search

hustvl / Vim

[ICML 2024] Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model
Apache License 2.0
2.83k stars 182 forks source link

封装Vim模型调用 #103

Open MPCheng-ZW opened 2 months ago

MPCheng-ZW commented 2 months ago

您好！

请问有可以封装好的Vim可以调用吗？

类似于mamba这种： import torch from mamba_ssm import Mamba

batch, length, dim = 2, 64, 16 x = torch.randn(batch, length, dim).to("cuda") model = Mamba(

This module uses roughly 3 expand d_model^2 parameters

d_model=dim, # Model dimension d_model
d_state=16,  # SSM state expansion factor
d_conv=4,    # Local convolution width
expand=2,    # Block expansion factor

).to("cuda") y = model(x) assert y.shape == x.shape

谢谢~

mahao18cm commented 4 weeks ago

model = create_model( 'vim_tiny_patch16_224_bimambav2_final_pool_mean_abs_pos_embed_with_midclstok_div2', pretrained=False, num_classes=1000, drop_rate=0.0, drop_path_rate=0.1, drop_block_rate=None, img_size=224 )