Merge pull request #730 from talesa/feature/more_parameterizable_xresnet1dplus

Allow for parameterizing block_szs in XResNet1dPlus

Author: oguiza

nbs/036_models.InceptionTimePlus.ipynb

@@ -155,7 +155,7 @@
     "        self.seq_len = seq_len\n",
     "        if custom_head is not None: \n",
     "            if isinstance(custom_head, nn.Module): head = custom_head\n",
-    "            head = custom_head(self.head_nf, c_out, seq_len)\n",
+    "            else: head = custom_head(self.head_nf, c_out, seq_len)\n",
     "        else: head = self.create_head(self.head_nf, c_out, seq_len, flatten=flatten, concat_pool=concat_pool, \n",
     "                                      fc_dropout=fc_dropout, bn=bn, y_range=y_range)\n",
     "            \n",

nbs/057_models.MINIROCKETPlus_Pytorch.ipynb

@@ -250,7 +250,7 @@
     "        self.head_nf = num_features\n",
     "        if custom_head is not None: \n",
     "            if isinstance(custom_head, nn.Module): head = custom_head\n",
-    "            head = custom_head(self.head_nf, c_out, 1)\n",
+    "            else: head = custom_head(self.head_nf, c_out, 1)\n",
     "        else:\n",
     "            layers = [Flatten()]\n",
     "            if bn:\n",
@@ -506,7 +506,7 @@
     "        self.head_nf = num_features\n",
     "        if custom_head is not None: \n",
     "            if isinstance(custom_head, nn.Module): head = custom_head\n",
-    "            head = custom_head(self.head_nf, c_out, 1)\n",
+    "            else: head = custom_head(self.head_nf, c_out, 1)\n",
     "        else:\n",
     "            layers = [Flatten()]\n",
     "            if bn:\n",

nbs/059_models.XResNet1dPlus.ipynb

@@ -45,7 +45,7 @@
     "class XResNet1dPlus(nn.Sequential):\n",
     "    @delegates(ResBlock1dPlus)\n",
     "    def __init__(self, block=ResBlock1dPlus, expansion=4, layers=[3,4,6,3], fc_dropout=0.0, c_in=3, c_out=None, n_out=1000, seq_len=None, stem_szs=(32,32,64),\n",
-    "                 widen=1.0, sa=False, act_cls=defaults.activation, ks=3, stride=2, coord=False, custom_head=None, **kwargs):\n",
+    "                 widen=1.0, sa=False, act_cls=defaults.activation, ks=3, stride=2, coord=False, custom_head=None, block_szs_base=(64,128,256,512), **kwargs):\n",
     "\n",
     "        store_attr('block,expansion,act_cls,ks')\n",
     "        n_out = c_out or n_out # added for compatibility\n",
@@ -55,14 +55,14 @@
     "                          act=act_cls)\n",
     "                for i in range(3)]\n",
     "\n",
-    "        block_szs = [int(o*widen) for o in [64,128,256,512] +[256]*(len(layers)-4)]\n",
+    "        block_szs = [int(o*widen) for o in (list(block_szs_base) + [int(block_szs_base[-1]/2)]*(len(layers)-4))]\n",
     "        block_szs = [64//expansion] + block_szs\n",
     "        blocks    = self._make_blocks(layers, block_szs, sa, coord, stride, **kwargs)\n",
     "        backbone = nn.Sequential(*stem, MaxPool(ks=ks, stride=stride, padding=ks//2, ndim=1), *blocks)\n",
     "        self.head_nf = block_szs[-1]*expansion\n",
     "        if custom_head is not None: \n",
     "            if isinstance(custom_head, nn.Module): head = custom_head\n",
-    "            head = custom_head(self.head_nf, n_out, seq_len)\n",
+    "            else: head = custom_head(self.head_nf, n_out, seq_len)\n",
     "        else: head = nn.Sequential(AdaptiveAvgPool(sz=1, ndim=1), Flatten(), nn.Dropout(fc_dropout), nn.Linear(block_szs[-1]*expansion, n_out))\n",
     "        super().__init__(OrderedDict([('backbone', backbone), ('head', head)]))\n",
     "        self._init_cnn(self)\n",

tsai/models/InceptionTimePlus.py

@@ -108,7 +108,7 @@ def __init__(self, c_in, c_out, seq_len=None, nf=32, nb_filters=None,
         self.seq_len = seq_len
         if custom_head is not None: 
             if isinstance(custom_head, nn.Module): head = custom_head
-            head = custom_head(self.head_nf, c_out, seq_len)
+            else: head = custom_head(self.head_nf, c_out, seq_len)
         else: head = self.create_head(self.head_nf, c_out, seq_len, flatten=flatten, concat_pool=concat_pool, 
                                       fc_dropout=fc_dropout, bn=bn, y_range=y_range)
 

tsai/models/MINIROCKETPlus_Pytorch.py

@@ -205,7 +205,7 @@ def __init__(self, c_in, c_out, seq_len, num_features=10_000, max_dilations_per_
         self.head_nf = num_features
         if custom_head is not None: 
             if isinstance(custom_head, nn.Module): head = custom_head
-            head = custom_head(self.head_nf, c_out, 1)
+            else: head = custom_head(self.head_nf, c_out, 1)
         else:
             layers = [Flatten()]
             if bn:
@@ -319,7 +319,7 @@ def __init__(self, c_in, c_out, seq_len, num_features=10_000, max_dilations_per_
         self.head_nf = num_features
         if custom_head is not None: 
             if isinstance(custom_head, nn.Module): head = custom_head
-            head = custom_head(self.head_nf, c_out, 1)
+            else: head = custom_head(self.head_nf, c_out, 1)
         else:
             layers = [Flatten()]
             if bn:

tsai/models/XResNet1dPlus.py

@@ -14,7 +14,7 @@
 class XResNet1dPlus(nn.Sequential):
     @delegates(ResBlock1dPlus)
     def __init__(self, block=ResBlock1dPlus, expansion=4, layers=[3,4,6,3], fc_dropout=0.0, c_in=3, c_out=None, n_out=1000, seq_len=None, stem_szs=(32,32,64),
-                 widen=1.0, sa=False, act_cls=defaults.activation, ks=3, stride=2, coord=False, custom_head=None, **kwargs):
+                 widen=1.0, sa=False, act_cls=defaults.activation, ks=3, stride=2, coord=False, custom_head=None, block_szs_base=(64,128,256,512), **kwargs):
 
         store_attr('block,expansion,act_cls,ks')
         n_out = c_out or n_out # added for compatibility
@@ -24,14 +24,14 @@ def __init__(self, block=ResBlock1dPlus, expansion=4, layers=[3,4,6,3], fc_dropo
                           act=act_cls)
                 for i in range(3)]
 
-        block_szs = [int(o*widen) for o in [64,128,256,512] +[256]*(len(layers)-4)]
+        block_szs = [int(o*widen) for o in (list(block_szs_base) + [int(block_szs_base[-1]/2)]*(len(layers)-4))]
         block_szs = [64//expansion] + block_szs
         blocks    = self._make_blocks(layers, block_szs, sa, coord, stride, **kwargs)
         backbone = nn.Sequential(*stem, MaxPool(ks=ks, stride=stride, padding=ks//2, ndim=1), *blocks)
         self.head_nf = block_szs[-1]*expansion
         if custom_head is not None: 
             if isinstance(custom_head, nn.Module): head = custom_head
-            head = custom_head(self.head_nf, n_out, seq_len)
+            else: head = custom_head(self.head_nf, n_out, seq_len)
         else: head = nn.Sequential(AdaptiveAvgPool(sz=1, ndim=1), Flatten(), nn.Dropout(fc_dropout), nn.Linear(block_szs[-1]*expansion, n_out))
         super().__init__(OrderedDict([('backbone', backbone), ('head', head)]))
         self._init_cnn(self)