[Relay] Add a PyTorch to Relay Parser#4497

Merged

icemelon merged 136 commits intoapache:masterfrom

alexwong:upstream-pytorch-relay

Feb 25, 2020

Commits on Feb 24, 2020

Add a PyTorch to Relay parser
alexwong
committed
Add alexnet, googlenet, mnasnet, shufflenet wip
alexwong
committed
Fix lint
alexwong
committed
Remove fix for shufflenet
alexwong
committed
Lower check
alexwong
committed
Pull changes from neo-ai/tvm changes
alexwong
committed
Remove commented out section
alexwong
committed
Use infer_shape everywhere
alexwong
committed
Change back to using trace instead of path in from_pytorch
alexwong
committed
Parse state_dict to add param names
alexwong
committed
Umbrella single_op under test_forwards
alexwong
committed
Remove print and cleanup call
alexwong
committed
Check if update to test broke CI
alexwong
committed
Retrigger CI
alexwong
committed
Add back in updated tests
alexwong
committed
Try splitting up tests
alexwong
committed
First pass at flexible typing, implemented for ones
alexwong
committed
Add int32 for all ops
alexwong
committed
Remove print statements
alexwong
committed
Fix lint
alexwong
committed
Broad except
alexwong
committed
Add other tensor types
alexwong
committed
Temporarily use old tests
alexwong
committed
Retrigger CI
alexwong
committed
Lower type names
alexwong
committed
Use numpy to convert in dense op
alexwong
committed
Fix lint
alexwong
committed
Remove print
alexwong
committed
Need to cleanup but verify int32 works for add
alexwong
committed
Rough tests for different types, a lot of types are not supported on CPU
alexwong
committed
Probably doesn't build, need to save work as I have to switch branches (constantly)
alexwong
committed
Parse param type
alexwong
committed
Remove print stmt in parser
alexwong
committed
Clean up some code
alexwong
committed
Working on flaot32 for bn
alexwong
committed
Add resnet18 double type
alexwong
committed
Fix lint
alexwong
committed
Temporarily move PT tests first
alexwong
committed
Temporarily add back refactored tests to fix mem issue
alexwong
committed
Add more type test and temp remove some tests
alexwong
committed
Comment out tests, hopefully CI prints a trace
alexwong
committed
Get stack trace
alexwong
committed
Remove operator dict key, rename op_name to node_id, remove dead code
alexwong
committed
Make relay map a list
alexwong
committed
Remove some hacky string stuff
alexwong
committed
Move to PyTorch 1.4
alexwong
committed
Remove input_type as param
alexwong
committed
Remove _get_fill_value, fix full ops
alexwong
committed
Remove unused code and combine ops for identity and none
alexwong
committed
Remove fn_param
alexwong
committed
Clean up main loop
alexwong
committed
Remove useless if/else for outputs
alexwong
committed
Remove ir_names, only used once
alexwong
committed
Remove some string hacking
alexwong
committed
Remove string parsing to get output name
alexwong
committed
Fix bug with output sizes of nodes
alexwong
committed
Use attributeNames in parse ops
alexwong
committed
Remove continue and add_op in parse_op
alexwong
committed
Do this everywhere, use assert instead of explciitly type casting
alexwong
committed
Remove unnecessary swap
alexwong
committed
Slight refactor for elemwise input parse
alexwong
committed
Use a copy of graph everywhere
alexwong
committed
Rename nid_to_node_name
alexwong
committed
Refactor parse import prereqs
alexwong
committed
Clean up input node kind check
alexwong
committed
Clean up conditionals
alexwong
committed
Clean up add_op
alexwong
committed
Cleanup type for ones and zeros op
alexwong
committed
Fix lint
alexwong
committed
Add torch install to CI
alexwong
committed
Actually use torch
alexwong
committed
Try moving import torch to only where it's needed
alexwong
committed
Import torch for CI
alexwong
committed
Use take op for select
alexwong
committed
Temporarily add ignore for jit inline pass for CI
alexwong
committed
Use CompleteTensorType, might be a PT 1.2 only thing
alexwong
committed
Use different types in elemwise op
alexwong
committed
Use float16 ones
alexwong
committed
Fix float16 test
alexwong
committed
Remove the temp docker changes
alexwong
committed
Remove temp test
alexwong
committed
Temporarily comment out original tests
alexwong
committed
Remove file
alexwong
committed
Empty cache after each test
alexwong
committed
Add some prints and lower input sizes
alexwong
committed
Try using no grad
alexwong
committed
Trying to globally set grad off
alexwong
committed
Use no grad for torchvision
alexwong
committed
Remove xfail tests
alexwong
committed
Remove VGG and AlexNet due to some issues
alexwong
committed
Combine pooling tests
alexwong
committed
Remove extra test file
alexwong
committed
Remove single op, remove larger pooling tests
alexwong
committed
Remove maxpool3
alexwong
committed
Remove debug prints
alexwong
committed
Remove inference call and add no_grad in measure latency
alexwong
committed
Use standard string start char
alexwong
committed
Remove redundant infer_shape in slice
alexwong
committed
Convert most to checks to just expr
alexwong
committed
Remove extra paren
alexwong
committed
More refactor of isinstance
alexwong
committed
Add helper for creating typed constants
alexwong
committed
Assert instead of return when no matching type
alexwong
committed
Remove network variants
alexwong
committed
Add no_grad when forward, remove deatch, fix lint
alexwong
committed
Change isinstance to expr in transpose
alexwong
committed
Use opnotimplemented, refactor
alexwong
committed
Fix full ops, remove duplicate tests
alexwong
committed
Never use shape field unless we know the type
alexwong
committed
Remove comma, retrigger CI
alexwong
committed
Add paren, retrigger CI
alexwong
committed
Use inline if-else for flags
alexwong
committed
Throw exception instead of assert
alexwong
committed
Remove version check for CI
alexwong
committed
Check version when doing inline pass
alexwong
committed
Fix lint
alexwong
committed
Lower more input sizes
alexwong
committed
Add new line, conv2d only accepts weight as expr
alexwong
committed
Use tvm.runtime.ndarray
alexwong
committed
Remove change to torch version install
alexwong
committed
Try no grad for mobilenet
alexwong
committed
Fix lint
alexwong
committed
Fix lint again
alexwong
committed
Revert to last passing
alexwong
committed
Delete test files
alexwong
committed
Ignore lint
alexwong
committed
Revert back
alexwong
committed
Comment out mobilenet
alexwong
committed
Clean up compare compiled and baseline outputs
alexwong
committed
Use IRModule
alexwong
committed
Add todos
alexwong
committed
Refactor use_bias
alexwong
committed
Add todo for fix conv op channels
alexwong
committed
Change input to data type
alexwong
committed
Remove todo
alexwong
committed
Handle channel multiplier > 1
alexwong
committed