[PT / TF] Add TextNet - FAST backbone

[felixdittrich92]

This PR:

• Adds TextNet (FAST detection model backbone) - https://arxiv.org/pdf/2111.02394
• Only the backbone for v0.8.0 (Rest after the next release)

Overview (tiny version)

"""
TORCH:

==============================================================
textnet     TextNet              (-1, 126)           0
├─0         Conv2d               (-1, 64, 16, 16)    1,728
├─1         BatchNorm2d          (-1, 64, 16, 16)    257
├─2         ReLU                 (-1, 64, 16, 16)    0
├─3         Sequential           (-1, 512, 1, 1)     0
|    └─0    Sequential           (-1, 64, 8, 8)      187,147
|    └─1    Sequential           (-1, 128, 4, 4)     505,868
|    └─2    Sequential           (-1, 256, 2, 2)     2,011,148
|    └─3    Sequential           (-1, 512, 1, 1)     8,019,980
├─4         Sequential           (-1, 126)           0
|    └─0    AdaptiveAvgPool2d    (-1, 512, 1, 1)     0
|    └─1    Flatten              (-1, 512)           0
|    └─2    Linear               (-1, 126)           64,638
==============================================================

Trainable params: 10,769,470
Non-trainable params: 0
Total params: 10,769,470
--------------------------------------------------------------
Model size (params + buffers): 41.18 Mb
Framework & CUDA overhead: 236.00 Mb
Total RAM usage: 277.18 Mb
--------------------------------------------------------------
Floating Point Operations on forward: 96.74 MFLOPs
Multiply-Accumulations on forward: 48.39 MMACs
Direct memory accesses on forward: 59.30 MDMAs



TF:
 Layer (type)                Output Shape              Param #
=================================================================
 conv2d (Conv2D)             (None, 16, 16, 64)        1728

 batch_normalization (Batch  (None, 16, 16, 64)        256
 Normalization)

 activation (Activation)     (None, 16, 16, 64)        0

 sequential (Sequential)     (None, 8, 8, 64)          187136

 sequential_1 (Sequential)   (None, 4, 4, 128)         505856

 sequential_2 (Sequential)   (None, 2, 2, 256)         2011136

 sequential_3 (Sequential)   (None, 1, 1, 512)         8019968

 sequential_4 (Sequential)   (None, 126)               64638

=================================================================
Total params: 10790718 (41.16 MB)
Trainable params: 10767678 (41.08 MB)
Non-trainable params: 23040 (90.00 KB)
_________________________________________________________________
None
"""

Any feedback is welcome 🤗

TODO:

• provide checkpoints
• check code again

[felixdittrich92]

CC @nikokks

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (ff9982b) 95.76% compared to head (82de763) 95.86%.
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1425      +/-   ##
==========================================
+ Coverage   95.76%   95.86%   +0.10%     
==========================================
  Files         155      161       +6     
  Lines        6941     7085     +144     
==========================================
+ Hits         6647     6792     +145     
+ Misses        294      293       -1     

Flag	Coverage Δ
unittests	95.86% <100.00%> (+0.10%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[felixdittrich92]

@odulcy-mindee Checkpoints are added if everything is fine feel free to upload that we can change the URLs :)

All models reached an Accuracy between 96-98%

[odulcy-mindee]

Wohohoho, thank you @felixdittrich92 ! 🚀

[odulcy-mindee]

Thank you @felixdittrich92, really great job here !