nachiket273 / VisTrans

Implementations of transformers based models for different vision tasks

VisTrans

Implementations of transformers based models for different vision tasks

Install

1. Install from PyPI

pip install vistrans

2. Install from Anaconda

conda install -c nachiket273 vistrans

Version 0.003 (06/30/2021)

---

Minor fixes to fix issues with existing models.

Version 0.002 (04/17/2021)

---

Pretrained Pytorch Bottleneck Transformers for Visual Recognition including following

• botnet50
• botnet101
• botnet152

Implementation based off Official Tensorflow Implementation

Usage

---

pip install vistrans

1) List Pretrained Models.
```Python
from vistrans import BotNet
BotNet.list_pretrained()

2. Create Pretrained Models.

from vistrans import BotNet
model = BotNet.create_pretrained(name, img_size, in_ch, num_classes,
                                 n_heads, pos_enc_type)

3. Create Custom Model

from vistrans import BotNet
model = BotNet.create_model(layers, img_size, in_ch, num_classes, groups,
                            norm_layer, n_heads, pos_enc_type)

Version 0.001 (03/04/2021)

---

Pretrained Pytorch Vision Transformer Models including following

• vit_s16_224
• vit_b16_224
• vit_b16_384
• vit_b32_384
• vit_l16_224
• vit_l16_384
• vit_l32_384

Implementation based off official jax repository and timm's implementation

Usage

---

1. List Pretrained Models.

from vistrans import VisionTransformer
VisionTransformer.list_pretrained()

2. Create Pretrained Models.

from vistrans import VisionTransformer
model = VisionTransformer.create_pretrained(name, img_size, in_ch, num_classes)

3. Create Custom Model

from vistrans import VisionTransformer
model = VisionTransformer.create_model(img_size, patch_size, in_ch, num_classes,
                                       embed_dim, depth, num_heads, mlp_ratio,
                                       drop_rate, attention_drop_rate, hybrid,
                                       norm_layer, bias)