Error when i create new model and split model

1

I need to process and refine the model for many different purposes, one automatically (split, concat, …) both in terms of model and weight. I can use some models (VGG, EffciennetB3, Xception, ViT), however in some variants of ViT (CrossViT, CrossFormer) I get error messages, still can’t find a fix.

import torch
from vit_pytorch.cross_vit import CrossViT
import torch.nn as nn
v = CrossViT(
image_size = 256,
num_classes = 1000,
depth = 4, # number of multi-scale encoding blocks
sm_dim = 192, # high res dimension
sm_patch_size = 16, # high res patch size (should be smaller than lg_patch_size)
sm_enc_depth = 2, # high res depth
sm_enc_heads = 8, # high res heads
sm_enc_mlp_dim = 2048, # high res feedforward dimension
lg_dim = 384, # low res dimension
lg_patch_size = 64, # low res patch size
lg_enc_depth = 3, # low res depth
lg_enc_heads = 8, # low res heads
lg_enc_mlp_dim = 2048, # low res feedforward dimensions
cross_attn_depth = 2, # cross attention rounds
cross_attn_heads = 8, # cross attention heads
dropout = 0.1,
emb_dropout = 0.1
)

v = nn.Sequential(*list(v.children()))
v1 = nn.Sequential(*list(v.children()[:5]))

img = torch.randn(1, 3, 256, 256)

pred = v(img) # (1, 1000)
print(pred)

“einops.EinopsError: Expected 4 dimensions, got 3”

from vit_pytorch.crossformer import CrossFormer

model = CrossFormer(
num_classes = 1000, # number of output classes
dim = (64, 128, 256, 512), # dimension at each stage
depth = (2, 2, 8, 2), # depth of transformer at each stage
global_window_size = (8, 4, 2, 1), # global window sizes at each stage
local_window_size = 7, # local window size (can be customized for each stage, but in paper, held constant at 7 for all stages)
)

img = torch.randn(1, 3, 224, 224)
model = nn.Sequential(*list(model.children()))
pred = model(img) # (1, 1000)

“NotImplementedError: Module [ModuleList] is missing the required “forward” function”

2

I don’t know where this error is raised from:

but the error message is pointing towards a dimension mismatch so you would need to check which tensors are used and that they have the expected number of dimensions.

This error is raise if you try to “call” an nn.ModuleList object like a module, which won’t work as you would need to iterate it.
Here is a small example:

modules = nn.ModuleList([
    nn.Linear(10, 10),
    nn.Linear(10, 10),
    nn.Linear(10, 10),
    nn.Linear(10, 10),
    nn.Linear(10, 10),
])

x = torch.randn(1, 10)

# breaks
out = modules(x)
# NotImplementedError: Module [ModuleList] is missing the required "forward" function

# works
for module in modules:
    x = module(x)