Summary

I cannot use caching while instantiating a Pytorch model via torchvision.models.__dict__, while it does work with torchvision.models.resnet18.

Steps to reproduce

import streamlit as st 
import torchvision

@st.cache(allow_output_mutation=True)
def fun():
    model = torchvision.models.__dict__['resnet18'](pretrained=True)
    return model

model = fun()

Expected behavior:

It should work just fine as this does work:

def fun():
    model = torchvision.models.resnet18(pretrained=True)
    return model

Actual behavior:

The app throws the following error:

UnhashableType: Cannot hash object of type PyCapsule

Is this a regression?

NA

Debug info

• Streamlit version: Streamlit, version 0.56.0
• Python version: Python 3.6.8 :: Anaconda, Inc.
• Using Conda .
• OS version: macOS 10.14.6
• Browser version: Safari 13.0.4

Additional information

import streamlit as st
from types import MappingProxyType

class C():
  a = 1

@st.cache(allow_output_mutation=True)
def f():
  return C.__dict__['a']

f()

raises UnhashableType: Cannot hash object of type mappingproxy

which I can resolve with adding hash_funcs={MappingProxyType: id} to st.cache arguments.