_size = max(1, MAX_TP_SIZE // num_workers)
         ds.options().experimental_threading.max_intra_op_parallelism = 1
-        if self.is_training:
+        if self.is_training or self.repeats > 1:
             # to prevent excessive drop_last batch behaviour w/ IterableDatasets
             # see warnings at https://pytorch.org/docs/stable/data.html#multi-process-data-loading
             ds = ds.repeat()  # allow wrap around and break iteration manually
@@ -143,7 +163,7 @@ class ParserTfds(Parser):
         #     This adds extra samples and will slightly alter validation results.
         #   2. determine loop ending condition in training w/ repeat enabled so that only full batch_size
         #     batches are produced (underlying tfds iter wraps around)
-        target_sample_count = math.ceil(self.num_samples / self._num_pipelines)
+        target_sample_count = math.ceil(max(1, self.repeats) * self.num_samples / self._num_pipelines)
         if self.is_training:
             # round up to nearest batch_size per worker-replica
             target_sample_count = math.ceil(target_sample_count / self.batch_size) * self.batch_size
@@ -160,8 +180,8 @@ class ParserTfds(Parser):
         if not self.is_training and self.dist_num_replicas and 0 < sample_count < target_sample_count:
             # Validation batch padding only done for distributed training where results are reduced across nodes.
             # For single process case, it won't matter if workers return different batch sizes.
-            # FIXME this needs more testing, possible for sharding / split api to cause differences of > 1?
-            assert target_sample_count - sample_count == 1  # should only be off by 1 or sharding is not optimal
+            # FIXME if using input_context or % based subsplits, sample count can vary by more than +/- 1 and this
+            # approach is not optimal
             yield img, sample['label']  # yield prev sample again
             sample_count += 1
 
@@ -176,7 +196,7 @@ class ParserTfds(Parser):
     def __len__(self):
         # this is just an estimate and does not factor in extra samples added to pad batches based on
         # complete worker & replica info (not available until init in dataloader).
-        return math.ceil(self.num_samples / self.dist_num_replicas)
+        return math.ceil(max(1, self.repeats) * self.num_samples / self.dist_num_replicas)
 
     def _filename(self, index, basename=False, absolute=False):
         assert False, "Not supported" # no random access to samples

timm/models/__init__.py

@@ -14,6 +14,7 @@ from .inception_v4 import *
 from .mobilenetv3 import *
 from .nasnet import *
 from .nfnet import *
+from .pit import *
 from .pnasnet import *
 from .regnet import *
 from .res2net import *
@@ -28,6 +29,7 @@ from .tnt import *
 from .tresnet import *
 from .vgg import *
 from .vision_transformer import *
+from .vision_transformer_hybrid import *
 from .vovnet import *
 from .xception import *
 from .xception_aligned import *

timm/models/helpers.py

@@ -198,20 +198,24 @@ def load_pretrained(model, default_cfg=None, num_classes=1000, in_chans=3, filte
                 _logger.warning(
                     f'Unable to convert pretrained {input_conv_name} weights, using random init for this layer.')
 
-    classifier_name = default_cfg.get('classifier', None)
+    classifiers = default_cfg.get('classifier', None)
     label_offset = default_cfg.get('label_offset', 0)
-    if classifier_name is not None:
+    if classifiers is not None:
+        if isinstance(classifiers, str):
+            classifiers = (classifiers,)
         if num_classes != default_cfg['num_classes']:
-            # completely discard fully connected if model num_classes doesn't match pretrained weights
-            del state_dict[classifier_name + '.weight']
-            del state_dict[classifier_name + '.bias']
+            for classifier_name in classifiers:
+                # completely discard fully connected if model num_classes doesn't match pretrained weights
+                del state_dict[classifier_name + '.weight']
+                del state_dict[classifier_name + '.bias']
             strict = False
         elif label_offset > 0:
-            # special case for pretrained weights with an extra background class in pretrained weights
-            classifier_weight = state_dict[classifier_name + '.weight']
-            state_dict[classifier_name + '.weight'] = classifier_weight[label_offset:]
-            classifier_bias = state_dict[classifier_name + '.bias']
-            state_dict[classifier_name + '.bias'] = classifier_bias[label_offset:]
+            for classifier_name in classifiers:
+                # special case for pretrained weights with an extra background class in pretrained weights
+                classifier_weight = state_dict[classifier_name + '.weight']
+                state_dict[classifier_name + '.weight'] = classifier_weight[label_offset:]
+                classifier_bias = state_dict[classifier_name + '.bias']
+                state_dict[classifier_name + '.bias'] = classifier_bias[label_offset:]
 
     model.load_state_dict(state_dict, strict=strict)
 

timm/models/layers/__init__.py

@@ -31,4 +31,4 @@ from .split_attn import SplitAttnConv2d
 from .split_batchnorm import SplitBatchNorm2d, convert_splitbn_model
 from .std_conv import StdConv2d, StdConv2dSame, ScaledStdConv2d, ScaledStdConv2dSame
 from .test_time_pool import TestTimePoolHead, apply_test_time_pool
-from .weight_init import trunc_normal_
+from .weight_init import trunc_normal_, variance_scaling_, lecun_normal_

timm/models/layers/weight_init.py

@@ -2,6 +2,8 @@ import torch
 import math
 import warnings
 
+from torch.nn.init import _calculate_fan_in_and_fan_out
+
 
 def _no_grad_trunc_normal_(tensor, mean, std, a, b):
     # Cut & paste from PyTorch official master until it's in a few official releases - RW
@@ -58,3 +60,30 @@ def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
         >>> nn.init.trunc_normal_(w)
     """
     return _no_grad_trunc_normal_(tensor, mean, std, a, b)
+
+
+def variance_scaling_(tensor, scale=1.0, mode='fan_in', distribution='normal'):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == 'fan_in':
+        denom = fan_in
+    elif mode == 'fan_out':
+        denom = fan_out
+    elif mode == 'fan_avg':
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_(tensor, std=math.sqrt(variance) / .87962566103423978)
+    elif distribution == "normal":
+        tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode='fan_in', distribution='truncated_normal')

timm/models/nfnet.py

@@ -100,14 +100,16 @@ default_cfgs = dict(
     nfnet_f7s=_dcfg(
         url='', pool_size=(15, 15), input_size=(3, 480, 480), test_input_size=(3, 608, 608)),
 
-    nfnet_l0a=_dcfg(
-        url='', pool_size=(7, 7), input_size=(3, 224, 224), test_input_size=(3, 288, 288)),
-    nfnet_l0b=_dcfg(
-        url='', pool_size=(7, 7), input_size=(3, 224, 224), test_input_size=(3, 288, 288)),
+    nfnet_l0=_dcfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/nfnet_l0_ra2-45c6688d.pth',
+        pool_size=(7, 7), input_size=(3, 224, 224), test_input_size=(3, 288, 288), crop_pct=1.0),
     eca_nfnet_l0=_dcfg(
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/ecanfnet_l0_ra2-e3e9ac50.pth',
         hf_hub='timm/eca_nfnet_l0',
         pool_size=(7, 7), input_size=(3, 224, 224), test_input_size=(3, 288, 288), crop_pct=1.0),
+    eca_nfnet_l1=_dcfg(
+        url='',
+        pool_size=(8, 8), input_size=(3, 256, 256), test_input_size=(3, 320, 320), crop_pct=1.0),
 
     nf_regnet_b0=_dcfg(
         url='', pool_size=(6, 6), input_size=(3, 192, 192), test_input_size=(3, 256, 256), first_conv='stem.conv'),
@@ -232,15 +234,15 @@ model_cfgs = dict(
     nfnet_f6s=_nfnet_cfg(depths=(7, 14, 42, 21), act_layer='silu'),
     nfnet_f7s=_nfnet_cfg(depths=(8, 16, 48, 24), act_layer='silu'),
 
-    # Experimental 'light' versions of nfnet-f that are little leaner
-    nfnet_l0a=_nfnet_cfg(
-        depths=(1, 2, 6, 3), channels=(256, 512, 1280, 1536), feat_mult=1.5, group_size=64, bottle_ratio=0.25,
-        attn_kwargs=dict(reduction_ratio=0.25, divisor=8), act_layer='silu'),
-    nfnet_l0b=_nfnet_cfg(
-        depths=(1, 2, 6, 3), channels=(256, 512, 1536, 1536), feat_mult=1.5, group_size=64, bottle_ratio=0.25,
+    # Experimental 'light' versions of NFNet-F that are little leaner
+    nfnet_l0=_nfnet_cfg(
+        depths=(1, 2, 6, 3), feat_mult=1.5, group_size=64, bottle_ratio=0.25,
         attn_kwargs=dict(reduction_ratio=0.25, divisor=8), act_layer='silu'),
     eca_nfnet_l0=_nfnet_cfg(
-        depths=(1, 2, 6, 3), channels=(256, 512, 1536, 1536), feat_mult=1.5, group_size=64, bottle_ratio=0.25,
+        depths=(1, 2, 6, 3), feat_mult=1.5, group_size=64, bottle_ratio=0.25,
+        attn_layer='eca', attn_kwargs=dict(), act_layer='silu'),
+    eca_nfnet_l1=_nfnet_cfg(
+        depths=(2, 4, 12, 6), feat_mult=2, group_size=64, bottle_ratio=0.25,
         attn_layer='eca', attn_kwargs=dict(), act_layer='silu'),
 
     # EffNet influenced RegNet defs.
@@ -789,29 +791,29 @@ def nfnet_f7s(pretrained=False, **kwargs):
 
 
 @register_model
-def nfnet_l0a(pretrained=False, **kwargs):
-    """ NFNet-L0a w/ SiLU
-    My experimental 'light' model w/ 1280 width stage 3, 1.5x final_conv mult, 64 group_size, .25 bottleneck & SE ratio
-    """
-    return _create_normfreenet('nfnet_l0a', pretrained=pretrained, **kwargs)
-
-
-@register_model
-def nfnet_l0b(pretrained=False, **kwargs):
+def nfnet_l0(pretrained=False, **kwargs):
     """ NFNet-L0b w/ SiLU
-    My experimental 'light' model w/ 1.5x final_conv mult, 64 group_size, .25 bottleneck & SE ratio
+    My experimental 'light' model w/ F0 repeats, 1.5x final_conv mult, 64 group_size, .25 bottleneck & SE ratio
     """
-    return _create_normfreenet('nfnet_l0b', pretrained=pretrained, **kwargs)
+    return _create_normfreenet('nfnet_l0', pretrained=pretrained, **kwargs)
 
 
 @register_model
 def eca_nfnet_l0(pretrained=False, **kwargs):
     """ ECA-NFNet-L0 w/ SiLU
-    My experimental 'light' model w/ 1.5x final_conv mult, 64 group_size, .25 bottleneck & ECA attn
+    My experimental 'light' model w/ F0 repeats, 1.5x final_conv mult, 64 group_size, .25 bottleneck & ECA attn
     """
     return _create_normfreenet('eca_nfnet_l0', pretrained=pretrained, **kwargs)
 
 
+@register_model
+def eca_nfnet_l1(pretrained=False, **kwargs):
+    """ ECA-NFNet-L1 w/ SiLU
+    My experimental 'light' model w/ F1 repeats, 2.0x final_conv mult, 64 group_size, .25 bottleneck & ECA attn
+    """
+    return _create_normfreenet('eca_nfnet_l1', pretrained=pretrained, **kwargs)
+
+
 @register_model
 def nf_regnet_b0(pretrained=False, **kwargs):
     """ Normalization-Free RegNet-B0

timm/models/pit.py

@@ -0,0 +1,388 @@
+""" Pooling-based Vision Transformer (PiT) in PyTorch
+
+A PyTorch implement of Pooling-based Vision Transformers as described in
+'Rethinking Spatial Dimensions of Vision Transformers' - https://arxiv.org/abs/2103.16302
+
+This code was adapted from the original version at https://github.com/naver-ai/pit, original copyright below.
+
+Modifications for timm by / Copyright 2020 Ross Wightman
+"""
+# PiT
+# Copyright 2021-present NAVER Corp.
+# Apache License v2.0
+
+import math
+import re
+from copy import deepcopy
+from functools import partial
+from typing import Tuple
+
+import torch
+from torch import nn
+
+from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+from .helpers import build_model_with_cfg, overlay_external_default_cfg
+from .layers import trunc_normal_, to_2tuple
+from .registry import register_model
+from .vision_transformer import Block
+
+
+def _cfg(url='', **kwargs):
+    return {
+        'url': url,
+        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
+        'crop_pct': .9, 'interpolation': 'bicubic',
+        'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
+        'first_conv': 'patch_embed.conv', 'classifier': 'head',
+        **kwargs
+    }
+
+
+default_cfgs = {
+    # deit models (FB weights)
+    'pit_ti_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-pit-weights/pit_ti_730.pth'),
+    'pit_xs_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-pit-weights/pit_xs_781.pth'),
+    'pit_s_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-pit-weights/pit_s_809.pth'),
+    'pit_b_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-pit-weights/pit_b_820.pth'),
+    'pit_ti_distilled_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-pit-weights/pit_ti_distill_746.pth',
+        classifier=('head', 'head_dist')),
+    'pit_xs_distilled_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-pit-weights/pit_xs_distill_791.pth',
+        classifier=('head', 'head_dist')),
+    'pit_s_distilled_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-pit-weights/pit_s_distill_819.pth',
+        classifier=('head', 'head_dist')),
+    'pit_b_distilled_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-pit-weights/pit_b_distill_840.pth',
+        classifier=('head', 'head_dist')),
+}
+
+
+class SequentialTuple(nn.Sequential):
+    """ This module exists to work around torchscript typing issues list -> list"""
+    def __init__(self, *args):
+        super(SequentialTuple, self).__init__(*args)
+
+    def forward(self, x: Tuple[torch.Tensor, torch.Tensor]) -> Tuple[torch.Tensor, torch.Tensor]:
+        for module in self:
+            x = module(x)
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+            self, base_dim, depth, heads, mlp_ratio, pool=None, drop_rate=.0, attn_drop_rate=.0, drop_path_prob=None):
+        super(Transformer, self).__init__()
+        self.layers = nn.ModuleList([])
+        embed_dim = base_dim * heads
+
+        self.blocks = nn.Sequential(*[
+            Block(
+                dim=embed_dim,
+                num_heads=heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=True,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=drop_path_prob[i],
+                norm_layer=partial(nn.LayerNorm, eps=1e-6)
+            )
+            for i in range(depth)])
+
+        self.pool = pool
+
+    def forward(self, x: Tuple[torch.Tensor, torch.Tensor]) -> Tuple[torch.Tensor, torch.Tensor]:
+        x, cls_tokens = x
+        B, C, H, W = x.shape
+        token_length = cls_tokens.shape[1]
+
+        x = x.flatten(2).transpose(1, 2)
+        x = torch.cat((cls_tokens, x), dim=1)
+
+        x = self.blocks(x)
+
+        cls_tokens = x[:, :token_length]
+        x = x[:, token_length:]
+        x = x.transpose(1, 2).reshape(B, C, H, W)
+
+        if self.pool is not None:
+            x, cls_tokens = self.pool(x, cls_tokens)
+        return x, cls_tokens
+
+
+class ConvHeadPooling(nn.Module):
+    def __init__(self, in_feature, out_feature, stride, padding_mode='zeros'):
+        super(ConvHeadPooling, self).__init__()
+
+        self.conv = nn.Conv2d(
+            in_feature, out_feature, kernel_size=stride + 1, padding=stride // 2, stride=stride,
+            padding_mode=padding_mode, groups=in_feature)
+        self.fc = nn.Linear(in_feature, out_feature)
+
+    def forward(self, x, cls_token) -> Tuple[torch.Tensor, torch.Tensor]:
+
+        x = self.conv(x)
+        cls_token = self.fc(cls_token)
+
+        return x, cls_token
+
+
+class ConvEmbedding(nn.Module):
+    def __init__(self, in_channels, out_channels, patch_size, stride, padding):
+        super(ConvEmbedding, self).__init__()
+        self.conv = nn.Conv2d(
+            in_channels, out_channels, kernel_size=patch_size, stride=stride, padding=padding, bias=True)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return x
+
+
+class PoolingVisionTransformer(nn.Module):
+    """ Pooling-based Vision Transformer
+
+    A PyTorch implement of 'Rethinking Spatial Dimensions of Vision Transformers'
+        - https://arxiv.org/abs/2103.16302
+    """
+    def __init__(self, img_size, patch_size, stride, base_dims, depth, heads,
+                 mlp_ratio, num_classes=1000, in_chans=3, distilled=False,
+                 attn_drop_rate=.0, drop_rate=.0, drop_path_rate=.0):
+        super(PoolingVisionTransformer, self).__init__()
+
+        padding = 0
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        height = math.floor((img_size[0] + 2 * padding - patch_size[0]) / stride + 1)
+        width = math.floor((img_size[1] + 2 * padding - patch_size[1]) / stride + 1)
+
+        self.base_dims = base_dims
+        self.heads = heads
+        self.num_classes = num_classes
+        self.num_tokens = 2 if distilled else 1
+
+        self.patch_size = patch_size
+        self.pos_embed = nn.Parameter(torch.randn(1, base_dims[0] * heads[0], height, width))
+        self.patch_embed = ConvEmbedding(in_chans, base_dims[0] * heads[0], patch_size, stride, padding)
+
+        self.cls_token = nn.Parameter(torch.randn(1, self.num_tokens, base_dims[0] * heads[0]))
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        transformers = []
+        # stochastic depth decay rule
+        dpr = [x.tolist() for x in torch.linspace(0, drop_path_rate, sum(depth)).split(depth)]
+        for stage in range(len(depth)):
+            pool = None
+            if stage < len(heads) - 1:
+                pool = ConvHeadPooling(
+                    base_dims[stage] * heads[stage], base_dims[stage + 1] * heads[stage + 1], stride=2)
+            transformers += [Transformer(
+                base_dims[stage], depth[stage], heads[stage], mlp_ratio, pool=pool,
+                drop_rate=drop_rate, attn_drop_rate=attn_drop_rate, drop_path_prob=dpr[stage])
+            ]
+        self.transformers = SequentialTuple(*transformers)
+        self.norm = nn.LayerNorm(base_dims[-1] * heads[-1], eps=1e-6)
+        self.embed_dim = base_dims[-1] * heads[-1]
+
+        # Classifier head
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+        self.head_dist = nn.Linear(self.embed_dim, self.num_classes) \
+            if num_classes > 0 and distilled else nn.Identity()
+
+        trunc_normal_(self.pos_embed, std=.02)
+        trunc_normal_(self.cls_token, std=.02)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+        self.head_dist = nn.Linear(self.embed_dim, self.num_classes) \
+            if num_classes > 0 and self.num_tokens == 2 else nn.Identity()
+
+    def forward_features(self, x):
+        x = self.patch_embed(x)
+        x = self.pos_drop(x + self.pos_embed)
+        cls_tokens = self.cls_token.expand(x.shape[0], -1, -1)
+        x, cls_tokens = self.transformers((x, cls_tokens))
+        cls_tokens = self.norm(cls_tokens)
+        return cls_tokens
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        x_cls = self.head(x[:, 0])
+        if self.num_tokens > 1:
+            x_dist = self.head_dist(x[:, 1])
+            if self.training and not torch.jit.is_scripting():
+                return x_cls, x_dist
+            else:
+                return (x_cls + x_dist) / 2
+        else:
+            return x_cls
+
+
+def checkpoint_filter_fn(state_dict, model):
+    """ preprocess checkpoints """
+    out_dict = {}
+    p_blocks = re.compile(r'pools\.(\d)\.')
+    for k, v in state_dict.items():
+        # FIXME need to update resize for PiT impl
+        # if k == 'pos_embed' and v.shape != model.pos_embed.shape:
+        #     # To resize pos embedding when using model at different size from pretrained weights
+        #     v = resize_pos_embed(v, model.pos_embed)
+        k = p_blocks.sub(lambda exp: f'transformers.{int(exp.group(1))}.pool.', k)
+        out_dict[k] = v
+    return out_dict
+
+
+def _create_pit(variant, pretrained=False, **kwargs):
+    default_cfg = deepcopy(default_cfgs[variant])
+    overlay_external_default_cfg(default_cfg, kwargs)
+    default_num_classes = default_cfg['num_classes']
+    default_img_size = default_cfg['input_size'][-2:]
+    img_size = kwargs.pop('img_size', default_img_size)
+    num_classes = kwargs.pop('num_classes', default_num_classes)
+
+    if kwargs.get('features_only', None):
+        raise RuntimeError('features_only not implemented for Vision Transformer models.')
+
+    model = build_model_with_cfg(
+        PoolingVisionTransformer, variant, pretrained,
+        default_cfg=default_cfg,
+        img_size=img_size,
+        num_classes=num_classes,
+        pretrained_filter_fn=checkpoint_filter_fn,
+        **kwargs)
+
+    return model
+
+
+@register_model
+def pit_b_224(pretrained, **kwargs):
+    model_kwargs = dict(
+        patch_size=14,
+        stride=7,
+        base_dims=[64, 64, 64],
+        depth=[3, 6, 4],
+        heads=[4, 8, 16],
+        mlp_ratio=4,
+        **kwargs
+    )
+    return _create_pit('pit_b_224', pretrained, **model_kwargs)
+
+
+@register_model
+def pit_s_224(pretrained, **kwargs):
+    model_kwargs = dict(
+        patch_size=16,
+        stride=8,
+        base_dims=[48, 48, 48],
+        depth=[2, 6, 4],
+        heads=[3, 6, 12],
+        mlp_ratio=4,
+        **kwargs
+    )
+    return _create_pit('pit_s_224', pretrained, **model_kwargs)
+
+
+@register_model
+def pit_xs_224(pretrained, **kwargs):
+    model_kwargs = dict(
+        patch_size=16,
+        stride=8,
+        base_dims=[48, 48, 48],
+        depth=[2, 6, 4],
+        heads=[2, 4, 8],
+        mlp_ratio=4,
+        **kwargs
+    )
+    return _create_pit('pit_xs_224', pretrained, **model_kwargs)
+
+
+@register_model
+def pit_ti_224(pretrained, **kwargs):
+    model_kwargs = dict(
+        patch_size=16,
+        stride=8,
+        base_dims=[32, 32, 32],
+        depth=[2, 6, 4],
+        heads=[2, 4, 8],
+        mlp_ratio=4,
+        **kwargs
+    )
+    return _create_pit('pit_ti_224', pretrained, **model_kwargs)
+
+
+@register_model
+def pit_b_distilled_224(pretrained, **kwargs):
+    model_kwargs = dict(
+        patch_size=14,
+        stride=7,
+        base_dims=[64, 64, 64],
+        depth=[3, 6, 4],
+        heads=[4, 8, 16],
+        mlp_ratio=4,
+        distilled=True,
+        **kwargs
+    )
+    return _create_pit('pit_b_distilled_224', pretrained, **model_kwargs)
+
+
+@register_model
+def pit_s_distilled_224(pretrained, **kwargs):
+    model_kwargs = dict(
+        patch_size=16,
+        stride=8,
+        base_dims=[48, 48, 48],
+        depth=[2, 6, 4],
+        heads=[3, 6, 12],
+        mlp_ratio=4,
+        distilled=True,
+        **kwargs
+    )
+    return _create_pit('pit_s_distilled_224', pretrained, **model_kwargs)
+
+
+@register_model
+def pit_xs_distilled_224(pretrained, **kwargs):
+    model_kwargs = dict(
+        patch_size=16,
+        stride=8,
+        base_dims=[48, 48, 48],
+        depth=[2, 6, 4],
+        heads=[2, 4, 8],
+        mlp_ratio=4,
+        distilled=True,
+        **kwargs
+    )
+    return _create_pit('pit_xs_distilled_224', pretrained, **model_kwargs)
+
+
+@register_model
+def pit_ti_distilled_224(pretrained, **kwargs):
+    model_kwargs = dict(
+        patch_size=16,
+        stride=8,
+        base_dims=[32, 32, 32],
+        depth=[2, 6, 4],
+        heads=[2, 4, 8],
+        mlp_ratio=4,
+        distilled=True,
+        **kwargs
+    )
+    return _create_pit('pit_ti_distilled_224', pretrained, **model_kwargs)

timm/models/regnet.py

@@ -57,12 +57,13 @@ model_cfgs = dict(
 )
 
 
-def _cfg(url=''):
+def _cfg(url='', **kwargs):
     return {
         'url': url, 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': (7, 7),
         'crop_pct': 0.875, 'interpolation': 'bicubic',
         'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
         'first_conv': 'stem.conv', 'classifier': 'head.fc',
+        **kwargs
     }
 
 
@@ -84,12 +85,16 @@ default_cfgs = dict(
     regnety_006=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_006-c67e57ec.pth'),
     regnety_008=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_008-dc900dbe.pth'),
     regnety_016=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_016-54367f74.pth'),
-    regnety_032=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/regnety_032_ra-7f2439f9.pth'),
+    regnety_032=_cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/regnety_032_ra-7f2439f9.pth',
+        crop_pct=1.0, test_input_size=(3, 288, 288)),
     regnety_040=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_040-f0d569f9.pth'),
     regnety_064=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_064-0a48325c.pth'),
     regnety_080=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_080-e7f3eb93.pth'),
     regnety_120=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_120-721ba79a.pth'),
-    regnety_160=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_160-d64013cd.pth'),
+    regnety_160=_cfg(
+        url='https://dl.fbaipublicfiles.com/deit/regnety_160-a5fe301d.pth',  # from Facebook DeiT GitHub repository
+        crop_pct=1.0, test_input_size=(3, 288, 288)),
     regnety_320=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_320-ba464b29.pth'),
 )
 
@@ -328,11 +333,20 @@ class RegNet(nn.Module):
         return x
 
 
+def _filter_fn(state_dict):
+    """ convert patch embedding weight from manual patchify + linear proj to conv"""
+    if 'model' in state_dict:
+        # For DeiT trained regnety_160 pretraiend model
+        state_dict = state_dict['model']
+    return state_dict
+
+
 def _create_regnet(variant, pretrained, **kwargs):
     return build_model_with_cfg(
         RegNet, variant, pretrained,
         default_cfg=default_cfgs[variant],
         model_cfg=model_cfgs[variant],
+        pretrained_filter_fn=_filter_fn,
         **kwargs)
 
 

timm/models/resnetv2.py

@@ -274,7 +274,9 @@ class ResNetStage(nn.Module):
         return x
 
 
-def create_stem(in_chs, out_chs, stem_type='', preact=True, conv_layer=None, norm_layer=None):
+def create_resnetv2_stem(
+        in_chs, out_chs=64, stem_type='', preact=True,
+        conv_layer=StdConv2d, norm_layer=partial(GroupNormAct, num_groups=32)):
     stem = OrderedDict()
     assert stem_type in ('', 'fixed', 'same', 'deep', 'deep_fixed', 'deep_same')
 
@@ -322,7 +324,8 @@ class ResNetV2(nn.Module):
 
         self.feature_info = []
         stem_chs = make_div(stem_chs * wf)
-        self.stem = create_stem(in_chans, stem_chs, stem_type, preact, conv_layer=conv_layer, norm_layer=norm_layer)
+        self.stem = create_resnetv2_stem(
+            in_chans, stem_chs, stem_type, preact, conv_layer=conv_layer, norm_layer=norm_layer)
         stem_feat = ('stem.conv3' if 'deep' in stem_type else 'stem.conv') if preact else 'stem.norm'
         self.feature_info.append(dict(num_chs=stem_chs, reduction=2, module=stem_feat))
 

timm/models/vision_transformer.py

@@ -5,6 +5,9 @@ A PyTorch implement of Vision Transformers as described in
 
 The official jax code is released and available at https://github.com/google-research/vision_transformer
 
+DeiT model defs and weights from https://github.com/facebookresearch/deit,
+paper `DeiT: Data-efficient Image Transformers` - https://arxiv.org/abs/2012.12877
+
 Acknowledgments:
 * The paper authors for releasing code and weights, thanks!
 * I fixed my class token impl based on Phil Wang's https://github.com/lucidrains/vit-pytorch ... check it out
@@ -12,9 +15,6 @@ for some einops/einsum fun
 * Simple transformer style inspired by Andrej Karpathy's https://github.com/karpathy/minGPT
 * Bert reference code checks against Huggingface Transformers and Tensorflow Bert
 
-DeiT model defs and weights from https://github.com/facebookresearch/deit,
-paper `DeiT: Data-efficient Image Transformers` - https://arxiv.org/abs/2012.12877
-
 Hacked together by / Copyright 2020 Ross Wightman
 """
 import math
@@ -29,9 +29,7 @@ import torch.nn.functional as F
 
 from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
 from .helpers import build_model_with_cfg, overlay_external_default_cfg
-from .layers import StdConv2dSame, DropPath, to_2tuple, trunc_normal_
-from .resnet import resnet26d, resnet50d
-from .resnetv2 import ResNetV2
+from .layers import DropPath, to_2tuple, trunc_normal_, lecun_normal_
 from .registry import register_model
 
 _logger = logging.getLogger(__name__)
@@ -98,20 +96,6 @@ default_cfgs = {
         hf_hub='timm/vit_huge_patch14_224_in21k',
         num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
 
-    # hybrid models (weights ported from official Google JAX impl)
-    'vit_base_resnet50_224_in21k': _cfg(
-        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_resnet50_224_in21k-6f7c7740.pth',
-        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=0.9, first_conv='patch_embed.backbone.stem.conv'),
-    'vit_base_resnet50_384': _cfg(
-        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_resnet50_384-9fd3c705.pth',
-        input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0, first_conv='patch_embed.backbone.stem.conv'),
-
-    # hybrid models (my experiments)
-    'vit_small_resnet26d_224': _cfg(),
-    'vit_small_resnet50d_s3_224': _cfg(),
-    'vit_base_resnet26d_224': _cfg(),
-    'vit_base_resnet50d_224': _cfg(),
-
     # deit models (FB weights)
     'vit_deit_tiny_patch16_224': _cfg(
         url='https://dl.fbaipublicfiles.com/deit/deit_tiny_patch16_224-a1311bcf.pth'),
@@ -123,14 +107,17 @@ default_cfgs = {
         url='https://dl.fbaipublicfiles.com/deit/deit_base_patch16_384-8de9b5d1.pth',
         input_size=(3, 384, 384), crop_pct=1.0),
     'vit_deit_tiny_distilled_patch16_224': _cfg(
-        url='https://dl.fbaipublicfiles.com/deit/deit_tiny_distilled_patch16_224-b40b3cf7.pth'),
+        url='https://dl.fbaipublicfiles.com/deit/deit_tiny_distilled_patch16_224-b40b3cf7.pth',
+        classifier=('head', 'head_dist')),
     'vit_deit_small_distilled_patch16_224': _cfg(
-        url='https://dl.fbaipublicfiles.com/deit/deit_small_distilled_patch16_224-649709d9.pth'),
+        url='https://dl.fbaipublicfiles.com/deit/deit_small_distilled_patch16_224-649709d9.pth',
+        classifier=('head', 'head_dist')),
     'vit_deit_base_distilled_patch16_224': _cfg(
-        url='https://dl.fbaipublicfiles.com/deit/deit_base_distilled_patch16_224-df68dfff.pth', ),
+        url='https://dl.fbaipublicfiles.com/deit/deit_base_distilled_patch16_224-df68dfff.pth',
+        classifier=('head', 'head_dist')),
     'vit_deit_base_distilled_patch16_384': _cfg(
         url='https://dl.fbaipublicfiles.com/deit/deit_base_distilled_patch16_384-d0272ac0.pth',
-        input_size=(3, 384, 384), crop_pct=1.0),
+        input_size=(3, 384, 384), crop_pct=1.0, classifier=('head', 'head_dist')),
 }
 
 
@@ -158,7 +145,6 @@ class Attention(nn.Module):
         super().__init__()
         self.num_heads = num_heads
         head_dim = dim // num_heads
-        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
         self.scale = qk_scale or head_dim ** -0.5
 
         self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
@@ -224,56 +210,20 @@ class PatchEmbed(nn.Module):
         return x
 
 
-class HybridEmbed(nn.Module):
-    """ CNN Feature Map Embedding
-    Extract feature map from CNN, flatten, project to embedding dim.
-    """
-    def __init__(self, backbone, img_size=224, feature_size=None, in_chans=3, embed_dim=768):
-        super().__init__()
-        assert isinstance(backbone, nn.Module)
-        img_size = to_2tuple(img_size)
-        self.img_size = img_size
-        self.backbone = backbone
-        if feature_size is None:
-            with torch.no_grad():
-                # FIXME this is hacky, but most reliable way of determining the exact dim of the output feature
-                # map for all networks, the feature metadata has reliable channel and stride info, but using
-                # stride to calc feature dim requires info about padding of each stage that isn't captured.
-                training = backbone.training
-                if training:
-                    backbone.eval()
-                o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))
-                if isinstance(o, (list, tuple)):
-                    o = o[-1]  # last feature if backbone outputs list/tuple of features
-                feature_size = o.shape[-2:]
-                feature_dim = o.shape[1]
-                backbone.train(training)
-        else:
-            feature_size = to_2tuple(feature_size)
-            if hasattr(self.backbone, 'feature_info'):
-                feature_dim = self.backbone.feature_info.channels()[-1]
-            else:
-                feature_dim = self.backbone.num_features
-        self.num_patches = feature_size[0] * feature_size[1]
-        self.proj = nn.Conv2d(feature_dim, embed_dim, 1)
-
-    def forward(self, x):
-        x = self.backbone(x)
-        if isinstance(x, (list, tuple)):
-            x = x[-1]  # last feature if backbone outputs list/tuple of features
-        x = self.proj(x).flatten(2).transpose(1, 2)
-        return x
-
-
 class VisionTransformer(nn.Module):
     """ Vision Transformer
 
-    A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale`  -
-        https://arxiv.org/abs/2010.11929
+    A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale`
+        - https://arxiv.org/abs/2010.11929
+
+    Includes distillation token & head support for `DeiT: Data-efficient Image Transformers`
+        - https://arxiv.org/abs/2012.12877
     """
+
     def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
-                 num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, representation_size=None,
-                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., hybrid_backbone=None, norm_layer=None):
+                 num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, representation_size=None, distilled=False,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., embed_layer=PatchEmbed, norm_layer=None,
+                 act_layer=None, weight_init=''):
         """
         Args:
             img_size (int, tuple): input image size
@@ -287,39 +237,40 @@ class VisionTransformer(nn.Module):
             qkv_bias (bool): enable bias for qkv if True
             qk_scale (float): override default qk scale of head_dim ** -0.5 if set
             representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set
+            distilled (bool): model includes a distillation token and head as in DeiT models
             drop_rate (float): dropout rate
             attn_drop_rate (float): attention dropout rate
             drop_path_rate (float): stochastic depth rate
-            hybrid_backbone (nn.Module): CNN backbone to use in-place of PatchEmbed module
+            embed_layer (nn.Module): patch embedding layer
             norm_layer: (nn.Module): normalization layer
+            weight_init: (str): weight init scheme
         """
         super().__init__()
         self.num_classes = num_classes
         self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        self.num_tokens = 2 if distilled else 1
         norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+        act_layer = act_layer or nn.GELU
 
-        if hybrid_backbone is not None:
-            self.patch_embed = HybridEmbed(
-                hybrid_backbone, img_size=img_size, in_chans=in_chans, embed_dim=embed_dim)
-        else:
-            self.patch_embed = PatchEmbed(
-                img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
+        self.patch_embed = embed_layer(
+            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
         num_patches = self.patch_embed.num_patches
 
         self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
-        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+        self.dist_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) if distilled else None
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim))
         self.pos_drop = nn.Dropout(p=drop_rate)
 
         dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
-        self.blocks = nn.ModuleList([
+        self.blocks = nn.Sequential(*[
             Block(
                 dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
-                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer)
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, act_layer=act_layer)
             for i in range(depth)])
         self.norm = norm_layer(embed_dim)
 
         # Representation layer
-        if representation_size:
+        if representation_size and not distilled:
             self.num_features = representation_size
             self.pre_logits = nn.Sequential(OrderedDict([
                 ('fc', nn.Linear(embed_dim, representation_size)),
@@ -328,110 +279,118 @@ class VisionTransformer(nn.Module):
         else:
             self.pre_logits = nn.Identity()
 
-        # Classifier head
+        # Classifier head(s)
         self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
+        self.head_dist = None
+        if distilled:
+            self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if num_classes > 0 else nn.Identity()
 
+        # Weight init
+        assert weight_init in ('jax', 'jax_nlhb', 'nlhb', '')
+        head_bias = -math.log(self.num_classes) if 'nlhb' in weight_init else 0.
         trunc_normal_(self.pos_embed, std=.02)
-        trunc_normal_(self.cls_token, std=.02)
-        self.apply(self._init_weights)
+        if self.dist_token is not None:
+            trunc_normal_(self.dist_token, std=.02)
+        if weight_init.startswith('jax'):
+            # leave cls token as zeros to match jax impl
+            for n, m in self.named_modules():
+                _init_vit_weights(m, n, head_bias=head_bias, jax_impl=True)
+        else:
+            trunc_normal_(self.cls_token, std=.02)
+            self.apply(_init_vit_weights)
 
     def _init_weights(self, m):
-        if isinstance(m, nn.Linear):
-            trunc_normal_(m.weight, std=.02)
-            if isinstance(m, nn.Linear) and m.bias is not None:
-                nn.init.constant_(m.bias, 0)
-        elif isinstance(m, nn.LayerNorm):
-            nn.init.constant_(m.bias, 0)
-            nn.init.constant_(m.weight, 1.0)
+        # this fn left here for compat with downstream users
+        _init_vit_weights(m)
 
     @torch.jit.ignore
     def no_weight_decay(self):
-        return {'pos_embed', 'cls_token'}
+        return {'pos_embed', 'cls_token', 'dist_token'}
 
     def get_classifier(self):
-        return self.head
+        if self.dist_token is None:
+            return self.head
+        else:
+            return self.head, self.head_dist
 
     def reset_classifier(self, num_classes, global_pool=''):
         self.num_classes = num_classes
         self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+        if self.num_tokens == 2:
+            self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if num_classes > 0 else nn.Identity()
 
     def forward_features(self, x):
-        B = x.shape[0]
         x = self.patch_embed(x)
-
-        cls_tokens = self.cls_token.expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
-        x = torch.cat((cls_tokens, x), dim=1)
-        x = x + self.pos_embed
-        x = self.pos_drop(x)
-
-        for blk in self.blocks:
-            x = blk(x)
-
-        x = self.norm(x)[:, 0]
-        x = self.pre_logits(x)
-        return x
+        cls_token = self.cls_token.expand(x.shape[0], -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+        if self.dist_token is None:
+            x = torch.cat((cls_token, x), dim=1)
+        else:
+            x = torch.cat((cls_token, self.dist_token.expand(x.shape[0], -1, -1), x), dim=1)
+        x = self.pos_drop(x + self.pos_embed)
+        x = self.blocks(x)
+        x = self.norm(x)
+        if self.dist_token is None:
+            return self.pre_logits(x[:, 0])
+        else:
+            return x[:, 0], x[:, 1]
 
     def forward(self, x):
         x = self.forward_features(x)
-        x = self.head(x)
+        if self.head_dist is not None:
+            x, x_dist = self.head(x[0]), self.head_dist(x[1])  # x must be a tuple
+            if self.training and not torch.jit.is_scripting():
+                # during inference, return the average of both classifier predictions
+                return x, x_dist
+            else:
+                return (x + x_dist) / 2
+        else:
+            x = self.head(x)
         return x
 
 
-class DistilledVisionTransformer(VisionTransformer):
-    """ Vision Transformer with distillation token.
-
-    Paper: `Training data-efficient image transformers & distillation through attention` -
-        https://arxiv.org/abs/2012.12877
-
-    This impl of distilled ViT is taken from https://github.com/facebookresearch/deit
+def _init_vit_weights(m, n: str = '', head_bias: float = 0., jax_impl: bool = False):
+    """ ViT weight initialization
+    * When called without n, head_bias, jax_impl args it will behave exactly the same
+      as my original init for compatibility with prev hparam / downstream use cases (ie DeiT).
+    * When called w/ valid n (module name) and jax_impl=True, will (hopefully) match JAX impl
     """
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.dist_token = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
-        num_patches = self.patch_embed.num_patches
-        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 2, self.embed_dim))
-        self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if self.num_classes > 0 else nn.Identity()
-
-        trunc_normal_(self.dist_token, std=.02)
-        trunc_normal_(self.pos_embed, std=.02)
-        self.head_dist.apply(self._init_weights)
-
-    def forward_features(self, x):
-        B = x.shape[0]
-        x = self.patch_embed(x)
-
-        cls_tokens = self.cls_token.expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
-        dist_token = self.dist_token.expand(B, -1, -1)
-        x = torch.cat((cls_tokens, dist_token, x), dim=1)
-
-        x = x + self.pos_embed
-        x = self.pos_drop(x)
-
-        for blk in self.blocks:
-            x = blk(x)
-
-        x = self.norm(x)
-        return x[:, 0], x[:, 1]
-
-    def forward(self, x):
-        x, x_dist = self.forward_features(x)
-        x = self.head(x)
-        x_dist = self.head_dist(x_dist)
-        if self.training:
-            return x, x_dist
+    if isinstance(m, nn.Linear):
+        if n.startswith('head'):
+            nn.init.zeros_(m.weight)
+            nn.init.constant_(m.bias, head_bias)
+        elif n.startswith('pre_logits'):
+            lecun_normal_(m.weight)
+            nn.init.zeros_(m.bias)
         else:
-            # during inference, return the average of both classifier predictions
-            return (x + x_dist) / 2
-
-
-def resize_pos_embed(posemb, posemb_new):
+            if jax_impl:
+                nn.init.xavier_uniform_(m.weight)
+                if m.bias is not None:
+                    if 'mlp' in n:
+                        nn.init.normal_(m.bias, std=1e-6)
+                    else:
+                        nn.init.zeros_(m.bias)
+            else:
+                trunc_normal_(m.weight, std=.02)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+    elif jax_impl and isinstance(m, nn.Conv2d):
+        # NOTE conv was left to pytorch default in my original init
+        lecun_normal_(m.weight)
+        if m.bias is not None:
+            nn.init.zeros_(m.bias)
+    elif isinstance(m, nn.LayerNorm):
+        nn.init.zeros_(m.bias)
+        nn.init.ones_(m.weight)
+
+
+def resize_pos_embed(posemb, posemb_new, num_tokens=1):
     # Rescale the grid of position embeddings when loading from state_dict. Adapted from
     # https://github.com/google-research/vision_transformer/blob/00883dd691c63a6830751563748663526e811cee/vit_jax/checkpoint.py#L224
     _logger.info('Resized position embedding: %s to %s', posemb.shape, posemb_new.shape)
     ntok_new = posemb_new.shape[1]
-    if True:
-        posemb_tok, posemb_grid = posemb[:, :1], posemb[0, 1:]
-        ntok_new -= 1
+    if num_tokens:
+        posemb_tok, posemb_grid = posemb[:, :num_tokens], posemb[0, num_tokens:]
+        ntok_new -= num_tokens
     else:
         posemb_tok, posemb_grid = posemb[:, :0], posemb[0]
     gs_old = int(math.sqrt(len(posemb_grid)))
@@ -457,13 +416,14 @@ def checkpoint_filter_fn(state_dict, model):
             v = v.reshape(O, -1, H, W)
         elif k == 'pos_embed' and v.shape != model.pos_embed.shape:
             # To resize pos embedding when using model at different size from pretrained weights
-            v = resize_pos_embed(v, model.pos_embed)
+            v = resize_pos_embed(v, model.pos_embed, getattr(model, 'num_tokens', 1))
         out_dict[k] = v
     return out_dict
 
 
-def _create_vision_transformer(variant, pretrained=False, distilled=False, **kwargs):
-    default_cfg = deepcopy(default_cfgs[variant])
+def _create_vision_transformer(variant, pretrained=False, default_cfg=None, **kwargs):
+    if default_cfg is None:
+        default_cfg = deepcopy(default_cfgs[variant])
     overlay_external_default_cfg(default_cfg, kwargs)
     default_num_classes = default_cfg['num_classes']
     default_img_size = default_cfg['input_size'][-2:]
@@ -480,9 +440,8 @@ def _create_vision_transformer(variant, pretrained=False, distilled=False, **kwa
     if kwargs.get('features_only', None):
         raise RuntimeError('features_only not implemented for Vision Transformer models.')
 
-    model_cls = DistilledVisionTransformer if distilled else VisionTransformer
     model = build_model_with_cfg(
-        model_cls, variant, pretrained,
+        VisionTransformer, variant, pretrained,
         default_cfg=default_cfg,
         img_size=img_size,
         num_classes=num_classes,
@@ -495,7 +454,11 @@ def _create_vision_transformer(variant, pretrained=False, distilled=False, **kwa
 
 @register_model
 def vit_small_patch16_224(pretrained=False, **kwargs):
-    """ My custom 'small' ViT model. Depth=8, heads=8= mlp_ratio=3."""
+    """ My custom 'small' ViT model. embed_dim=768, depth=8, num_heads=8, mlp_ratio=3.
+    NOTE:
+        * this differs from the DeiT based 'small' definitions with embed_dim=384, depth=12, num_heads=6
+        * this model does not have a bias for QKV (unlike the official ViT and DeiT models)
+    """
     model_kwargs = dict(
         patch_size=16, embed_dim=768, depth=8, num_heads=8, mlp_ratio=3.,
         qkv_bias=False, norm_layer=nn.LayerNorm, **kwargs)
@@ -640,76 +603,6 @@ def vit_huge_patch14_224_in21k(pretrained=False, **kwargs):
     return model
 
 
-@register_model
-def vit_base_resnet50_224_in21k(pretrained=False, **kwargs):
-    """ R50+ViT-B/16 hybrid model from original paper (https://arxiv.org/abs/2010.11929).
-    ImageNet-21k weights @ 224x224, source https://github.com/google-research/vision_transformer.
-    """
-    # create a ResNetV2 w/o pre-activation, that uses StdConv and GroupNorm and has 3 stages, no head
-    backbone = ResNetV2(
-        layers=(3, 4, 9), num_classes=0, global_pool='', in_chans=kwargs.get('in_chans', 3),
-        preact=False, stem_type='same', conv_layer=StdConv2dSame)
-    model_kwargs = dict(
-        embed_dim=768, depth=12, num_heads=12, hybrid_backbone=backbone,
-        representation_size=768, **kwargs)
-    model = _create_vision_transformer('vit_base_resnet50_224_in21k', pretrained=pretrained, **model_kwargs)
-    return model
-
-
-@register_model
-def vit_base_resnet50_384(pretrained=False, **kwargs):
-    """ R50+ViT-B/16 hybrid from original paper (https://arxiv.org/abs/2010.11929).
-    ImageNet-1k weights fine-tuned from in21k @ 384x384, source https://github.com/google-research/vision_transformer.
-    """
-    # create a ResNetV2 w/o pre-activation, that uses StdConv and GroupNorm and has 3 stages, no head
-    backbone = ResNetV2(
-        layers=(3, 4, 9), num_classes=0, global_pool='', in_chans=kwargs.get('in_chans', 3),
-        preact=False, stem_type='same', conv_layer=StdConv2dSame)
-    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, hybrid_backbone=backbone, **kwargs)
-    model = _create_vision_transformer('vit_base_resnet50_384', pretrained=pretrained, **model_kwargs)
-    return model
-
-
-@register_model
-def vit_small_resnet26d_224(pretrained=False, **kwargs):
-    """ Custom ViT small hybrid w/ ResNet26D stride 32. No pretrained weights.
-    """
-    backbone = resnet26d(pretrained=pretrained, in_chans=kwargs.get('in_chans', 3), features_only=True, out_indices=[4])
-    model_kwargs = dict(embed_dim=768, depth=8, num_heads=8, mlp_ratio=3, hybrid_backbone=backbone, **kwargs)
-    model = _create_vision_transformer('vit_small_resnet26d_224', pretrained=pretrained, **model_kwargs)
-    return model
-
-
-@register_model
-def vit_small_resnet50d_s3_224(pretrained=False, **kwargs):
-    """ Custom ViT small hybrid w/ ResNet50D 3-stages, stride 16. No pretrained weights.
-    """
-    backbone = resnet50d(pretrained=pretrained, in_chans=kwargs.get('in_chans', 3), features_only=True, out_indices=[3])
-    model_kwargs = dict(embed_dim=768, depth=8, num_heads=8, mlp_ratio=3, hybrid_backbone=backbone, **kwargs)
-    model = _create_vision_transformer('vit_small_resnet50d_s3_224', pretrained=pretrained, **model_kwargs)
-    return model
-
-
-@register_model
-def vit_base_resnet26d_224(pretrained=False, **kwargs):
-    """ Custom ViT base hybrid w/ ResNet26D stride 32. No pretrained weights.
-    """
-    backbone = resnet26d(pretrained=pretrained, in_chans=kwargs.get('in_chans', 3), features_only=True, out_indices=[4])
-    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, hybrid_backbone=backbone, **kwargs)
-    model = _create_vision_transformer('vit_base_resnet26d_224', pretrained=pretrained, **model_kwargs)
-    return model
-
-
-@register_model
-def vit_base_resnet50d_224(pretrained=False, **kwargs):
-    """ Custom ViT base hybrid w/ ResNet50D stride 32. No pretrained weights.
-    """
-    backbone = resnet50d(pretrained=pretrained, in_chans=kwargs.get('in_chans', 3), features_only=True, out_indices=[4])
-    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, hybrid_backbone=backbone, **kwargs)
-    model = _create_vision_transformer('vit_base_resnet50d_224', pretrained=pretrained, **model_kwargs)
-    return model
-
-
 @register_model
 def vit_deit_tiny_patch16_224(pretrained=False, **kwargs):
     """ DeiT-tiny model @ 224x224 from paper (https://arxiv.org/abs/2012.12877).
@@ -791,4 +684,4 @@ def vit_deit_base_distilled_patch16_384(pretrained=False, **kwargs):
     model_kwargs = dict(patch_size=16, embed_dim=768, depth=12, num_heads=12, **kwargs)
     model = _create_vision_transformer(
         'vit_deit_base_distilled_patch16_384', pretrained=pretrained, distilled=True, **model_kwargs)
-    return model
+    return model

timm/models/vision_transformer_hybrid.py

@@ -0,0 +1,313 @@
+""" Hybrid Vision Transformer (ViT) in PyTorch
+
+A PyTorch implement of the Hybrid Vision Transformers as described in
+'An Image Is Worth 16 x 16 Words: Transformers for Image Recognition at Scale'
+    - https://arxiv.org/abs/2010.11929
+
+NOTE This relies on code in vision_transformer.py. The hybrid model definitions were moved here to
+keep file sizes sane.
+
+Hacked together by / Copyright 2020 Ross Wightman
+"""
+from copy import deepcopy
+from functools import partial
+
+import torch
+import torch.nn as nn
+
+from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+from .layers import StdConv2dSame, StdConv2d, to_2tuple
+from .resnet import resnet26d, resnet50d
+from .resnetv2 import ResNetV2, create_resnetv2_stem
+from .registry import register_model
+from timm.models.vision_transformer import _create_vision_transformer
+
+
+def _cfg(url='', **kwargs):
+    return {
+        'url': url,
+        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
+        'crop_pct': .9, 'interpolation': 'bicubic',
+        'mean': (0.5, 0.5, 0.5), 'std': (0.5, 0.5, 0.5),
+        'first_conv': 'patch_embed.backbone.stem.conv', 'classifier': 'head',
+        **kwargs
+    }
+
+
+default_cfgs = {
+    # hybrid in-21k models (weights ported from official Google JAX impl where they exist)
+    'vit_base_r50_s16_224_in21k': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_resnet50_224_in21k-6f7c7740.pth',
+        num_classes=21843, crop_pct=0.9),
+
+    # hybrid in-1k models (weights ported from official JAX impl)
+    'vit_base_r50_s16_384': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_resnet50_384-9fd3c705.pth',
+        input_size=(3, 384, 384), crop_pct=1.0),
+
+    # hybrid in-1k models (mostly untrained, experimental configs w/ resnetv2 stdconv backbones)
+    'vit_tiny_r_s16_p8_224': _cfg(),
+    'vit_small_r_s16_p8_224': _cfg(),
+    'vit_small_r20_s16_p2_224': _cfg(),
+    'vit_small_r20_s16_224': _cfg(),
+    'vit_small_r26_s32_224': _cfg(),
+    'vit_base_r20_s16_224': _cfg(),
+    'vit_base_r26_s32_224': _cfg(),
+    'vit_base_r50_s16_224': _cfg(),
+    'vit_large_r50_s32_224': _cfg(),
+
+    # hybrid models (using timm resnet backbones)
+    'vit_small_resnet26d_224': _cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD),
+    'vit_small_resnet50d_s16_224': _cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD),
+    'vit_base_resnet26d_224': _cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD),
+    'vit_base_resnet50d_224': _cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD),
+}
+
+
+class HybridEmbed(nn.Module):
+    """ CNN Feature Map Embedding
+    Extract feature map from CNN, flatten, project to embedding dim.
+    """
+    def __init__(self, backbone, img_size=224, patch_size=1, feature_size=None, in_chans=3, embed_dim=768):
+        super().__init__()
+        assert isinstance(backbone, nn.Module)
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.backbone = backbone
+        if feature_size is None:
+            with torch.no_grad():
+                # NOTE Most reliable way of determining output dims is to run forward pass
+                training = backbone.training
+                if training:
+                    backbone.eval()
+                o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))
+                if isinstance(o, (list, tuple)):
+                    o = o[-1]  # last feature if backbone outputs list/tuple of features
+                feature_size = o.shape[-2:]
+                feature_dim = o.shape[1]
+                backbone.train(training)
+        else:
+            feature_size = to_2tuple(feature_size)
+            if hasattr(self.backbone, 'feature_info'):
+                feature_dim = self.backbone.feature_info.channels()[-1]
+            else:
+                feature_dim = self.backbone.num_features
+        assert feature_size[0] % patch_size[0] == 0 and feature_size[1] % patch_size[1] == 0
+        self.num_patches = feature_size[0] // patch_size[0] * feature_size[1] // patch_size[1]
+        self.proj = nn.Conv2d(feature_dim, embed_dim, kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, x):
+        x = self.backbone(x)
+        if isinstance(x, (list, tuple)):
+            x = x[-1]  # last feature if backbone outputs list/tuple of features
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        return x
+
+
+def _create_vision_transformer_hybrid(variant, backbone, pretrained=False, **kwargs):
+    default_cfg = deepcopy(default_cfgs[variant])
+    embed_layer = partial(HybridEmbed, backbone=backbone)
+    kwargs.setdefault('patch_size', 1)  # default patch size for hybrid models if not set
+    return _create_vision_transformer(
+        variant, pretrained=pretrained, default_cfg=default_cfg, embed_layer=embed_layer, **kwargs)
+
+
+def _resnetv2(layers=(3, 4, 9), **kwargs):
+    """ ResNet-V2 backbone helper"""
+    padding_same = kwargs.get('padding_same', True)
+    if padding_same:
+        stem_type = 'same'
+        conv_layer = StdConv2dSame
+    else:
+        stem_type = ''
+        conv_layer = StdConv2d
+    if len(layers):
+        backbone = ResNetV2(
+            layers=layers, num_classes=0, global_pool='', in_chans=kwargs.get('in_chans', 3),
+            preact=False, stem_type=stem_type, conv_layer=conv_layer)
+    else:
+        backbone = create_resnetv2_stem(
+            kwargs.get('in_chans', 3), stem_type=stem_type, preact=False, conv_layer=conv_layer)
+    return backbone
+
+
+@register_model
+def vit_base_r50_s16_224_in21k(pretrained=False, **kwargs):
+    """ R50+ViT-B/16 hybrid model from original paper (https://arxiv.org/abs/2010.11929).
+    ImageNet-21k weights @ 224x224, source https://github.com/google-research/vision_transformer.
+    """
+    backbone = _resnetv2(layers=(3, 4, 9), **kwargs)
+    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, representation_size=768, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_base_r50_s16_224_in21k', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_base_resnet50_224_in21k(pretrained=False, **kwargs):
+    # NOTE this is forwarding to model def above for backwards compatibility
+    return vit_base_r50_s16_224_in21k(pretrained=pretrained, **kwargs)
+
+
+@register_model
+def vit_base_r50_s16_384(pretrained=False, **kwargs):
+    """ R50+ViT-B/16 hybrid from original paper (https://arxiv.org/abs/2010.11929).
+    ImageNet-1k weights fine-tuned from in21k @ 384x384, source https://github.com/google-research/vision_transformer.
+    """
+    backbone = _resnetv2((3, 4, 9), **kwargs)
+    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_base_r50_s16_384', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_base_resnet50_384(pretrained=False, **kwargs):
+    # NOTE this is forwarding to model def above for backwards compatibility
+    return vit_base_r50_s16_384(pretrained=pretrained, **kwargs)
+
+
+@register_model
+def vit_tiny_r_s16_p8_224(pretrained=False, **kwargs):
+    """ R+ViT-Ti/S16 w/ 8x8 patch hybrid @ 224 x 224.
+    """
+    backbone = _resnetv2(layers=(), **kwargs)
+    model_kwargs = dict(patch_size=8, embed_dim=192, depth=12, num_heads=3, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_tiny_r_s16_p8_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_small_r_s16_p8_224(pretrained=False, **kwargs):
+    """ R+ViT-S/S16 w/ 8x8 patch hybrid @ 224 x 224.
+    """
+    backbone = _resnetv2(layers=(), **kwargs)
+    model_kwargs = dict(patch_size=8, embed_dim=384, depth=12, num_heads=6, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_small_r_s16_p8_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+
+    return model
+
+
+@register_model
+def vit_small_r20_s16_p2_224(pretrained=False, **kwargs):
+    """ R52+ViT-S/S16 w/ 2x2 patch hybrid @ 224 x 224.
+    """
+    backbone = _resnetv2((2, 4), **kwargs)
+    model_kwargs = dict(patch_size=2, embed_dim=384, depth=12, num_heads=6, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_small_r20_s16_p2_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_small_r20_s16_224(pretrained=False, **kwargs):
+    """ R20+ViT-S/S16 hybrid.
+    """
+    backbone = _resnetv2((2, 2, 2), **kwargs)
+    model_kwargs = dict(embed_dim=384, depth=12, num_heads=6, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_small_r20_s16_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_small_r26_s32_224(pretrained=False, **kwargs):
+    """ R26+ViT-S/S32 hybrid.
+    """
+    backbone = _resnetv2((2, 2, 2, 2), **kwargs)
+    model_kwargs = dict(embed_dim=384, depth=12, num_heads=6, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_small_r26_s32_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_base_r20_s16_224(pretrained=False, **kwargs):
+    """ R20+ViT-B/S16 hybrid.
+    """
+    backbone = _resnetv2((2, 2, 2), **kwargs)
+    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_base_r20_s16_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_base_r26_s32_224(pretrained=False, **kwargs):
+    """ R26+ViT-B/S32 hybrid.
+    """
+    backbone = _resnetv2((2, 2, 2, 2), **kwargs)
+    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_base_r26_s32_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_base_r50_s16_224(pretrained=False, **kwargs):
+    """ R50+ViT-B/S16 hybrid from original paper (https://arxiv.org/abs/2010.11929).
+    """
+    backbone = _resnetv2((3, 4, 9), **kwargs)
+    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_base_r50_s16_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_large_r50_s32_224(pretrained=False, **kwargs):
+    """ R50+ViT-L/S32 hybrid.
+    """
+    backbone = _resnetv2((3, 4, 6, 3), **kwargs)
+    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_large_r50_s32_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_small_resnet26d_224(pretrained=False, **kwargs):
+    """ Custom ViT small hybrid w/ ResNet26D stride 32. No pretrained weights.
+    """
+    backbone = resnet26d(pretrained=pretrained, in_chans=kwargs.get('in_chans', 3), features_only=True, out_indices=[4])
+    model_kwargs = dict(embed_dim=768, depth=8, num_heads=8, mlp_ratio=3, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_small_resnet26d_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_small_resnet50d_s16_224(pretrained=False, **kwargs):
+    """ Custom ViT small hybrid w/ ResNet50D 3-stages, stride 16. No pretrained weights.
+    """
+    backbone = resnet50d(pretrained=pretrained, in_chans=kwargs.get('in_chans', 3), features_only=True, out_indices=[3])
+    model_kwargs = dict(embed_dim=768, depth=8, num_heads=8, mlp_ratio=3, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_small_resnet50d_s16_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_base_resnet26d_224(pretrained=False, **kwargs):
+    """ Custom ViT base hybrid w/ ResNet26D stride 32. No pretrained weights.
+    """
+    backbone = resnet26d(pretrained=pretrained, in_chans=kwargs.get('in_chans', 3), features_only=True, out_indices=[4])
+    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_base_resnet26d_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_base_resnet50d_224(pretrained=False, **kwargs):
+    """ Custom ViT base hybrid w/ ResNet50D stride 32. No pretrained weights.
+    """
+    backbone = resnet50d(pretrained=pretrained, in_chans=kwargs.get('in_chans', 3), features_only=True, out_indices=[4])
+    model_kwargs = dict(embed_dim=768, depth=12, num_heads=12, **kwargs)
+    model = _create_vision_transformer_hybrid(
+        'vit_base_resnet50d_224', backbone=backbone, pretrained=pretrained, **model_kwargs)
+    return model

timm/optim/__init__.py

@@ -10,4 +10,4 @@ from .radam import RAdam
 from .rmsprop_tf import RMSpropTF
 from .sgdp import SGDP
 
-from .optim_factory import create_optimizer
+from .optim_factory import create_optimizer, create_optimizer_v2, optimizer_kwargs

timm/optim/optim_factory.py

@@ -1,8 +1,11 @@
 """ Optimizer Factory w/ Custom Weight Decay
 Hacked together by / Copyright 2020 Ross Wightman
 """
+from typing import Optional
+
 import torch
-from torch import optim as optim
+import torch.nn as nn
+import torch.optim as optim
 
 from .adafactor import Adafactor
 from .adahessian import Adahessian
@@ -37,9 +40,63 @@ def add_weight_decay(model, weight_decay=1e-5, skip_list=()):
         {'params': decay, 'weight_decay': weight_decay}]
 
 
+def optimizer_kwargs(cfg):
+    """ cfg/argparse to kwargs helper
+    Convert optimizer args in argparse args or cfg like object to keyword args for updated create fn.
+    """
+    kwargs = dict(
+        optimizer_name=cfg.opt,
+        learning_rate=cfg.lr,
+        weight_decay=cfg.weight_decay,
+        momentum=cfg.momentum)
+    if getattr(cfg, 'opt_eps', None) is not None:
+        kwargs['eps'] = cfg.opt_eps
+    if getattr(cfg, 'opt_betas', None) is not None:
+        kwargs['betas'] = cfg.opt_betas
+    if getattr(cfg, 'opt_args', None) is not None:
+        kwargs.update(cfg.opt_args)
+    return kwargs
+
+
 def create_optimizer(args, model, filter_bias_and_bn=True):
-    opt_lower = args.opt.lower()
-    weight_decay = args.weight_decay
+    """ Legacy optimizer factory for backwards compatibility.
+    NOTE: Use create_optimizer_v2 for new code.
+    """
+    return create_optimizer_v2(
+        model,
+        **optimizer_kwargs(cfg=args),
+        filter_bias_and_bn=filter_bias_and_bn,
+    )
+
+
+def create_optimizer_v2(
+        model: nn.Module,
+        optimizer_name: str = 'sgd',
+        learning_rate: Optional[float] = None,
+        weight_decay: float = 0.,
+        momentum: float = 0.9,
+        filter_bias_and_bn: bool = True,
+        **kwargs):
+    """ Create an optimizer.
+
+    TODO currently the model is passed in and all parameters are selected for optimization.
+    For more general use an interface that allows selection of parameters to optimize and lr groups, one of:
+      * a filter fn interface that further breaks params into groups in a weight_decay compatible fashion
+      * expose the parameters interface and leave it up to caller
+
+    Args:
+        model (nn.Module): model containing parameters to optimize
+        optimizer_name: name of optimizer to create
+        learning_rate: initial learning rate
+        weight_decay: weight decay to apply in optimizer
+        momentum:  momentum for momentum based optimizers (others may use betas via kwargs)
+        filter_bias_and_bn:  filter out bias, bn and other 1d params from weight decay
+        **kwargs: extra optimizer specific kwargs to pass through
+
+    Returns:
+        Optimizer
+    """
+    opt_lower = optimizer_name.lower()
     if weight_decay and filter_bias_and_bn:
         skip = {}
         if hasattr(model, 'no_weight_decay'):
@@ -48,26 +105,18 @@ def create_optimizer(args, model, filter_bias_and_bn=True):
         weight_decay = 0.
     else:
         parameters = model.parameters()
-
     if 'fused' in opt_lower:
         assert has_apex and torch.cuda.is_available(), 'APEX and CUDA required for fused optimizers'
 
-    opt_args = dict(lr=args.lr, weight_decay=weight_decay)
-    if hasattr(args, 'opt_eps') and args.opt_eps is not None:
-        opt_args['eps'] = args.opt_eps
-    if hasattr(args, 'opt_betas') and args.opt_betas is not None:
-        opt_args['betas'] = args.opt_betas
-    if hasattr(args, 'opt_args') and args.opt_args is not None:
-        opt_args.update(args.opt_args)
-
+    opt_args = dict(lr=learning_rate, weight_decay=weight_decay, **kwargs)
     opt_split = opt_lower.split('_')
     opt_lower = opt_split[-1]
     if opt_lower == 'sgd' or opt_lower == 'nesterov':
         opt_args.pop('eps', None)
-        optimizer = optim.SGD(parameters, momentum=args.momentum, nesterov=True, **opt_args)
+        optimizer = optim.SGD(parameters, momentum=momentum, nesterov=True, **opt_args)
     elif opt_lower == 'momentum':
         opt_args.pop('eps', None)
-        optimizer = optim.SGD(parameters, momentum=args.momentum, nesterov=False, **opt_args)
+        optimizer = optim.SGD(parameters, momentum=momentum, nesterov=False, **opt_args)
     elif opt_lower == 'adam':
         optimizer = optim.Adam(parameters, **opt_args)
     elif opt_lower == 'adamw':
@@ -78,30 +127,30 @@ def create_optimizer(args, model, filter_bias_and_bn=True):
         optimizer = RAdam(parameters, **opt_args)
     elif opt_lower == 'adamp':        
         optimizer = AdamP(parameters, wd_ratio=0.01, nesterov=True, **opt_args)
-    elif opt_lower == 'sgdp':        
-        optimizer = SGDP(parameters, momentum=args.momentum, nesterov=True, **opt_args)
+    elif opt_lower == 'sgdp':
+        optimizer = SGDP(parameters, momentum=momentum, nesterov=True, **opt_args)
     elif opt_lower == 'adadelta':
         optimizer = optim.Adadelta(parameters, **opt_args)
     elif opt_lower == 'adafactor':
-        if not args.lr:
+        if not learning_rate:
             opt_args['lr'] = None
         optimizer = Adafactor(parameters, **opt_args)
     elif opt_lower == 'adahessian':
         optimizer = Adahessian(parameters, **opt_args)
     elif opt_lower == 'rmsprop':
-        optimizer = optim.RMSprop(parameters, alpha=0.9, momentum=args.momentum, **opt_args)
+        optimizer = optim.RMSprop(parameters, alpha=0.9, momentum=momentum, **opt_args)
     elif opt_lower == 'rmsproptf':
-        optimizer = RMSpropTF(parameters, alpha=0.9, momentum=args.momentum, **opt_args)
+        optimizer = RMSpropTF(parameters, alpha=0.9, momentum=momentum, **opt_args)
     elif opt_lower == 'novograd':
         optimizer = NovoGrad(parameters, **opt_args)
     elif opt_lower == 'nvnovograd':
         optimizer = NvNovoGrad(parameters, **opt_args)
     elif opt_lower == 'fusedsgd':
         opt_args.pop('eps', None)
-        optimizer = FusedSGD(parameters, momentum=args.momentum, nesterov=True, **opt_args)
+        optimizer = FusedSGD(parameters, momentum=momentum, nesterov=True, **opt_args)
     elif opt_lower == 'fusedmomentum':
         opt_args.pop('eps', None)
-        optimizer = FusedSGD(parameters, momentum=args.momentum, nesterov=False, **opt_args)
+        optimizer = FusedSGD(parameters, momentum=momentum, nesterov=False, **opt_args)
     elif opt_lower == 'fusedadam':
         optimizer = FusedAdam(parameters, adam_w_mode=False, **opt_args)
     elif opt_lower == 'fusedadamw':

timm/version.py

@@ -1 +1 @@
-__version__ = '0.4.6'
+__version__ = '0.4.7'

train.py

@@ -33,7 +33,7 @@ from timm.models import create_model, safe_model_name, resume_checkpoint, load_c
     convert_splitbn_model, model_parameters
 from timm.utils import *
 from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy, JsdCrossEntropy
-from timm.optim import create_optimizer
+from timm.optim import create_optimizer_v2, optimizer_kwargs
 from timm.scheduler import create_scheduler
 from timm.utils import ApexScaler, NativeScaler
 
@@ -142,6 +142,8 @@ 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('--epoch-repeats', type=float, default=0., metavar='N',
+                    help='epoch repeat multiplier (number of times to repeat dataset epoch per train epoch).')
 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',
@@ -258,6 +260,8 @@ 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('--experiment', default='', type=str, metavar='NAME',
+                    help='name of train experiment, name of sub-folder for output')
 parser.add_argument('--eval-metric', default='top1', type=str, metavar='EVAL_METRIC',
                     help='Best metric (default: "top1"')
 parser.add_argument('--tta', type=int, default=0, metavar='N',
@@ -385,7 +389,7 @@ def main():
         assert not args.sync_bn, 'Cannot use SyncBatchNorm with torchscripted model'
         model = torch.jit.script(model)
 
-    optimizer = create_optimizer(args, model)
+    optimizer = create_optimizer_v2(model, **optimizer_kwargs(cfg=args))
 
     # setup automatic mixed-precision (AMP) loss scaling and op casting
     amp_autocast = suppress  # do nothing
@@ -451,7 +455,9 @@ def main():
 
     # create the train and eval datasets
     dataset_train = create_dataset(
-        args.dataset, root=args.data_dir, split=args.train_split, is_training=True, batch_size=args.batch_size)
+        args.dataset,
+        root=args.data_dir, split=args.train_split, is_training=True,
+        batch_size=args.batch_size, repeats=args.epoch_repeats)
     dataset_eval = create_dataset(
         args.dataset, root=args.data_dir, split=args.val_split, is_training=False, batch_size=args.batch_size)
 
@@ -541,13 +547,15 @@ def main():
     saver = None
     output_dir = ''
     if args.local_rank == 0:
-        output_base = args.output if args.output else './output'
-        exp_name = '-'.join([
-            datetime.now().strftime("%Y%m%d-%H%M%S"),
-            safe_model_name(args.model),
-            str(data_config['input_size'][-1])
-        ])
-        output_dir = get_outdir(output_base, 'train', exp_name)
+        if args.experiment:
+            exp_name = args.experiment
+        else:
+            exp_name = '-'.join([
+                datetime.now().strftime("%Y%m%d-%H%M%S"),
+                safe_model_name(args.model),
+                str(data_config['input_size'][-1])
+            ])
+        output_dir = get_outdir(args.output if args.output else './output/train', exp_name)
         decreasing = True if eval_metric == 'loss' else False
         saver = CheckpointSaver(
             model=model, optimizer=optimizer, args=args, model_ema=model_ema, amp_scaler=loss_scaler,

validate.py

@@ -152,7 +152,7 @@ def validate(args):
     param_count = sum([m.numel() for m in model.parameters()])
     _logger.info('Model %s created, param count: %d' % (args.model, param_count))
 
-    data_config = resolve_data_config(vars(args), model=model, use_test_size=True)
+    data_config = resolve_data_config(vars(args), model=model, use_test_size=True, verbose=True)
     test_time_pool = False
     if not args.no_test_pool:
         model, test_time_pool = apply_test_time_pool(model, data_config, use_test_size=True)