The issue stems from incomplete downloads when Circle CI downloads the wheels in order to install the dependencies. The download is halted, the file is incomplete which results in a different hash. With this PR, the CirlceCI ~/.cache/pip directory which contains the downloaded wheels is cached between runs, which means that the files won't be re-downloaded as long as the latest version is available in the cache. A cache is created for each workflow. Unfortunately, CircleCI does not make it available to update the cache nor to delete the cache. If a new version of either Torch or TensorFlow is released, the cache won't be updated until we update either setup.py, the cache version, or until the cache expires 15 days after its creation. If the cache works well enough, I'll include ~/.cache/torch in the cache, so that all the downloads done during the tests are saved in the cache as well. This should prevent other flaky tests from happening due to missing models on the S3.