Skip to content

Commit 2e92f31

Browse files
author
ayasyrev
committed
add resnet50, move layers to module.
1 parent fb83065 commit 2e92f31

File tree

16 files changed

+2653
-984
lines changed

16 files changed

+2653
-984
lines changed

‎00_constructor.ipynb‎

Lines changed: 54 additions & 137 deletions
Original file line numberDiff line numberDiff line change
@@ -38,7 +38,8 @@
3838
"# export\n",
3939
"import torch.nn as nn\n",
4040
"import torch\n",
41-
"from collections import OrderedDict"
41+
"from collections import OrderedDict\n",
42+
"from model_constructor.layers import *"
4243
]
4344
},
4445
{
@@ -47,150 +48,66 @@
4748
"metadata": {},
4849
"outputs": [],
4950
"source": [
50-
"test_eq(1, 1**2)"
51-
]
52-
},
53-
{
54-
"cell_type": "code",
55-
"execution_count": null,
56-
"metadata": {},
57-
"outputs": [],
58-
"source": [
59-
"# test_eq(1, 2)"
60-
]
61-
},
62-
{
63-
"cell_type": "markdown",
64-
"metadata": {},
65-
"source": [
66-
"# Base ConvLayer"
67-
]
68-
},
69-
{
70-
"cell_type": "code",
71-
"execution_count": null,
72-
"metadata": {},
73-
"outputs": [],
74-
"source": [
75-
"# export\n",
76-
"_act_fn = nn.ReLU(inplace=True)\n",
51+
"# _act_fn = nn.ReLU(inplace=True)\n",
7752
"\n",
78-
"class ConvLayer(nn.Sequential):\n",
79-
" \"\"\"Basic conv layers block\"\"\"\n",
80-
" def __init__(self, ni, nf, ks=3, stride=1, \n",
81-
" act=True, act_fn=_act_fn, \n",
82-
" bn_layer=True, bn_1st=False, zero_bn=False, \n",
83-
" padding=None, bias=True, groups=1):\n",
53+
"# class ConvLayer(nn.Sequential):\n",
54+
"# \"\"\"Basic conv layers block\"\"\"\n",
55+
"# def __init__(self, ni, nf, ks=3, stride=1, \n",
56+
"# act=True, act_fn=_act_fn, \n",
57+
"# bn_layer=True, bn_1st=False, zero_bn=False, \n",
58+
"# padding=None, bias=True, groups=1):\n",
8459
"\n",
85-
" self.act = act\n",
86-
" if padding==None: padding = ks//2 \n",
87-
" layers = [('conv', nn.Conv2d(ni, nf, ks, stride=stride, padding=padding, bias=bias, groups=groups))]\n",
88-
" act_bn = [('act_fn', act_fn)] if act else []\n",
89-
" if bn_layer:\n",
90-
" bn = nn.BatchNorm2d(nf)\n",
91-
" nn.init.constant_(bn.weight, 0. if zero_bn else 1.) \n",
92-
" act_bn += [('bn', bn)]\n",
93-
" if bn_1st: act_bn.reverse()\n",
94-
" layers += act_bn\n",
95-
" super().__init__(OrderedDict(layers))"
60+
"# self.act = act\n",
61+
"# if padding==None: padding = ks//2 \n",
62+
"# layers = [('conv', nn.Conv2d(ni, nf, ks, stride=stride, padding=padding, bias=bias, groups=groups))]\n",
63+
"# act_bn = [('act_fn', act_fn)] if act else []\n",
64+
"# if bn_layer:\n",
65+
"# bn = nn.BatchNorm2d(nf)\n",
66+
"# nn.init.constant_(bn.weight, 0. if zero_bn else 1.) \n",
67+
"# act_bn += [('bn', bn)]\n",
68+
"# if bn_1st: act_bn.reverse()\n",
69+
"# layers += act_bn\n",
70+
"# super().__init__(OrderedDict(layers))"
9671
]
9772
},
9873
{
9974
"cell_type": "code",
10075
"execution_count": null,
10176
"metadata": {},
102-
"outputs": [
103-
{
104-
"data": {
105-
"text/plain": [
106-
"ConvLayer(\n",
107-
" (conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
108-
" (act_fn): ReLU(inplace=True)\n",
109-
" (bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
110-
")"
111-
]
112-
},
113-
"execution_count": null,
114-
"metadata": {},
115-
"output_type": "execute_result"
116-
}
117-
],
77+
"outputs": [],
11878
"source": [
119-
"conv_layer = ConvLayer(32, 64)\n",
120-
"conv_layer"
79+
"# conv_layer = ConvLayer(32, 64)\n",
80+
"# conv_layer"
12181
]
12282
},
12383
{
12484
"cell_type": "code",
12585
"execution_count": null,
12686
"metadata": {},
127-
"outputs": [
128-
{
129-
"data": {
130-
"text/plain": [
131-
"ConvLayer(\n",
132-
" (conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
133-
" (bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
134-
")"
135-
]
136-
},
137-
"execution_count": null,
138-
"metadata": {},
139-
"output_type": "execute_result"
140-
}
141-
],
87+
"outputs": [],
14288
"source": [
143-
"conv_layer = ConvLayer(32, 64, act=False)\n",
144-
"conv_layer"
89+
"# conv_layer = ConvLayer(32, 64, act=False)\n",
90+
"# conv_layer"
14591
]
14692
},
14793
{
14894
"cell_type": "code",
14995
"execution_count": null,
15096
"metadata": {},
151-
"outputs": [
152-
{
153-
"data": {
154-
"text/plain": [
155-
"ConvLayer(\n",
156-
" (conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
157-
" (act_fn): ReLU(inplace=True)\n",
158-
")"
159-
]
160-
},
161-
"execution_count": null,
162-
"metadata": {},
163-
"output_type": "execute_result"
164-
}
165-
],
97+
"outputs": [],
16698
"source": [
167-
"conv_layer = ConvLayer(32, 64, bn_layer=False)\n",
168-
"conv_layer"
99+
"# conv_layer = ConvLayer(32, 64, bn_layer=False)\n",
100+
"# conv_layer"
169101
]
170102
},
171103
{
172104
"cell_type": "code",
173105
"execution_count": null,
174106
"metadata": {},
175-
"outputs": [
176-
{
177-
"data": {
178-
"text/plain": [
179-
"ConvLayer(\n",
180-
" (conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
181-
" (bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
182-
" (act_fn): ReLU(inplace=True)\n",
183-
")"
184-
]
185-
},
186-
"execution_count": null,
187-
"metadata": {},
188-
"output_type": "execute_result"
189-
}
190-
],
107+
"outputs": [],
191108
"source": [
192-
"conv_layer = ConvLayer(32, 64, bn_1st=True)\n",
193-
"conv_layer"
109+
"# conv_layer = ConvLayer(32, 64, bn_1st=True)\n",
110+
"# conv_layer"
194111
]
195112
},
196113
{
@@ -206,12 +123,11 @@
206123
"metadata": {},
207124
"outputs": [],
208125
"source": [
209-
"# export\n",
210-
"class Flatten(nn.Module):\n",
211-
" '''flat x to vector'''\n",
212-
" def __init__(self):\n",
213-
" super().__init__()\n",
214-
" def forward(self, x): return x.view(x.size(0), -1)"
126+
"# class Flatten(nn.Module):\n",
127+
"# '''flat x to vector'''\n",
128+
"# def __init__(self):\n",
129+
"# super().__init__()\n",
130+
"# def forward(self, x): return x.view(x.size(0), -1)"
215131
]
216132
},
217133
{
@@ -220,14 +136,13 @@
220136
"metadata": {},
221137
"outputs": [],
222138
"source": [
223-
"# export\n",
224-
"class Noop(nn.Module): # alternative name Merge\n",
225-
" '''Dummy module for vizualize skip conn'''\n",
226-
" def __init__(self):\n",
227-
" super().__init__()\n",
139+
"# class Noop(nn.Module): # alternative name Merge\n",
140+
"# '''Dummy module for vizualize skip conn'''\n",
141+
"# def __init__(self):\n",
142+
"# super().__init__()\n",
228143
" \n",
229-
" def forward(self, x):\n",
230-
" return x"
144+
"# def forward(self, x):\n",
145+
"# return x"
231146
]
232147
},
233148
{
@@ -440,7 +355,7 @@
440355
"cell_type": "markdown",
441356
"metadata": {},
442357
"source": [
443-
"## Block"
358+
"## BasicBlock"
444359
]
445360
},
446361
{
@@ -452,10 +367,9 @@
452367
"# export\n",
453368
"class BasicBlock(nn.Module):\n",
454369
" \"\"\"Basic block (simplified) as in pytorch resnet\"\"\"\n",
455-
" expansion = 1\n",
456370
" def __init__(self, ni, nf, stride=1, bn_1st=False, zero_bn=False,\n",
457371
"# groups=1, base_width=64, dilation=1, norm_layer=None\n",
458-
" **kwargs):\n",
372+
" expansion = 1, **kwargs):\n",
459373
" super().__init__()\n",
460374
" self.downsample = not ni==nf or stride==2\n",
461375
" self.conv = nn.Sequential(OrderedDict([\n",
@@ -464,7 +378,7 @@
464378
" if self.downsample:\n",
465379
" self.downsample = ConvLayer(ni, nf, ks=1, stride=stride, act=False, **kwargs)\n",
466380
" self.merge = Noop()\n",
467-
" self.act_conn = _act_fn\n",
381+
" self.act_conn = act_fn\n",
468382
" \n",
469383
" def forward(self, x):\n",
470384
" identity = x\n",
@@ -558,7 +472,7 @@
558472
{
559473
"data": {
560474
"text/plain": [
561-
"torch.Size([64, 64, 32, 32])"
475+
"torch.Size([64, 64, 16, 16])"
562476
]
563477
},
564478
"execution_count": null,
@@ -654,7 +568,7 @@
654568
" body_in=64, body_out=512, \n",
655569
" layer_szs=[64,128,256,], blocks=[2,2,2,2],\n",
656570
" expansion=1, **kwargs): # Downsample Module as parameter\n",
657-
" layer_szs = [body_in] + layer_szs + [body_out]\n",
571+
" layer_szs = [body_in//expansion] + layer_szs + [body_out]\n",
658572
" num_layers = len(layer_szs)-1\n",
659573
" layers = [(f\"layer_{i}\", self._make_layer(block, layer_szs[i], layer_szs[i+1], blocks[i], 1 if i==0 else 2, **kwargs))\n",
660574
" for i in range(num_layers)]\n",
@@ -971,7 +885,8 @@
971885
" # block_szs = [64,128,128,256,256,512]\n",
972886
" super().__init__(OrderedDict([\n",
973887
" ('stem', stem(c_in=c_in,stem_out=body_in, **kwargs)),\n",
974-
" ('body', body(block, body_in, body_out, layer_szs=layer_szs, blocks=blocks, **kwargs)),\n",
888+
" ('body', body(block, body_in, body_out, \n",
889+
" layer_szs=layer_szs, blocks=blocks, expansion=expansion, **kwargs)),\n",
975890
" ('head', head(body_out*expansion, num_classes, **kwargs))\n",
976891
" ]))\n",
977892
" init_model(self)"
@@ -1919,7 +1834,8 @@
19191834
"cell_type": "markdown",
19201835
"metadata": {},
19211836
"source": [
1922-
"# fin"
1837+
"# model_constructor\n",
1838+
"by ayasyrev"
19231839
]
19241840
},
19251841
{
@@ -1939,7 +1855,8 @@
19391855
"output_type": "stream",
19401856
"text": [
19411857
"Converted 00_constructor.ipynb.\n",
1942-
"Converted 01_resnet.ipynb.\n",
1858+
"Converted 01_layers.ipynb.\n",
1859+
"Converted 02_resnet.ipynb.\n",
19431860
"Converted index.ipynb.\n"
19441861
]
19451862
}

0 commit comments

Comments
 (0)