Lazy Config

.isort.cfg

@@ -1,5 +1,5 @@
 [settings]
-known_third_party = cv2,einops,numpy,setuptools,torch,torchvision
+known_third_party = cloudpickle,cv2,einops,hydra,numpy,omegaconf,pytest,setuptools,torch,torchvision,yaml
 multi_line_output=3
 include_trailing_comma=True
 force_grid_wrap=0

configs/Swin/readme.md

@@ -0,0 +1,2 @@
+#Config File for Swin Transformer
+

configs/Swin/swin_base_patch4_window7_224.py

@@ -0,0 +1,22 @@
+from vformer.config import LazyCall as L
+from vformer.models import SwinTransformer
+
+model = L(SwinTransformer)(
+    img_size=224,
+    in_channels=3,
+    patch_size=4,
+    n_classes=1000,
+    embedding_dim=128,
+    depths=(2, 2, 18, 2),
+    num_heads=(4, 8, 16, 32),
+    window_size=7,
+    mlp_ratio=4,
+    qkv_bias=True,
+    qk_scale=None,
+    p_dropout=0.1,
+    attn_dropout=0.1,
+    drop_path_rate=0.1,
+    ape=True,
+    decoder_config=None,
+    patch_norm=True,
+)

configs/Vanilla/readme.md

@@ -0,0 +1 @@
+#Config file for VanillaViT

configs/Vanilla/vit_tiny_patch_16_224.py

@@ -0,0 +1,18 @@
+from vformer.config import LazyCall as L
+from vformer.models import VanillaViT
+
+model = L(VanillaViT)(
+    img_size=224,
+    in_channels=3,
+    patch_size=16,
+    embedding_dim=192,
+    head_dim=192,
+    depth=12,
+    attn_heads=3,
+    encoder_mlp_dim=192,
+    decoder_config=None,
+    pool="cls",
+    p_dropout_encoder=0.1,
+    p_dropout_embedding=0.1,
+    n_classes=1000,
+)

requirements.txt

@@ -2,15 +2,18 @@ arrow==1.1.1
 attrs==21.2.0
 backports.entry-points-selectable==1.1.0
 binaryornot==0.4.4
+black==22.3.0
 certifi==2021.5.30
 cfgv==3.3.1
 chardet==4.0.0
 charset-normalizer==2.0.4
 click==8.0.1
+cloudpickle==2.0.0
 cookiecutter==1.7.3
 distlib==0.3.2
 einops==0.3.2
 filelock==3.0.12
+hydra-core==1.1.1
 identify==2.2.13
 idna==3.2
 iniconfig==1.1.1
@@ -19,6 +22,7 @@ jinja2-time==0.2.0
 MarkupSafe==2.0.1
 nodeenv==1.6.0
 olefile
+omegaconf
 packaging==21.0
 Pillow
 platformdirs==2.3.0
@@ -40,3 +44,5 @@ torchvision>=0.11.0
 typing-extensions
 urllib3==1.26.6
 virtualenv==20.7.2
+PYyaml==5.4.1
+

setup.py

@@ -1,5 +1,7 @@
 import codecs
+import glob
 import os
+import shutil
 
 from setuptools import find_packages, setup
 

tests/dir1/dir1_a.py

@@ -0,0 +1,2 @@
+dir1a_str = "base_a_1"
+dir1a_dict = {"a": 1, "b": 2}

tests/dir1/dir1_b.py

@@ -0,0 +1,10 @@
+from vformer.config import LazyConfig
+
+# equivalent to relative import
+dir1a_str, dir1a_dict = LazyConfig.load_rel("dir1_a.py", ("dir1a_str", "dir1a_dict"))
+
+dir1b_str = dir1a_str + "_from_b"
+dir1b_dict = dir1a_dict
+
+# Every import is a reload: not modified by other config files
+assert dir1a_dict.a == 1

tests/root_cfg.py

@@ -0,0 +1,12 @@
+from itertools import count
+
+from vformer.config import LazyCall as L
+
+from .dir1.dir1_a import dir1a_dict, dir1a_str
+
+dir1a_dict.a = "modified"
+
+# modification above won't affect future imports
+from .dir1.dir1_b import dir1b_dict, dir1b_str
+
+lazyobj = L(count)(x=dir1a_str, y=dir1b_str)

tests/test_config.py

@@ -0,0 +1,192 @@
+import os
+import tempfile
+from itertools import count
+
+import pytest
+import torch
+from omegaconf import DictConfig
+
+from vformer.config import LazyCall
+from vformer.config import LazyCall as L
+from vformer.config import LazyConfig, instantiate
+from vformer.models import PVTSegmentation, SwinTransformer, VanillaViT, ViViTModel2
+
+
+def test_lazy():
+    # classification models
+    vanilla_config = LazyCall(VanillaViT)(img_size=224, patch_size=7, n_classes=10)
+    swin_config = LazyCall(SwinTransformer)(
+        img_size=224,
+        patch_size=4,
+        in_channels=3,
+        n_classes=10,
+        embedding_dim=96,
+        depths=[2, 2, 6, 2],
+        num_heads=[3, 6, 12, 24],
+        window_size=7,
+        p_dropout=0.2,
+    )
+    vivit_config = LazyCall(ViViTModel2)(
+        img_size=224,
+        in_channels=3,
+        patch_size=16,
+        embedding_dim=192,
+        depth=4,
+        num_heads=3,
+        head_dim=64,
+        num_frames=1,
+        n_classes=10,
+    )
+
+    # dense models
+    pvt_config = LazyCall(PVTSegmentation)()
+    pvt_config["img_size"] = 224
+    rand_img_tensor = torch.randn(4, 3, 224, 224)
+    rand_vdo_tensor = torch.randn([32, 16, 3, 224, 224])
+
+    vanilla_vit = instantiate(vanilla_config)
+    swin_vit = instantiate(swin_config)
+    vivit = instantiate(vivit_config)
+
+    pvt = instantiate(pvt_config)
+
+    assert vanilla_vit(rand_img_tensor).shape == (4, 10)
+    assert swin_vit(rand_img_tensor).shape == (4, 10)
+    assert pvt(rand_img_tensor).shape == (4, 1, 224, 224)
+    assert vivit(rand_vdo_tensor).shape == (32, 10)
+
+
+def test_raise_errors():
+    a = "strings"
+    with pytest.raises(TypeError):
+        LazyConfig(a)
+    with pytest.raises(TypeError):
+        LazyCall(2)
+
+    cfg = [1, 2, 3, 4]
+    cfg2 = instantiate(cfg)
+    assert cfg2 == cfg, "it should return same object"
+
+    with pytest.raises(AssertionError):
+        instantiate({"_target_": "test"})
+
+
+def test_load():
+    root_filename = os.path.join(os.path.dirname(__file__), "root_cfg.py")
+    cfg = LazyConfig.load(root_filename)
+
+    assert cfg.dir1a_dict.a == "modified"
+
+    assert cfg.dir1b_dict.a == 1
+    assert cfg.lazyobj.x == "base_a_1"
+
+    cfg.lazyobj.x = "new_x"
+    # reload
+    cfg = LazyConfig.load(root_filename)
+    assert cfg.lazyobj.x == "base_a_1"
+
+
+def test_save_load():
+    root_filename = os.path.join(os.path.dirname(__file__), "root_cfg.py")
+
+    cfg = LazyConfig.load(root_filename)
+    with tempfile.TemporaryDirectory(prefix="vformer") as d:
+        fname = os.path.join(d, "test_config.yaml")
+        LazyConfig.save(cfg, fname)
+        cfg2 = LazyConfig.load(fname)
+
+    assert cfg2.lazyobj._target_ == "itertools.count"
+    assert cfg.lazyobj._target_ == count
+    cfg2.lazyobj.pop("_target_")
+    cfg.lazyobj.pop("_target_")
+    # the rest are equal
+    assert cfg == cfg2
+
+
+def test_failed_save():
+    cfg = DictConfig({"x": lambda: 3}, flags={"allow_objects": True})
+    with tempfile.TemporaryDirectory(prefix="vformer") as d:
+        fname = os.path.join(d, "test_config.yaml")
+        LazyConfig.save(cfg, fname)
+        assert os.path.exists(fname) == True
+        assert os.path.exists(fname + ".pkl") == True
+
+
+def test_overrides():
+    root_filename = os.path.join(os.path.dirname(__file__), "root_cfg.py")
+
+    cfg = LazyConfig.load(root_filename)
+    LazyConfig.apply_overrides(cfg, ["lazyobj.x=123", 'dir1b_dict.a="123"'])
+    assert cfg.dir1b_dict.a == "123"
+    assert cfg.lazyobj.x == 123
+
+
+def test_invalid_overrides():
+    root_filename = os.path.join(os.path.dirname(__file__), "root_cfg.py")
+
+    cfg = LazyConfig.load(root_filename)
+    with pytest.raises(KeyError):
+        LazyConfig.apply_overrides(cfg, ["lazyobj.x.xxx=123"])
+
+
+def test_to_py():
+    root_filename = os.path.join(os.path.dirname(__file__), "root_cfg.py")
+
+    cfg = LazyConfig.load(root_filename)
+    cfg.lazyobj.x = {
+        "a": 1,
+        "b": 2,
+        "c": L(count)(x={"r": "a", "s": 2.4, "t": [1, 2, 3, "z"]}),
+    }
+    cfg.list = ["a", 1, "b", 3.2]
+    py_str = LazyConfig.to_py(cfg)
+    expected = """cfg.dir1a_dict.a = "modified"
+cfg.dir1a_dict.b = 2
+cfg.dir1b_dict.a = 1
+cfg.dir1b_dict.b = 2
+cfg.lazyobj = itertools.count(
+    x={
+        "a": 1,
+        "b": 2,
+        "c": itertools.count(x={"r": "a", "s": 2.4, "t": [1, 2, 3, "z"]}),
+    },
+    y="base_a_1_from_b",
+)
+cfg.list = ["a", 1, "b", 3.2]
+"""
+    assert py_str == expected
+
+    root_filename = os.path.join(os.path.dirname(__file__), "testing.yaml")
+    cfg = LazyConfig.load(root_filename)
+    obj = LazyConfig.to_py(cfg)
+
+
+def test_check_configs():
+    config_dir = os.path.join(
+        os.path.dirname(os.path.dirname(__file__)),
+        "configs",
+        "Vanilla",
+        "vit_tiny_patch_16_224.py",
+    )
+
+    cfg = LazyConfig.load(config_dir)
+    cfg.model.img_size = 224
+    cfg.model.in_channels = 3
+    cfg.model.n_classes = 1000
+
+    new_model = instantiate(cfg.model)
+    assert new_model(torch.randn(4, 3, 224, 224)).shape == (4, 1000)
+
+    cfg.model.num_classes = 10
+    with pytest.raises(TypeError):
+        new_model = instantiate((cfg.model))
+
+    config_dir = os.path.join(
+        os.path.dirname(os.path.dirname(__file__)),
+        "configs",
+        "Swin",
+        "swin_base_patch4_window7_224.py",
+    )
+    cfg = LazyConfig.load(config_dir)
+    new_model = instantiate(cfg.model)
+    assert new_model(torch.randn(4, 3, 224, 224)).shape == (4, 1000)

tests/testing.yaml

@@ -0,0 +1,2 @@
+x : 3
+y : 4

vformer/attention/cross.py

@@ -29,6 +29,7 @@ def __init__(self, cls_dim, patch_dim, num_heads=8, head_dim=64):
         inner_dim = num_heads * head_dim
         self.num_heads = num_heads
         self.scale = head_dim**-0.5
+
         self.fl = (
             nn.Linear(cls_dim, patch_dim) if not cls_dim == patch_dim else nn.Identity()
         )

vformer/config/__init__.py

@@ -0,0 +1,2 @@
+from .config_utils import instantiate
+from .lazy import LazyCall, LazyConfig