[Pytorch] [Quantization] Error during quantization

masahi · July 4, 2020, 1:38am

ok I bisected, if you undo this commit it should work

github.com/apache/tvm

[Relay] Fix Calibration Pass to Support Modules with Multiple Functions

master ← seanlatias:master

opened 04:47PM - 10 Jun 20 UTC

seanlatias

+8 -4

With the existing calibration pass, it only returns the module that is built fro…m the main function. However, the input can have more than one functions (e.g., after using BYOC). This PR provides a quick fix by setting the function dictionary of `from_expr`. Following is a case where the existing pass would go wrong but is fixed by this PR. ```python def test_extern_dnnl_mobilenet(): if not tvm.get_global_func("relay.ext.dnnl", True): print("skip because DNNL codegen is not available") return dtype = 'float32' ishape = (1, 3, 224, 224) mod, params = relay.testing.mobilenet.get_workload( batch_size=1, dtype='float32') mod = transform.AnnotateTarget(["dnnl"])(mod) mod = transform.MergeCompilerRegions()(mod) mod = transform.PartitionGraph()(mod) with relay.quantize.qconfig(skip_conv_layers=[]): mod = relay.quantize.quantize(mod, params) ``` However, with this current fix, we still only run the calibration pass on the main function instead of all input functions. In the future, we may need to consider applying the analysis to all functions. I also did a quick check on other passes to see if there exist similar issues but I think this is the only pass. @comaniac @zhiics @anijain2305 please take a look. Thanks.

cc @seanlatias