Initial implementation of SavedModel I/O #13
Conversation
tests/graph_test.py
Outdated
| # Remove the directory after the test. | ||
| # Comment out this line to prevent deleting temps. | ||
| shutil.rmtree(self.temp_dir) | ||
| pass |
There was a problem hiding this comment.
You do if you comment out the previous line ;-) I'll add a comment to that effect.
There was a problem hiding this comment.
Ah, I see. I totally missed the comment :)
tests/rewrite_test.py
Outdated
| name="Input") | ||
| result_tensor = input_tensor + 42 | ||
| g = gde.Graph(tf_g) | ||
| gde.rewrite.change_batch_size(g, 3, [g[input_tensor.op.name]]) | ||
| print("Graph def is:\n{}".format(g.to_graph_def())) |
There was a problem hiding this comment.
did you mean to keep this print statement?
There was a problem hiding this comment.
Ugh, I thought I got all those. Thanks for catching that one!
tests/rewrite_test.py
Outdated
| np.array([42., 43.]).reshape([2, 1]))) | ||
|
|
||
| # Remove temp dir if the test is successful | ||
| shutil.rmtree(temp_dir) |
There was a problem hiding this comment.
did you want to keep the director if the test was not successful? If not then it might be good to put this in a finally block.
There was a problem hiding this comment.
Good point. I had been using the directory's contents for debugging, but it's not a good idea to check in code that pollutes /tmp
tests/rewrite_test.py
Outdated
| SavedModel | ||
| """ | ||
| temp_dir = tempfile.mkdtemp() | ||
| print("Temp dir is {}".format(temp_dir)) |
graph_def_editor/graph.py
Outdated
| return self._signature_defs | ||
|
|
||
|
|
||
|
|
There was a problem hiding this comment.
nit: i think extra newline here
There was a problem hiding this comment.
Oops, I neglected to go through the PEP8 linter warnings before checking in. Fixed that newline and a bunch of other, similar warnings.
This PR contains my implementation of issue #5 -- reading and writing SavedModel files.
There are limitations on the use of variables when writing to SavedModel files as described in #5.
I added tests of the new functionality.
I modified the example
batch_size_example.pyto read and write SavedModel files. After the modification, I noticed that, although the example script correctly returns a single row in its result, the script infers a size of (64, 1001) for thesoftmax_tensoroutput, i.e.:This result appears to be technically correct --- the graph contains a
batch_normalizationmeta-operator that hard-codes a batch size of 64. Some follow-on work will be needed to make this example set the batch size properly all the way to the end.