access2rohit commented on a change in pull request #12637: [MXNET-912]
Refactoring ctc loss operator
URL: https://github.com/apache/incubator-mxnet/pull/12637#discussion_r223193574
##########
File path: tests/python/unittest/test_operator.py
##########
@@ -4619,6 +4647,85 @@ def check_ctc_loss_grad(blank_label): # from tf
label_lens = np.array([5, 4], dtype=np.int32)
loss_truth = np.array([-loss_log_prob_0, -loss_log_prob_1], np.float32)
+ with default_context():
+ data = mx.nd.array(inputs)
+ label = mx.nd.array(labels)
+ data.attach_grad()
+ with mx.autograd.record():
+ l = mx.ndarray.CTCLoss(data, label,
+ use_data_lengths=True,
+ use_label_lengths=True,
+ data_lengths=mx.nd.array(seq_lens),
+ label_lengths=mx.nd.array(label_lens),
+ blank_label=blank_label)
+ l.backward()
+ assert_almost_equal(l.asnumpy(), loss_truth, atol=1e-5, rtol=1e-5)
+ assert_almost_equal(data.grad.asnumpy(), grad_truth, atol=1e-5,
rtol=1e-5)
+
+ # check contrib operator for backward compatibility
+ def check_contrib_ctc_loss_grad(blank_label): # from tf
+ vocab_size = 5
+ max_label_len = 5
+ padding_mask = -1+ (blank_label=='first')
+
+ targets_0 = [0, 1, 2, 1, 0]
+ loss_log_prob_0 = -3.34211
+ input_prob_matrix_0 = np.asarray(
+ [[0.633766, 0.221185, 0.0917319, 0.0129757, 0.0142857, 0.0260553],
+ [0.111121, 0.588392, 0.278779, 0.0055756, 0.00569609, 0.010436],
+ [0.0357786, 0.633813, 0.321418, 0.00249248, 0.00272882,
0.0037688],
+ [0.0663296, 0.643849, 0.280111, 0.00283995, 0.0035545,
0.00331533],
+ [0.458235, 0.396634, 0.123377, 0.00648837, 0.00903441,
0.00623107]],
Review comment:
I see .... previous tests were also manual. My point was if it is possible
to write the same functionality in python like a naive approach (it wouldn't be
optimal of course, but for sanity check). Since we would want to improve our
tests. On the other hand implementation of such logic takes time similar in
amount to implementing the operator itself them we can park it in our backlog
as todo but having more versatile tests would definitely help :)
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
