Image Classification PyTorch

Library for quick training models for image classification.

Table of content

• Installation
• Quick start
• Prediction
• Parameters
• Data preparation
• Models
• License

---

Installation

pip install image-classification-pytorch

---

Quick Start

import image_classification_pytorch as icp

# add model
# your can add several models for consistent training
tf_efficientnet_b4_ns = {'model_type': 'tf_efficientnet_b4_ns', 
                         'im_size': 380, 
                         'im_size_test': 380, 
                         'batch_size': 8, 
                         'mean': [0.485, 0.456, 0.406], 
                         'std': [0.229, 0.224, 0.225]}
                         
models = [tf_efficientnet_b4_ns]

# create trainer
trainer = icp.ICPTrainer(models=models, data_dir='my_data')
# start training
trainer.fit_test()

Example

Simple example of training and prediction

---

Prediction

Put folders with samples into a folder (data_dir). You can use class labels for folder names.

Example of folders structure

├── inference                    # data_dir folder
    ├── dogs                     # Folder Class 1
    ├── cats                     # Folder Class 2

Use the same parameters as for training.

import image_classification_pytorch as icp

icp.ICPInference(data_dir='inference',
                 img_size=380,
                 show_accuracy=True,
                 checkpoint='tb_logs/tf_efficientnet_b4_ns/version_4/checkpoints/tf_efficientnet_b4_ns__epoch=2_val_loss=0.922_val_acc=0.830_val_f1_epoch=0.000.ckpt',
                 std=[0.229, 0.224, 0.225],
                 mean=[0.485, 0.456, 0.406],
                 confidence_threshold=1).predict()

After prediction you can see such folders structure

├── inference                    # data_dir folder
    ├── dogs                     # Initial dogs folder 
    ├── dogs_gt___dogs           # In this folder should be dogs pictures (ground truth(gt) dogs) and they predicted as dogs
    ├── dogs_gt___cats           # In this folder should be dogs pictures (ground truth(gt) dogs) but they predicted as cats
    ├── cats                     # Initial cats folder
    ├── cats_gt___cats           # In this folder should be cats pictures (ground truth(gt) cats) and they predicted as cats

As you can see all cats predicted as cats and some dogs predicted as cats.

---

Data Preparation

Prepare data for training in the following format

├── animals                      # Data folder
    ├── dogs                     # Folder Class 1
    ├── cats                     # Folder Class 2
    ├── gray_bears               # Folder Class 3
    ├── zebras                   # Folder Class 4
    ├── ...

---

Detailed Quick Start

import image_classification_pytorch as icp

# add model
# your can add several models for consistent training
tf_efficientnet_b4_ns = {'model_type': 'tf_efficientnet_b4_ns', 
                         'im_size': 380, 
                         'im_size_test': 380, 
                         'batch_size': 8, 
                         'mean': [0.485, 0.456, 0.406], 
                         'std': [0.229, 0.224, 0.225]}
                         
ens_adv_inception_resnet_v2 = {'model_type': 'ens_adv_inception_resnet_v2',
                               'im_size': 256,
                               'im_size_test': 256,
                               'batch_size': 8,
                               'mean': [0.5, 0.5, 0.5],
                               'std': [0.5, 0.5, 0.5]}
                         
models = [tf_efficientnet_b4_ns, ens_adv_inception_resnet_v2]

# create trainer
trainer = ICPTrainer(models=models, 
                     data_dir='my_data',
                     images_ext='jpg',
                     init_lr=1e-5,
                     max_epochs=500,
                     augment_p=0.7,
                     progress_bar_refresh_rate=10,
                     early_stop_patience=6,
                     optimizer=Adam(),
                     scheduler=ExponentialLR())

# start training
trainer.fit_test()

---

Models

323 models out of the box. Used models from PyTorch Image Models (timm)

Models

1. adv_inception_v3
2. cspdarknet53
3. cspresnet50
4. cspresnext50
5. densenet121
6. densenet161
7. densenet169
8. densenet201
9. densenetblur121d
10. dla102
11. dla102x
12. dla102x2
13. dla169
14. dla34
15. dla46_c
16. dla46x_c
17. dla60_res2net
18. dla60_res2next
19. dla60
20. dla60x_c
21. dla60x
22. dm_nfnet_f0
23. dm_nfnet_f1
24. dm_nfnet_f2
25. dm_nfnet_f3
26. dm_nfnet_f4
27. dm_nfnet_f5
28. dm_nfnet_f6
29. dpn107
30. dpn131
31. dpn68
32. dpn68b
33. dpn92
34. dpn98
35. ecaresnet101d_pruned
36. ecaresnet101d
37. ecaresnet269d
38. ecaresnet26t
39. ecaresnet50d_pruned
40. ecaresnet50d
41. ecaresnet50t
42. ecaresnetlight
43. efficientnet_b0
44. efficientnet_b1_pruned
45. efficientnet_b1
46. efficientnet_b2
47. efficientnet_b2a
48. efficientnet_b3_pruned
49. efficientnet_b3
50. efficientnet_b3a
51. efficientnet_em
52. efficientnet_es
53. efficientnet_lite0
54. ens_adv_inception_resnet_v2
55. ese_vovnet19b_dw
56. ese_vovnet39b
57. fbnetc_100
58. gernet_l
59. gernet_m
60. gernet_s
61. gluon_inception_v3
62. gluon_resnet101_v1b
63. gluon_resnet101_v1c
64. gluon_resnet101_v1d
65. gluon_resnet101_v1s
66. gluon_resnet152_v1b
67. gluon_resnet152_v1c
68. gluon_resnet152_v1d
69. gluon_resnet152_v1s
70. gluon_resnet18_v1b
71. gluon_resnet34_v1b
72. gluon_resnet50_v1b
73. gluon_resnet50_v1c
74. gluon_resnet50_v1d
75. gluon_resnet50_v1s
76. gluon_resnext101_32x4d
77. gluon_resnext101_64x4d
78. gluon_resnext50_32x4d
79. gluon_senet154
80. gluon_seresnext101_32x4d
81. gluon_seresnext101_64x4d
82. gluon_seresnext50_32x4d
83. gluon_xception65
84. hrnet_w18_small_v2
85. hrnet_w18_small
86. hrnet_w18
87. hrnet_w30
88. hrnet_w32
89. hrnet_w40
90. hrnet_w44
91. hrnet_w48
92. hrnet_w64
93. ig_resnext101_32x16d
94. ig_resnext101_32x32d
95. ig_resnext101_32x48d
96. ig_resnext101_32x8d
97. inception_resnet_v2
98. inception_v3
99. inception_v4
100. legacy_senet154
101. legacy_seresnet101
102. legacy_seresnet152
103. legacy_seresnet18
104. legacy_seresnet34
105. legacy_seresnet50
106. legacy_seresnext101_32x4d
107. legacy_seresnext26_32x4d
108. legacy_seresnext50_32x4d
109. mixnet_l
110. mixnet_m
111. mixnet_s
112. mixnet_xl
113. mnasnet_100
114. mobilenetv2_100
115. mobilenetv2_110d
116. mobilenetv2_120d
117. mobilenetv2_140
118. mobilenetv3_large_100
119. mobilenetv3_rw
120. nasnetalarge
121. nf_regnet_b1
122. nf_resnet50
123. nfnet_l0c
124. pnasnet5large
125. regnetx_002
126. regnetx_004
127. regnetx_006
128. regnetx_008
129. regnetx_016
130. regnetx_032
131. regnetx_040
132. regnetx_064
133. regnetx_080
134. regnetx_120
135. regnetx_160
136. regnetx_320
137. regnety_002
138. regnety_004
139. regnety_006
140. regnety_008
141. regnety_016
142. regnety_032
143. regnety_040
144. regnety_064
145. regnety_080
146. regnety_120
147. regnety_160
148. regnety_320
149. repvgg_a2
150. repvgg_b0
151. repvgg_b1
152. repvgg_b1g4
153. repvgg_b2
154. repvgg_b2g4
155. repvgg_b3
156. repvgg_b3g4
157. res2net101_26w_4s
158. res2net50_14w_8s
159. res2net50_26w_4s
160. res2net50_26w_6s
161. res2net50_26w_8s
162. res2net50_48w_2s
163. res2next50
164. resnest101e
165. resnest14d
166. resnest200e
167. resnest269e
168. resnest26d
169. resnest50d_1s4x24d
170. resnest50d_4s2x40d
171. resnest50d
172. resnet101d
173. resnet152d
174. resnet18
175. resnet18d
176. resnet200d
177. resnet26
178. resnet26d
179. resnet34
180. resnet34d
181. resnet50
182. resnet50d
183. resnetblur50
184. resnetv2_101x1_bitm_in21k
185. resnetv2_101x1_bitm
186. resnetv2_101x3_bitm_in21k
187. resnetv2_101x3_bitm
188. resnetv2_152x2_bitm_in21k
189. resnetv2_152x2_bitm
190. resnetv2_152x4_bitm_in21k
191. resnetv2_152x4_bitm
192. resnetv2_50x1_bitm_in21k
193. resnetv2_50x1_bitm
194. resnetv2_50x3_bitm_in21k
195. resnetv2_50x3_bitm
196. resnext101_32x8d
197. resnext50_32x4d
198. resnext50d_32x4d
199. rexnet_100
200. rexnet_130
201. rexnet_150
202. rexnet_200
203. selecsls42b
204. selecsls60
205. selecsls60b
206. semnasnet_100
207. seresnet152d
208. seresnet50
209. seresnext26d_32x4d
210. seresnext26t_32x4d
211. seresnext50_32x4d
212. skresnet18
213. skresnet34
214. skresnext50_32x4d
215. spnasnet_100
216. ssl_resnet18
217. ssl_resnet50
218. ssl_resnext101_32x16d
219. ssl_resnext101_32x4d
220. ssl_resnext101_32x8d
221. ssl_resnext50_32x4d
222. swsl_resnet18
223. swsl_resnet50
224. swsl_resnext101_32x16d
225. swsl_resnext101_32x4d
226. swsl_resnext101_32x8d
227. swsl_resnext50_32x4d
228. tf_efficientnet_b0_ap
229. tf_efficientnet_b0_ns
230. tf_efficientnet_b0
231. tf_efficientnet_b1_ap
232. tf_efficientnet_b1_ns
233. tf_efficientnet_b1
234. tf_efficientnet_b2_ap
235. tf_efficientnet_b2_ns
236. tf_efficientnet_b2
237. tf_efficientnet_b3_ap
238. tf_efficientnet_b3_ns
239. tf_efficientnet_b3
240. tf_efficientnet_b4_ap
241. tf_efficientnet_b4_ns
242. tf_efficientnet_b4
243. tf_efficientnet_b5_ap
244. tf_efficientnet_b5_ns
245. tf_efficientnet_b5
246. tf_efficientnet_b6_ap
247. tf_efficientnet_b6_ns
248. tf_efficientnet_b6
249. tf_efficientnet_b7_ap
250. tf_efficientnet_b7_ns
251. tf_efficientnet_b7
252. tf_efficientnet_b8_ap
253. tf_efficientnet_b8
254. tf_efficientnet_cc_b0_4e
255. tf_efficientnet_cc_b0_8e
256. tf_efficientnet_cc_b1_8e
257. tf_efficientnet_el
258. tf_efficientnet_em
259. tf_efficientnet_es
260. tf_efficientnet_l2_ns_475
261. tf_efficientnet_l2_ns
262. tf_efficientnet_lite0
263. tf_efficientnet_lite1
264. tf_efficientnet_lite2
265. tf_efficientnet_lite3
266. tf_efficientnet_lite4
267. tf_inception_v3
268. tf_mixnet_l
269. tf_mixnet_m
270. tf_mixnet_s
271. tf_mobilenetv3_large_075
272. tf_mobilenetv3_large_100
273. tf_mobilenetv3_large_minimal_100
274. tf_mobilenetv3_small_075
275. tf_mobilenetv3_small_100
276. tf_mobilenetv3_small_minimal_100
277. tresnet_l_448
278. tresnet_l
279. tresnet_m_448
280. tresnet_m
281. tresnet_xl_448
282. tresnet_xl
283. tv_densenet121
284. tv_resnet101
285. tv_resnet152
286. tv_resnet34
287. tv_resnet50
288. tv_resnext50_32x4d
289. vgg11_bn
290. vgg11
291. vgg13_bn
292. vgg13
293. vgg16_bn
294. vgg16
295. vgg19_bn
296. vgg19
297. vit_base_patch16_224_in21k
298. vit_base_patch16_224
299. vit_base_patch16_384
300. vit_base_patch32_224_in21k
301. vit_base_patch32_384
302. vit_base_resnet50_224_in21k
303. vit_base_resnet50_384
304. vit_deit_base_distilled_patch16_224
305. vit_deit_base_distilled_patch16_384
306. vit_deit_base_patch16_224
307. vit_deit_base_patch16_384
308. vit_deit_small_distilled_patch16_224
309. vit_deit_small_patch16_224
310. vit_deit_tiny_distilled_patch16_224
311. vit_deit_tiny_patch16_224
312. vit_large_patch16_224_in21k
313. vit_large_patch16_224
314. vit_large_patch16_384
315. vit_large_patch32_224_in21k
316. vit_large_patch32_384
317. vit_small_patch16_224
318. wide_resnet101_2
319. wide_resnet50_2
320. xception
321. xception41
322. xception65
323. xception71

---

Get List Models

import timm
from pprint import pprint
model_names = timm.list_models(pretrained=True)
pprint(model_names)
>>> ['adv_inception_v3',
 'cspdarknet53',
 'cspresnext50',
 'densenet121',
 'densenet161',
 'densenet169',
 'densenet201',
 'densenetblur121d',
 'dla34',
 'dla46_c',
...
]

Get Model Parameters

import timm
from pprint import pprint
m = timm.create_model('efficientnet_b0', pretrained=True)
pprint(m.default_cfg)

Timm documentation here

---

License

Project is distributed under MIT License