TFLite failures resulted from TF latest version upgrade resolved #6774
Conversation
|
cc @leandron |
| if quantized: | ||
| if package_version.parse(tf.VERSION) < package_version.parse("2.1.0"): | ||
| pytest.skip("Testcase requires tflite version >= 2.1.0") | ||
| data_in = tf.keras.layers.Input(shape=data.shape[1:]) | ||
| relu = tf.keras.layers.ReLU()(data_in) | ||
| keras_model = tf.keras.models.Model(inputs=data_in, outputs=relu) | ||
| input_name = data_in.name.split(":")[0] | ||
|
|
||
| # To create quantized values with dynamic range of activations, needs representative dataset | ||
| def representative_data_gen(): | ||
| for i in range(1): | ||
| yield [data] | ||
|
|
||
| tflite_model_quant = _quantize_keras_model(keras_model, representative_data_gen) | ||
|
|
||
| tflite_output = run_tflite_graph(tflite_model_quant, data) | ||
| tvm_output = run_tvm_graph(tflite_model_quant, data, input_name) | ||
| tvm.testing.assert_allclose( | ||
| np.squeeze(tvm_output[0]), np.squeeze(tflite_output[0]), rtol=1e-5, atol=1e-5 | ||
| ) | ||
| else: | ||
| with tf.Graph().as_default(): | ||
| in_data = array_ops.placeholder(shape=data.shape, dtype=data.dtype) | ||
| with tf.Graph().as_default(): | ||
| in_data = array_ops.placeholder(shape=data.shape, dtype="float32", name="in_0") | ||
|
|
||
| if quantized: | ||
| inq_data = tf.quantization.fake_quant_with_min_max_args( | ||
| in_data, min=-10, max=10, name="inq_0" | ||
| ) | ||
| input_range = {"inq_0": (-10, 10)} | ||
| out = nn_ops.relu(inq_data) | ||
| out = tf.quantization.fake_quant_with_min_max_args(out, min=0, max=6, name="out") | ||
| compare_tflite_with_tvm( | ||
| data, "inq_0:0", [inq_data], [out], quantized=True, input_range=input_range | ||
| ) | ||
| else: |
There was a problem hiding this comment.
Could you give a hint as to the reason for these changes ? is there any reason why the existing mechanism doesn't work in your research ?
There was a problem hiding this comment.
Yes, generally, we want to move towards keras quantization as TF gets more mature. What issues do you see?
There was a problem hiding this comment.
The issue here was the test case was not behaving as expected. When we need a quantized graph with quantized inference, then we should provide quant input(Uint8 or int8, based on settings chosen). But here it was feeding float input to the graph, so the failure. Also i think current change is to keep a uniform steps as followed similar to case of other operators. Any specific reason to move towards Keras quantization than TF quant ?
Please let me know, if still need to keep Keras quant. Thanks!
ANSHUMAN87
force-pushed
the
tflite-ver-upgrade
branch
from
aed9439
to
04c9b16
Compare
|
@leandron please confirm if it fixes your problem and we can proceed with a merge versus revert |
I re-ran the tests and it seems this PR fixes them all. Thanks @ANSHUMAN87. |
Thanks @leandron for confirmation. |
|
Is there a committer around who might be able to merge this in now ? |
|
Thanks @u99127 , @tqchen , @leandron , @anijain2305 ! |
…che#6774) * TFLite failures resulted from TF latest version upgrade resolved * [1] Review comments handled
…che#6774) * TFLite failures resulted from TF latest version upgrade resolved * [1] Review comments handled
…che#6774) * TFLite failures resulted from TF latest version upgrade resolved * [1] Review comments handled
This PR addressed below 2 points:
TFLite frontend version support upgraded from 2.1 to 2.3
Resolved test failures reported because of Tensorflow version upgrade in CI from 2.1.0 to 2.3.1 (Please refer)
cc @tqchen