Update xlayers

CHANGE-LOG.md

@@ -1,12 +1,12 @@
 # This file shows the major updates of this repo.
 
-- [2020.04.11] [4ef9531] Add change log as `CHANGE-LOG.md`.
-- [2019.12.20] [69ca086] Release NAS-Bench-201.
-- [2019.09.28] [f8f3f38] TAS and SETN codes were publicly released.
-- [2019.01.31] [13e908f] GDAS codes were publicly released.
-- [2020.07.01] [a45808b] Upgrade NAS-API to the 2.0 version.
-- [2020.09.16] [7052265] Create NATS-BENCH.
+- [2020.04.11] [4ef9531](https://github.com/D-X-Y/AutoDL-Projects/tree/4ef9531) Add change log as `CHANGE-LOG.md`.
+- [2019.12.20] [69ca086](https://github.com/D-X-Y/AutoDL-Projects/tree/69ca086) Release NAS-Bench-201.
+- [2019.09.28] [f8f3f38](https://github.com/D-X-Y/AutoDL-Projects/tree/f8f3f38) TAS and SETN codes were publicly released.
+- [2019.01.31] [13e908f](https://github.com/D-X-Y/AutoDL-Projects/tree/13e908f) GDAS codes were publicly released.
+- [2020.07.01] [a45808b](https://github.com/D-X-Y/AutoDL-Projects/tree/a45808b) Upgrade NAS-API to the 2.0 version.
+- [2020.09.16] [7052265](https://github.com/D-X-Y/AutoDL-Projects/tree/7052265) Create NATS-BENCH.
 - [2020.10.15] [446262a](https://github.com/D-X-Y/AutoDL-Projects/tree/446262a) Update NATS-BENCH to version 1.0
 - [2020.12.20] [dae387a](https://github.com/D-X-Y/AutoDL-Projects/tree/dae387a) Update NATS-BENCH to version 1.1
 - [2021.05.18] [98fadf8](https://github.com/D-X-Y/AutoDL-Projects/tree/98fadf8) Before moving to `xautodl`
-- [2021.05.21] [b4e8eae](https://github.com/D-X-Y/AutoDL-Projects/tree/b4e8eae) `xautodl` is close to ready
+- [2021.05.21] [5b09f05](https://github.com/D-X-Y/AutoDL-Projects/tree/5b09f05) `xautodl` is close to ready

exps/LFNA/lfna.py

@@ -106,8 +106,13 @@ def pretrain_v2(base_model, meta_model, criterion, xenv, args, logger):
     logger.log("Using the optimizer: {:}".format(optimizer))
 
     meta_model.set_best_dir(logger.path(None) / "ckps-pretrain-v2")
+    final_best_name = "final-pretrain-{:}.pth".format(args.rand_seed)
+    if meta_model.has_best(final_best_name):
+        meta_model.load_best(final_best_name)
+        return
+
     meta_model.set_best_name("pretrain-{:}.pth".format(args.rand_seed))
-    last_success_epoch = 0
+    last_success_epoch, early_stop_thresh = 0, args.pretrain_early_stop_thresh
     per_epoch_time, start_time = AverageMeter(), time.time()
     for iepoch in range(args.epochs):
         left_time = "Time Left: {:}".format(
@@ -164,14 +169,21 @@ def pretrain_v2(base_model, meta_model, criterion, xenv, args, logger):
             )
             + ", batch={:}".format(len(total_meta_losses))
             + ", success={:}, best_score={:.4f}".format(success, -best_score)
-            + " {:}".format(left_time)
+            + ", LS={:}/{:}".format(last_success_epoch, early_stop_thresh)
+            + ", {:}".format(left_time)
         )
-        if iepoch - last_success_epoch >= args.early_stop_thresh * 5:
+        if success:
+            last_success_epoch = iepoch
+        if iepoch - last_success_epoch >= early_stop_thresh:
             logger.log("Early stop the pre-training at {:}".format(iepoch))
             break
         per_epoch_time.update(time.time() - start_time)
         start_time = time.time()
     meta_model.load_best()
+    # save to the final model
+    meta_model.set_best_name(final_best_name)
+    success, _ = meta_model.save_best(best_score + 1e-6)
+    assert success
 
 
 def pretrain_v1(base_model, meta_model, criterion, xenv, args, logger):
@@ -189,7 +201,7 @@ def pretrain_v1(base_model, meta_model, criterion, xenv, args, logger):
     meta_model.set_best_dir(logger.path(None) / "ckps-pretrain-v1")
     meta_model.set_best_name("pretrain-{:}.pth".format(args.rand_seed))
     per_epoch_time, start_time = AverageMeter(), time.time()
-    last_success_epoch = 0
+    last_success_epoch, early_stop_thresh = 0, args.pretrain_early_stop_thresh
     for iepoch in range(args.epochs):
         left_time = "Time Left: {:}".format(
             convert_secs2time(per_epoch_time.avg * (args.epochs - iepoch), True)
@@ -232,9 +244,12 @@ def pretrain_v1(base_model, meta_model, criterion, xenv, args, logger):
             )
             + ", batch={:}".format(len(losses))
             + ", success={:}, best_score={:.4f}".format(success, -best_score)
+            + ", LS={:}/{:}".format(last_success_epoch, early_stop_thresh)
             + " {:}".format(left_time)
         )
-        if iepoch - last_success_epoch >= args.early_stop_thresh * 5:
+        if success:
+            last_success_epoch = iepoch
+        if iepoch - last_success_epoch >= early_stop_thresh:
             logger.log("Early stop the pre-training at {:}".format(iepoch))
             break
         per_epoch_time.update(time.time() - start_time)
@@ -521,7 +536,7 @@ if __name__ == "__main__":
     parser.add_argument(
         "--refine_lr",
         type=float,
-        default=0.005,
+        default=0.001,
         help="The learning rate for the optimizer, during refine",
     )
     parser.add_argument(
@@ -533,6 +548,12 @@ if __name__ == "__main__":
         default=20,
         help="The #epochs for early stop.",
     )
+    parser.add_argument(
+        "--pretrain_early_stop_thresh",
+        type=int,
+        default=200,
+        help="The #epochs for early stop.",
+    )
     parser.add_argument(
         "--seq_length", type=int, default=10, help="The sequence length."
     )

exps/LFNA/lfna_meta_model.py

@@ -70,6 +70,7 @@ class LFNA_Meta(super_core.SuperModule):
                     dropout,
                     norm_affine=False,
                     order=super_core.LayerOrder.PostNorm,
+                    use_mask=True,
                 )
             )
         layers.append(super_core.SuperLinear(time_embedding * 2, time_embedding))
@@ -162,11 +163,14 @@ class LFNA_Meta(super_core.SuperModule):
     def _obtain_time_embed(self, timestamps):
         # timestamps is a batch of sequence of timestamps
         batch, seq = timestamps.shape
+        meta_timestamps, meta_embeds = self.meta_timestamps, self.super_meta_embed
         timestamp_q_embed = self._tscalar_embed(timestamps)
-        timestamp_k_embed = self._tscalar_embed(self.meta_timestamps.view(1, -1))
-        timestamp_v_embed = self.super_meta_embed.unsqueeze(dim=0)
+        timestamp_k_embed = self._tscalar_embed(meta_timestamps.view(1, -1))
+        timestamp_v_embed = meta_embeds.unsqueeze(dim=0)
+        # create the mask
+        mask = torch.unsqueeze(timestamps, dim=-1) <= meta_timestamps.view(1, 1, -1)
         timestamp_embeds = self._trans_att(
-            timestamp_q_embed, timestamp_k_embed, timestamp_v_embed
+            timestamp_q_embed, timestamp_k_embed, timestamp_v_embed, mask
         )
         # relative_timestamps = timestamps - timestamps[:, :1]
         # relative_pos_embeds = self._tscalar_embed(relative_timestamps)
@@ -186,8 +190,12 @@ class LFNA_Meta(super_core.SuperModule):
         layer_embed = self._super_layer_embed.view(1, 1, num_layer, -1).expand(
             batch, seq, -1, -1
         )
-        joint_embed = torch.cat((meta_embed, layer_embed), dim=-1)
-        batch_weights = self._generator(joint_embed)
+        joint_embed = torch.cat(
+            (meta_embed, layer_embed), dim=-1
+        )  # batch, seq, num-layers, input-dim
+        batch_weights = self._generator(
+            joint_embed
+        )  # batch, seq, num-layers, num-weights
         batch_containers = []
         for seq_weights in torch.split(batch_weights, 1):
             seq_containers = []

xautodl/xlayers/super_attention.py

@@ -31,12 +31,15 @@ class SuperSelfAttention(SuperModule):
         qkv_bias: BoolSpaceType = False,
         attn_drop: Optional[float] = None,
         proj_drop: Optional[float] = None,
+        use_mask=False,
     ):
         super(SuperSelfAttention, self).__init__()
         self._input_dim = input_dim
         self._proj_dim = proj_dim
         self._num_heads = num_heads
         self._qkv_bias = qkv_bias
+        self._use_mask = use_mask
+        self._infinity = 1e9
 
         self.q_fc = SuperLinear(input_dim, input_dim, bias=qkv_bias)
         self.k_fc = SuperLinear(input_dim, input_dim, bias=qkv_bias)
@@ -113,6 +116,12 @@ class SuperSelfAttention(SuperModule):
             .permute(0, 2, 1, 3)
         )
         attn_v1 = (q_v1 @ k_v1.transpose(-2, -1)) * math.sqrt(head_dim)
+        if self._use_mask:
+            mask = torch.triu(
+                torch.ones((N, N), dtype=torch.bool, device=input.device), 1
+            )
+            mask = torch.unsqueeze(torch.unsqueeze(mask, dim=0), dim=0)
+            attn_v1 = attn_v1.masked_fill(mask, -self._infinity)
         attn_v1 = attn_v1.softmax(dim=-1)  # B * #head * N * N
         attn_v1 = self.attn_drop(attn_v1)
         feats_v1 = (attn_v1 @ v_v1).permute(0, 2, 1, 3).reshape(B, N, -1)
@@ -147,8 +156,14 @@ class SuperSelfAttention(SuperModule):
         return outs
 
     def extra_repr(self) -> str:
-        return "input_dim={:}, proj_dim={:}, num_heads={:}".format(
-            self._input_dim, self._proj_dim, self._num_heads
+        return (
+            "input_dim={:}, proj_dim={:}, num_heads={:}, mask={:}, infinity={:}".format(
+                self._input_dim,
+                self._proj_dim,
+                self._num_heads,
+                self._use_mask,
+                self._infinity,
+            )
         )
 
 
@@ -181,6 +196,7 @@ class SuperQKVAttention(SuperModule):
         self.attn_drop = nn.Dropout(attn_drop or 0.0)
         self.proj = SuperLinear(proj_dim, proj_dim)
         self.proj_drop = nn.Dropout(proj_drop or 0.0)
+        self._infinity = 1e9
 
     @property
     def num_heads(self):
@@ -232,7 +248,9 @@ class SuperQKVAttention(SuperModule):
         if "proj" in abstract_child:
             self.proj.apply_candidate(abstract_child["proj"])
 
-    def forward_qkv(self, q_tensor, k_tensor, v_tensor, num_head: int) -> torch.Tensor:
+    def forward_qkv(
+        self, q_tensor, k_tensor, v_tensor, num_head: int, mask=None
+    ) -> torch.Tensor:
         q = self.q_fc(q_tensor)
         B, N, C = q.shape
 
@@ -257,6 +275,9 @@ class SuperQKVAttention(SuperModule):
         )
         # compute the attention map
         attn_v1 = (q_v1 @ k_v1.transpose(-2, -1)) * math.sqrt(head_dim)
+        if mask is not None:
+            mask = torch.unsqueeze(mask, dim=1)
+            attn_v1 = attn_v1.masked_fill(mask, -self._infinity)
         attn_v1 = attn_v1.softmax(dim=-1)  # B * #head * N * S
         attn_v1 = self.attn_drop(attn_v1)
 
@@ -281,26 +302,29 @@ class SuperQKVAttention(SuperModule):
             feats = torch.cat([feats_v1, feats_v2], dim=-1)
         return feats
 
-    def forward_candidate(self, q_tensor, k_tensor, v_tensor) -> torch.Tensor:
+    def forward_candidate(
+        self, q_tensor, k_tensor, v_tensor, mask=None
+    ) -> torch.Tensor:
         # check the num_heads:
         if not spaces.is_determined(self._num_heads):
             num_heads = self.abstract_child["_num_heads"].value
         else:
             num_heads = spaces.get_determined_value(self._num_heads)
-        feats = self.forward_qkv(q_tensor, k_tensor, v_tensor, num_heads)
+        feats = self.forward_qkv(q_tensor, k_tensor, v_tensor, num_heads, mask)
         outs = self.proj(feats)
         outs = self.proj_drop(outs)
         return outs
 
-    def forward_raw(self, q_tensor, k_tensor, v_tensor) -> torch.Tensor:
-        feats = self.forward_qkv(q_tensor, k_tensor, v_tensor, self.num_heads)
+    def forward_raw(self, q_tensor, k_tensor, v_tensor, mask=None) -> torch.Tensor:
+        feats = self.forward_qkv(q_tensor, k_tensor, v_tensor, self.num_heads, mask)
         outs = self.proj(feats)
         outs = self.proj_drop(outs)
         return outs
 
     def extra_repr(self) -> str:
-        return "input_dim={:}, proj_dim={:}, num_heads={:}".format(
+        return "input_dim={:}, proj_dim={:}, num_heads={:}, infinity={:}".format(
             (self.in_q_dim, self.in_k_dim, self.in_v_dim),
             self._proj_dim,
             self._num_heads,
+            self._infinity,
         )

xautodl/xlayers/super_module.py

@@ -117,16 +117,32 @@ class SuperModule(abc.ABC, nn.Module):
         else:
             return False, self._meta_info[BEST_SCORE_KEY]
 
-    def load_best(self):
-        if BEST_DIR_KEY not in self._meta_info or BEST_SCORE_KEY not in self._meta_info:
-            raise ValueError("Please call save_best at first")
-        best_save_path = os.path.join(
-            self._meta_info[BEST_DIR_KEY],
-            "best-{:}.pth".format(self.__class__.__name__),
-        )
+    def load_best(self, best_save_path=None):
+        if best_save_path is None:
+            if (
+                BEST_DIR_KEY not in self._meta_info
+                or BEST_SCORE_KEY not in self._meta_info
+            ):
+                raise ValueError("Please call save_best at first")
+            best_save_name = self._meta_info.get(
+                BEST_NAME_KEY, "best-{:}.pth".format(self.__class__.__name__)
+            )
+            best_save_path = os.path.join(self._meta_info[BEST_DIR_KEY], best_save_name)
         state_dict = torch.load(best_save_path)
         self.load_state_dict(state_dict)
 
+    def has_best(self, best_name=None):
+        if BEST_DIR_KEY not in self._meta_info:
+            raise ValueError("Please set BEST_DIR_KEY at first")
+        if best_name is None:
+            best_save_name = self._meta_info.get(
+                BEST_NAME_KEY, "best-{:}.pth".format(self.__class__.__name__)
+            )
+        else:
+            best_save_name = best_name
+        best_save_path = os.path.join(self._meta_info[BEST_DIR_KEY], best_save_name)
+        return os.path.exists(best_save_path)
+
     @property
     def abstract_search_space(self):
         raise NotImplementedError

xautodl/xlayers/super_transformer.py

@@ -45,6 +45,7 @@ class SuperTransformerEncoderLayer(SuperModule):
         norm_affine: bool = True,
         act_layer: Callable[[], nn.Module] = nn.GELU,
         order: LayerOrder = LayerOrder.PreNorm,
+        use_mask: bool = False,
     ):
         super(SuperTransformerEncoderLayer, self).__init__()
         mha = SuperSelfAttention(
@@ -54,6 +55,7 @@ class SuperTransformerEncoderLayer(SuperModule):
             qkv_bias=qkv_bias,
             attn_drop=drop,
             proj_drop=drop,
+            use_mask=use_mask,
         )
         mlp = SuperMLPv2(
             d_model,