[models] Complete the list of pretrained available models in TensorFlow and PyTorch

[felixdittrich92]

🚀 The feature

Ensure that each model has a pretrained available checkpoint

TensorFlow:

classification:

• [x] ViT
• [x] MAGC
• [x] MobileNetV3

detection:

• [x] DB_MobileNetV3
• [X] LinkNet (all)

recognition:

• [x] SAR
• [x] MASTER
• [x] VitSTR (small)
• [x] VitSTR (base)
• [x] PARSeq
• [x] CRNN (mobilenet_v3_large)

PyTorch:

classification:

• [x] ViT
• [x] ResNet34_wide (@felixdittrich92)

detection:

• [ ] DB_ResNet34
• [ ] LinkNet (all)

recognition:

• [x] SAR
• [x] MASTER
• [x] VitSTR (small)
• [x] VitSTR (base)
• [x] PARSeq

Motivation, pitch

• Users wants to test each available model and needs the pretrained versions to finetune on there own data

[felixdittrich92]

@charlesmindee i think you are the only one who can do it (data) :smiley:

[nikokks]

@charlesmindee, what kind of data do you use to train your text recogntions models ? Do you use custom data ? and do you use special data for french documents ?

[frgfm]

@nikokks for text recognition, we use cropped words from various types of documents, which sum up to about 11M images. The bias is indeed around the choice of characters (European mostly) and the language (some sequences of characters are more frequent than others).

However, a while back, we started a fully synthetic approach (cf. CharacterGenerator and WordGenerator). So far, the results are not on par, but still very decent for a fully synthetic approach :)

[odulcy-mindee]

Training started on Linknet_Resnet50 (Tensorflow), we'll post the results soon :wink:

[odulcy-mindee]

Detection models in Tensorflow are done! Now, we start recognition models in Tensorflow. First one is master!

[felixdittrich92]

Closing because we track it in #1364