...
Code Block | ||
---|---|---|
| ||
def convert_model(sym, arg_params, aux_params, target_dtype="float16", target_dtype_ops=None, fp32_ops=None, widest_dtype_ops=None, conditional_fp32_ops=None, excluded_sym_names=None): """API for converting a model from FP32 model to a mixed precision model. MXNet tries to convert the FP32 model to mixed precision model by adding cast layers using amp_cast and amp_multicast operators. The decision on which cast layer to add is based on hardcoded lists for Automatic Mixed Precision in MXNet. These lists can be overridden by the user by providing their own lists using : targe_precision_ops, fp32_ops, widest_precision_ops, conditional_fp32_ops Parameters ---------- sym : str or Symbol Defines the structure of a neural network for FP32 types. arg_params : dict Dictionary of name to `NDArray`. aux_params : dict Dictionary of name to `NDArray`. target_dtype : str Currently only supports float16. The target dtype indicates to add cast layers when possible so that lower precision computation can be leveraged. target_dtype_ops : list of strs Override the list of operator names casted to target_dtype. If None, uses the framework's default list to be casted to target dtype. fp32_ops : list of strs Override the lists of operator names casted to FP32. If None, uses the framework's default list to be casted to FP32. widestconditional_dtypefp32_ops : list of strs A (string, string, list of op names provided by user which should run in widest precision among its inputs. If None, uses the framework's default list of widest_precision_ops. conditional_fp32_ops : list of (string, string, list of string) Override the list of operatorsstring) Override the list of operators to be casted to FP32. The format of the list is (name of the function, name of the parameter, list of values of the parameter that make the operator to be casted to FP32. The format of the list is (name of the function, name of the parameter, list of values of the parameter that make the operator to be casted to fp32) excluded_sym_names : list of strs A list of strings that represent the names of symbols that users want to exclude from being quantized. """ fp32) excluded_sym_names : list of strs A list of strings that represent the names of symbols that users want to exclude from being quantized. """ |
target_dtype should decide target_dtype should decide which lists need to be overridden.
For example, in the future bfloat16 support may be added in which case these lists for operators running in bfloat16 will also be added to AMP.
In this case, target_dtype will allow users to choose the right dtype for the mixed precision model.
...
Code Block | ||
---|---|---|
| ||
def convert_block(block, target_dtype="float16", target_dtype_ops=None, fp32_ops=None, widest_dtype_ops=None, conditional_fp32_ops=None, excluded_sym_names=None, input_names=['data']): """Given a hybrid block/symbol block representing a neural network of data type FP32 and target_dtype, return a block with mixed precision support Parameters ---------- block : HybridBlock or SymbolBlock object FP32 HybridBlock or SymbolBlock object target_dtype : str or numpy currently only supports float16. The target dtype indicates to add cast layers when possible so that lower precision computation can be leveraged. target_precision_ops : list of strs Override the list of operator names casted to target_dtype. If None, uses the framework's default list to be casted to target dtype. fp32_ops : list of strs Override the lists of operator names casted to FP32. If None, uses the framework's default list to be casted to FP32. widestconditional_precisionfp32_ops : list of (string, string, list of strsstring) Override the list of operatorfunctions namescasted whichto shouldFP32. run in widest precision among its The format of the list is input arguments. (name of Ifthe Nonefunction, name usesof the framework's default list of widest_precision_ops. parameter, conditional_fp32_ops : list of (string, string, list of string) Override the list of functions casted to FP32.values of the parameter that make the operator to be casted to fp32) excluded_sym_names : list of Thestrs format of the list is (nameA list of thestrings function, name ofthat represent the parameter, listnames of values of the parametersymbols that makeusers the operatorwant to beexclude casted to from being fp32)quantized. excluded_syminput_names : list of strs A list of strings that representrepresenting the names of symbols that users want to exclude from being quantized. input_names : list of strs A list of strings representing the names of input variables """ |
User experience will be similar to the export API experience today. Users will have to call hybridize followed by one forward pass before calling convert_model.
Backend Changes
NNVM Pass
input variables
""" |
User experience will be similar to the export API experience today. Users will have to call hybridize followed by one forward pass before calling convert_model.
Backend Changes
NNVM Pass
Add a NNVM pass for the backend. Add a NNVM pass for the backend. This would use the amp lists based on the target_dtype.
This pass will perform graph traversal and add amp_cast and amp_multicast layers for FP16 and FP32 ops based on the op whitelists and excluded_sym_names. Some of the ideas have been borrowed from quantization pass added as part of quantization support [2].
...
For Gluon code, we need to add an internal API to retrieve sym, arg_params and aux_params from a hybrid_block. Following this, convert_model can be used to convert a symbol json, model params and auxiliary params. After conversion, the symbolic model (json, arg_params, aux_params) can be imported back into gluon with SymbolBlock.imports. The returned symbolblock is ready to use for inference.
Frontend Bindings
Need to add amp convert_model API support for different bindings like C++, Scala etc.
Frontend Bindings
...
Performance
Setup
EC2 Instance: p3.8xlarge
CUDNN: 7.4.2
CUDA: 10.0
Commit Hash: b3b952f9d5490ee2707209ab866e6c3f094e2046 (PoC changes made on top of this built from source)
Mixed Precision Models:
Resnet50_v1: JSON File, Params File
imagenet1k-resnet-152: JSON File, Params File
Results
Model (Samples/sec) | Batch Size | Original Model (Samples/sec) | Mixed Precision Model (Samples/sec) | Original Model with Implicit Type Conversion (MXNET_CUDA_TENSOR_OP_MATH_ALLOW_CONVERSION=1) (Samples/sec) |
---|---|---|---|---|
imagenet1k-resnet-152 | 1 | 85 | 72 | 72 |
2 | 140 | 140 | 142 | |
4 | 240 | 270 | 228 | |
8 | 320 | 470 | 261 | |
16 | 405 | 680 | 315 | |
resnet50_v1 | 1 | 215 | 165 | 205 |
2 | 370 | 330 | 365 | |
4 | 560 | 600 | 545 | |
8 | 760 | 980 | 635 | |
16 | 935 | 1400 | 790 |
FAQ
Will the arg_params and aux_params be casted to fp16 ?
Depends on the whitelists provided. The default whitelists have been selected in a way to avoid casting of the params, for commonly used convnet networks. If the whitelist is such that the type inference decides that certain param needs to be float16 then it will be castedInputs of ops in FP16 will be casted. Other params may or may not be casted based on the type inference logic.
How is this different from casting inputs to FP16 and casting params to FP16 in Gluon ?
...