torchscript model size increases when loading and saving it again

[iseeyuan]

🐛 Bug

For the same jit model file, if I load and save it again, the size increases from 1.8 MB to 3.1 MB.

To Reproduce

Steps to reproduce the behavior:

1. Please note that the model link provided below is an internal link
   cp /mnt/vol/gfsai-oregon/aml/mobile-vision/body_tracking/ts_f179073222_300_10_3_opt_09162020/detectandtrack.jit ~/models/.
2. Python

m = torch.jit.load("/Users/myuan/models/detectandtrack.jit")
m.save("/Users/myuan/models/detectandtrack_save_again.jit")

Expected behavior

detectandtrack_save_again.jit is 3 MB in size, which is larger than the original detectandtrack.jit.

Environment

• PyTorch Version (1.6):
• OS (Mac, CentOS):
• How you installed PyTorch (Install, internal build):
• Build command you used (if compiling from source):
• Python version: 3.7
• CUDA/cuDNN version: N/A
• GPU models and configuration: N/A
• Any other relevant information:
  The data folder in detectandtrack_save_again.jit is larger (from 1.7 MB to 3.8 MB) with more items (from 243 to 296 files).

Additional context

The model was generated with this script which takes in a trained model, runs optimize_for_mobile and saves the mobile model: https://www.internalfb.com/intern/diffusion/FBS/browse/master/fbcode/mobile-vision/projects/detectron2go/tools/generate_tracker.py?commit=1b05890d965b84d9390bcd55dfbd625073537392&lines=83-104
The model generation can be reproduced with the test plan in: D23739593

cc @gmagogsfm

[nikithamalgifb]

@ iseeyuan: Hi, I am working on this issue. Can you please help me with reproducing the issue? I wanted to understand how I could get a hold of the above model?

[dzhulgakov]

@nikithamalgifb - see the description in #46749 (comment) specifically https://colab.research.google.com/drive/1qWUq6PQ4fj67TBbD3xg_BvYJve9E91AZ?usp=sharing

[nikithamalgifb]

Removing the JIT tag, since this is not a JIT issue