package torch

You can search for identifiers within the package.
in-package search v0.2.0
Torch bindings for OCaml
Install

Dune Dependency

github.com Readme MIT License Edit opam file Versions (18)
Authors

J

Jane Street Group, LLC
Maintainers

J

Jane Street developers
Sources

torch-v0.16.0.tar.gz
sha256=ccd9ef3b630bdc7c41e363e71d8ecb86c316460cbf79afe67546c6ff22c19da4
doc/src/torch.vision/inception.ml.html

Source file `inception.ml`

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
(* Inception V3 *)
open Base
open Torch

let sub = Var_store.sub
let max_pool2d ~k ~stride:s = Tensor.max_pool2d ~ksize:(k, k) ~stride:(s, s)

let cv_bn vs ~k ~pad ?(stride = 1) ~input_dim out_dim =
  let conv =
    Layer.conv2d_
      (sub vs "conv")
      ~ksize:k
      ~padding:pad
      ~stride
      ~use_bias:false
      ~input_dim
      out_dim
  in
  let bn = Layer.batch_norm2d (sub vs "bn") ~eps:0.001 out_dim in
  Layer.of_fn_ (fun xs ~is_training ->
    Layer.forward conv xs |> Layer.forward_ bn ~is_training |> Tensor.relu)
;;

let cv_bn2 vs ~k ~pad ?(stride = 1, 1) ~input_dim out_dim =
  let conv =
    Layer.conv2d
      (sub vs "conv")
      ~ksize:k
      ~padding:pad
      ~stride
      ~use_bias:false
      ~input_dim
      out_dim
  in
  let bn = Layer.batch_norm2d (sub vs "bn") ~eps:0.001 out_dim in
  Layer.of_fn_ (fun xs ~is_training ->
    Layer.forward conv xs |> Layer.forward_ bn ~is_training |> Tensor.relu)
;;

let inception_a vs ~input_dim ~pool =
  let b1 = cv_bn (sub vs "branch1x1") ~k:1 ~pad:0 ~input_dim 64 in
  let b2_1 = cv_bn (sub vs "branch5x5_1") ~k:1 ~pad:0 ~input_dim 48 in
  let b2_2 = cv_bn (sub vs "branch5x5_2") ~k:5 ~pad:2 ~input_dim:48 64 in
  let b3_1 = cv_bn (sub vs "branch3x3dbl_1") ~k:1 ~pad:0 ~input_dim 64 in
  let b3_2 = cv_bn (sub vs "branch3x3dbl_2") ~k:3 ~pad:1 ~input_dim:64 96 in
  let b3_3 = cv_bn (sub vs "branch3x3dbl_3") ~k:3 ~pad:1 ~input_dim:96 96 in
  let bpool = cv_bn (sub vs "branch_pool") ~k:1 ~pad:0 ~input_dim pool in
  Layer.of_fn_ (fun xs ~is_training ->
    let apply_ = Layer.forward_ ~is_training in
    let b1 = apply_ b1 xs in
    let b2 = apply_ b2_1 xs |> apply_ b2_2 in
    let b3 = apply_ b3_1 xs |> apply_ b3_2 |> apply_ b3_3 in
    let bpool =
      Tensor.avg_pool2d xs ~ksize:(3, 3) ~stride:(1, 1) ~padding:(1, 1) |> apply_ bpool
    in
    Tensor.concat [ b1; b2; b3; bpool ] ~dim:1)
;;

let inception_b vs ~input_dim =
  let b1 = cv_bn (sub vs "branch3x3") ~k:3 ~pad:0 ~stride:2 ~input_dim 384 in
  let b2_1 = cv_bn (sub vs "branch3x3dbl_1") ~k:1 ~pad:0 ~input_dim 64 in
  let b2_2 = cv_bn (sub vs "branch3x3dbl_2") ~k:3 ~pad:1 ~input_dim:64 96 in
  let b2_3 = cv_bn (sub vs "branch3x3dbl_3") ~k:3 ~pad:0 ~stride:2 ~input_dim:96 96 in
  Layer.of_fn_ (fun xs ~is_training ->
    let apply_ = Layer.forward_ ~is_training in
    let b1 = apply_ b1 xs in
    let b2 = apply_ b2_1 xs |> apply_ b2_2 |> apply_ b2_3 in
    let bpool = max_pool2d xs ~k:3 ~stride:2 in
    Tensor.concat [ b1; b2; bpool ] ~dim:1)
;;

let inception_c vs ~input_dim ~c7 =
  let b1 = cv_bn (sub vs "branch1x1") ~k:1 ~pad:0 ~input_dim 192 in
  let b2_1 = cv_bn (sub vs "branch7x7_1") ~k:1 ~pad:0 ~input_dim c7 in
  let b2_2 = cv_bn2 (sub vs "branch7x7_2") ~k:(1, 7) ~pad:(0, 3) ~input_dim:c7 c7 in
  let b2_3 = cv_bn2 (sub vs "branch7x7_3") ~k:(7, 1) ~pad:(3, 0) ~input_dim:c7 192 in
  let b3_1 = cv_bn (sub vs "branch7x7dbl_1") ~k:1 ~pad:0 ~input_dim c7 in
  let b3_2 = cv_bn2 (sub vs "branch7x7dbl_2") ~k:(7, 1) ~pad:(3, 0) ~input_dim:c7 c7 in
  let b3_3 = cv_bn2 (sub vs "branch7x7dbl_3") ~k:(1, 7) ~pad:(0, 3) ~input_dim:c7 c7 in
  let b3_4 = cv_bn2 (sub vs "branch7x7dbl_4") ~k:(7, 1) ~pad:(3, 0) ~input_dim:c7 c7 in
  let b3_5 = cv_bn2 (sub vs "branch7x7dbl_5") ~k:(1, 7) ~pad:(0, 3) ~input_dim:c7 192 in
  let bpool = cv_bn (sub vs "branch_pool") ~k:1 ~pad:0 ~input_dim 192 in
  Layer.of_fn_ (fun xs ~is_training ->
    let apply_ = Layer.forward_ ~is_training in
    let b1 = apply_ b1 xs in
    let b2 = apply_ b2_1 xs |> apply_ b2_2 |> apply_ b2_3 in
    let b3 = apply_ b3_1 xs |> apply_ b3_2 |> apply_ b3_3 |> apply_ b3_4 |> apply_ b3_5 in
    let bpool =
      Tensor.avg_pool2d xs ~ksize:(3, 3) ~stride:(1, 1) ~padding:(1, 1) |> apply_ bpool
    in
    Tensor.concat [ b1; b2; b3; bpool ] ~dim:1)
;;

let inception_d vs ~input_dim =
  let b1_1 = cv_bn (sub vs "branch3x3_1") ~k:1 ~pad:0 ~input_dim 192 in
  let b1_2 = cv_bn (sub vs "branch3x3_2") ~k:3 ~pad:0 ~stride:2 ~input_dim:192 320 in
  let b2_1 = cv_bn (sub vs "branch7x7x3_1") ~k:1 ~pad:0 ~input_dim 192 in
  let b2_2 = cv_bn2 (sub vs "branch7x7x3_2") ~k:(1, 7) ~pad:(0, 3) ~input_dim:192 192 in
  let b2_3 = cv_bn2 (sub vs "branch7x7x3_3") ~k:(7, 1) ~pad:(3, 0) ~input_dim:192 192 in
  let b2_4 = cv_bn (sub vs "branch7x7x3_4") ~k:3 ~pad:0 ~stride:2 ~input_dim:192 192 in
  Layer.of_fn_ (fun xs ~is_training ->
    let apply_ = Layer.forward_ ~is_training in
    let b1 = apply_ b1_1 xs |> apply_ b1_2 in
    let b2 = apply_ b2_1 xs |> apply_ b2_2 |> apply_ b2_3 |> apply_ b2_4 in
    let bpool = max_pool2d xs ~k:3 ~stride:2 in
    Tensor.concat [ b1; b2; bpool ] ~dim:1)
;;

let inception_e vs ~input_dim =
  let b1 = cv_bn (sub vs "branch1x1") ~k:1 ~pad:0 ~input_dim 320 in
  let b2_1 = cv_bn (sub vs "branch3x3_1") ~k:1 ~pad:0 ~input_dim 384 in
  let b2_2a = cv_bn2 (sub vs "branch3x3_2a") ~k:(1, 3) ~pad:(0, 1) ~input_dim:384 384 in
  let b2_2b = cv_bn2 (sub vs "branch3x3_2b") ~k:(3, 1) ~pad:(1, 0) ~input_dim:384 384 in
  let b3_1 = cv_bn (sub vs "branch3x3dbl_1") ~k:1 ~pad:0 ~input_dim 448 in
  let b3_2 = cv_bn (sub vs "branch3x3dbl_2") ~k:3 ~pad:1 ~input_dim:448 384 in
  let b3_3a =
    cv_bn2 (sub vs "branch3x3dbl_3a") ~k:(1, 3) ~pad:(0, 1) ~input_dim:384 384
  in
  let b3_3b =
    cv_bn2 (sub vs "branch3x3dbl_3b") ~k:(3, 1) ~pad:(1, 0) ~input_dim:384 384
  in
  let bpool = cv_bn (sub vs "branch_pool") ~k:1 ~pad:0 ~input_dim 192 in
  Layer.of_fn_ (fun xs ~is_training ->
    let apply_ = Layer.forward_ ~is_training in
    let b1 = apply_ b1 xs in
    let b2 = apply_ b2_1 xs in
    let b2 = Tensor.concat [ apply_ b2_2a b2; apply_ b2_2b b2 ] ~dim:1 in
    let b3 = apply_ b3_1 xs |> apply_ b3_2 in
    let b3 = Tensor.concat [ apply_ b3_3a b3; apply_ b3_3b b3 ] ~dim:1 in
    let bpool =
      Tensor.avg_pool2d xs ~ksize:(3, 3) ~stride:(1, 1) ~padding:(1, 1) |> apply_ bpool
    in
    Tensor.concat [ b1; b2; b3; bpool ] ~dim:1)
;;

let v3 ?num_classes vs =
  let conv1 = cv_bn (sub vs "Conv2d_1a_3x3") ~k:3 ~pad:0 ~stride:2 ~input_dim:3 32 in
  let conv2a = cv_bn (sub vs "Conv2d_2a_3x3") ~k:3 ~pad:0 ~input_dim:32 32 in
  let conv2b = cv_bn (sub vs "Conv2d_2b_3x3") ~k:3 ~pad:1 ~input_dim:32 64 in
  let conv3b = cv_bn (sub vs "Conv2d_3b_1x1") ~k:1 ~pad:0 ~input_dim:64 80 in
  let conv4a = cv_bn (sub vs "Conv2d_4a_3x3") ~k:3 ~pad:0 ~input_dim:80 192 in
  let mixed_5b = inception_a (sub vs "Mixed_5b") ~input_dim:192 ~pool:32 in
  let mixed_5c = inception_a (sub vs "Mixed_5c") ~input_dim:256 ~pool:64 in
  let mixed_5d = inception_a (sub vs "Mixed_5d") ~input_dim:288 ~pool:64 in
  let mixed_6a = inception_b (sub vs "Mixed_6a") ~input_dim:288 in
  let mixed_6b = inception_c (sub vs "Mixed_6b") ~input_dim:768 ~c7:128 in
  let mixed_6c = inception_c (sub vs "Mixed_6c") ~input_dim:768 ~c7:160 in
  let mixed_6d = inception_c (sub vs "Mixed_6d") ~input_dim:768 ~c7:160 in
  let mixed_6e = inception_c (sub vs "Mixed_6e") ~input_dim:768 ~c7:192 in
  let mixed_7a = inception_d (sub vs "Mixed_7a") ~input_dim:768 in
  let mixed_7b = inception_e (sub vs "Mixed_7b") ~input_dim:1280 in
  let mixed_7c = inception_e (sub vs "Mixed_7c") ~input_dim:2048 in
  let final =
    match num_classes with
    | None -> Layer.id
    | Some num_classes -> Layer.linear (sub vs "fc") ~input_dim:2048 num_classes
  in
  Layer.of_fn_ (fun xs ~is_training ->
    let apply_ = Layer.forward_ ~is_training in
    let batch_size = Tensor.shape xs |> List.hd_exn in
    apply_ conv1 xs
    |> apply_ conv2a
    |> apply_ conv2b
    |> Tensor.relu
    |> max_pool2d ~k:3 ~stride:2
    |> apply_ conv3b
    |> apply_ conv4a
    |> max_pool2d ~k:3 ~stride:2
    |> apply_ mixed_5b
    |> apply_ mixed_5c
    |> apply_ mixed_5d
    |> apply_ mixed_6a
    |> apply_ mixed_6b
    |> apply_ mixed_6c
    |> apply_ mixed_6d
    |> apply_ mixed_6e
    |> apply_ mixed_7a
    |> apply_ mixed_7b
    |> apply_ mixed_7c
    |> Tensor.adaptive_avg_pool2d ~output_size:[ 1; 1 ]
    |> Tensor.dropout ~p:0.5 ~is_training
    |> Tensor.view ~size:[ batch_size; -1 ]
    |> Layer.forward final)
;;
package torch

Install

Dune Dependency

Authors

Maintainers

Sources

doc/src/torch.vision/inception.ml.html

Source file inception.ml

Source file `inception.ml`
