Refine Transformer

xautodl/xmodels/__init__.py

@@ -4,4 +4,4 @@
 # The models in this folder is written with xlayers #
 #####################################################
 
-from .transformers import get_transformer
+from .core import *

xautodl/xmodels/core.py

@@ -0,0 +1,67 @@
+#######################################################
+# Use module in xlayers to construct different models #
+#######################################################
+from typing import List, Text, Dict, Any
+import torch
+
+__all__ = ["get_model"]
+
+
+from xautodl.xlayers.super_core import SuperSequential
+from xautodl.xlayers.super_core import SuperLinear
+from xautodl.xlayers.super_core import SuperDropout
+from xautodl.xlayers.super_core import super_name2norm
+from xautodl.xlayers.super_core import super_name2activation
+
+
+def get_model(config: Dict[Text, Any], **kwargs):
+    model_type = config.get("model_type", "simple_mlp").lower()
+    if model_type == "simple_mlp":
+        act_cls = super_name2activation[kwargs["act_cls"]]
+        norm_cls = super_name2norm[kwargs["norm_cls"]]
+        mean, std = kwargs.get("mean", None), kwargs.get("std", None)
+        if "hidden_dim" in kwargs:
+            hidden_dim1 = kwargs.get("hidden_dim")
+            hidden_dim2 = kwargs.get("hidden_dim")
+        else:
+            hidden_dim1 = kwargs.get("hidden_dim1", 200)
+            hidden_dim2 = kwargs.get("hidden_dim2", 100)
+        model = SuperSequential(
+            norm_cls(mean=mean, std=std),
+            SuperLinear(kwargs["input_dim"], hidden_dim1),
+            act_cls(),
+            SuperLinear(hidden_dim1, hidden_dim2),
+            act_cls(),
+            SuperLinear(hidden_dim2, kwargs["output_dim"]),
+        )
+    elif model_type == "norm_mlp":
+        act_cls = super_name2activation[kwargs["act_cls"]]
+        norm_cls = super_name2norm[kwargs["norm_cls"]]
+        sub_layers, last_dim = [], kwargs["input_dim"]
+        for i, hidden_dim in enumerate(kwargs["hidden_dims"]):
+            sub_layers.append(SuperLinear(last_dim, hidden_dim))
+            if hidden_dim > 1:
+                sub_layers.append(norm_cls(hidden_dim, elementwise_affine=False))
+            sub_layers.append(act_cls())
+            last_dim = hidden_dim
+        sub_layers.append(SuperLinear(last_dim, kwargs["output_dim"]))
+        model = SuperSequential(*sub_layers)
+    elif model_type == "dual_norm_mlp":
+        act_cls = super_name2activation[kwargs["act_cls"]]
+        norm_cls = super_name2norm[kwargs["norm_cls"]]
+        sub_layers, last_dim = [], kwargs["input_dim"]
+        for i, hidden_dim in enumerate(kwargs["hidden_dims"]):
+            if i > 0:
+                sub_layers.append(norm_cls(last_dim, elementwise_affine=False))
+            sub_layers.append(SuperLinear(last_dim, hidden_dim))
+            sub_layers.append(SuperDropout(kwargs["dropout"]))
+            sub_layers.append(SuperLinear(hidden_dim, hidden_dim))
+            sub_layers.append(act_cls())
+            last_dim = hidden_dim
+        sub_layers.append(SuperLinear(last_dim, kwargs["output_dim"]))
+        model = SuperSequential(*sub_layers)
+    elif model_type == "quant_transformer":
+        raise NotImplementedError
+    else:
+        raise TypeError("Unkonwn model type: {:}".format(model_type))
+    return model

xautodl/xmodels/transformers.py

@@ -20,20 +20,6 @@ def pair(t):
     return t if isinstance(t, tuple) else (t, t)
 
 
-def _init_weights(m):
-    if isinstance(m, nn.Linear):
-        weight_init.trunc_normal_(m.weight, std=0.02)
-        if isinstance(m, nn.Linear) and m.bias is not None:
-            nn.init.constant_(m.bias, 0)
-    elif isinstance(m, xlayers.SuperLinear):
-        weight_init.trunc_normal_(m._super_weight, std=0.02)
-        if m._super_bias is not None:
-            nn.init.constant_(m._super_bias, 0)
-    elif isinstance(m, xlayers.SuperLayerNorm1D):
-        nn.init.constant_(m.weight, 1.0)
-        nn.init.constant_(m.bias, 0)
-
-
 name2config = {
     "vit-cifar10-p4-d4-h4-c32": dict(
         type="vit",
@@ -155,7 +141,7 @@ class SuperViT(xlayers.SuperModule):
         )
 
         weight_init.trunc_normal_(self.cls_token, std=0.02)
-        self.apply(_init_weights)
+        self.apply(weight_init.init_transformer)
 
     @property
     def abstract_search_space(self):

xautodl/xmodels/transformers_quantum.py

@@ -0,0 +1,124 @@
+#####################################################
+# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.06 #
+#####################################################
+# Vision Transformer: arxiv.org/pdf/2010.11929.pdf  #
+#####################################################
+import copy, math
+from functools import partial
+from typing import Optional, Text, List
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from xautodl import spaces
+from xautodl import xlayers
+from xautodl.xlayers import weight_init
+
+
+class SuperQuaT(xlayers.SuperModule):
+    """The super transformer for transformer."""
+
+    def __init__(
+        self,
+        image_size,
+        patch_size,
+        num_classes,
+        dim,
+        depth,
+        heads,
+        mlp_multiplier=4,
+        channels=3,
+        dropout=0.0,
+        att_dropout=0.0,
+    ):
+        super(SuperQuaT, self).__init__()
+        image_height, image_width = pair(image_size)
+        patch_height, patch_width = pair(patch_size)
+
+        if image_height % patch_height != 0 or image_width % patch_width != 0:
+            raise ValueError("Image dimensions must be divisible by the patch size.")
+
+        num_patches = (image_height // patch_height) * (image_width // patch_width)
+        patch_dim = channels * patch_height * patch_width
+        self.to_patch_embedding = xlayers.SuperSequential(
+            xlayers.SuperReArrange(
+                "b c (h p1) (w p2) -> b (h w) (p1 p2 c)",
+                p1=patch_height,
+                p2=patch_width,
+            ),
+            xlayers.SuperLinear(patch_dim, dim),
+        )
+
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
+        self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
+        self.dropout = nn.Dropout(dropout)
+
+        # build the transformer encode layers
+        layers = []
+        for ilayer in range(depth):
+            layers.append(
+                xlayers.SuperTransformerEncoderLayer(
+                    dim,
+                    heads,
+                    False,
+                    mlp_multiplier,
+                    dropout=dropout,
+                    att_dropout=att_dropout,
+                )
+            )
+        self.backbone = xlayers.SuperSequential(*layers)
+        self.cls_head = xlayers.SuperSequential(
+            xlayers.SuperLayerNorm1D(dim), xlayers.SuperLinear(dim, num_classes)
+        )
+
+        weight_init.trunc_normal_(self.cls_token, std=0.02)
+        self.apply(_init_weights)
+
+    @property
+    def abstract_search_space(self):
+        raise NotImplementedError
+
+    def apply_candidate(self, abstract_child: spaces.VirtualNode):
+        super(SuperQuaT, self).apply_candidate(abstract_child)
+        raise NotImplementedError
+
+    def forward_candidate(self, input: torch.Tensor) -> torch.Tensor:
+        raise NotImplementedError
+
+    def forward_raw(self, input: torch.Tensor) -> torch.Tensor:
+        tensors = self.to_patch_embedding(input)
+        batch, seq, _ = tensors.shape
+
+        cls_tokens = self.cls_token.expand(batch, -1, -1)
+        feats = torch.cat((cls_tokens, tensors), dim=1)
+        feats = feats + self.pos_embedding[:, : seq + 1, :]
+        feats = self.dropout(feats)
+
+        feats = self.backbone(feats)
+
+        x = feats[:, 0]  # the features for cls-token
+
+        return self.cls_head(x)
+
+
+def get_transformer(config):
+    if isinstance(config, str) and config.lower() in name2config:
+        config = name2config[config.lower()]
+    if not isinstance(config, dict):
+        raise ValueError("Invalid Configuration: {:}".format(config))
+    model_type = config.get("type", "vit").lower()
+    if model_type == "vit":
+        model = SuperQuaT(
+            image_size=config.get("image_size"),
+            patch_size=config.get("patch_size"),
+            num_classes=config.get("num_classes"),
+            dim=config.get("dim"),
+            depth=config.get("depth"),
+            heads=config.get("heads"),
+            dropout=config.get("dropout"),
+            att_dropout=config.get("att_dropout"),
+        )
+    else:
+        raise ValueError("Unknown model type: {:}".format(model_type))
+    return model