ANSHUMAN87 commented on a change in pull request #7107:
URL: https://github.com/apache/tvm/pull/7107#discussion_r542585227
##########
File path: tutorials/frontend/deploy_sparse.py
##########
@@ -314,13 +310,20 @@ def run_sparse(mod, params, shape_dict, target, ctx,
bs_r, sparsity, gen_weights
# you'll need to uncomment the last line first.
def benchmark():
mod, params, shape_dict = import_graphdef(name, batch_size, seq_len)
- run_dense(mod, params, shape_dict, target, ctx)
+ input_shape = shape_dict["input_1"]
+ dummy_data = np.random.uniform(size=input_shape, low=0,
high=input_shape[1]).astype("int32")
+ dense_output = run_dense(mod, params, shape_dict, target, ctx, dummy_data)
if measure_sparse:
gen_weights = "prune" not in name
- run_sparse(mod, params, shape_dict, target, ctx, bs_r, sparsity,
gen_weights)
+ sparse_output = run_sparse(
+ mod, params, shape_dict, target, ctx, bs_r, sparsity, gen_weights,
dummy_data
+ )
+ np.testing.assert_allclose(
+ dense_output, sparse_output, equal_nan=True, verbose=True,
atol=1e-5, rtol=1e-5
+ )
-# benchmark()
+benchmark()
Review comment:
Got your point. Perhaps we disable the benchmarking part by default and
compare only the output!
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
