id
int64
0
190k
prompt
stringlengths
21
13.4M
docstring
stringlengths
1
12k
23,643
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def isqrt_approx_bfloat16(x, mantissa_bit=7, exponent_bit=8): # From fast inverse squre root. # gamma = 0.0450466 first order Taylor error. # magic_number = 3/2 * 2^mantissa_bit * (2^(exponent_bit-1) - 1 - gamma) # gamma = 0.0450466 # magic_n = np.array(round(3.0/2.0 * 2.0**mantissa_bit * (2.0**(exponent_bit-1) - 1.0 - gamma)), dtype=np.int16) # From https://www.mdpi.com/1099-4300/23/1/86/pdf authored by # Walczy, C.J.; Moroz, L.V..; Cie´sli´nski, J.L. # This work is licensed under the Creative Commons Attribution 4.0 # International License. To view a copy of this license, visit # http://creativecommons.org/licenses/by/4.0/ or send a letter to # Creative Commons, PO Box 1866, Mountain View, CA 94042, USA. magic_n = np.round(2**mantissa_bit * (3 * (2.0**(exponent_bit - 1) - 1.0) - 1) / 2 + np.round(2**mantissa_bit * (3.7315712401613957182292407381942955 - 2) / 4 - 0.5)) magic_n = np.array(magic_n, dtype=np.int16) x2 = (x * 0.5).to(torch.bfloat16) number = x.to(torch.float32).cpu().detach().numpy() threehalfs = 1.5 y = np.float32(number) i = y.view(np.int32) i = (magic_n - np.int32(i >> 17)) << 16 y = i.view(np.float32) y = torch.from_numpy(y).to(x.device).to(x.dtype) out = (x2 * y).to(torch.bfloat16) out = (out * y).to(torch.bfloat16) out = (threehalfs - out).to(torch.bfloat16) out = (y * out).to(torch.bfloat16) # y = y * (threehalfs - (x2 * y * y)) # y = y * (threehalfs - (x2 * y * y)) return out def isqrt_approx_walcyzk(x, mantissa_bit=23, exponent_bit=8): assert x.dtype == torch.float32 # From fast inverse squre root. # gamma = 0.0450466 first order Taylor error. # magic_number = 3/2 * 2^mantissa_bit * (2^(exponent_bit-1) - 1 - gamma) # gamma = 0.0450466 # magic_n = np.array(round(3.0/2.0 * 2.0**mantissa_bit * (2.0**(exponent_bit-1) - 1.0 - gamma)), dtype=np.int16) # From https://www.mdpi.com/1099-4300/23/1/86/pdf authored by # Walczy, C.J.; Moroz, L.V..; Cie´sli´nski, J.L. # This work is licensed under the Creative Commons Attribution 4.0 # International License. To view a copy of this license, visit # http://creativecommons.org/licenses/by/4.0/ or send a letter to # Creative Commons, PO Box 1866, Mountain View, CA 94042, USA. # magic_n = np.round(2**mantissa_bit *(3* (2.0**(exponent_bit-1) - 1.0) - 1) /2 + np.round(2**mantissa_bit * (3.7315712401613957182292407381942955 - 2)/4 - 0.5)) magic_n = 0x5f376908 # np.array(magic_n, dtype=np.int32) number = x.cpu().numpy() threehalfs1 = 1.50087896 threehalfs2 = 1.50000057 x2 = number * 0.5 y = np.float32(number) i = y.view(np.int32) i = magic_n - np.int32(i >> 1) y = i.view(np.float32) y = y * (threehalfs1 - (x2 * y * y)) y = y * (threehalfs2 - (x2 * y * y)) return torch.from_numpy(y).to(x.device).to(x.dtype) def reciprocal_approx_isqrt(x, input_dtype=torch.bfloat16, output_dtype=torch.bfloat16, isqrt_dtype=torch.bfloat16): assert x.dtype == input_dtype sign = torch.sign(x).to(output_dtype) out = torch.abs(x) if isqrt_dtype == torch.float32: out = isqrt_approx_walcyzk(out.to(isqrt_dtype)) elif isqrt_dtype == torch.bfloat16: out = isqrt_approx_bfloat16(out.to(isqrt_dtype)) else: raise NotImplementedError("only support isqrt fp32 and bfloat16.") out = out.to(output_dtype) out = out * out out = out * sign assert out.dtype == output_dtype return out
null
23,644
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version from torch.nn import init from torch.nn.modules.utils import _pair from torch.nn.modules.utils import _triple from torch.nn.parameter import Parameter _FUSED_CLS = [ QuantizedConvBatchNorm2d, QuantizedConvBatchNorm3d, QuantizedConvTransposeBatchNorm2d, QuantizedConvTransposeBatchNorm3d ] def update_bn_stats(mod): if type(mod) in _FUSED_CLS: mod.update_bn_stats()
null
23,645
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version from torch.nn import init from torch.nn.modules.utils import _pair from torch.nn.modules.utils import _triple from torch.nn.parameter import Parameter _FUSED_CLS = [ QuantizedConvBatchNorm2d, QuantizedConvBatchNorm3d, QuantizedConvTransposeBatchNorm2d, QuantizedConvTransposeBatchNorm3d ] def freeze_bn_stats(mod): if type(mod) in _FUSED_CLS: mod.freeze_bn_stats()
null
23,646
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version from torch.nn import init from torch.nn.modules.utils import _pair from torch.nn.modules.utils import _triple from torch.nn.parameter import Parameter _FUSED_CLS = [ QuantizedConvBatchNorm2d, QuantizedConvBatchNorm3d, QuantizedConvTransposeBatchNorm2d, QuantizedConvTransposeBatchNorm3d ] def fuse_conv_bn(mod): if type(mod) in _FUSED_CLS: mod.merge_bn_to_conv()
null
23,647
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version from torch.nn import init from torch.nn.modules.utils import _pair from torch.nn.modules.utils import _triple from torch.nn.parameter import Parameter _FUSED_CLS = [ QuantizedConvBatchNorm2d, QuantizedConvBatchNorm3d, QuantizedConvTransposeBatchNorm2d, QuantizedConvTransposeBatchNorm3d ] def clear_non_native_bias(mod): if type(mod) in _FUSED_CLS: mod.clear_non_native_bias()
null
23,648
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.modules import fix_ops from pytorch_nndct.nn.quantization.ops import tqt_ops class FakeQuantizer(nn.Module): def __init__(self, bitwidth): def forward(self, x): def _save_to_state_dict(self, destination, prefix, keep_vars): def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): def enable_quant(mod): if isinstance(mod, FakeQuantizer): mod.enable_quant()
null
23,649
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.modules import fix_ops from pytorch_nndct.nn.quantization.ops import tqt_ops class FakeQuantizer(nn.Module): """Simulate the quantize and dequantize operations in training time. In general, the output of this module is given by x_out = (clamp(round(x / scale + zero_point), quant_min, quant_max) - zero_point) * scale See https://arxiv.org/pdf/1903.08066.pdf In nndct, we use symmetric quantization and power-of-2 scaling. That is, zero_point = 0, quant_min = -2^(bitwidth - 1), quant_max = 2^(bitwidth - 1) - 1 """ _version = 2 def __init__(self, bitwidth): super(FakeQuantizer, self).__init__() # quant_enabled is registered as buffer to support their replication in DDP. # Data type is uint8 because NCCL does not support bool tensors. self.register_buffer('quant_enabled', torch.tensor([1], dtype=torch.uint8)) self.register_buffer('bitwidth', torch.tensor([bitwidth], dtype=torch.uint8)) self.register_buffer('domain', torch.tensor([2**(bitwidth - 1)]).float()) def forward(self, x): raise NotImplementedError( 'Do not use FakeQuantizer directly, please use its derivatives.') # PyTorch has been using _save_to_state_dict since 1.2.0. # See https://github.com/pytorch/pytorch/blob/v1.2.0/torch/nn/modules/module.py. def _save_to_state_dict(self, destination, prefix, keep_vars): super(FakeQuantizer, self)._save_to_state_dict(destination, prefix, keep_vars) destination.pop(prefix + 'quant_enabled') destination.pop(prefix + 'domain') def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): # We save 'bitwidth' to state_dict but not load it. # In low-bit tranining, bitwidth incrementally decreases from 8 -> 6 -> 4. # So the bitwidth should be get from quantizer's initialization argument # instead of state dict # For checkpoint BC with version 1. replace_map = {'num_bits': 'bitwidth'} version = local_metadata.get('version', None) if version is None or version < 2: keys = list(state_dict.keys()) for key in keys: key_parts = key.split('.') weight_name = key_parts[-1] if weight_name in replace_map: key_parts[-1] = replace_map[weight_name] new_key = '.'.join(key_parts) assert new_key not in state_dict state_dict[new_key] = state_dict[key] state_dict.pop(key) # Check if bitwidth in the state dict but not load it. missing_bitwidth = False bitwidth_key = prefix + 'bitwidth' if bitwidth_key not in state_dict: missing_bitwidth = True else: # The value of bitwidth should be set at initilization. state_dict.pop(bitwidth_key) super(FakeQuantizer, self)._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) ignored_params = ['bitwidth', 'quant_enabled', 'domain'] ignored_keys = [prefix + name for name in ignored_params] for key in ignored_keys: if key in missing_keys: if key == bitwidth_key and missing_bitwidth: continue missing_keys.remove(key) else: print('[WARNING] Unexpected key in state dict:', key) def disable_quant(mod): if isinstance(mod, FakeQuantizer): mod.disable_quant()
null
23,650
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.modules import fix_ops from pytorch_nndct.nn.quantization.ops import tqt_ops class TQTQuantizer(FakeQuantizer): def __init__(self, bitwidth, tensor_type, method = None): super(TQTQuantizer, self).__init__(bitwidth) valid_tensor_types = ['weight', 'act'] if tensor_type not in valid_tensor_types: raise ValueError( "'tensor_type' must be one of {}".format(valid_tensor_types)) self.tensor_type = tensor_type # See TorchQuantizer::quantize() in quantization/torchquantizer.py if method is not None: self.method = method else: self.method = 3 if tensor_type == 'weight' else 2 self.quantize_fn_cls = tqt_ops.TQTQuantize self.log_threshold = nn.Parameter(torch.tensor([0.0])) self.register_buffer('warmup_enabled', torch.tensor([1], dtype=torch.uint8)) self._forward_fn = self._quantize_with_warmup def _init_threshold(self, x): """See Table 2 in https://arxiv.org/pdf/1903.08066.pdf""" def _max(x): return np.max(np.abs(x)) def _3sd(x): y = x.astype(np.float32) if x.dtype == np.float16 else x return np.abs(np.mean(y + 1e-6)) + 3 * np.std(y) def _kl_j(x): """ Ref paper (Algorithm 1): "Quantizing Convolutional Neural Networks for Low-Power High-Throughput Inference Engines" - Sean Settle et al. https://arxiv.org/pdf/1805.07941.pdf """ def calculate_kl_j(x, y): return np.sum((x - y) * np.log2(x / y)) mn = 0 mx = np.max(np.abs(x)) y = x.astype(np.float32) if x.dtype == np.float16 else x hist, bin_edges = np.histogram((np.abs(y)), 'sqrt', range=(mn, mx), density=True) hist = hist.astype(x.dtype) bin_edges = bin_edges.astype(x.dtype) pdf = hist / np.sum(hist) cdf = np.cumsum(pdf) n = pow(2, self.bitwidth.item() - 1) threshold = [] d = [] if n + 1 > len(bin_edges) - 1: return bin_edges[(-1)] else: for i in range(n + 1, len(bin_edges), 1): threshold_tmp = (i + 0.5) * (bin_edges[1] - bin_edges[0]) threshold = np.concatenate((threshold, [threshold_tmp])) p = np.copy(cdf) p[i - 1:] = 1 x = np.linspace(0.0, 1.0, n) xp = np.linspace(0.0, 1.0, i) fp = p[:i] p_interp = np.interp(x, xp, fp) x = np.linspace(0.0, 1.0, i) xp = np.linspace(0.0, 1.0, n) fp = p_interp q_interp = np.interp(x, xp, fp) q = np.copy(p) q[:i] = q_interp d_tmp = calculate_kl_j(cdf[np.nonzero(cdf)], q[np.nonzero(cdf)]) d = np.concatenate((d, [d_tmp])) return threshold[np.argmin(d)] init_scheme = {'weight': _3sd, 'act': _kl_j} #init_scheme = {'weight': _max, 'act': _kl_j} data = x.detach().cpu().numpy() th = init_scheme[self.tensor_type](data) # TODO(yuwang): Check if th < 0. return torch.tensor([th], dtype=x.dtype, device=x.device) def _forward_pass_input(self, x, log_threshold, domain, method): return x def _quantize(self, x, log_threshold, domain, method): return self.quantize_fn_cls.apply(x, log_threshold, domain, method) def _quantize_with_warmup(self, x, log_threshold, domain, method): self.disable_warmup() log_threshold.data[0] = torch.log2(self._init_threshold(x))[0] return self._quantize(x, log_threshold, domain, method) def forward(self, x): #if self.quant_enabled[0] == 0: # return x #if self.warmup_enabled[0] == 1: # self.warmup_enabled[0] = 0 # threshold = self._init_threshold(x) # self.log_threshold.data = torch.log2(threshold) #return self.quantize_fn_cls.apply(x, self.log_threshold, self.domain, # self.method) return self._forward_fn(x, self.log_threshold, self.domain, self.method) def enable_quant(self, enabled=True): self.quant_enabled[0] = 1 if enabled else 0 if enabled: self._forward_fn = self._quantize_with_warmup if self.warmup_enabled[ 0] == 1 else self._quantize else: self._forward_fn = self._forward_pass_input return self def disable_quant(self): return self.enable_quant(False) def enable_warmup(self, enabled=True): self.warmup_enabled[0] = 1 if enabled else 0 self._forward_fn = self._quantize_with_warmup if enabled else self._quantize return self def disable_warmup(self): return self.enable_warmup(False) def is_warmup_enabled(self): return self.warmup_enabled[0] == 1 def freeze_quant(self, frozen=True): self.log_threshold.requires_grad = (not frozen) def unfreeze_quant(self): self.freeze_quant(False) def extra_repr(self): return 'quant_enabled={}, bitwidth={}, method={}'.format( self.quant_enabled, self.bitwidth, self.method) def _save_to_state_dict(self, destination, prefix, keep_vars): super(TQTQuantizer, self)._save_to_state_dict(destination, prefix, keep_vars) def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): super(TQTQuantizer, self)._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) self._forward_fn = self._quantize_with_warmup if self.warmup_enabled[ 0] == 1 else self._quantize if self.quant_enabled[0] == 0: self._forward_fn = self._forward_pass_input def export_quant_info(self): """Export trained threshold to TorchQuantizer's quant info [bitwidth, fp]. (1) TQT: qx = clip(round(fx / scale)) * scale, scale = 2^ceil(log2t) / 2^(b-1) (2) NndctFixNeron: qx = clip(round(fx * scale)) * (1 / scale), scale = 2^fp Let (1) equals (2), we can get (3): 2^(b-1) / 2^ceil(log2t) = 2^fp => fp = b - 1 - ceil(log2t) For more details, see nndct/include/cuda/nndct_fix_kernels.cuh::_fix_neuron_v2_device """ bitwidth = self.bitwidth.item() ceil_log2t = torch.ceil(self.log_threshold).item() return [[bitwidth, int(bitwidth - 1 - ceil_log2t)]] def import_quant_info(self, qinfo): bitwidth, fp = qinfo self.bitwidth[0] = bitwidth self.log_threshold.data = torch.tensor([bitwidth - 1 - fp], dtype=self.log_threshold.dtype) self.disable_warmup() def enable_warmup(mod): if isinstance(mod, TQTQuantizer): mod.enable_warmup()
null
23,651
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.modules import fix_ops from pytorch_nndct.nn.quantization.ops import tqt_ops class TQTQuantizer(FakeQuantizer): def __init__(self, bitwidth, tensor_type, method = None): super(TQTQuantizer, self).__init__(bitwidth) valid_tensor_types = ['weight', 'act'] if tensor_type not in valid_tensor_types: raise ValueError( "'tensor_type' must be one of {}".format(valid_tensor_types)) self.tensor_type = tensor_type # See TorchQuantizer::quantize() in quantization/torchquantizer.py if method is not None: self.method = method else: self.method = 3 if tensor_type == 'weight' else 2 self.quantize_fn_cls = tqt_ops.TQTQuantize self.log_threshold = nn.Parameter(torch.tensor([0.0])) self.register_buffer('warmup_enabled', torch.tensor([1], dtype=torch.uint8)) self._forward_fn = self._quantize_with_warmup def _init_threshold(self, x): """See Table 2 in https://arxiv.org/pdf/1903.08066.pdf""" def _max(x): return np.max(np.abs(x)) def _3sd(x): y = x.astype(np.float32) if x.dtype == np.float16 else x return np.abs(np.mean(y + 1e-6)) + 3 * np.std(y) def _kl_j(x): """ Ref paper (Algorithm 1): "Quantizing Convolutional Neural Networks for Low-Power High-Throughput Inference Engines" - Sean Settle et al. https://arxiv.org/pdf/1805.07941.pdf """ def calculate_kl_j(x, y): return np.sum((x - y) * np.log2(x / y)) mn = 0 mx = np.max(np.abs(x)) y = x.astype(np.float32) if x.dtype == np.float16 else x hist, bin_edges = np.histogram((np.abs(y)), 'sqrt', range=(mn, mx), density=True) hist = hist.astype(x.dtype) bin_edges = bin_edges.astype(x.dtype) pdf = hist / np.sum(hist) cdf = np.cumsum(pdf) n = pow(2, self.bitwidth.item() - 1) threshold = [] d = [] if n + 1 > len(bin_edges) - 1: return bin_edges[(-1)] else: for i in range(n + 1, len(bin_edges), 1): threshold_tmp = (i + 0.5) * (bin_edges[1] - bin_edges[0]) threshold = np.concatenate((threshold, [threshold_tmp])) p = np.copy(cdf) p[i - 1:] = 1 x = np.linspace(0.0, 1.0, n) xp = np.linspace(0.0, 1.0, i) fp = p[:i] p_interp = np.interp(x, xp, fp) x = np.linspace(0.0, 1.0, i) xp = np.linspace(0.0, 1.0, n) fp = p_interp q_interp = np.interp(x, xp, fp) q = np.copy(p) q[:i] = q_interp d_tmp = calculate_kl_j(cdf[np.nonzero(cdf)], q[np.nonzero(cdf)]) d = np.concatenate((d, [d_tmp])) return threshold[np.argmin(d)] init_scheme = {'weight': _3sd, 'act': _kl_j} #init_scheme = {'weight': _max, 'act': _kl_j} data = x.detach().cpu().numpy() th = init_scheme[self.tensor_type](data) # TODO(yuwang): Check if th < 0. return torch.tensor([th], dtype=x.dtype, device=x.device) def _forward_pass_input(self, x, log_threshold, domain, method): return x def _quantize(self, x, log_threshold, domain, method): return self.quantize_fn_cls.apply(x, log_threshold, domain, method) def _quantize_with_warmup(self, x, log_threshold, domain, method): self.disable_warmup() log_threshold.data[0] = torch.log2(self._init_threshold(x))[0] return self._quantize(x, log_threshold, domain, method) def forward(self, x): #if self.quant_enabled[0] == 0: # return x #if self.warmup_enabled[0] == 1: # self.warmup_enabled[0] = 0 # threshold = self._init_threshold(x) # self.log_threshold.data = torch.log2(threshold) #return self.quantize_fn_cls.apply(x, self.log_threshold, self.domain, # self.method) return self._forward_fn(x, self.log_threshold, self.domain, self.method) def enable_quant(self, enabled=True): self.quant_enabled[0] = 1 if enabled else 0 if enabled: self._forward_fn = self._quantize_with_warmup if self.warmup_enabled[ 0] == 1 else self._quantize else: self._forward_fn = self._forward_pass_input return self def disable_quant(self): return self.enable_quant(False) def enable_warmup(self, enabled=True): self.warmup_enabled[0] = 1 if enabled else 0 self._forward_fn = self._quantize_with_warmup if enabled else self._quantize return self def disable_warmup(self): return self.enable_warmup(False) def is_warmup_enabled(self): return self.warmup_enabled[0] == 1 def freeze_quant(self, frozen=True): self.log_threshold.requires_grad = (not frozen) def unfreeze_quant(self): self.freeze_quant(False) def extra_repr(self): return 'quant_enabled={}, bitwidth={}, method={}'.format( self.quant_enabled, self.bitwidth, self.method) def _save_to_state_dict(self, destination, prefix, keep_vars): super(TQTQuantizer, self)._save_to_state_dict(destination, prefix, keep_vars) def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): super(TQTQuantizer, self)._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) self._forward_fn = self._quantize_with_warmup if self.warmup_enabled[ 0] == 1 else self._quantize if self.quant_enabled[0] == 0: self._forward_fn = self._forward_pass_input def export_quant_info(self): """Export trained threshold to TorchQuantizer's quant info [bitwidth, fp]. (1) TQT: qx = clip(round(fx / scale)) * scale, scale = 2^ceil(log2t) / 2^(b-1) (2) NndctFixNeron: qx = clip(round(fx * scale)) * (1 / scale), scale = 2^fp Let (1) equals (2), we can get (3): 2^(b-1) / 2^ceil(log2t) = 2^fp => fp = b - 1 - ceil(log2t) For more details, see nndct/include/cuda/nndct_fix_kernels.cuh::_fix_neuron_v2_device """ bitwidth = self.bitwidth.item() ceil_log2t = torch.ceil(self.log_threshold).item() return [[bitwidth, int(bitwidth - 1 - ceil_log2t)]] def import_quant_info(self, qinfo): bitwidth, fp = qinfo self.bitwidth[0] = bitwidth self.log_threshold.data = torch.tensor([bitwidth - 1 - fp], dtype=self.log_threshold.dtype) self.disable_warmup() def disable_warmup(mod): if isinstance(mod, TQTQuantizer): mod.disable_warmup()
null
23,652
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.modules import fix_ops from pytorch_nndct.nn.quantization.ops import tqt_ops class TQTQuantizer(FakeQuantizer): def __init__(self, bitwidth, tensor_type, method = None): super(TQTQuantizer, self).__init__(bitwidth) valid_tensor_types = ['weight', 'act'] if tensor_type not in valid_tensor_types: raise ValueError( "'tensor_type' must be one of {}".format(valid_tensor_types)) self.tensor_type = tensor_type # See TorchQuantizer::quantize() in quantization/torchquantizer.py if method is not None: self.method = method else: self.method = 3 if tensor_type == 'weight' else 2 self.quantize_fn_cls = tqt_ops.TQTQuantize self.log_threshold = nn.Parameter(torch.tensor([0.0])) self.register_buffer('warmup_enabled', torch.tensor([1], dtype=torch.uint8)) self._forward_fn = self._quantize_with_warmup def _init_threshold(self, x): """See Table 2 in https://arxiv.org/pdf/1903.08066.pdf""" def _max(x): return np.max(np.abs(x)) def _3sd(x): y = x.astype(np.float32) if x.dtype == np.float16 else x return np.abs(np.mean(y + 1e-6)) + 3 * np.std(y) def _kl_j(x): """ Ref paper (Algorithm 1): "Quantizing Convolutional Neural Networks for Low-Power High-Throughput Inference Engines" - Sean Settle et al. https://arxiv.org/pdf/1805.07941.pdf """ def calculate_kl_j(x, y): return np.sum((x - y) * np.log2(x / y)) mn = 0 mx = np.max(np.abs(x)) y = x.astype(np.float32) if x.dtype == np.float16 else x hist, bin_edges = np.histogram((np.abs(y)), 'sqrt', range=(mn, mx), density=True) hist = hist.astype(x.dtype) bin_edges = bin_edges.astype(x.dtype) pdf = hist / np.sum(hist) cdf = np.cumsum(pdf) n = pow(2, self.bitwidth.item() - 1) threshold = [] d = [] if n + 1 > len(bin_edges) - 1: return bin_edges[(-1)] else: for i in range(n + 1, len(bin_edges), 1): threshold_tmp = (i + 0.5) * (bin_edges[1] - bin_edges[0]) threshold = np.concatenate((threshold, [threshold_tmp])) p = np.copy(cdf) p[i - 1:] = 1 x = np.linspace(0.0, 1.0, n) xp = np.linspace(0.0, 1.0, i) fp = p[:i] p_interp = np.interp(x, xp, fp) x = np.linspace(0.0, 1.0, i) xp = np.linspace(0.0, 1.0, n) fp = p_interp q_interp = np.interp(x, xp, fp) q = np.copy(p) q[:i] = q_interp d_tmp = calculate_kl_j(cdf[np.nonzero(cdf)], q[np.nonzero(cdf)]) d = np.concatenate((d, [d_tmp])) return threshold[np.argmin(d)] init_scheme = {'weight': _3sd, 'act': _kl_j} #init_scheme = {'weight': _max, 'act': _kl_j} data = x.detach().cpu().numpy() th = init_scheme[self.tensor_type](data) # TODO(yuwang): Check if th < 0. return torch.tensor([th], dtype=x.dtype, device=x.device) def _forward_pass_input(self, x, log_threshold, domain, method): return x def _quantize(self, x, log_threshold, domain, method): return self.quantize_fn_cls.apply(x, log_threshold, domain, method) def _quantize_with_warmup(self, x, log_threshold, domain, method): self.disable_warmup() log_threshold.data[0] = torch.log2(self._init_threshold(x))[0] return self._quantize(x, log_threshold, domain, method) def forward(self, x): #if self.quant_enabled[0] == 0: # return x #if self.warmup_enabled[0] == 1: # self.warmup_enabled[0] = 0 # threshold = self._init_threshold(x) # self.log_threshold.data = torch.log2(threshold) #return self.quantize_fn_cls.apply(x, self.log_threshold, self.domain, # self.method) return self._forward_fn(x, self.log_threshold, self.domain, self.method) def enable_quant(self, enabled=True): self.quant_enabled[0] = 1 if enabled else 0 if enabled: self._forward_fn = self._quantize_with_warmup if self.warmup_enabled[ 0] == 1 else self._quantize else: self._forward_fn = self._forward_pass_input return self def disable_quant(self): return self.enable_quant(False) def enable_warmup(self, enabled=True): self.warmup_enabled[0] = 1 if enabled else 0 self._forward_fn = self._quantize_with_warmup if enabled else self._quantize return self def disable_warmup(self): return self.enable_warmup(False) def is_warmup_enabled(self): return self.warmup_enabled[0] == 1 def freeze_quant(self, frozen=True): self.log_threshold.requires_grad = (not frozen) def unfreeze_quant(self): self.freeze_quant(False) def extra_repr(self): return 'quant_enabled={}, bitwidth={}, method={}'.format( self.quant_enabled, self.bitwidth, self.method) def _save_to_state_dict(self, destination, prefix, keep_vars): super(TQTQuantizer, self)._save_to_state_dict(destination, prefix, keep_vars) def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): super(TQTQuantizer, self)._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) self._forward_fn = self._quantize_with_warmup if self.warmup_enabled[ 0] == 1 else self._quantize if self.quant_enabled[0] == 0: self._forward_fn = self._forward_pass_input def export_quant_info(self): """Export trained threshold to TorchQuantizer's quant info [bitwidth, fp]. (1) TQT: qx = clip(round(fx / scale)) * scale, scale = 2^ceil(log2t) / 2^(b-1) (2) NndctFixNeron: qx = clip(round(fx * scale)) * (1 / scale), scale = 2^fp Let (1) equals (2), we can get (3): 2^(b-1) / 2^ceil(log2t) = 2^fp => fp = b - 1 - ceil(log2t) For more details, see nndct/include/cuda/nndct_fix_kernels.cuh::_fix_neuron_v2_device """ bitwidth = self.bitwidth.item() ceil_log2t = torch.ceil(self.log_threshold).item() return [[bitwidth, int(bitwidth - 1 - ceil_log2t)]] def import_quant_info(self, qinfo): bitwidth, fp = qinfo self.bitwidth[0] = bitwidth self.log_threshold.data = torch.tensor([bitwidth - 1 - fp], dtype=self.log_threshold.dtype) self.disable_warmup() def freeze_quant(mod): if isinstance(mod, TQTQuantizer): mod.freeze_quant()
null
23,653
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.modules import fix_ops from pytorch_nndct.nn.quantization.ops import tqt_ops class TQTQuantizer(FakeQuantizer): def __init__(self, bitwidth, tensor_type, method = None): super(TQTQuantizer, self).__init__(bitwidth) valid_tensor_types = ['weight', 'act'] if tensor_type not in valid_tensor_types: raise ValueError( "'tensor_type' must be one of {}".format(valid_tensor_types)) self.tensor_type = tensor_type # See TorchQuantizer::quantize() in quantization/torchquantizer.py if method is not None: self.method = method else: self.method = 3 if tensor_type == 'weight' else 2 self.quantize_fn_cls = tqt_ops.TQTQuantize self.log_threshold = nn.Parameter(torch.tensor([0.0])) self.register_buffer('warmup_enabled', torch.tensor([1], dtype=torch.uint8)) self._forward_fn = self._quantize_with_warmup def _init_threshold(self, x): """See Table 2 in https://arxiv.org/pdf/1903.08066.pdf""" def _max(x): return np.max(np.abs(x)) def _3sd(x): y = x.astype(np.float32) if x.dtype == np.float16 else x return np.abs(np.mean(y + 1e-6)) + 3 * np.std(y) def _kl_j(x): """ Ref paper (Algorithm 1): "Quantizing Convolutional Neural Networks for Low-Power High-Throughput Inference Engines" - Sean Settle et al. https://arxiv.org/pdf/1805.07941.pdf """ def calculate_kl_j(x, y): return np.sum((x - y) * np.log2(x / y)) mn = 0 mx = np.max(np.abs(x)) y = x.astype(np.float32) if x.dtype == np.float16 else x hist, bin_edges = np.histogram((np.abs(y)), 'sqrt', range=(mn, mx), density=True) hist = hist.astype(x.dtype) bin_edges = bin_edges.astype(x.dtype) pdf = hist / np.sum(hist) cdf = np.cumsum(pdf) n = pow(2, self.bitwidth.item() - 1) threshold = [] d = [] if n + 1 > len(bin_edges) - 1: return bin_edges[(-1)] else: for i in range(n + 1, len(bin_edges), 1): threshold_tmp = (i + 0.5) * (bin_edges[1] - bin_edges[0]) threshold = np.concatenate((threshold, [threshold_tmp])) p = np.copy(cdf) p[i - 1:] = 1 x = np.linspace(0.0, 1.0, n) xp = np.linspace(0.0, 1.0, i) fp = p[:i] p_interp = np.interp(x, xp, fp) x = np.linspace(0.0, 1.0, i) xp = np.linspace(0.0, 1.0, n) fp = p_interp q_interp = np.interp(x, xp, fp) q = np.copy(p) q[:i] = q_interp d_tmp = calculate_kl_j(cdf[np.nonzero(cdf)], q[np.nonzero(cdf)]) d = np.concatenate((d, [d_tmp])) return threshold[np.argmin(d)] init_scheme = {'weight': _3sd, 'act': _kl_j} #init_scheme = {'weight': _max, 'act': _kl_j} data = x.detach().cpu().numpy() th = init_scheme[self.tensor_type](data) # TODO(yuwang): Check if th < 0. return torch.tensor([th], dtype=x.dtype, device=x.device) def _forward_pass_input(self, x, log_threshold, domain, method): return x def _quantize(self, x, log_threshold, domain, method): return self.quantize_fn_cls.apply(x, log_threshold, domain, method) def _quantize_with_warmup(self, x, log_threshold, domain, method): self.disable_warmup() log_threshold.data[0] = torch.log2(self._init_threshold(x))[0] return self._quantize(x, log_threshold, domain, method) def forward(self, x): #if self.quant_enabled[0] == 0: # return x #if self.warmup_enabled[0] == 1: # self.warmup_enabled[0] = 0 # threshold = self._init_threshold(x) # self.log_threshold.data = torch.log2(threshold) #return self.quantize_fn_cls.apply(x, self.log_threshold, self.domain, # self.method) return self._forward_fn(x, self.log_threshold, self.domain, self.method) def enable_quant(self, enabled=True): self.quant_enabled[0] = 1 if enabled else 0 if enabled: self._forward_fn = self._quantize_with_warmup if self.warmup_enabled[ 0] == 1 else self._quantize else: self._forward_fn = self._forward_pass_input return self def disable_quant(self): return self.enable_quant(False) def enable_warmup(self, enabled=True): self.warmup_enabled[0] = 1 if enabled else 0 self._forward_fn = self._quantize_with_warmup if enabled else self._quantize return self def disable_warmup(self): return self.enable_warmup(False) def is_warmup_enabled(self): return self.warmup_enabled[0] == 1 def freeze_quant(self, frozen=True): self.log_threshold.requires_grad = (not frozen) def unfreeze_quant(self): self.freeze_quant(False) def extra_repr(self): return 'quant_enabled={}, bitwidth={}, method={}'.format( self.quant_enabled, self.bitwidth, self.method) def _save_to_state_dict(self, destination, prefix, keep_vars): super(TQTQuantizer, self)._save_to_state_dict(destination, prefix, keep_vars) def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): super(TQTQuantizer, self)._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) self._forward_fn = self._quantize_with_warmup if self.warmup_enabled[ 0] == 1 else self._quantize if self.quant_enabled[0] == 0: self._forward_fn = self._forward_pass_input def export_quant_info(self): """Export trained threshold to TorchQuantizer's quant info [bitwidth, fp]. (1) TQT: qx = clip(round(fx / scale)) * scale, scale = 2^ceil(log2t) / 2^(b-1) (2) NndctFixNeron: qx = clip(round(fx * scale)) * (1 / scale), scale = 2^fp Let (1) equals (2), we can get (3): 2^(b-1) / 2^ceil(log2t) = 2^fp => fp = b - 1 - ceil(log2t) For more details, see nndct/include/cuda/nndct_fix_kernels.cuh::_fix_neuron_v2_device """ bitwidth = self.bitwidth.item() ceil_log2t = torch.ceil(self.log_threshold).item() return [[bitwidth, int(bitwidth - 1 - ceil_log2t)]] def import_quant_info(self, qinfo): bitwidth, fp = qinfo self.bitwidth[0] = bitwidth self.log_threshold.data = torch.tensor([bitwidth - 1 - fp], dtype=self.log_threshold.dtype) self.disable_warmup() def unfreeze_quant(mod): if isinstance(mod, TQTQuantizer): mod.unfreeze_quant()
null
23,654
import torch from pytorch_nndct.utils.hw_dtype import fp32 from pytorch_nndct.nn.quantization.ops import quantize_ops def get_exponent(tensor): with torch.no_grad(): t = torch.nan_to_num(tensor, nan=0, posinf=0, neginf=0) ''' smallest positive subnormal, largest subnormal, smallest positive normal, largest normal >>> t = torch.tensor([0, 1.4012984643*10**-45, 1.1754942107*10**-38, 1.1754943508 *10**-38, 3.4028234664*10**38]) >>> torch.frexp(t) torch.return_types.frexp( mantissa=tensor([0.0000, 0.5000, 1.0000, 0.5000, 1.0000]), exponent=tensor([0, -148, -126, -125, 128], dtype=torch.int32)) ''' _, exp = torch.frexp(t) # The exponent of subnormal is less than fp32.emin exp[(t==0) | (exp < fp32.emin)] = fp32.emin return exp - 1.0 fp32 = DType('fp32', 'rn', -126, 127, 24) def cast_to_fp(tensor, exp_bias, m_bits, round_mode, min_val, max_val): exp = get_exponent(tensor) # 2**(1 - exp_bias - m_bits) is the smallest representable value scale = torch.pow(2.0, torch.clamp(exp, 1 - exp_bias, fp32.emax + 1) - m_bits) return quantize_ops.quantize(tensor, scale, round_mode, min_val, max_val)
null
23,655
import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.quantization.ops import quantize_ops from pytorch_nndct.utils import onnx_utils def _get_exponent_v1(tensor, epsilon=2**-23): t = tensor.abs() # we use fp32's 1.mantissa_bits max_t, _ = t.max(t.dim() - 1, keepdim=True) max_exp = (max_t + epsilon).log2().floor() t_exp = (t + epsilon).log2().floor() return max_exp, t_exp
null
23,656
import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.quantization.ops import quantize_ops from pytorch_nndct.utils import onnx_utils def _get_exponent_v2(tensor): _, exp = torch.frexp(tensor) # we use fp32's 1.mantissa_bits max_exp, _ = exp.max(exp.dim() - 1, keepdim=True) return max_exp - 1, 1
null
23,657
import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.quantization.ops import quantize_ops from pytorch_nndct.utils import onnx_utils def _get_exponent_v3(tensor): tensor_shape = list(tensor.shape) exponent = torch.ops.vai.calculate_shared_exponent(tensor, tensor_shape[-1]) tensor_shape[-1] = 1 return exponent.reshape(tensor_shape), 1
null
23,658
import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.quantization.ops import quantize_ops from pytorch_nndct.utils import onnx_utils def _min_max_at_exp(exp, bit_width): # sign bits: 1, exponent bits: 8, no implicit leading 1 mantissa_bits = bit_width - 9 # The min/max representable value with exp # x = {sign, exp, mant} = {0 01001101 0000001} # e = int(01001101) - 127 # M = int(0.000001) = 2^(-6) # x = (-1)^sign*2^e*M # = 2^e * 2^(-6) = 2^(e - 6) min_v = torch.pow(2.0, exp - (mantissa_bits - 1)) max_v = torch.pow(2.0, exp + 1) - min_v return min_v, max_v
null
23,659
import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.quantization.ops import quantize_ops from pytorch_nndct.utils import onnx_utils def transform_to_block_wise(input, block_size=8, axis=1): def transform_block_to_shape(input, shape, axis=1): class BFPQuantizeV1(torch.autograd.Function): def forward(ctx, t, bit_width, rounding_mode='round_even'): def quantize_to_bfp_v1(tensor, bit_width, block_size, axis, rounding_mode): shape = tensor.shape tensor = transform_to_block_wise(tensor, block_size, axis) tensor = BFPQuantizeV1.apply(tensor, bit_width, rounding_mode) return transform_block_to_shape(tensor, shape, axis)
null
23,660
import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.quantization.ops import quantize_ops from pytorch_nndct.utils import onnx_utils def transpose_to_block_wise(input, block_size=8, axis=1): def transpose_block_to_shape(input, shape, axis=1): class BFPQuantizeV2(BFPQuantize): def forward(ctx, t, bit_width, block_size, rounding_mode='round_even'): def quantize_to_bfp_v2(tensor, bit_width, block_size, axis, rounding_mode): shape = tensor.shape tensor = transpose_to_block_wise(tensor, block_size, axis) tensor = BFPQuantizeV2.apply(tensor, bit_width, block_size, rounding_mode) return transpose_block_to_shape(tensor, shape, axis)
null
23,661
import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.quantization.ops import quantize_ops from pytorch_nndct.utils import onnx_utils def pad_to_block_last(tensor, block_size=8, axis=1): def depad_and_transpose(tensor, shape, axis=1): class BFPQuantizeV3(BFPQuantize): def forward(ctx, t, bit_width, block_size, rounding_mode='round_even'): def quantize_to_bfp_v3(tensor, bit_width, block_size, axis, rounding_mode): shape = tensor.shape tensor = pad_to_block_last(tensor, block_size, axis) tensor = BFPQuantizeV3.apply(tensor, bit_width, block_size, rounding_mode) return depad_and_transpose(tensor, shape, axis)
null
23,662
import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.quantization.ops import quantize_ops from pytorch_nndct.utils import onnx_utils def transform_to_block_wise(input, block_size=8, axis=1): def transform_block_to_shape(input, shape, axis=1): def quantize_to_bfp_prime(tensor, bit_width, block_size, axis, rounding_mode, epsilon=torch.pow(torch.tensor(2.0), -23)): shape = tensor.shape tensor = transform_to_block_wise(tensor, block_size, axis) tensor = _to_bfp_prime(tensor, bit_width, rounding_mode, epsilon) return transform_block_to_shape(tensor, shape, axis)
null
23,663
import torch import torch.nn as nn import torch.nn.functional as F from pytorch_nndct.nn.quantization.ops import quantize_ops from pytorch_nndct.utils import onnx_utils def pad_to_block_last(tensor, block_size=8, axis=1): def depad_and_transpose(tensor, shape, axis=1): class BFPPrimeSharedQuantize(BFPQuantize): def forward(ctx, t, bit_width, block_size, sub_block_size, sub_block_shift_bits, rounding_mode='round_to_nearest'): def quantize_to_bfp_prime_shared(tensor, bit_width, block_size, sub_block_size, sub_block_shift_bits, axis, rounding_mode): if block_size % sub_block_size != 0: raise ValueError( f('The block_size must be divisible by sub_block_size.' '(block_size={block_size}, sub_block_size={sub_block_size})')) shape = tensor.shape tensor = pad_to_block_last(tensor, block_size, axis) tensor = BFPPrimeSharedQuantize.apply(tensor, bit_width, block_size, sub_block_size, sub_block_shift_bits, rounding_mode) return depad_and_transpose(tensor, shape, axis)
null
23,664
import math import numpy as np import torch from pytorch_nndct.nn.modules import fix_ops The provided code snippet includes necessary dependencies for implementing the `_cdf_measure` function. Write a Python function `def _cdf_measure(x, y, measure_name='Kullback-Leibler-J')` to solve the following problem: Ref paper: "Non-parametric Information-Theoretic Measures of One-Dimensional Distribution Functions from Continuous Time Series" - Paolo D Alberto et al. https://epubs.siam.org/doi/abs/10.1137/1.9781611972795.59 https://epubs.siam.org/doi/pdf/10.1137/1.9781611972795.59 measure_names_symm = ['Camberra', 'Chi-Squared', 'Cramer-von Mises', 'Euclidean', 'Hellinger', 'Jin-L', 'Jensen-Shannon', 'Kolmogorov-Smirnov', 'Kullback-Leibler-J', 'Variational'] measure_names_asym = ['Jin-K', 'Kullback-Leibler-I'] measure_names_excl = ['Bhattacharyya', 'Phi', 'Xi'] Here is the function: def _cdf_measure(x, y, measure_name='Kullback-Leibler-J'): """ Ref paper: "Non-parametric Information-Theoretic Measures of One-Dimensional Distribution Functions from Continuous Time Series" - Paolo D Alberto et al. https://epubs.siam.org/doi/abs/10.1137/1.9781611972795.59 https://epubs.siam.org/doi/pdf/10.1137/1.9781611972795.59 measure_names_symm = ['Camberra', 'Chi-Squared', 'Cramer-von Mises', 'Euclidean', 'Hellinger', 'Jin-L', 'Jensen-Shannon', 'Kolmogorov-Smirnov', 'Kullback-Leibler-J', 'Variational'] measure_names_asym = ['Jin-K', 'Kullback-Leibler-I'] measure_names_excl = ['Bhattacharyya', 'Phi', 'Xi'] """ if measure_name == 'Bhattacharyya': return np.sum(np.sqrt(x * y)) else: if measure_name == 'Camberra': return np.sum(np.abs(x - y) / (x + y)) else: if measure_name == 'Chi-Squared': return np.sum(np.power(x - y, 2.0) / x) else: if measure_name == 'Cramer-von Mises': return np.sum(np.power(x - y, 2.0)) else: if measure_name == 'Euclidean': return np.power(np.sum(np.power(x - y, 2.0)), 0.5) else: if measure_name == 'Hellinger': return np.power(np.sum(np.sqrt(x) - np.sqrt(y)), 2.0) / 2.0 else: if measure_name == 'Jin-K': return _cdf_measure(x, (x + y) / 2.0, 'Kullback-Leibler-I') else: if measure_name == 'Jin-L': return _cdf_measure( x, (x + y) / 2.0, 'Kullback-Leibler-I') + _cdf_measure( y, (x + y) / 2.0, 'Kullback-Leibler-I') if measure_name == 'Jensen-Shannon': return ( _cdf_measure(x, (x + y) / 2.0, 'Kullback-Leibler-I') + _cdf_measure(y, (x + y) / 2.0, 'Kullback-Leibler-I')) / 2.0 if measure_name == 'Kolmogorov-Smirnov': return np.max(np.abs(x - y)) if measure_name == 'Kullback-Leibler-I': return np.sum(x * np.log2(x / y)) if measure_name == 'Kullback-Leibler-J': return np.sum((x - y) * np.log2(x / y)) if measure_name == 'Phi': return np.max( np.abs(x - y) / np.sqrt(np.minimum((x + y) / 2.0, 1 - (x + y) / 2.0))) if measure_name == 'Variational': return np.sum(np.abs(x - y)) if measure_name == 'Xi': return np.max( np.abs(x - y) / np.sqrt((x + y) / 2.0 * (1 - (x + y) / 2.0))) return _cdf_measure(x, y, 'Kullback-Leibler-J')
Ref paper: "Non-parametric Information-Theoretic Measures of One-Dimensional Distribution Functions from Continuous Time Series" - Paolo D Alberto et al. https://epubs.siam.org/doi/abs/10.1137/1.9781611972795.59 https://epubs.siam.org/doi/pdf/10.1137/1.9781611972795.59 measure_names_symm = ['Camberra', 'Chi-Squared', 'Cramer-von Mises', 'Euclidean', 'Hellinger', 'Jin-L', 'Jensen-Shannon', 'Kolmogorov-Smirnov', 'Kullback-Leibler-J', 'Variational'] measure_names_asym = ['Jin-K', 'Kullback-Leibler-I'] measure_names_excl = ['Bhattacharyya', 'Phi', 'Xi']
23,665
from functools import wraps import torch def pre_and_post_process_f16_tensor(func): @wraps(func) def wrapper(*args, **kwargs): tensor_type_list = [] tensor_type_dict = {} out_need_convert = False for arg in args: if isinstance(arg, torch.Tensor): tensor_type_list.append(arg.dtype) if arg.dtype == torch.float16: arg.data = arg.data.to(torch.float32) out_need_convert = True for key, value in kwargs.items(): if isinstance(value, torch.Tensor): tensor_type_dict[key] = value.dtype if value.dtype == torch.float16: value.data = value.data.to(torch.float32) out_need_convert = True out = func(*args, **kwargs) if out_need_convert: if isinstance(out, torch.Tensor): if out.dtype == torch.float32: out.data = out.data.to(torch.float16) #for i in range(len(args)): # if isinstance(args[i], torch.Tensor): # if args[i].dtype != tensor_type_list[i]: # args[i].data = args[i].data.to(tensor_type_list[i]) index = 0 for arg in args: if isinstance(arg, torch.Tensor): if arg.dtype != tensor_type_list[index]: arg.data = arg.data.to(tensor_type_list[index]) index = index + 1 for key, value in kwargs.items(): if isinstance(value, torch.Tensor): if value.dtype != tensor_type_dict[key]: value.data = value.data.to(tensor_type_dict[key]) return out return wrapper
null
23,666
import torch import numpy as np from torch.nn.utils.rnn import PackedSequence import torch.nn as nn def deephi_pack_padded_sequence(input, lengths, batch_first=False): if isinstance(lengths, list): lengths = torch.LongTensor(lengths) data = input if not batch_first: batch_sizes = np.ones(input.size(0)) * input.size(1) else: batch_sizes = np.ones(input.size(1)) * input.size(0) return PackedSequence(data, batch_sizes) def pack_padded_sequence(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return nn.utils.rnn.pack_padded_sequence(*args, **kwargs) return deephi_pack_padded_sequence(*args, **kwargs)
null
23,667
import torch import numpy as np from torch.nn.utils.rnn import PackedSequence import torch.nn as nn def deephi_pad_packed_sequence(sequence, batch_first=False, padding_value=0.0, total_length=None): output = sequence if not batch_first: lengths = np.ones(sequence.size(0)) * sequence.size(1) else: lengths = np.ones(sequecne.size(1)) * sequence.size(0) return output, torch.LongTensor(lengths) def pad_packed_sequence(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return nn.utils.rnn.pad_packed_sequence(*args, **kwargs) return deephi_pad_packed_sequence(*args, **kwargs)
null
23,668
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util _torch_layouts = {2: 'OI', 4: 'OIHW'} _nndct_layouts = {2: 'OI', 4: 'OHWI'} def transpose_tensor(tensor, src_layouts, dst_layouts): if not isinstance(tensor, Tensor): raise errors.OptimizerDataFormatError( "'tensor' must be Tensor, but given {}".format(type(tensor))) if tensor.ndim != 4 and tensor.ndim != 2: return tensor src_layout = src_layouts[tensor.ndim] dst_layout = dst_layouts[tensor.ndim] axis = [src_layout.index(d) for d in dst_layout] tensor.transpose(axis) return tensor def torch_to_nndct(tensor): return transpose_tensor(tensor, _torch_layouts, _nndct_layouts)
null
23,669
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util _torch_layouts = {2: 'OI', 4: 'OIHW'} _nndct_layouts = {2: 'OI', 4: 'OHWI'} def transpose_tensor(tensor, src_layouts, dst_layouts): if not isinstance(tensor, Tensor): raise errors.OptimizerDataFormatError( "'tensor' must be Tensor, but given {}".format(type(tensor))) if tensor.ndim != 4 and tensor.ndim != 2: return tensor src_layout = src_layouts[tensor.ndim] dst_layout = dst_layouts[tensor.ndim] axis = [src_layout.index(d) for d in dst_layout] tensor.transpose(axis) return tensor def nndct_to_torch(tensor): return transpose_tensor(tensor, _nndct_layouts, _torch_layouts)
null
23,670
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util _torch_layouts = {2: 'OI', 4: 'OIHW'} _nndct_layouts = {2: 'OI', 4: 'OHWI'} def transpose_tensor(tensor, src_layouts, dst_layouts): if not isinstance(tensor, Tensor): raise errors.OptimizerDataFormatError( "'tensor' must be Tensor, but given {}".format(type(tensor))) if tensor.ndim != 4 and tensor.ndim != 2: return tensor src_layout = src_layouts[tensor.ndim] dst_layout = dst_layouts[tensor.ndim] axis = [src_layout.index(d) for d in dst_layout] tensor.transpose(axis) return tensor def torch_tensor_from_nndct(tensor): replicated_tensor = copy.deepcopy(tensor) return torch.from_numpy( transpose_tensor(replicated_tensor, _nndct_layouts, _torch_layouts).data)
null
23,671
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util def dummy_inputs(input_specs): inputs = [] for spec in input_specs: inputs.append(torch.rand(*spec.shape).type(spec.dtype).cuda()) return tuple(inputs)
null
23,672
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util def unwrap_parallel_module(module): if isinstance(module, (DataParallel, DistributedDataParallel)): model = module.module else: model = module return model
null
23,673
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util import os if not os.path.exists(BASE_DIR): os.makedirs(BASE_DIR) def is_debug_mode(): return os.environ.get('VAI_OPTIMIZER_DEBUG', None) == '1'
null
23,674
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util def write_graph_script(graph, filename=None): if not filename: _, filename = tempfile.mkstemp(suffix='.py', text=True) writer = get_script_writer(enable_quant=False) writer.write(graph, filename) return filename def parse_to_graph(module, inputs, debug=False): if debug: from nndct_shared.utils import NndctOption, NndctDebugLogger NndctDebugLogger("vai_opt_debug.log") NndctOption.nndct_parse_debug.value = 5 parser = parse.TorchParser() graph = parser(module._get_name(), module, inputs) if debug: from nndct_shared.utils import saving saving.save_graph(graph, hdf5_path='{}.hdf5'.format(graph.name)) write_graph_script(graph, '{}_baseline.py'.format(graph.name)) return graph
null
23,675
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util def random_str(str_length=4): letters = string.ascii_lowercase return ''.join(random.choice(letters) for i in range(str_length)) def write_graph_script(graph, filename=None): if not filename: _, filename = tempfile.mkstemp(suffix='.py', text=True) writer = get_script_writer(enable_quant=False) writer.write(graph, filename) return filename def rebuild_model(graph, filename=None): filename = write_graph_script(graph, filename) #module_name = graph.name py_module_name = "_".join(["nndct", random_str()]) spec = importlib.util.spec_from_file_location(py_module_name, filename) py_module = importlib.util.module_from_spec(spec) sys.modules[py_module_name] = py_module spec.loader.exec_module(py_module) rebuilt_model = py_module.__dict__[graph.name]() api_utils.connect_module_with_graph(rebuilt_model, graph) return rebuilt_model, filename
null
23,676
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util def map_rebuilt_module_to_node(model, graph): module_to_node = {} for name, module in model.named_children(): # module_name -> node_id idx = int(name.split(TorchSymbol.MODULE_NAME_SEPERATOR)[-1]) node = graph.get_node_by_idx(idx) module_to_node[name] = node return module_to_node
null
23,677
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util def map_original_module_to_node(model, graph): module_to_node = {} for node in graph.nodes: module = mod_util.get_module_by_node(model, node) if module: module_to_node[id(module)] = node.name return module_to_node def excluded_node_names(model, graph, excludes): excluded_nodes = [] module_to_node = map_original_module_to_node(model, graph) for exclude in excludes: if isinstance(exclude, str): excluded_nodes.append(exclude) elif isinstance(exclude, torch.nn.Module): for module in exclude.modules(): module_id = id(module) if module_id in module_to_node: excluded_nodes.append(module_to_node[module_id]) else: raise errors.OptimizerInvalidArgumentError( 'Excludes must be either string or torch.nn.Module') return excluded_nodes
null
23,678
import copy import importlib import os import random import string import sys import tempfile import torch from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.pruning import errors from pytorch_nndct import parse from pytorch_nndct.export import get_script_writer from pytorch_nndct.parse import parse_utils from pytorch_nndct.qproc import utils as api_utils from pytorch_nndct.utils import TorchSymbol from pytorch_nndct.utils import module_util as mod_util import os if not os.path.exists(BASE_DIR): os.makedirs(BASE_DIR) def get_actual_device(gpu: int) -> int: gpu = int(gpu) if 'CUDA_VISIBLE_DEVICES' in os.environ: available_devices = os.environ['CUDA_VISIBLE_DEVICES'].split(",") if gpu >= len(available_devices): raise ValueError( f"CUDA_VISIBLE_DEVICES={os.environ['CUDA_VISIBLE_DEVICES']}, arg gpu must be less than {len(available_devices)}" ) return int(available_devices[gpu]) else: return gpu
null
23,679
import collections import copy import json import numpy as np import os import random import torch import torch.multiprocessing as mp import types from typing import List from nndct_shared.base import NNDCT_OP as OpTypes from nndct_shared.pruning import errors from nndct_shared.pruning import logging from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.pruning import utils as spu from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.utils import common from pytorch_nndct import parse from pytorch_nndct.pruning import utils from pytorch_nndct.utils import module_util as mod_util from pytorch_nndct.utils import profiler from pytorch_nndct.utils import torch_const from pytorch_nndct.utils.calibration import calibrate_sens, calibrate_spec class IterativePruningRunner(PruningRunner): def __init__(self, model, input_signature): super(IterativePruningRunner, self).__init__(model, input_signature) self._sens_path = os.path.join(_VAI_DIR, self._graph.name + '.sens') def _load_analysis_result(self): return sens.load_sens(self._sens_path) if os.path.exists( self._sens_path) else None def ana(self, eval_fn, args=(), gpus=None, excludes=None, forced=False, with_group_conv: bool = False): """Performs model analysis. Arguments: eval_fn: Callable object that takes a `torch.nn.Module` object as its first argument and returns the evaluation score. args: A tuple of arguments that will be passed to eval_fn. gpus: A tuple or list of gpu indices used for model analysis. If not set, the default gpu will be used. excludes: A list of node name or torch module to be excluded from pruning. """ net_sens = None if forced else self._load_analysis_result() excluded_nodes = self._get_exclude_nodes(excludes) if excludes else [] analyser = sens.ModelAnalyser(self._graph, excluded_nodes, with_group_conv) if net_sens: analyser.recover_state(calibrate_sens(net_sens, self._graph)) uncompleted_steps = analyser.uncompleted_steps() if len(uncompleted_steps) == 0: logging.info( 'Skip model analysis and use cached result, if you do not want to use it, set forced=True' ) return gpu = gpus[0] if gpus else 0 self._ana_pre_check(gpu, eval_fn, args, excluded_nodes, with_group_conv) num_parallel = len(gpus) if gpus else 1 if num_parallel > 1: self._parallel_ana(gpus, eval_fn, args, analyser, excluded_nodes, with_group_conv) else: self._ana(gpu, eval_fn, args, analyser) def _ana_pre_check(self, gpu, eval_fn, args, excludes, with_group_conv: bool = False): """Prune model but not test it to check if all pruning steps can pass.""" logging.info('Pre-checking for analysis...') groups = pruning_lib.group_nodes(self._graph, excludes, with_group_conv) spec = pruning_lib.PruningSpec.from_node_groups(groups, 0.9) model, pruned_graph, pruning_info = self._prune(spec, mode='sparse') current_env = copy.deepcopy(os.environ) os.environ["CUDA_VISIBLE_DEVICES"] = str(utils.get_actual_device(gpu)) model.eval() model.cuda() eval_fn(model, *args) os.environ = current_env def _ana(self, gpu, eval_fn, args, analyser): uncompleted_steps = copy.copy(analyser.uncompleted_steps()) total_steps = analyser.steps() try: for step in uncompleted_steps: spec = analyser.spec(step) model, _, _ = self._prune(spec, mode='sparse') current_env = copy.deepcopy(os.environ) os.environ["CUDA_VISIBLE_DEVICES"] = str(utils.get_actual_device(gpu)) model.eval() model.cuda() eval_res = eval_fn(model, *args) os.environ = current_env analyser.record( step, eval_res.item() if isinstance(eval_res, torch.Tensor) else eval_res) analyser.save(self._sens_path) logging.info('Analysis complete %d/%d' % (step + 1, total_steps)) finally: analyser.save(self._sens_path) def _parallel_ana(self, gpus, eval_fn, args, analyser, excludes, with_group_conv: bool = False): graph = utils.parse_to_graph(self._model, self._input_signature) total_steps = analyser.steps() uncompleted_steps = copy.copy(analyser.uncompleted_steps()) ctx = mp.get_context('spawn') error_queues = [] input_queue = ctx.Queue() output_queue = ctx.Queue() for step in uncompleted_steps: input_queue.put(step) try: processes = [] # To avoid CUDA-out-of-memory error. Main process should execute one copy of task # using the same GPU on which we do torch tracing input_signature_device = self._input_signature.device for rank in range(len(gpus)): # For the GPU on which we do torch tracing, the task should be executed by main process instead of sub-processes if input_signature_device.type == "cuda" and input_signature_device.index == int( gpus[rank]): continue current_env = copy.deepcopy(os.environ) p = ctx.Process( target=AnaTask( self._model, self._input_signature, excludes, gpus[rank], eval_fn, eval_fn_args=args, with_group_conv=with_group_conv), args=(input_queue, output_queue)) p.start() os.environ = current_env processes.append(p) # Main process executes one copy of task if input_signature_device.type == "cuda": current_env = copy.deepcopy(os.environ) ana_task = AnaTask( self._model, self._input_signature, excludes, input_signature_device.index, eval_fn, eval_fn_args=args) ana_task(input_queue, output_queue) os.environ = current_env for p in processes: p.join() finally: while not output_queue.empty(): cur_step, score = output_queue.get() analyser.record(cur_step, score) analyser.save(self._sens_path) def prune(self, removal_ratio=None, threshold=None, spec_path=None, excludes=None, mode='sparse', pruning_info_path=None, channel_divisible=2): """Prune the network by given removal_ratio or threshold. Arguments: removal_ratio: The expected percentage of macs reduction. This is just a hint value, the actual macs drop not necessarily strictly to this value. threshold: Relative proportion of model performance loss that can be tolerated. spec_path: Pre-defined pruning specification. excludes: Modules that need to excludes from pruning. mode: One of ['sparse', 'slim']. channel_divisible: The number of remaining channels in the pruned layer can be divided by channel_divisble. Return: A `torch.nn.Module` object with addtional pruning info. """ if not isinstance(removal_ratio, float): raise errors.OptimizerInvalidArgumentError( 'Expected float "ratio", but got {}({})'.format( removal_ratio, type(removal_ratio))) net_sens = self._load_analysis_result() if net_sens is None: raise errors.OptimizerNoAnaResultsError( "Must call ana() before model pruning.") excluded_nodes = self._get_exclude_nodes(excludes) if excludes else [] training_flag_info = {} for name, module in self._model.named_modules(): training_flag_info[name] = module.training if removal_ratio: logging.info('Pruning ratio = {}'.format(removal_ratio)) spec = self._spec_from_ratio(net_sens, removal_ratio, excluded_nodes) spec.channel_divisible = channel_divisible elif threshold: logging.info('Pruning threshold = {}'.format(threshold)) spec = self._spec_from_threshold(net_sens, threshold, excluded_nodes) spec.channel_divisible = channel_divisible elif spec_path: logging.info('Pruning specification = {}'.format(spec_path)) spec = pruning_lib.PruningSpec.deserialize( json.load(open(spec_path, 'r'))) else: raise errors.OptimizerInvalidArgumentError( 'One of [ratio, threshold, spec_path] must be given.') spec_path = os.path.join( _VAI_DIR, '{}_ratio_{}.spec'.format(self._graph.name, removal_ratio)) with open(spec_path, 'w') as f: json.dump(spec.serialize(), f, indent=2) logging.info('Pruning spec saves in {}'.format(spec_path)) model, pruned_graph, pruning_info = self._prune(spec, mode) pruning_info_path = pruning_info_path if pruning_info_path else spec_path.split( '.spec')[0] + '_pruning_info.json' self.generate_pruning_info(pruning_info_path, pruning_info) model._graph = pruned_graph model._pruning_info = pruning_info if mode == 'sparse': model._register_state_dict_hook(_remove_mask) model.slim_state_dict = types.MethodType(slim_state_dict, model) else: model.sparse_state_dict = types.MethodType(sparse_state_dict, model) for name, module in model.named_modules(): module.training = training_flag_info[name] logging.info('Pruning summary:') slim_model = self._prune(spec, 'slim')[0] self._summary(slim_model) return model def transform(self, spec_path, script_path='pruned_model.py'): with open(spec_path, 'r') as f: spec = json.load(f) _, pruned_graph, _ = self._prune(spec, 'slim') pruned_model, _ = utils.rebuild_model(pruned_graph) return pruned_model def _spec_from_threshold(self, net_sens, threshold, excludes): groups = net_sens.prunable_groups_by_threshold(threshold, excludes) return pruning_lib.PruningSpec(groups) def _spec_from_ratio(self, net_sens, ratio, excludes): logging.info('Searching for appropriate ratio for each layer...') macs, _ = profiler.model_complexity(self._model, self._input_signature) expected_macs = (1 - ratio) * macs macs_tolerance = 1e-2 min_th = 1e-5 max_th = 1 - min_th num_attempts = 0 max_attempts = 100 prev_spec = None cur_spec = None while num_attempts < max_attempts: prev_spec = cur_spec num_attempts += 1 threshold = (min_th + max_th) / 2 cur_spec = self._spec_from_threshold(net_sens, threshold, excludes) if prev_spec and prev_spec == cur_spec: continue pruned_model = self._prune(cur_spec, 'slim')[0] current_macs, _ = profiler.model_complexity(pruned_model, self._input_signature) error = abs(expected_macs - current_macs) / expected_macs if error < macs_tolerance: break if current_macs < expected_macs: max_th = threshold else: min_th = threshold return cur_spec class OneStepPruningRunner(PruningRunner): """EagleEye.""" def __init__(self, model, input_signature): super(OneStepPruningRunner, self).__init__(model, input_signature) self._searcher_saved_path = os.path.join(_VAI_DIR, self._graph.name + '.search') def search(self, gpus=['0'], calibration_fn=None, calib_args=(), num_subnet=200, removal_ratio=0.5, excludes=[], eval_fn=None, eval_args=(), forced=False, with_group_conv: bool = False): """ Perform pruned candidates search. Arguments: gpus: A tuple or list of gpu indices used for model analysis. If not set, the default gpu will be used. calibration_fn: Callable object that takes a `torch.nn.Module` object as its first argument. It's for calibrating BN layer's statistics. calib_args: A tuple of arguments that will be passed to calibration_fn. num_subnet: The number of subnets needed to search matching the macs requirement. removal_ratio: The expected percentage of macs reduction. excludes: Modules that need to exclude from pruning. eval_fn: Callable object that takes a `torch.nn.Module` object as its first argument and returns the evaluation score. eval_args: A tuple of arguments that will be passed to eval_fn. """ self._searcher_saved_path = os.path.join( _VAI_DIR, '{}_ratio_{}.search'.format(self._graph.name, removal_ratio)) excluded_nodes = self._get_exclude_nodes(excludes) if excludes else [] orig_macs, orig_params = profiler.model_complexity( self._model, self._input_signature, readable=False) searcher = None if not forced and os.path.exists(self._searcher_saved_path): searcher = search.load_searcher(self._searcher_saved_path) if len(searcher._subnets) >= num_subnet: logging.info( 'Skip subnet search and use cached result, if you do not want to use it, set forced=True' ) return if not searcher: groups = pruning_lib.group_nodes(self._graph, excluded_nodes, with_group_conv) searcher = search.SubnetSearcher(groups) base_score = eval_fn(self._model, *eval_args) if isinstance(base_score, torch.Tensor): base_score = base_score.item() searcher.set_supernet(base_score, orig_macs) self._search_random_precheck(calibration_fn, eval_fn, removal_ratio, excluded_nodes, calib_args, eval_args, orig_macs, with_group_conv) num_parallel = len(gpus) if gpus else 1 if num_parallel > 1: self._parallel_search_random(gpus, calibration_fn, eval_fn, num_subnet, removal_ratio, excluded_nodes, calib_args, eval_args, searcher, orig_macs, with_group_conv) else: self._search_random(calibration_fn, eval_fn, num_subnet, removal_ratio, excluded_nodes, calib_args, eval_args, searcher, orig_macs, with_group_conv) def _parallel_search_random(self, gpus, calibration_fn, eval_fn, num_subnet, removal_ratio, excluded_nodes, calib_args, eval_args, searcher, orig_macs, with_group_conv: bool = False): ctx = mp.get_context('spawn') output_queue = ctx.Queue() input_queue = ctx.Queue() for index in range(len(searcher._subnets), num_subnet): input_queue.put(index) try: processes = [] # To avoid CUDA-out-of-memory error. Main process should execute one copy of task # using the same GPU on which we do torch tracing input_signature_device = self._input_signature.device for rank in range(len(gpus)): # For the GPU on which we do torch tracing, the task should be executed by main process instead of sub-processes if input_signature_device.type == "cuda" and input_signature_device.index == int( gpus[rank]): continue current_env = copy.deepcopy(os.environ) p = ctx.Process( target=RandomSearchTask( self._model, self._input_signature, removal_ratio, num_subnet, orig_macs, excluded_nodes, gpus[rank], eval_fn, calibration_fn, eval_args=eval_args, calib_args=calib_args, with_group_conv=with_group_conv), args=(input_queue, output_queue)) p.start() os.environ = current_env processes.append(p) # Main process executes one copy of task if input_signature_device.type == "cuda": current_env = copy.deepcopy(os.environ) search_task = RandomSearchTask( self._model, self._input_signature, removal_ratio, num_subnet, orig_macs, excluded_nodes, input_signature_device.index, eval_fn, calibration_fn, eval_args=eval_args, calib_args=calib_args, with_group_conv=with_group_conv) search_task(input_queue, output_queue) os.environ = current_env for p in processes: p.join() finally: while not output_queue.empty(): searcher.add_subnet(*output_queue.get()) search.save_searcher(searcher, self._searcher_saved_path) def _search_random_precheck(self, calibration_fn, eval_fn, removal_ratio, excluded_nodes, calib_args, eval_args, orig_macs, with_group_conv: bool = False): groups = pruning_lib.group_nodes(self._graph, excluded_nodes, with_group_conv) spec = pruning_lib.PruningSpec.from_node_groups(groups, 0.9) pruned_model, pruned_graph, pruning_info = self._prune(spec, mode='slim') macs, params = profiler.model_complexity( pruned_model, self._input_signature, readable=False) cur_removal_ratio = 1 - macs / orig_macs calibration_fn(pruned_model, *(calib_args)) score = eval_fn(pruned_model, *(eval_args)) if isinstance(score, torch.Tensor): score = score.item() logging.info('ratios={}, score={}, cur_removal_ratio={}'.format( removal_ratio, score, cur_removal_ratio)) def _search_random(self, calibration_fn, eval_fn, num_subnet, removal_ratio, excluded_nodes, calib_args, eval_args, searcher, orig_macs, with_group_conv: bool = False): groups = pruning_lib.group_nodes(self._graph, excluded_nodes, with_group_conv) ratio_min = 0.05 ratio_max = 0.9 epsilon = 0.01 index = len(searcher._subnets) + 1 try: while True: ratios = [] spec = pruning_lib.PruningSpec() for group in groups: if removal_ratio >= 0.5: mu = random.uniform(removal_ratio / 2, removal_ratio) else: mu = random.uniform(0, removal_ratio / 2) sigma = random.uniform(0, 0.3) pruning_ratio = random.gauss(mu, sigma) pruning_ratio = np.clip(pruning_ratio, ratio_min, ratio_max) ratios.append(pruning_ratio) spec.add_group( pruning_lib.PrunableGroup(group.nodes, pruning_ratio, group.num_groups)) pruned_model, pruned_graph, pruning_info = self._prune( spec, mode='slim') macs, params = profiler.model_complexity( pruned_model, self._input_signature, readable=False) cur_removal_ratio = 1 - macs / orig_macs if abs(removal_ratio - cur_removal_ratio) < epsilon: calibration_fn(pruned_model, *(calib_args)) score = eval_fn(pruned_model, *(eval_args)) if isinstance(score, torch.Tensor): score = score.item() searcher.add_subnet(ratios, score, 1 - cur_removal_ratio) search.save_searcher(searcher, self._searcher_saved_path) logging.info('Search complete %d/%d' % (index, num_subnet)) #logging.info('Index={}, ratios={}, score={}, cur_removal_ratio={}'.format( # index, ratios, score, cur_removal_ratio)) index += 1 #logging.info('cur_removal_ratio={}'.format(cur_removal_ratio)) if index > num_subnet: break finally: search.save_searcher(searcher, self._searcher_saved_path) logging.info('Search %d subnets!' % (len(searcher._subnets))) def prune(self, mode='slim', removal_ratio=None, index=None, pruning_info_path=None, channel_divisible=2): """Get pruned candidate subnet of the specific index. Arguments: mode: One of ['sparse', 'slim']. removal_ratio: Pruning ratio for model. index: Subnet index. By default, the optimal subnet is selected automatically. Return: A `torch.nn.Module` object with addtional pruning info. """ assert removal_ratio, 'Pruning ratio for model is needed!' self._searcher_saved_path = os.path.join( _VAI_DIR, '{}_ratio_{}.search'.format(self._graph.name, removal_ratio)) searcher = search.load_searcher(self._searcher_saved_path) groups = searcher.groups if index: subnet = searcher.subnet(index) else: subnet = searcher.best_subnet() logging.info('Sparsity={}, score={}'.format(1 - subnet.macs, subnet.score)) assert len(groups) == len(subnet.ratios) spec = searcher.spec(subnet.ratios) spec_path = self._searcher_saved_path.split( '.search')[0] + '_{}.spec'.format(index if index else 'best') with open(spec_path, 'w') as f: json.dump(spec.serialize(), f, indent=2) logging.info('Pruning spec saves in {}'.format(spec_path)) spec.channel_divisible = channel_divisible model, pruned_graph, pruning_info = self._prune(spec, mode=mode) pruning_info_path = pruning_info_path if pruning_info_path else self._searcher_saved_path.split( '.search')[0] + '_pruning_info.json' self.generate_pruning_info(pruning_info_path, pruning_info) model._graph = pruned_graph model._pruning_info = pruning_info model._register_state_dict_hook(_remove_mask) if mode == 'sparse': model.slim_state_dict = types.MethodType(slim_state_dict, model) else: model.sparse_state_dict = types.MethodType(sparse_state_dict, model) return model def get_pruning_runner(model, input_signature, method): assert method in ['iterative', 'one_step'] cls = IterativePruningRunner if method == 'iterative' else OneStepPruningRunner return cls(model, input_signature)
null
23,680
import collections import copy import json import numpy as np import os import random import torch import torch.multiprocessing as mp import types from typing import List from nndct_shared.base import NNDCT_OP as OpTypes from nndct_shared.pruning import errors from nndct_shared.pruning import logging from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.pruning import utils as spu from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.utils import common from pytorch_nndct import parse from pytorch_nndct.pruning import utils from pytorch_nndct.utils import module_util as mod_util from pytorch_nndct.utils import profiler from pytorch_nndct.utils import torch_const from pytorch_nndct.utils.calibration import calibrate_sens, calibrate_spec def sparse_state_dict(self, destination=None, prefix='', keep_vars=False): state_dict = self.state_dict(destination, prefix, keep_vars) for node in self._graph.nodes: pruning_info = self._pruning_info.get(node.name, None) if not pruning_info: continue for tensor in node.op.params.values(): # ResNet::layer1.1.conv1.weight -> layer1.1.conv1.weight key = tensor.name.lstrip(self._graph.name + torch_const.TorchGraphSymbol.GRAPH_SCOPE_SYM) value = utils.pad_to_sparse_tensor(state_dict[key], pruning_info) state_dict[key] = value return state_dict
null
23,681
import collections import copy import json import numpy as np import os import random import torch import torch.multiprocessing as mp import types from typing import List from nndct_shared.base import NNDCT_OP as OpTypes from nndct_shared.pruning import errors from nndct_shared.pruning import logging from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.pruning import utils as spu from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.utils import common from pytorch_nndct import parse from pytorch_nndct.pruning import utils from pytorch_nndct.utils import module_util as mod_util from pytorch_nndct.utils import profiler from pytorch_nndct.utils import torch_const from pytorch_nndct.utils.calibration import calibrate_sens, calibrate_spec _CONVTRANSPOSE_OPS = [OpTypes.CONVTRANSPOSE2D, OpTypes.CONVTRANSPOSE3D] def _prune_tensor(tensor, out_dims, in_dims): """Remove pruned channels of given tensor and returns a slim tensor.""" dim_size = len(tensor.shape) ndarray = tensor.detach().cpu().numpy() out_axis, in_axis = 0, 1 if out_dims: ndarray = np.delete(ndarray, out_dims, axis=out_axis) if in_dims and dim_size > in_axis: ndarray = np.delete(ndarray, in_dims, axis=in_axis) return torch.from_numpy(ndarray) The provided code snippet includes necessary dependencies for implementing the `slim_state_dict` function. Write a Python function `def slim_state_dict(self, destination=None, prefix='', keep_vars=False)` to solve the following problem: Returns a slim state dict in which the weight names are same with the original model and the tensors are pruned to slim ones. Here is the function: def slim_state_dict(self, destination=None, prefix='', keep_vars=False): """Returns a slim state dict in which the weight names are same with the original model and the tensors are pruned to slim ones. """ assert prefix == '' if destination is None: destination = collections.OrderedDict() destination._metadata = collections.OrderedDict() state_dict = self.state_dict(None, prefix, keep_vars) for node in self._graph.nodes: node_pruning = self._pruning_info.get(node.name, None) for param, tensor in node.op.params.items(): # ResNet::layer1.1.conv1.weight -> layer1.1.conv1.weight key = utils.state_dict_key_from_tensor(tensor) value = state_dict[key] if not node_pruning: destination[key] = value else: out_dims = node_pruning.removed_outputs in_dims = node_pruning.removed_inputs if (node.op.type in _CONVTRANSPOSE_OPS and param == node.op.ParamName.WEIGHTS): out_dims, in_dims = in_dims, out_dims destination[key] = _prune_tensor(value, out_dims, in_dims) # The model during train/inference procedure may be differnt. # 'self._graph' is generated according to inference model, # so some keys in state_dict may not in destination. for key in state_dict: if key not in destination: destination[key] = state_dict[key] return destination
Returns a slim state dict in which the weight names are same with the original model and the tensors are pruned to slim ones.
23,682
import collections import copy import json import numpy as np import os import random import torch import torch.multiprocessing as mp import types from typing import List from nndct_shared.base import NNDCT_OP as OpTypes from nndct_shared.pruning import errors from nndct_shared.pruning import logging from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.pruning import utils as spu from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.utils import common from pytorch_nndct import parse from pytorch_nndct.pruning import utils from pytorch_nndct.utils import module_util as mod_util from pytorch_nndct.utils import profiler from pytorch_nndct.utils import torch_const from pytorch_nndct.utils.calibration import calibrate_sens, calibrate_spec def generate_indices_group(indices: List[int], dim_size: int, groups: int) -> List[List[int]]: indices_set = set(indices) interval: int = dim_size // groups start_idx = 0 end_idx = interval ret: List[List[int]] = [] while start_idx < dim_size: idx_group: List[int] = [] for i in range(start_idx, end_idx): if i in indices_set: idx_group.append(i - start_idx) ret.append(idx_group) start_idx = end_idx end_idx += interval return ret The provided code snippet includes necessary dependencies for implementing the `_sparsify_tensor` function. Write a Python function `def _sparsify_tensor(tensor: torch.Tensor, out_dims: List[int], in_dims: List[int], groups: int = 1)` to solve the following problem: Fill 0 in removed channels. Here is the function: def _sparsify_tensor(tensor: torch.Tensor, out_dims: List[int], in_dims: List[int], groups: int = 1): """Fill 0 in removed channels.""" device = tensor.device tensor = tensor.detach().cpu().clone(memory_format=torch.contiguous_format) dim_size = len(tensor.shape) if groups == 1: if out_dims: tensor[out_dims, ...] = 0 # data format in pytorch: OIHW if in_dims and dim_size > 1: tensor[:, in_dims, ...] = 0 else: out_dims_group = generate_indices_group(out_dims, tensor.size(0), groups) in_dims_group = generate_indices_group( in_dims, tensor.size(1) * groups, groups) if dim_size > 1 else [[]] * groups parts = tensor.split(tensor.size(0) // groups, dim=0) sparse_parts: List[torch.Tensor] = [] for part, o, i in zip(parts, out_dims_group, in_dims_group): sparse_parts.append(_sparsify_tensor(part, o, i)) tensor = torch.cat(sparse_parts, dim=0) return tensor.to(device)
Fill 0 in removed channels.
23,683
import collections import copy import json import numpy as np import os import random import torch import torch.multiprocessing as mp import types from typing import List from nndct_shared.base import NNDCT_OP as OpTypes from nndct_shared.pruning import errors from nndct_shared.pruning import logging from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.pruning import utils as spu from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.utils import common from pytorch_nndct import parse from pytorch_nndct.pruning import utils from pytorch_nndct.utils import module_util as mod_util from pytorch_nndct.utils import profiler from pytorch_nndct.utils import torch_const from pytorch_nndct.utils.calibration import calibrate_sens, calibrate_spec def _apply_mask(module, inputs): for tensor_name in module._tensor_names: weight = getattr(module, tensor_name) mask = getattr(module, tensor_name + '_mask') weight.data = mask.to(dtype=weight.dtype) * weight
null
23,684
import collections import copy import json import numpy as np import os import random import torch import torch.multiprocessing as mp import types from typing import List from nndct_shared.base import NNDCT_OP as OpTypes from nndct_shared.pruning import errors from nndct_shared.pruning import logging from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.pruning import utils as spu from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.utils import common from pytorch_nndct import parse from pytorch_nndct.pruning import utils from pytorch_nndct.utils import module_util as mod_util from pytorch_nndct.utils import profiler from pytorch_nndct.utils import torch_const from pytorch_nndct.utils.calibration import calibrate_sens, calibrate_spec def _remove_mask(self, destination=None, prefix='', keep_vars=False): keys = list(destination.keys()) for key in keys: if '_mask' in key: del destination[key] return destination
null
23,685
from collections import ChainMap, defaultdict from enum import Enum from tqdm import tqdm import torch from nndct_shared.base.key_names import FrameworkType from nndct_shared.nndct_graph import (Graph, Tensor, Block, Node, reorder_multi_subgraph_nodes) from nndct_shared.utils import NndctDebugLogger, NndctOption, NndctScreenLogger, QError, QWarning, QNote from pytorch_nndct.utils import build_aten_torch_ops_table, TorchGraphSymbol from .op_dispatcher import * from .parse_utils import * from .parser_post_process import change_addmm_to_linear from .torch_op_def import TorchUnknownOperation from .trace_helper import TorchGraphHandler from typing import Mapping from .rich_in_out_helper import FlattenInOutModelForTrace, StandardInputData class TorchUnknownOperation(Operation): def __init__(self, nndct_op_type): super().__init__(nndct_op_type) def unknown_op_type_check(graph: Graph): unkown_ops = set() custom_ops = set() for node in graph.all_nodes(): if isinstance(node.op, TorchUnknownOperation): unkown_ops.add(node.op.type) elif node.has_custom_op(): custom_ops.add(node.op.type) for op in custom_ops: NndctScreenLogger().warning2user(QWarning.FLOAT_OP, f"The quantizer recognize new op `{op}` as a float operator by default.") # if custom_ops: # NndctScreenLogger().info(f"You can make these new ops quantizable by add them to custom_quant_ops, \ # e.g. quantizer= torch_quantizer(..., custom_quant_ops=['{list(custom_ops)[0]}',...])") NndctScreenLogger().check2user(QError.UNSUPPORTED_OPS, f"Unsupported Ops: {unkown_ops}.", len(unkown_ops)==0)
null
23,686
from nndct_shared.nndct_graph.base_tensor import Tensor from pytorch_nndct.utils import TorchGraphSymbol from .rich_in_out_helper import FlattenInOutModelForTrace def convert_np_type_to_pytorch_type(np_type): return { 'int64': 'torch.int64', 'int32': 'torch.int32', 'float32': 'torch.float', 'float64': 'torch.double' }.get(np_type, np_type)
null
23,687
from nndct_shared.nndct_graph.base_tensor import Tensor from pytorch_nndct.utils import TorchGraphSymbol from .rich_in_out_helper import FlattenInOutModelForTrace def convert_dtype_between_np_and_pytorch(dtype): return { 'int64': 'torch.int64', 'int32': 'torch.int32', 'float32': 'torch.float', 'float64': 'torch.double', 'torch.int64': 'int64', 'torch.long': 'int64', 'torch.int32': 'int32', 'torch.int': 'int32', 'torch.float32': 'float32', 'torch.float': 'float32', 'torch.float64': 'float64', 'torch.double': 'float64', }.get(dtype, dtype)
null
23,688
from nndct_shared.nndct_graph.base_tensor import Tensor from pytorch_nndct.utils import TorchGraphSymbol from .rich_in_out_helper import FlattenInOutModelForTrace _GRAPH_SCOPE_SYM = TorchGraphSymbol.GRAPH_SCOPE_SYM class FlattenInOutModelForTrace(torch.nn.Module): def getModelName(cls): return 'FlattenInOutModelForTrace' def getOriginModelNameFormString(cls, data_str): names = re.findall(r"nndct_st_([\w_]+)_ed", data_str) if len(names) > 0: return names[0] else: return None def check_need_recovery_name(cls, name): return 'nndct_st_' in name or 'FlattenInOutModelForTrace' in name def recovery_tensor_name(cls, name): return re.sub(r"nndct_st_[\w_]+_ed\.", '', name) def recovery_node_scope_name(cls, scope_name): real_class_name = re.findall(r"FlattenInOutModelForTrace/(.*)?\[nndct_st_[\w_]+_ed\]", scope_name) if len(real_class_name) > 0: scope_name = re.sub(r"FlattenInOutModelForTrace/(.*)?\[nndct_st_[\w_]+_ed\]", real_class_name[0], scope_name) real_model_name = re.findall(r".*FlattenInOutModelForTrace::/(.*::)nndct_st_[\w_]+_ed", scope_name) if len(real_model_name) > 0: scope_name = re.sub(r".*FlattenInOutModelForTrace::/(.*::)nndct_st_[\w_]+_ed", real_model_name[0], scope_name) return scope_name def __init__(self, inner_model, input_schema) -> None: super().__init__() self.module_name = "nndct_st_" + inner_model._get_name() + "_ed" setattr(self, self.module_name, inner_model) self.input_schema = input_schema self.training = inner_model.training def forward(self, *flatten_input): input = self.input_schema(flatten_input) output = getattr(self, self.module_name)(*input['args'], **input['kwargs']) flatten_output, _ = flatten_to_tuple(output) return flatten_output The provided code snippet includes necessary dependencies for implementing the `get_full_name` function. Write a Python function `def get_full_name(graph_name: str, name: str) -> str` to solve the following problem: get the full name of node/tensor in graph Args: graph_name (str): graph name Returns: str: full name Here is the function: def get_full_name(graph_name: str, name: str) -> str: """get the full name of node/tensor in graph Args: graph_name (str): graph name Returns: str: full name """ name = _GRAPH_SCOPE_SYM.join([graph_name, name]) name = FlattenInOutModelForTrace.recovery_tensor_name(name) name = FlattenInOutModelForTrace.recovery_node_scope_name(name) return name
get the full name of node/tensor in graph Args: graph_name (str): graph name Returns: str: full name
23,689
from nndct_shared.nndct_graph.base_tensor import Tensor from pytorch_nndct.utils import TorchGraphSymbol from .rich_in_out_helper import FlattenInOutModelForTrace def get_short_name(full_name: str) -> str: """get the name of node/tensor in graph without graph name Args: full_name (str): full name of node/tensor Returns: str: short name """ return full_name.split(_GRAPH_SCOPE_SYM)[-1] The provided code snippet includes necessary dependencies for implementing the `get_formal_name` function. Write a Python function `def get_formal_name(hier_name: str) -> str` to solve the following problem: replace `.` with `_` Args: hier_name (str): "layer_0.layer_1" Returns: str: "layer_0_layer_1" Here is the function: def get_formal_name(hier_name: str) -> str: """replace `.` with `_` Args: hier_name (str): "layer_0.layer_1" Returns: str: "layer_0_layer_1" """ return get_short_name(hier_name.replace(".", "_"))
replace `.` with `_` Args: hier_name (str): "layer_0.layer_1" Returns: str: "layer_0_layer_1"
23,690
from nndct_shared.nndct_graph.base_tensor import Tensor from pytorch_nndct.utils import TorchGraphSymbol from .rich_in_out_helper import FlattenInOutModelForTrace class TorchScriptModuleHandler(object): def __init__(self): def build_torch_graph(self, graph_name, script_module, *args): def rename_graph_inputs(graph): def _optimize_raw_graph(self, graph): def _is_param_const_node(fw_node, raw_graph): def _build_raw_graph(self, graph_name, fw_graph, params=None, blobs=None): def _create_ret_value(self, graph, raw_graph): def _create_inputs_value(self, graph, raw_graph): def _create_attrs_value(self, graph, raw_graph): def _create_params_value(self, graph, script_module): def _add_node(self, fw_node, raw_graph): def _execute_optimize(self, raw_graph): class TorchGraphHandler(object): def __init__(self): def _check_control_flow(fw_graph): def _get_fw_graph_from_module(self, module, input_data, train): def build_torch_graph(self, graph_name, module, input_args, train=False): def _trace_graph_from_model(self, module, input_args, train): def _get_graph_from_script(self, module, input): def _get_param_names(self, graph): def _opt_raw_graph(self, raw_graph, is_control_flow_graph): def _opt_raw_block(self, raw_block): def _optimize_for_jit(self, raw_block): def _optimize(self, raw_block): def _create_attrs_value(self, graph): def _create_inputs_value(self, graph): def _build_block_graph(self, fw_block): def _create_raw_graph(self, graph_name, fw_graph): def _is_param_const_node(self, fw_node): def add_torch_node(self, fw_node): def _check_stub_topology(self, raw_graph): def _path_str(path): def create_graph_handler(module): import torch if isinstance(module, torch.jit.ScriptModule): from .script_helper import TorchScriptModuleHandler return TorchScriptModuleHandler() elif isinstance(module, torch.nn.Module): from .trace_helper import TorchGraphHandler return TorchGraphHandler() else: raise NotImplementedError()
null
23,691
from nndct_shared.nndct_graph.base_tensor import Tensor from pytorch_nndct.utils import TorchGraphSymbol from .rich_in_out_helper import FlattenInOutModelForTrace class Tensor(object): """A wrapper of np.ndarray used in two ways: - The outputs of an operation. - The parameters of an operation. In the former case, you can use `tensor.node` to get the node that outputs this tensor. In the latter case, `tensor.node` is None. For getting raw ndarray, call `tensor.data`. """ def __init__(self, name=None, shape=None, dtype=None, device=None, requires_grad=None, data=None, node=None, layout=None): self._node = weakref.ref(node) if node else node self._name = name self._shape = shape self._data = data self._dtype_map = { np.dtype('float64'): 'float64', np.dtype('float32'): 'float32', np.dtype('float16'): 'float16', np.dtype('complex64'): 'complex64', np.dtype('int64'): 'int64', np.dtype('int32'): 'int32', np.dtype('int16'): 'int16', np.dtype('int8'): 'int8', np.dtype('bool'): 'bool', np.dtype('uint8'): 'uint8' } if dtype in self._dtype_map: self._dtype = self._dtype_map[dtype] else: self._dtype = dtype self._device = device self._requires_grad = requires_grad self._offset = 0 self._uses = [] self._attr_uses = [] self._device = device def __deepcopy__(self, memo): raise NotImplementedError("Deep copy is prohibited, use `clone_from` instead.") def clone_from(self, src_tensor): self._shape = src_tensor._shape self._data = copy.deepcopy(src_tensor._data) self._dtype = src_tensor._dtype self._device = src_tensor._device self._requires_grad = src_tensor._requires_grad def from_ndarray(self, data): if not isinstance(data, np.ndarray): raise TypeError("'data' must be a numpy ndarray") self._data = np.copy(data) self._dtype = self._dtype_map[self._data.dtype] self._shape = list(self._data.shape) def from_tensor(self, tensor): self._dtype = tensor.dtype self._shape = tensor.shape def from_des(self, shape, dtype): self._shape = shape self._dtype = dtype def transpose(self, axes=None): trans_data = None if self._data is not None: trans_data = self._data.transpose(axes) trans_data = np.ascontiguousarray(trans_data) trans_shape = list(trans_data.shape) else: trans_shape = [self._shape[i] for i in axes] self._data = trans_data self._shape = trans_shape def clean_data(self): self._data = None def __str__(self): return "Tensor: {}(shape={}, dtype={})".format( self._name if self._name else "", self._shape, self._dtype) def description(self): desp = {} desp['name'] = self._name desp['shape'] = self._shape desp['dtype'] = self._dtype desp['node'] = self.node.name if self.node else None return desp def is_real_tensor(self): return self.is_complete_tensor() or self.dtype == "tensor" def is_list_type(self): return "list" in self.dtype def is_complete_tensor(self) -> bool: # not necessary to hold real data for completeTensor return True if self.shape and self.dtype else False def is_param_tensor(self) -> bool: return True if self._node is None else False def shape(self): return self._shape def shape(self, shape): self._shape = shape def ndim(self): return len(self._shape) if self._shape else None def dtype(self): return self._dtype def dtype(self, dtype): self._dtype = dtype def data(self): return self._data def data(self, value): if isinstance(value, np.ndarray): self.from_ndarray(value) elif isinstance(value, Tensor): self.from_tensor(value) elif isinstance(value, (int, float, bool)): #raise ValueError(f"Accept [int, float, bool] type data, but {type(value)} is given") self._data = value self._shape = [] def node(self): return self._node() if self._node is not None else None def node(self, value): self._node = weakref.ref(value) if value else value def name(self): return self._name def name(self, name): self._name = name def device(self): return self._device def device(self, device): self._device = device def requires_grad(self): return self._requires_grad def requires_grad(self, need_grad): self._requires_grad = need_grad def offset(self): return self._offset def offset(self, offset): self._offset = offset def uses(self): return self._uses def owning_graph(self): return self.node.owning_graph def attr_uses(self): return self._attr_uses def replace_first_use_with(self, new_tensor): assert self.owning_graph is new_tensor.owning_graph u = self.uses[0] u.user.in_tensors[u.offset] = new_tensor new_tensor.uses.append(u) self.uses.pop(0) def replace_uses_with(self, new_tensor): assert self is not new_tensor while len(self.uses) > 0: self.replace_first_use_with(new_tensor) self.replace_attr_uses_with(new_tensor) def replace_attr_with_new_tensor_v2(self, attr_use, new_tensor): def _replace(attr_name, attr_value): if isinstance(attr_value, list): new_attr_value = [] for value in attr_value: if self is value: new_value = _replace(attr_name, value) new_attr_value.append(new_value) elif isinstance(value, (tuple, list)): new_value = _replace(attr_name, value) new_attr_value.append(new_value) else: new_attr_value.append(value) return new_attr_value elif isinstance(attr_value, tuple): new_attr_value = [] for value in list(attr_value): if self is value: new_value = _replace(attr_name, value) new_attr_value.append(new_value) elif isinstance(value, (tuple, list)): new_value = _replace(attr_name, value) new_attr_value.append(new_value) else: new_attr_value.append(value) new_attr_value = tuple(new_attr_value) return new_attr_value else: if self is not attr_value: if attr_name == attr_use.attr_name: self.attr_uses.remove(attr_use) return attr_value else: if attr_use not in new_tensor.attr_uses: new_tensor.attr_uses.append(attr_use) self.attr_uses.remove(attr_use) return new_tensor if isinstance(attr_use.attr_name, str): attr_value = attr_use.user.get_config(attr_use.attr_name) else: attr_value = attr_use.user.get_attr(attr_use.attr_name) new_attr_value = _replace(attr_use.attr_name, attr_value) if isinstance(attr_use.attr_name, str): attr_use.user.set_config(attr_use.attr_name, new_attr_value) else: attr_use.user.update_attr(attr_use.attr_name, new_attr_value) def replace_attr_with_new_tensor(self, attr_use, new_tensor): def _replace(attr_name, attr_value): if isinstance(attr_value, list): if self in attr_value: for i in range(len(attr_value)): if attr_value[i] is self: attr_value[i] = new_tensor if attr_use not in new_tensor.attr_uses: new_tensor.attr_uses.append(attr_use) self.attr_uses.remove(attr_use) return else: for val in attr_value: _replace(attr_name, val) if self is not attr_value: if attr_name == attr_use.attr_name: self.attr_uses.remove(attr_use) return if isinstance(attr_name, str): attr_use.user.set_config(attr_name, new_tensor) else: attr_use.user.update_attr(attr_name, new_tensor) if isinstance(attr_use.attr_name, str): attr_value = attr_use.user.get_config(attr_use.attr_name) else: attr_value = attr_use.user.get_attr(attr_use.attr_name) _replace(attr_use.attr_name, attr_value) def replace_first_attr_use_with(self, new_tensor): attr_use = self.attr_uses[0] self.replace_attr_with_new_tensor_v2(attr_use, new_tensor) def replace_attr_uses_with(self, new_tensor): while len(self.attr_uses) > 0: self.replace_first_attr_use_with(new_tensor) def python_dtype(value): type_map = { "torch.int": "int", "torch.long": "int", "torch.short": "int", "torch.float": "float", "torch.half": "float", "torch.double": "float", "torch.bool": "bool", int: "int", float: "float", bool: "bool", str: "str", "int64": "int", "int32": "int", "float32": "float", "float64": "float", "float16": "float" } if isinstance(value, Tensor): return type_map.get(value.dtype, value.dtype) else: return type_map[type(value)]
null
23,692
import torch import torch import math import functools from nndct_shared.utils import (AddXopError, NndctOption, NndctScreenLogger, option_util, QError, QWarning) HANDLED_FUNCTIONS = {} The provided code snippet includes necessary dependencies for implementing the `implements` function. Write a Python function `def implements(torch_function_list)` to solve the following problem: Register a torch function override for ScalarTensor Here is the function: def implements(torch_function_list): """Register a torch function override for ScalarTensor""" @functools.wraps(torch_function_list) def decorator(func): for torch_fn in torch_function_list: HANDLED_FUNCTIONS[torch_fn] = func return func return decorator
Register a torch function override for ScalarTensor
23,693
import torch import torch import math import functools from nndct_shared.utils import (AddXopError, NndctOption, NndctScreenLogger, option_util, QError, QWarning) def convert_tensor_to_tracetensor(data): if isinstance(data, torch.Tensor): return TraceTensor(data.detach().cpu().numpy()).to(data.device).to(data.dtype) if not isinstance(data,(tuple, list)): return data new_data = [] for item in data: new_data.append(convert_tensor_to_tracetensor(item)) if isinstance(data, tuple): new_data = tuple(new_data) return new_data def convert_tracetensor_to_tensor(data): if isinstance(data, TraceTensor): return torch.Tensor(data.detach().cpu().numpy()).to(data.device).to(data.dtype) if not isinstance(data,(tuple, list)): return data new_data = [] for item in data: new_data.append(convert_tracetensor_to_tensor(item)) if isinstance(data, tuple): new_data = tuple(new_data) return new_data def torch_function_old_version(func, types, args, kwargs): args_t = convert_tracetensor_to_tensor(args) ret = func(*args_t, **kwargs) ret = convert_tensor_to_tracetensor(ret) return ret
null
23,694
import torch import torch import math import functools from nndct_shared.utils import (AddXopError, NndctOption, NndctScreenLogger, option_util, QError, QWarning) class TraceTensor(torch.Tensor): def __torch_function__(cls, func, types, args=(), kwargs=None): def split(tensor:TraceTensor, split_size_or_sections:'int|list[int]|tuple[int]', dim:int=0): def aten_split_with_sizes(tensor, split_size_or_sections, dim=0): if isinstance(split_size_or_sections, (list,tuple)): sum = 0 for x in split_size_or_sections: sum = sum + x assert(sum == tensor.shape[dim]) tensor_list_return = [] last_end_index = 0 for k in range(len(split_size_or_sections)): tmp = [slice(None, None) for x in tensor.shape] start = last_end_index end = start + split_size_or_sections[k] last_end_index = end tmp[dim] = slice(start,end) tensor_list_return.append(tensor[tmp]) return tuple(tensor_list_return) def aten_split(tensor, split_size_or_sections, dim=0): if not isinstance(split_size_or_sections, int): assert(False) group_num = tensor.shape[dim] / split_size_or_sections group_num = math.ceil(group_num) if group_num == 0: group_num = 1 tensor_list_return = [] for k in range(group_num): tmp = [slice(None, None) for x in tensor.shape] start = k * split_size_or_sections if k == (group_num - 1): end = tensor.shape[dim] else: end = (k + 1) * split_size_or_sections tmp[dim] = slice(start,end) tensor_list_return.append(tensor[tmp]) return tuple(tensor_list_return) if isinstance(split_size_or_sections, (list, tuple)): return aten_split_with_sizes(tensor, split_size_or_sections, dim) if isinstance(split_size_or_sections, int): return aten_split(tensor, split_size_or_sections, dim) assert(False)
null
23,695
import torch import torch import math import functools from nndct_shared.utils import (AddXopError, NndctOption, NndctScreenLogger, option_util, QError, QWarning) class TraceTensor(torch.Tensor): def __torch_function__(cls, func, types, args=(), kwargs=None): if kwargs is None: kwargs = {} if func not in HANDLED_FUNCTIONS: if hasattr(torch.Tensor, '__torch_function__'): return super().__torch_function__(func, types, args, kwargs) return torch_function_old_version(func, types, args, kwargs) return HANDLED_FUNCTIONS[func](*args, **kwargs) def chunk(input:TraceTensor, chunks:int, dim:int=0): assert(chunks != 0) split_size = math.ceil(input.shape[dim] / chunks) if split_size != 0: chunks = math.ceil(input.shape[dim] / split_size) tensor_list_return = [] for k in range(chunks): tmp = [slice(None, None) for x in input.shape] start = k * split_size if k == (chunks - 1): end = input.shape[dim] else: end = (k + 1) * split_size tmp[dim] = slice(start,end) tensor_list_return.append(input[tmp]) return tuple(tensor_list_return)
null
23,696
import torch import torch import math import functools from nndct_shared.utils import (AddXopError, NndctOption, NndctScreenLogger, option_util, QError, QWarning) def logging_warn(message): NndctScreenLogger().warning2user(QWarning.FLOAT_OP, message) class TraceTensor(torch.Tensor): def __torch_function__(cls, func, types, args=(), kwargs=None): if kwargs is None: kwargs = {} if func not in HANDLED_FUNCTIONS: if hasattr(torch.Tensor, '__torch_function__'): return super().__torch_function__(func, types, args, kwargs) return torch_function_old_version(func, types, args, kwargs) return HANDLED_FUNCTIONS[func](*args, **kwargs) def check_big_pooling(kernel_size, stride, padding): if not NndctOption.nndct_pooling_split_mode.value: return False if isinstance(kernel_size, int): kernel_size = (kernel_size, kernel_size) if isinstance(stride, int): stride = (stride, stride) if isinstance(padding, int): padding = (padding, padding) #------------------------ # only support pooling == 0 check_padding_flag = True if padding[0] != 0 or padding[1] != 0: check_padding_flag = False #------------------------ # only support without overlap check_stride_flag = True if stride != None: if stride[0] != kernel_size[0] or stride[1] != kernel_size[1]: check_stride_flag = False #------------------------ # only support kernel_size < 512 check_kernel_size_flag = False if kernel_size[0] * kernel_size[1] >= 512: check_kernel_size_flag = True #------------------------ base_check_result = check_stride_flag and check_kernel_size_flag and check_padding_flag #------------------------ if base_check_result: # pre-check new kernel_size is equal to orgin kernel_size, prevent infinite recursion new_kernel_size1 = [] new_kernel_size2 = [] for size in kernel_size: a, b = change2TwoNumberMultiplication(size) new_kernel_size1.append(a) new_kernel_size2.append(b) new_kernel_size1 = tuple(new_kernel_size1) new_kernel_size2 = tuple(new_kernel_size2) if new_kernel_size1[0] == kernel_size[0] and new_kernel_size1[1] == kernel_size[1]: return False if new_kernel_size2[0] == kernel_size[0] and new_kernel_size2[1] == kernel_size[1]: return False #------------------------ return base_check_result def avg_pool2d(input, kernel_size, stride=None, padding=0, ceil_mode=False, count_include_pad=True, divisor_override=None): # kernel_size (Union[int, Tuple[int, int]]) – the size of the window # stride (Union[int, Tuple[int, int]]) – the stride of the window. Default value is kernel_size # padding (Union[int, Tuple[int, int]]) – implicit zero padding to be added on both sides # ceil_mode (bool) – when True, will use ceil instead of floor to compute the output shape # count_include_pad (bool) – when True, will include the zero-padding in the averaging calculation # divisor_override (Optional[int]) – if specified, it will be used as divisor, otherwise size of the pooling region will be used. #------------------------------------------------------- # big kernel size split into two smaller kernels if check_big_pooling(kernel_size, stride, padding): logging_warn('big pooling split') new_kernel_size1 = [] new_kernel_size2 = [] for size in kernel_size: a, b = change2TwoNumberMultiplication(size) new_kernel_size1.append(a) new_kernel_size2.append(b) new_kernel_size1 = tuple(new_kernel_size1) new_kernel_size2 = tuple(new_kernel_size2) new_stride1 = new_kernel_size1 new_stride2 = new_kernel_size2 input = torch.nn.functional.avg_pool2d(input, new_kernel_size1, new_stride1, padding, ceil_mode, count_include_pad, divisor_override) input = torch.nn.functional.avg_pool2d(input, new_kernel_size2, new_stride2, padding, ceil_mode, count_include_pad, divisor_override) return input else: normal_data = input.as_subclass(torch.Tensor) normal_data = torch.nn.functional.avg_pool2d(normal_data, kernel_size, stride, padding, ceil_mode, count_include_pad, divisor_override) return normal_data.as_subclass(TraceTensor) def adaptive_avg_pool2d(input, output_size): # output_size (Union[int, None, Tuple[Optional[int], Optional[int]]]) – the target output size of the image of the form H x W. Can be a tuple (H, W) or a single H for a square image H x H. H and W can be either a int, or None which means the size will be the same as that of the input. if output_size == 1: kernel_size = [int(input.shape[-2]), int(input.shape[-1])] if check_big_pooling(kernel_size, kernel_size, 0): logging_warn('replace adapt pooling to pooling') return torch.nn.functional.avg_pool2d(input, kernel_size) normal_data = input.as_subclass(torch.Tensor) normal_data = torch.nn.functional.adaptive_avg_pool2d(normal_data, output_size) return normal_data.as_subclass(TraceTensor)
null
23,697
import torch import torch import math import functools from nndct_shared.utils import (AddXopError, NndctOption, NndctScreenLogger, option_util, QError, QWarning) def logging_warn(message): NndctScreenLogger().warning2user(QWarning.FLOAT_OP, message) class TraceTensor(torch.Tensor): def __torch_function__(cls, func, types, args=(), kwargs=None): if kwargs is None: kwargs = {} if func not in HANDLED_FUNCTIONS: if hasattr(torch.Tensor, '__torch_function__'): return super().__torch_function__(func, types, args, kwargs) return torch_function_old_version(func, types, args, kwargs) return HANDLED_FUNCTIONS[func](*args, **kwargs) def check_big_pooling(kernel_size, stride, padding): if not NndctOption.nndct_pooling_split_mode.value: return False if isinstance(kernel_size, int): kernel_size = (kernel_size, kernel_size) if isinstance(stride, int): stride = (stride, stride) if isinstance(padding, int): padding = (padding, padding) #------------------------ # only support pooling == 0 check_padding_flag = True if padding[0] != 0 or padding[1] != 0: check_padding_flag = False #------------------------ # only support without overlap check_stride_flag = True if stride != None: if stride[0] != kernel_size[0] or stride[1] != kernel_size[1]: check_stride_flag = False #------------------------ # only support kernel_size < 512 check_kernel_size_flag = False if kernel_size[0] * kernel_size[1] >= 512: check_kernel_size_flag = True #------------------------ base_check_result = check_stride_flag and check_kernel_size_flag and check_padding_flag #------------------------ if base_check_result: # pre-check new kernel_size is equal to orgin kernel_size, prevent infinite recursion new_kernel_size1 = [] new_kernel_size2 = [] for size in kernel_size: a, b = change2TwoNumberMultiplication(size) new_kernel_size1.append(a) new_kernel_size2.append(b) new_kernel_size1 = tuple(new_kernel_size1) new_kernel_size2 = tuple(new_kernel_size2) if new_kernel_size1[0] == kernel_size[0] and new_kernel_size1[1] == kernel_size[1]: return False if new_kernel_size2[0] == kernel_size[0] and new_kernel_size2[1] == kernel_size[1]: return False #------------------------ return base_check_result def max_pool2d(input, kernel_size, stride=None, padding=0, dilation=1, ceil_mode=False, return_indices=False): # kernel_size (Union[int, Tuple[int, int]]) – the size of the window to take a max over # stride (Union[int, Tuple[int, int]]) – the stride of the window. Default value is kernel_size # padding (Union[int, Tuple[int, int]]) – implicit zero padding to be added on both sides # dilation (Union[int, Tuple[int, int]]) – a parameter that controls the stride of elements in the window # return_indices (bool) – if True, will return the max indices along with the outputs. Useful for torch.nn.MaxUnpool2d later # ceil_mode (bool) – when True, will use ceil instead of floor to compute the output shape #----------------------------------------------- # big kernel size split into two smaller kernels if check_big_pooling(kernel_size, stride, padding): logging_warn('big pooling split') new_kernel_size1 = [] new_kernel_size2 = [] for size in kernel_size: a, b = change2TwoNumberMultiplication(size) new_kernel_size1.append(a) new_kernel_size2.append(b) new_kernel_size1 = tuple(new_kernel_size1) new_kernel_size2 = tuple(new_kernel_size2) new_stride1 = new_kernel_size1 new_stride2 = new_kernel_size2 input = torch.nn.functional.max_pool2d(input, new_kernel_size1, new_stride1, padding, dilation, ceil_mode, return_indices) input = torch.nn.functional.max_pool2d(input, new_kernel_size2, new_stride2, padding, dilation, ceil_mode, return_indices) return input else: normal_data = input.as_subclass(torch.Tensor) normal_data = torch.nn.functional.max_pool2d(normal_data, kernel_size, stride, padding, dilation, ceil_mode, return_indices) return normal_data.as_subclass(TraceTensor) def adaptive_max_pool2d(input, output_size, return_indices=False): # output_size (Union[int, None, Tuple[Optional[int], Optional[int]]]) – the target output size of the image of the form can be either a int, or None which means the size will be the same as that of the input. # return_indices (bool) – if True, will return the indices along with the outputs. Useful to pass to nn.MaxUnpool2d. Default: False if output_size == 1: kernel_size = [int(input.shape[-2]), int(input.shape[-1])] if check_big_pooling(kernel_size, kernel_size, 0): logging_warn('replace adapt pooling to pooling') return torch.nn.functional.max_pool2d(input, kernel_size) normal_data = input.as_subclass(torch.Tensor) normal_data = torch.nn.functional.adaptive_max_pool2d(normal_data, output_size, return_indices) return normal_data.as_subclass(TraceTensor)
null
23,698
import torch import torch import math import functools from nndct_shared.utils import (AddXopError, NndctOption, NndctScreenLogger, option_util, QError, QWarning) def logging_warn(message): class TraceTensor(torch.Tensor): def __torch_function__(cls, func, types, args=(), kwargs=None): def check_big_pooling(kernel_size, stride, padding): def avg_pool2d(input, kernel_size, stride=None, padding=0, ceil_mode=False, count_include_pad=True, divisor_override=None): def mean(input, dim, keepdim=False, *, dtype=None, out=None): # Tensor.mean(dim=None, keepdim=False, *, dtype=None) if isinstance(dim, (list, tuple)) and len(dim) == 2: new_dim = [] for k in dim: if k < 0: k = k + input.dim() new_dim.append(k) dim = new_dim if (input.dim() == 3 and tuple(dim) == (1, 2)) or (input.dim() == 4 and tuple(dim) == (2, 3)): kernel_size = [int(input.shape[-2]), int(input.shape[-1])] if check_big_pooling(kernel_size, kernel_size, 0): logging_warn('replace mean to avg pooling') data = torch.nn.functional.avg_pool2d(input, kernel_size) if not keepdim: data = torch.squeeze(data, dim=-1) data = torch.squeeze(data, dim=-1) return data normal_data = input.as_subclass(torch.Tensor) normal_data = torch.mean(normal_data, dim, keepdim=keepdim, dtype=dtype, out=out) return normal_data.as_subclass(TraceTensor)
null
23,699
import torch import torch import math import functools from nndct_shared.utils import (AddXopError, NndctOption, NndctScreenLogger, option_util, QError, QWarning) def clear_override_import_redundant_op(graph): # remove alias alias_node_list = graph.findAllNodes("aten::alias") for alias_node in alias_node_list: out_val = alias_node.output() inp_val = alias_node.inputsAt(0) out_val.replaceAllUsesWith(inp_val) alias_node.destroy()
null
23,700
from .op_dispatcher import * from .parse_utils import * def change_addmm_to_linear(raw_graph): for node in raw_graph.nodes: if node.op.type in [NNDCT_OP.ADDMM]: weight = node.op.get_config('mat2') bias = node.op.get_config('input') if (weight and weight.node == None) and (bias and bias.node == None): linear_op = TorchLinear() weight_size = weight.shape linear_op.set_param(linear_op.ParamName.WEIGHTS, weight) if bias is None: linear_op.set_config("bias", False) else: linear_op.set_config("bias", True) linear_op.set_param(linear_op.ParamName.BIAS, bias) linear_op.set_config('out_features', weight_size[1]) linear_op.set_config('in_features', weight_size[0]) addmm_weights = linear_op.params[linear_op.ParamName.WEIGHTS].data addmm_weights = addmm_weights.transpose(1,0) linear_op.set_param_from_data( linear_op.ParamName.WEIGHTS, addmm_weights) node.op = linear_op
null
23,701
import re import torch import collections from typing import List, Dict from dataclasses import dataclass class HandleListType(Schema): schemas: List[Schema] sizes: List[int] def __call__(self, values): values = self._split(values, self.sizes) if len(values) != len(self.schemas): raise ValueError( f"Values has length {len(values)} but schemas " f"has length {len(self.schemas)}!" ) values = [m(v) for m, v in zip(self.schemas, values)] return list(values) def flatten(cls, obj): res = [flatten_to_tuple(k) for k in obj] values, sizes = cls._concat([k[0] for k in res]) return values, cls([k[1] for k in res], sizes) class HandleTupleType(HandleListType): def __call__(self, values): return tuple(super().__call__(values)) class IdentitySchema(Schema): def __call__(self, values): try: return values[0] except: return values def flatten(cls, obj): return (obj,), cls() class HandleDictType(HandleListType): keys: List[str] def __call__(self, values): values = super().__call__(values) return dict(zip(self.keys, values)) def flatten(cls, obj): for k in obj.keys(): if not isinstance(k, str): raise KeyError("Only support flattening dictionaries if keys are str.") keys = sorted(obj.keys()) values = [obj[k] for k in keys] ret, schema = HandleListType.flatten(values) return ret, cls(schema.schemas, schema.sizes, keys) class HandleOrderDictType(HandleListType): keys: List[str] def __call__(self, values): values = super().__call__(values) return dict(zip(self.keys, values)) def flatten(cls, obj): for k in obj.keys(): if not isinstance(k, str): raise KeyError("Only support flattening dictionaries if keys are str.") keys = obj.keys() values = [obj[k] for k in keys] ret, schema = HandleListType.flatten(values) return ret, cls(schema.schemas, schema.sizes, keys) def flatten_to_tuple(data): support_list = [ ((str, bytes), IdentitySchema), (list, HandleListType), (tuple, HandleTupleType), (collections.OrderedDict, HandleOrderDictType), (collections.abc.Mapping, HandleDictType), ] for support_type, schema in support_list: if isinstance(data, support_type): handle = schema break else: handle = IdentitySchema return handle.flatten(data)
null
23,702
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper class DeviceType(AutoName): CPU = auto() DPU = auto() class DeviceInfo(object): def __init__(self, device_type): assert isinstance(device_type, DeviceType) self._type = device_type self._device_partition_check_msg = None def get_device_type(self): return self._type def set_filter_message(self, msg): self._device_partition_check_msg = msg def get_filter_message(self): return self._device_partition_check_msg def clear_filter_message(self): self._device_partition_check_msg = None def check_nonlinear(engine, node): op_nonlinear_map = { NNDCT_OP.CONV2D: [NNDCT_OP.RELU, NNDCT_OP.RELU6, NNDCT_OP.PRELU, NNDCT_OP.LEAKY_RELU, NNDCT_OP.HSWISH, NNDCT_OP.HSIGMOID], NNDCT_OP.CONVTRANSPOSE2D: [NNDCT_OP.RELU, NNDCT_OP.RELU6, NNDCT_OP.PRELU, NNDCT_OP.LEAKY_RELU], NNDCT_OP.DEPTHWISE_CONV2D: [NNDCT_OP.RELU, NNDCT_OP.RELU6, NNDCT_OP.PRELU, NNDCT_OP.LEAKY_RELU], NNDCT_OP.DEPTHWISE_CONVTRANSPOSE2D: [], } msg = "" nonlinear_types = op_nonlinear_map.get(node.op.type) if nonlinear_types is None: nonlinear_types = [] children_nodes = node.owning_graph.children(node) if len(children_nodes) == 1 and children_nodes[0].op.type in nonlinear_types: if children_nodes[0].op.type == NNDCT_OP.LEAKY_RELU: alpha = children_nodes[0].node_attr(children_nodes[0].op.AttrName.ALPHA) dpu_alpha = 26.0 / 256 if alpha != dpu_alpha: msg = f"Its alpa is {alpha}, but DPU only support {dpu_alpha}." return False, msg children_nodes[0].target_device = DeviceInfo(DeviceType.DPU) return True, msg
null
23,703
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper def check_kernel(kernels, kernel_limit): def check_stride(strides, stride_limit): def check_load_jump_write(ic, channel_parallel, dilation=None): def check_pad(pad, kernel): def check_conv_weights_bank_depth(target, engine, kernel_shape): class DPUTargetHelper(object): def get_basic_info(dpu_target): def get_full_info(dpu_target): def parse_range(num_range): def has_attr(message, member): def get_name(dpu_target): def get_type(dpu_target): def get_conv_engine(dpu_target): def get_alu_engine(dpu_target): def get_pool_engine(dpu_target): def get_eltwise_engine(dpu_target): def has_alu_engine(dpu_target): def has_pool_engine(dpu_target): def has_dwconv_engine(dpu_target): def has_eltwise_engine(dpu_target): def get_bank_group(dpu_target): def get_load_engine(dpu_target): def get_dwconv_engine(dpu_target): def filter_conv2d(node, target): msg = "" ksize = node.node_attr(node.op.AttrName.KERNEL) strides = node.node_attr(node.op.AttrName.STRIDE) dilation = node.node_attr(node.op.AttrName.DILATION) padding = node.node_attr(node.op.AttrName.PAD) conv_engine = DPUTargetHelper.get_conv_engine(target) channel_parallel = conv_engine.input_channel_parallel ic = node.in_tensors[0].shape[3] oc = node.out_tensors[0].shape[3] dilated_ksize = list(ksize) for i in range(len(dilated_ksize)): dilated_ksize[i] = (ksize[i] - 1) * dilation[i] + 1 kernel_limit = DPUTargetHelper.parse_range("1-16") if DPUTargetHelper.has_attr(conv_engine, "conv_limit") and conv_engine.conv_limit.kernel_size: kernel_limit = DPUTargetHelper.parse_range(conv_engine.conv_limit.kernel_size) ret, msg = check_kernel(ksize, kernel_limit) if not ret: return ret, msg ret, msg = check_conv_weights_bank_depth(target, conv_engine, node.op.get_param(node.op.ParamName.WEIGHTS).shape) if not ret: return ret, msg stride_limit = DPUTargetHelper.parse_range("1-4") if DPUTargetHelper.has_attr(conv_engine, "conv_limit") and conv_engine.conv_limit.stride: stride_limit = DPUTargetHelper.parse_range(conv_engine.conv_limit.stride) iw = node.in_tensors[0].shape[2] ih = node.in_tensors[0].shape[1] if iw != ksize[0] or ih != ksize[1]: ret, msg = check_stride(strides, stride_limit) if not ret: return ret, msg ret, msg = check_load_jump_write(ic, channel_parallel, dilation) if not ret: return ret, msg ret, msg = check_pad(padding, dilated_ksize) if not ret: return ret, msg # ret, msg = check_nonlinear(conv_engine, node) # if not ret: # return ret, msg return True, msg
null
23,704
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper def check_load_jump_write(ic, channel_parallel, dilation=None): msg = "" dilation = dilation if dilation is not None else [1, 1] cp_limit = 256 * channel_parallel if ic > cp_limit: msg = f"DPU only supports 'input_channel'({ic}) less than ({cp_limit})" return False, msg return True, msg def check_save_jump_read(oc, channel_parallel): msg = "" cp_limit = 256 * channel_parallel if oc > cp_limit: msg = f"DPU only support 'output_channel'({oc}) less than {cp_limit}" return False, msg return True, msg def check_pad(pad, kernel): msg = "" if any([p < 0 for p in pad]): msg = f"DPU only support non-negative 'pad'({pad})" return False, msg if pad[0] > kernel[0]: msg = f"DPU only supports 'pad_left'({pad[0]}) less than 'kernel_width'({kernel[0]})" return False, msg if pad[1] > kernel[0]: msg = f"DPU only supports 'pad_right'({pad[1]}) less than 'kernel_width'({kernel[0]})" return False, msg if pad[2] > kernel[1]: msg = f"DPU only supports 'pad_top'({pad[2]}) less than 'kernel_width'({kernel[1]})" return False, msg if pad[3] > kernel[1]: msg = f"DPU only supports 'pad_bottom'({pad[3]}) less than 'kernel_width'({kernel[1]})" return False, msg return True, msg def check_conv_weights_bank_depth(target, engine, kernel_shape): msg = "" weight_bank_name = engine.weight_bank bank_groups = DPUTargetHelper.get_bank_group(target) weights_bank = None for bank_group in bank_groups: if bank_group.name == weight_bank_name: weights_bank = bank_group break if weights_bank is None: msg = f"{target.get_name()}'s bank group configure is error, there's no weights bank for the engine." return False, msg output_channel_parallel = engine.output_channel_parallel k_oc, k_h, k_w, k_ic = kernel_shape weight_depth = k_w * k_h * math.ceil(k_ic * 1.0 / weights_bank.bank_width) * math.ceil(output_channel_parallel * 1.0 / weights_bank.bank_num) if weight_depth > weights_bank.bank_depth: msg = f"Weights({kernel_shape}) is too large to be loaded into parameter buffer. 'kernel_h * kernel_w * ⌈input_channel / weights_bank_width⌉ * ⌈output_channel_parallel / weights_bank_num⌉({weight_depth})' is supporsed to be less equal than {weights_bank.bank_depth}." return False, msg return True, msg def check_transposed_kernel(kernel, stride, limit): msg = "" if not (kernel // stride in limit and (kernel % stride == 0 or (kernel // stride + 1) in limit)): msg = f"'kernel / stride'({kernel} / {stride}) is not in DPU supported range{limit}." return False, msg return True, msg class DPUTargetHelper(object): def get_basic_info(dpu_target): return f"name: {dpu_target.name}\ntype: {dpu_target.type}\nisa_version: {dpu_target.isa_version}" def get_full_info(dpu_target): return dpu_target._legacy_dpu_target_def def parse_range(num_range): new_range = [] token = "," single_pattern = re.compile(r"""\s*(\d+)\s*""") range_pattern = re.compile(r"""\s*(\d+)\s*-\s*(\d+)\s*""") for num_item in num_range.split(token): if "-" in num_item: result = range_pattern.match(num_item) lower = int(result.group(1)) upper = int(result.group(2)) new_range.extend(list(range(lower, upper + 1))) else: result = single_pattern.match(num_item) num = int(result.group(1)) new_range.append(num) return new_range def has_attr(message, member): assert hasattr(message, "ByteSize") return hasattr(message, member) and getattr(message, member).ByteSize() > 0 def get_name(dpu_target): return dpu_target.get_name() def get_type(dpu_target): return dpu_target.get_type() def get_conv_engine(dpu_target): return dpu_target.get_conv_engine() def get_alu_engine(dpu_target): return dpu_target.get_alu_engine() def get_pool_engine(dpu_target): return dpu_target.get_pool_engine() def get_eltwise_engine(dpu_target): return dpu_target.get_eltwise_engine() def has_alu_engine(dpu_target): return hasattr(dpu_target, "alu_engine") and dpu_target.get_alu_engine().ByteSize() > 0 def has_pool_engine(dpu_target): return hasattr(dpu_target, "pool_engine") and dpu_target.get_pool_engine().ByteSize() > 0 def has_dwconv_engine(dpu_target): return hasattr(dpu_target, "dwconv_engine") and dpu_target.get_dwconv_engine().ByteSize() > 0 def has_eltwise_engine(dpu_target): return hasattr(dpu_target, "eltwise_engine") and dpu_target.get_eltwise_engine().ByteSize() > 0 def get_bank_group(dpu_target): return dpu_target.get_bank_group() def get_load_engine(dpu_target): return dpu_target.get_load_engine() def get_dwconv_engine(dpu_target): return dpu_target.get_dwconv_engine() def filter_transpose_conv2d(node, target): msg = "" ksize = node.node_attr(node.op.AttrName.KERNEL) strides = node.node_attr(node.op.AttrName.STRIDE) dilation = node.node_attr(node.op.AttrName.DILATION) padding = node.node_attr(node.op.AttrName.PAD) output_padding = node.node_config("output_padding") if any([pad != 0 for pad in output_padding]): msg = "DPU does not support output_padding." return False, msg conv_engine = DPUTargetHelper.get_conv_engine(target) channel_parallel = conv_engine.input_channel_parallel ic = node.in_tensors[0].shape[3] oc = node.out_tensors[0].shape[3] kernel_limit = DPUTargetHelper.parse_range("1-16") if DPUTargetHelper.has_attr(conv_engine, "conv_limit") and conv_engine.conv_limit.kernel_size: kernel_limit = DPUTargetHelper.parse_range(conv_engine.conv_limit.kernel_size) ret, msg = check_transposed_kernel(ksize[0], strides[0], kernel_limit) if not ret: return ret, msg ret, msg = check_transposed_kernel(ksize[1], strides[1], kernel_limit) if not ret: return ret, msg ret, msg = check_conv_weights_bank_depth(target, conv_engine, node.op.get_param(node.op.ParamName.WEIGHTS).shape) if not ret: return ret, msg ret, msg = check_load_jump_write(ic, channel_parallel, dilation) if not ret: return ret, msg ret, msg = check_save_jump_read(oc, channel_parallel) if not ret: return ret, msg ret, msg = check_pad(padding, ksize) if not ret: return ret, msg # ret, msg = check_nonlinear(conv_engine, node) # if not ret: # return ret, msg return True, msg
null
23,705
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper class DPUTargetHelper(object): def get_basic_info(dpu_target): return f"name: {dpu_target.name}\ntype: {dpu_target.type}\nisa_version: {dpu_target.isa_version}" def get_full_info(dpu_target): return dpu_target._legacy_dpu_target_def def parse_range(num_range): new_range = [] token = "," single_pattern = re.compile(r"""\s*(\d+)\s*""") range_pattern = re.compile(r"""\s*(\d+)\s*-\s*(\d+)\s*""") for num_item in num_range.split(token): if "-" in num_item: result = range_pattern.match(num_item) lower = int(result.group(1)) upper = int(result.group(2)) new_range.extend(list(range(lower, upper + 1))) else: result = single_pattern.match(num_item) num = int(result.group(1)) new_range.append(num) return new_range def has_attr(message, member): assert hasattr(message, "ByteSize") return hasattr(message, member) and getattr(message, member).ByteSize() > 0 def get_name(dpu_target): return dpu_target.get_name() def get_type(dpu_target): return dpu_target.get_type() def get_conv_engine(dpu_target): return dpu_target.get_conv_engine() def get_alu_engine(dpu_target): return dpu_target.get_alu_engine() def get_pool_engine(dpu_target): return dpu_target.get_pool_engine() def get_eltwise_engine(dpu_target): return dpu_target.get_eltwise_engine() def has_alu_engine(dpu_target): return hasattr(dpu_target, "alu_engine") and dpu_target.get_alu_engine().ByteSize() > 0 def has_pool_engine(dpu_target): return hasattr(dpu_target, "pool_engine") and dpu_target.get_pool_engine().ByteSize() > 0 def has_dwconv_engine(dpu_target): return hasattr(dpu_target, "dwconv_engine") and dpu_target.get_dwconv_engine().ByteSize() > 0 def has_eltwise_engine(dpu_target): return hasattr(dpu_target, "eltwise_engine") and dpu_target.get_eltwise_engine().ByteSize() > 0 def get_bank_group(dpu_target): return dpu_target.get_bank_group() def get_load_engine(dpu_target): return dpu_target.get_load_engine() def get_dwconv_engine(dpu_target): return dpu_target.get_dwconv_engine() def filter_conv3d(node, target): msg = "" if DPUTargetHelper.get_type(target) != "DPUCVDX8G": msg = f"{DPUTargetHelper.get_name(target)} does not support {node.op.type}. Only DPUCVDX8G support this." return False, msg return True, msg
null
23,706
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper class DPUTargetHelper(object): def get_basic_info(dpu_target): return f"name: {dpu_target.name}\ntype: {dpu_target.type}\nisa_version: {dpu_target.isa_version}" def get_full_info(dpu_target): return dpu_target._legacy_dpu_target_def def parse_range(num_range): new_range = [] token = "," single_pattern = re.compile(r"""\s*(\d+)\s*""") range_pattern = re.compile(r"""\s*(\d+)\s*-\s*(\d+)\s*""") for num_item in num_range.split(token): if "-" in num_item: result = range_pattern.match(num_item) lower = int(result.group(1)) upper = int(result.group(2)) new_range.extend(list(range(lower, upper + 1))) else: result = single_pattern.match(num_item) num = int(result.group(1)) new_range.append(num) return new_range def has_attr(message, member): assert hasattr(message, "ByteSize") return hasattr(message, member) and getattr(message, member).ByteSize() > 0 def get_name(dpu_target): return dpu_target.get_name() def get_type(dpu_target): return dpu_target.get_type() def get_conv_engine(dpu_target): return dpu_target.get_conv_engine() def get_alu_engine(dpu_target): return dpu_target.get_alu_engine() def get_pool_engine(dpu_target): return dpu_target.get_pool_engine() def get_eltwise_engine(dpu_target): return dpu_target.get_eltwise_engine() def has_alu_engine(dpu_target): return hasattr(dpu_target, "alu_engine") and dpu_target.get_alu_engine().ByteSize() > 0 def has_pool_engine(dpu_target): return hasattr(dpu_target, "pool_engine") and dpu_target.get_pool_engine().ByteSize() > 0 def has_dwconv_engine(dpu_target): return hasattr(dpu_target, "dwconv_engine") and dpu_target.get_dwconv_engine().ByteSize() > 0 def has_eltwise_engine(dpu_target): return hasattr(dpu_target, "eltwise_engine") and dpu_target.get_eltwise_engine().ByteSize() > 0 def get_bank_group(dpu_target): return dpu_target.get_bank_group() def get_load_engine(dpu_target): return dpu_target.get_load_engine() def get_dwconv_engine(dpu_target): return dpu_target.get_dwconv_engine() def filter_depthwise_conv3d(node, target): msg = "" if DPUTargetHelper.get_type(target) != "DPUCVDX8G": msg = f"{DPUTargetHelper.get_name(target)} does not support {node.op.type}. Only DPUCVDX8G support this." return False, msg return True, msg
null
23,707
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper class DPUTargetHelper(object): def get_basic_info(dpu_target): return f"name: {dpu_target.name}\ntype: {dpu_target.type}\nisa_version: {dpu_target.isa_version}" def get_full_info(dpu_target): return dpu_target._legacy_dpu_target_def def parse_range(num_range): new_range = [] token = "," single_pattern = re.compile(r"""\s*(\d+)\s*""") range_pattern = re.compile(r"""\s*(\d+)\s*-\s*(\d+)\s*""") for num_item in num_range.split(token): if "-" in num_item: result = range_pattern.match(num_item) lower = int(result.group(1)) upper = int(result.group(2)) new_range.extend(list(range(lower, upper + 1))) else: result = single_pattern.match(num_item) num = int(result.group(1)) new_range.append(num) return new_range def has_attr(message, member): assert hasattr(message, "ByteSize") return hasattr(message, member) and getattr(message, member).ByteSize() > 0 def get_name(dpu_target): return dpu_target.get_name() def get_type(dpu_target): return dpu_target.get_type() def get_conv_engine(dpu_target): return dpu_target.get_conv_engine() def get_alu_engine(dpu_target): return dpu_target.get_alu_engine() def get_pool_engine(dpu_target): return dpu_target.get_pool_engine() def get_eltwise_engine(dpu_target): return dpu_target.get_eltwise_engine() def has_alu_engine(dpu_target): return hasattr(dpu_target, "alu_engine") and dpu_target.get_alu_engine().ByteSize() > 0 def has_pool_engine(dpu_target): return hasattr(dpu_target, "pool_engine") and dpu_target.get_pool_engine().ByteSize() > 0 def has_dwconv_engine(dpu_target): return hasattr(dpu_target, "dwconv_engine") and dpu_target.get_dwconv_engine().ByteSize() > 0 def has_eltwise_engine(dpu_target): return hasattr(dpu_target, "eltwise_engine") and dpu_target.get_eltwise_engine().ByteSize() > 0 def get_bank_group(dpu_target): return dpu_target.get_bank_group() def get_load_engine(dpu_target): return dpu_target.get_load_engine() def get_dwconv_engine(dpu_target): return dpu_target.get_dwconv_engine() def filter_transpose_conv3d(node, target): msg = "" if DPUTargetHelper.get_type(target) != "DPUCVDX8G": msg = f"{DPUTargetHelper.get_name(target)} does not support {node.op.type}. Only DPUCVDX8G support this." return False, msg output_padding = node.node_config("output_padding") if any([pad != 0 for pad in output_padding]): msg = "DPU does not support output_padding." return False, msg return True, msg
null
23,708
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper class DPUTargetHelper(object): def get_basic_info(dpu_target): def get_full_info(dpu_target): def parse_range(num_range): def has_attr(message, member): def get_name(dpu_target): def get_type(dpu_target): def get_conv_engine(dpu_target): def get_alu_engine(dpu_target): def get_pool_engine(dpu_target): def get_eltwise_engine(dpu_target): def has_alu_engine(dpu_target): def has_pool_engine(dpu_target): def has_dwconv_engine(dpu_target): def has_eltwise_engine(dpu_target): def get_bank_group(dpu_target): def get_load_engine(dpu_target): def get_dwconv_engine(dpu_target): def filter_transpose_depthwise_conv3d(node, target): msg = "" if DPUTargetHelper.get_type(target) != "DPUCVDX8G": msg = f"{DPUTargetHelper.get_name(target)} does not support {node.op.type}. Only DPUCVDX8G support this." return False, msg output_padding = node.node_config("output_padding") if any([pad != 0 for pad in output_padding]): msg = "DPU does not support output_padding." return False, msg return True, msg
null
23,709
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper def check_kernel(kernels, kernel_limit): msg = "" if any([k not in kernel_limit for k in kernels]): msg = f"'kernel'({kernels[0]} x {kernels[1]}) is not in DPU supported range({kernel_limit})." return False, msg return True, msg def check_stride(strides, stride_limit): msg = "" if any([s not in stride_limit for s in stride_limit]): msg = f"'stride'({strides}) is not in DPU supported range({stride_limit})." return False, msg return True, msg def check_pad(pad, kernel): msg = "" if any([p < 0 for p in pad]): msg = f"DPU only support non-negative 'pad'({pad})" return False, msg if pad[0] > kernel[0]: msg = f"DPU only supports 'pad_left'({pad[0]}) less than 'kernel_width'({kernel[0]})" return False, msg if pad[1] > kernel[0]: msg = f"DPU only supports 'pad_right'({pad[1]}) less than 'kernel_width'({kernel[0]})" return False, msg if pad[2] > kernel[1]: msg = f"DPU only supports 'pad_top'({pad[2]}) less than 'kernel_width'({kernel[1]})" return False, msg if pad[3] > kernel[1]: msg = f"DPU only supports 'pad_bottom'({pad[3]}) less than 'kernel_width'({kernel[1]})" return False, msg return True, msg def check_pad_with_limit(pad, kernel, pad_limit): msg = "" if any([p < 0 for p in pad]): msg = f"DPU only support non-negative 'pad'({pad})" return False, msg pad_idx_kernel_map = { "pad_left": [0, 0], # [pad_idx, kernel_idx] "pad_right": [1, 0], "pad_top": [2, 1], "pad_bottom": [3, 1] } for key, pad_idx_kernel_idx in pad_idx_kernel_map.items(): pad_idx, kernel_idx = pad_idx_kernel_idx if pad_limit[key]: if pad[pad_idx] not in pad_limit[key]: msg = f"{key}({pad[pad_idx]}) is not in range." return False, msg else: if pad[pad_idx] > kernel[kernel_idx]: msg = f"DPU only supports {key}({pad[pad_idx]}) less than 'kernel'({kernel[kernel_idx]})." return False, msg return True, msg def check_pool_engine(target): msg = "" if not (DPUTargetHelper.has_pool_engine(target) or DPUTargetHelper.has_alu_engine(target)): msg = f"{DPUTargetHelper.get_name(target)} does not have pool-engine." return False, msg return True, msg class DPUTargetHelper(object): def get_basic_info(dpu_target): return f"name: {dpu_target.name}\ntype: {dpu_target.type}\nisa_version: {dpu_target.isa_version}" def get_full_info(dpu_target): return dpu_target._legacy_dpu_target_def def parse_range(num_range): new_range = [] token = "," single_pattern = re.compile(r"""\s*(\d+)\s*""") range_pattern = re.compile(r"""\s*(\d+)\s*-\s*(\d+)\s*""") for num_item in num_range.split(token): if "-" in num_item: result = range_pattern.match(num_item) lower = int(result.group(1)) upper = int(result.group(2)) new_range.extend(list(range(lower, upper + 1))) else: result = single_pattern.match(num_item) num = int(result.group(1)) new_range.append(num) return new_range def has_attr(message, member): assert hasattr(message, "ByteSize") return hasattr(message, member) and getattr(message, member).ByteSize() > 0 def get_name(dpu_target): return dpu_target.get_name() def get_type(dpu_target): return dpu_target.get_type() def get_conv_engine(dpu_target): return dpu_target.get_conv_engine() def get_alu_engine(dpu_target): return dpu_target.get_alu_engine() def get_pool_engine(dpu_target): return dpu_target.get_pool_engine() def get_eltwise_engine(dpu_target): return dpu_target.get_eltwise_engine() def has_alu_engine(dpu_target): return hasattr(dpu_target, "alu_engine") and dpu_target.get_alu_engine().ByteSize() > 0 def has_pool_engine(dpu_target): return hasattr(dpu_target, "pool_engine") and dpu_target.get_pool_engine().ByteSize() > 0 def has_dwconv_engine(dpu_target): return hasattr(dpu_target, "dwconv_engine") and dpu_target.get_dwconv_engine().ByteSize() > 0 def has_eltwise_engine(dpu_target): return hasattr(dpu_target, "eltwise_engine") and dpu_target.get_eltwise_engine().ByteSize() > 0 def get_bank_group(dpu_target): return dpu_target.get_bank_group() def get_load_engine(dpu_target): return dpu_target.get_load_engine() def get_dwconv_engine(dpu_target): return dpu_target.get_dwconv_engine() def filter_pool(node, target): msg = "" ret, msg = check_pool_engine(target) if not ret: return ret, msg avg_pool_type = [NNDCT_OP.AVG_POOL, NNDCT_OP.ADAPTIVEAVGPOOL2D] max_pool_type = [NNDCT_OP.MAX_POOL] ksize = node.node_attr(node.op.AttrName.KERNEL) strides = node.node_attr(node.op.AttrName.STRIDE) padding = node.node_attr(node.op.AttrName.PAD) if DPUTargetHelper.has_alu_engine(target): alu_engine = DPUTargetHelper.get_alu_engine(target) support_list = alu_engine.alu_type has_max = any([t == alu_engine.max_pool for t in support_list]) has_avg = any([t == alu_engine.avg_pool for t in support_list]) has_max_reduce = any([t == alu_engine.max_reduce for t in support_list]) else: pool_engine = DPUTargetHelper.get_pool_engine(target) support_list = pool_engine.pool_type has_max = any([t == pool_engine.max for t in support_list]) has_avg = any([t == pool_engine.avg for t in support_list]) has_max_reduce = any([t == pool_engine.max_reduce for t in support_list]) if not ((node.op.type in max_pool_type and (has_max or has_max_reduce)) or (node.op.type in avg_pool_type and has_avg)): msg = f"{DPUTargetHelper.get_name(target)} does not support {node.op.type}." return False, msg kernel_limit = DPUTargetHelper.parse_range("1-8") stride_limit = DPUTargetHelper.parse_range("1-8") pad_limit = {} if DPUTargetHelper.has_alu_engine(target): alu_engine = DPUTargetHelper.get_alu_engine(target) if DPUTargetHelper.has_attr(alu_engine, "alu_limit"): alu_limit = alu_engine.alu_limit if alu_limit.kernel_size: kernel_limit = DPUTargetHelper.parse_range(alu_limit.kernel_size) if alu_limit.stride: stride_limit = DPUTargetHelper.parse_range(alu_limit.stride) if DPUTargetHelper.has_attr(alu_engine, "pad_limit"): alu_pad_limit = alu_engine.pad_limit if alu_pad_limit.pad_left: pad_limit["pad_left"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_left) if alu_pad_limit.pad_right: pad_limit["pad_right"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_right) if alu_pad_limit.pad_top: pad_limit["pad_top"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_top) if alu_pad_limit.pad_bottom: pad_limit["pad_bottom"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_bottom) elif node.op.type in avg_pool_type: if ksize[0] != ksize[1]: msg = f"DPU only supports avgpool with square kernel, but this op has kernel {ksize[0]} x {ksize[1]}." return False, msg pool_engine = DPUTargetHelper.get_pool_engine(target) if DPUTargetHelper.has_attr(pool_engine, "avg_limit"): avg_limit = pool_engine.avg_limit if avg_limit.kernel_size: kernel_limit = DPUTargetHelper.parse_range(avg_limit.kernel_size) if avg_limit.stride: stride_limit = DPUTargetHelper.parse_range(avg_limit.stride) elif node.op.type in max_pool_type: pool_engine = DPUTargetHelper.get_pool_engine(target) if DPUTargetHelper.has_attr(pool_engine, "max_limit"): max_limit = pool_engine.max_limit if max_limit.kernel_size: kernel_limit = DPUTargetHelper.parse_range(max_limit.kernel_size) if max_limit.stride: stride_limit = DPUTargetHelper.parse_range(max_limit.stride) if node.op.type in max_pool_type and has_max_reduce and ksize[0] not in kernel_limit: if ksize[0] > 100: msg = f"'kernel_width'({ksize[0]}) is not in DPU supported range [1, 100]" return False, msg if ksize[1] < 1 or ksizse[1] > 2: msg = f"'kernel_height'({ksize[1]}) is not in DPU supported range [1, 2]" return False, msg else: ret, msg = check_kernel(ksize, kernel_limit) if not ret: return ret, msg iw = node.in_tensors[0].shape[2] ih = node.in_tensors[0].shape[1] if iw != ksize[0] or ih != ksize[1]: ret, msg = check_stride(strides, stride_limit) if not ret: return ret, msg if pad_limit: ret, msg = check_pad_with_limit(padding, ksize, pad_limit) if not ret: return ret, msg else: ret, msg = check_pad(padding, ksize) if not ret: return ret, msg return True, msg
null
23,710
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper def check_eltwise_engine(target): msg = "" if not DPUTargetHelper.has_eltwise_engine(target): msg = f"{DPUTargetHelper.get_name(target)} does not have eltwise-engine" return False, msg return True, msg def filter_depthwise_conv2d(node, target): msg = "" ret, msg = check_dwconv_engine(target) if not ret: return ret, msg ksize = node.node_attr(node.op.AttrName.KERNEL) strides = node.node_attr(node.op.AttrName.STRIDE) dilation = node.node_attr(node.op.AttrName.DILATION) padding = node.node_attr(node.op.AttrName.PAD) ic = node.in_tensors[0].shape[3] oc = node.out_tensors[0].shape[3] dilated_ksize = list(ksize) for i in range(len(dilated_ksize)): dilated_ksize[i] = (ksize[i] - 1) * dilation[i] + 1 kernel_limit = DPUTargetHelper.parse_range("1-16") stride_limit = DPUTargetHelper.parse_range("1-4") pad_limit = {} if DPUTargetHelper.has_alu_engine(target): alu_engine = DPUTargetHelper.get_alu_engine(target) channel_parallel = alu_engine.channel_parallel if DPUTargetHelper.has_attr(alu_engine, "alu_limit"): alu_limit = alu_engine.alu_limit if alu_limit.kernel_size: kernel_limit = DPUTargetHelper.parse_range(alu_limit.kernel_size) if alu_limit.stride: stride_limit = DPUTargetHelper.parse_range(alu_limit.stride) if DPUTargetHelper.has_attr(alu_engine, "pad_limit"): alu_pad_limit = alu_engine.pad_limit if alu_pad_limit.pad_left: pad_limit["pad_left"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_left) if alu_pad_limit.pad_right: pad_limit["pad_right"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_right) if alu_pad_limit.pad_top: pad_limit["pad_top"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_top) if alu_pad_limit.pad_bottom: pad_limit["pad_bottom"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_bottom) else: dwconv_engine = DPUTargetHelper.get_dwconv_engine(target) channel_parallel = dwconv_engine.channel_parallel if DPUTargetHelper.has_attr(dwconv_engine, "dwconv_limit"): dwconv_limit = dwconv_engine.dwconv_limit if dwconv_limit.kernel_size: kernel_limit = DPUTargetHelper.parse_range(dwconv_limit.kernel_size) if dwconv_limit.stride: stride_limit = DPUTargetHelper.parse_range(dwconv_limit.stride) if DPUTargetHelper.get_type(target) == "DPUCAHX8H": if strides[0] > ksize[0]: msg = f"The stride_w({strides[0]}) > kernel_w({ksize[0]}), but {DPUTargetHelper.get_name(target)} only support stride_w <= kernel_w." return False, msg ret, msg = check_kernel(ksize, kernel_limit) if not ret: return ret, msg iw = node.in_tensors[0].shape[2] ih = node.in_tensors[0].shape[1] if not(iw == ksize[0] and ih == ksize[1]): ret, msg = check_stride(strides, stride_limit) if not ret: return ret, msg ret, msg = check_load_jump_write(ic, channel_parallel, dilation) if not ret: return ret, msg ret, msg = check_save_jump_read(oc, channel_parallel) if not ret: return ret, msg if DPUTargetHelper.has_alu_engine(target): ret, msg = check_dwconv_weights_bank_depth(target, DPUTargetHelper.get_alu_engine(target), node.op.get_param(node.op.ParamName.WEIGHTS).shape) if not ret: return ret, msg else: ret, msg = check_dwconv_weights_bank_depth(target, DPUTargetHelper.get_dwconv_engine(target), node.op.get_param(node.op.ParamName.WEIGHTS).shape) if not ret: return ret, msg if pad_limit: ret, msg = check_pad_with_limit(padding, dilated_ksize, pad_limit) if not ret: return ret, msg else: ret, msg = check_pad(padding, dilated_ksize) if not ret: return ret, msg # if DPUTargetHelper.has_alu_engine(target): # ret, msg = check_nonlinear(DPUTargetHelper.get_alu_engine(target), node) # if not ret: # return ret, msg # else: # ret, msg = check_nonlinear(DPUTargetHelper.get_dwconv_engine(target), node) # if not ret: # return ret, msg return True, msg def check_dim_of_inputs_of_mul(node): msg = "" if_replaceabel = True input_shape = node.in_tensors[0].shape const_node = node.in_tensors[1].node const_data = const_node.node_attr(const_node.op.AttrName.DATA) if not isinstance(const_data, list): const_data = [const_data] const_data = np.array(const_data) const_shape = const_data.shape if_replaceabel = (len(input_shape) <= 4 and len(const_shape) == 1 and (input_shape[-1] == const_shape[0] or const_shape[0] == 1)) if not if_replaceabel: msg = f"mul's input has the tensor dimension {input_shape} and weights has the tensor dimenstion {const_shape}." return False, msg return True, msg def create_dwconv2d_from_mul(node): dwconv2d = TorchConv2d(NNDCT_OP.DEPTHWISE_CONV2D) dwconv2d.set_config('kernel_size', [1, 1]) dwconv2d.set_config('stride', [1, 1]) dwconv2d.set_config('padding', [0, 0]) input_channel = node.in_tensors[0].shape[-1] weight_tensor = Tensor("weight") weight_tensor.from_ndarray(np.random.randn(1, 1, 1, input_channel)) dwconv2d.set_param(dwconv2d.ParamName.WEIGHTS, weight_tensor) return dwconv2d class DPUTargetHelper(object): def get_basic_info(dpu_target): return f"name: {dpu_target.name}\ntype: {dpu_target.type}\nisa_version: {dpu_target.isa_version}" def get_full_info(dpu_target): return dpu_target._legacy_dpu_target_def def parse_range(num_range): new_range = [] token = "," single_pattern = re.compile(r"""\s*(\d+)\s*""") range_pattern = re.compile(r"""\s*(\d+)\s*-\s*(\d+)\s*""") for num_item in num_range.split(token): if "-" in num_item: result = range_pattern.match(num_item) lower = int(result.group(1)) upper = int(result.group(2)) new_range.extend(list(range(lower, upper + 1))) else: result = single_pattern.match(num_item) num = int(result.group(1)) new_range.append(num) return new_range def has_attr(message, member): assert hasattr(message, "ByteSize") return hasattr(message, member) and getattr(message, member).ByteSize() > 0 def get_name(dpu_target): return dpu_target.get_name() def get_type(dpu_target): return dpu_target.get_type() def get_conv_engine(dpu_target): return dpu_target.get_conv_engine() def get_alu_engine(dpu_target): return dpu_target.get_alu_engine() def get_pool_engine(dpu_target): return dpu_target.get_pool_engine() def get_eltwise_engine(dpu_target): return dpu_target.get_eltwise_engine() def has_alu_engine(dpu_target): return hasattr(dpu_target, "alu_engine") and dpu_target.get_alu_engine().ByteSize() > 0 def has_pool_engine(dpu_target): return hasattr(dpu_target, "pool_engine") and dpu_target.get_pool_engine().ByteSize() > 0 def has_dwconv_engine(dpu_target): return hasattr(dpu_target, "dwconv_engine") and dpu_target.get_dwconv_engine().ByteSize() > 0 def has_eltwise_engine(dpu_target): return hasattr(dpu_target, "eltwise_engine") and dpu_target.get_eltwise_engine().ByteSize() > 0 def get_bank_group(dpu_target): return dpu_target.get_bank_group() def get_load_engine(dpu_target): return dpu_target.get_load_engine() def get_dwconv_engine(dpu_target): return dpu_target.get_dwconv_engine() def filter_eltwise(node, target): msg = "" if node.op.type == NNDCT_OP.MULTIPLY and node.in_tensors[1].node.op.type in [NNDCT_OP.CONST, NNDCT_OP.TENSOR]: prefix_msg = "Try to convert mul to DepthwiseConv2d failed." ret, check_msg = check_dim_of_inputs_of_mul(node) if not ret: msg = prefix_msg + check_msg else: dwconv2d = create_dwconv2d_from_mul(node) mul_op = node.op node.op = dwconv2d input_tensor = node.in_tensors[0] out_tensor = node.out_tensors[0] old_shape = input_tensor.shape new_shape = list(old_shape) for i in range(4 - len(old_shape)): new_shape.insert(0, 1) input_tensor.shape = new_shape out_tensor.shape = new_shape ret, check_msg = filter_depthwise_conv2d(node, target) node.op = mul_op input_tensor.shape = old_shape out_tensor.shape = old_shape if not ret: msg = prefix_msg + check_msg else: return True, msg ret, check_msg = check_eltwise_engine(target) if not ret: return ret, msg + check_msg eltwise_engine = DPUTargetHelper.get_eltwise_engine(target) support_list = eltwise_engine.elew_type if node.op.type == NNDCT_OP.ADD: has_add = any([t == eltwise_engine.add for t in support_list]) if not has_add: msg = f"{DPUTargetHelper.get_name(target)} does not support eltwise ADD." return False, msg elif node.op.type == NNDCT_OP.MULTIPLY: has_mul = any([t == eltwise_engine.mult for t in support_list]) if not has_mul: msg += f"{DPUTargetHelper.get_name(target)} does not support eltwise MUL." return False, msg else: msg = f"{DPUTargetHelper.get_name(target)} does not support {node.op.type}." return False, msg return True, msg
null
23,711
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper def check_bilinear_upsample_scale(node): msg = "" input_shape = node.in_tensors[0].shape output_shape = node.out_tensors[0].shape i_h = input_shape[1] i_w = input_shape[2] o_h = output_shape[1] o_w = output_shape[2] scale_f = [1.0, 1.0] # [scale_w, scale_h] scale = [] scale_f[0] = float(o_w) / float(i_w) scale_f[1] = float(o_h) / float(i_h) half_pixel_centers = node.node_attr(node.op.AttrName.HALF_PIXEL_CENTERS) if half_pixel_centers: allowed_scale = [2, 4] else: allowed_scale = [2, 4, 8] for s_f in scale_f: if not (math.ceil(s_f) == s_f and math.floor(s_f) == s_f and any([s== s_f for s in allowed_scale])): msg = f"{node.op.type} output / input scale is {scale_f}" return False, msg scale.append(int(s_f)) if not all([scale[0] == s for s in scale]): msg = "scale_w is not equal with scale_h" return False, msg ret, msg = check_bilinear_upsample_fake_weight(node, scale[1], scale[0]) if not ret: return ret, msg node.set_node_attr(node.op.AttrName.SCALE, [float(scale[0]), float(scale[1])]) return True, msg def create_transpose_dwconv2d_from_bilinear_upsample(node): transpose_dwconv2d = TorchConvTranspose2d(NNDCT_OP.DEPTHWISE_CONVTRANSPOSE2D) scale_w, scale_h = node.node_attr(node.op.AttrName.SCALE) scale_w = int(scale_w) scale_h = int(scale_h) half_pixel_centers = node.node_attr(node.op.AttrName.HALF_PIXEL_CENTERS) input_shape = node.in_tensors[0].shape output_shape = node.out_tensors[0].shape kernel_h = 2 * scale_h kernel_w = 2 * scale_w transpose_dwconv2d.set_config('output_padding', [0, 0]) transpose_dwconv2d.set_config('kernel_size', [kernel_h, kernel_w]) transpose_dwconv2d.set_config('stride', [scale_h, scale_w]) input_w = input_shape[2] + 2 input_h = input_shape[1] + 2 if half_pixel_centers: pad_l = int(math.floor(float(scale_w) / 2.0 - 0.5)) pad_r = output_shape[2] + int(kernel_w) - 2 - (input_w - 1) * scale_w - pad_l pad_t = int(math.floor(float(scale_h) / 2.0 - 0.5)) pad_b = output_shape[1] + int(kernel_h) - 2 - (input_h - 1) * scale_h - pad_t else: pad_l = scale_w - 1 pad_r = scale_w - 1 pad_t = scale_h - 1 pad_b = scale_h - 1 padding = [int(kernel_w) - 1 - pad_l, int(kernel_w) - 1 - pad_r, int(kernel_h) - 1 - pad_t, int(kernel_h) - 1 - pad_b] transpose_dwconv2d.set_attr(transpose_dwconv2d.AttrName.PAD_MODE, 0) transpose_dwconv2d.set_attr(transpose_dwconv2d.AttrName.PAD, padding) return transpose_dwconv2d def check_load_jump_write(ic, channel_parallel, dilation=None): msg = "" dilation = dilation if dilation is not None else [1, 1] cp_limit = 256 * channel_parallel if ic > cp_limit: msg = f"DPU only supports 'input_channel'({ic}) less than ({cp_limit})" return False, msg return True, msg def filter_transpose_depthwise_conv2d(node, target): msg = "" ret, msg = check_dwconv_engine(target) if not ret: return ret, msg ksize = node.node_attr(node.op.AttrName.KERNEL) strides = node.node_attr(node.op.AttrName.STRIDE) padding = node.node_attr(node.op.AttrName.PAD) dilation = node.node_attr(node.op.AttrName.DILATION) ic = node.in_tensors[0].shape[3] oc = node.out_tensors[0].shape[3] output_padding = node.node_config("output_padding") if any([pad != 0 for pad in output_padding]): msg = "DPU does not support output_padding." return False, msg kernel_limit = DPUTargetHelper.parse_range("1-16") stride_limit = DPUTargetHelper.parse_range("1-4") pad_limit = {} if DPUTargetHelper.has_alu_engine(target): alu_engine = DPUTargetHelper.get_alu_engine(target) channel_parallel = alu_engine.channel_parallel if DPUTargetHelper.has_attr(alu_engine, "alu_limit"): alu_limit = alu_engine.alu_limit if alu_limit.kernel_size: kernel_limit = DPUTargetHelper.parse_range(alu_limit.kernel_size) if alu_limit.stride: stride_limit = DPUTargetHelper.parse_range(alu_limit.stride) if DPUTargetHelper.has_attr(alu_engine, "pad_limit"): alu_pad_limit = alu_engine.pad_limit if alu_pad_limit.pad_left: pad_limit["pad_left"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_left) if alu_pad_limit.pad_right: pad_limit["pad_right"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_right) if alu_pad_limit.pad_top: pad_limit["pad_top"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_top) if alu_pad_limit.pad_bottom: pad_limit["pad_bottom"] = DPUTargetHelper.parse_range(alu_pad_limit.pad_bottom) else: dwconv_engine = DPUTargetHelper.get_dwconv_engine(target) channel_parallel = dwconv_engine.channel_parallel if DPUTargetHelper.has_attr(dwconv_engine, "dwconv_limit"): dwconv_limit = dwconv_engine.dwconv_limit if dwconv_limit.kernel_size: kernel_limit = DPUTargetHelper.parse_range(dwconv_limit.kernel_size) if dwconv_limit.stride: stride_limit = DPUTargetHelper.parse_range(dwconv_limit.stride) ret, msg = check_transposed_kernel(ksize[0], strides[0], kernel_limit) if not ret: return ret, msg ret, msg = check_transposed_kernel(ksize[1], strides[1], kernel_limit) if not ret: return ret, msg ret, msg = check_stride([1, 1], stride_limit) if not ret: return ret, msg ret, msg = check_load_jump_write(ic, channel_parallel, dilation) if not ret: return ret, msg ret, msg = check_save_jump_read(oc, channel_parallel) if pad_limit: ret, msg = check_pad_with_limit(padding, ksize, pad_limit) if not ret: return ret, msg else: ret, msg = check_pad(padding, ksize) if not ret: return ret, msg # if DPUTargetHelper.has_alu_engine(target): # ret, msg = check_nonlinear(DPUTargetHelper.get_alu_engine(target), node) # if not ret: # return ret, msg # else: # ret, msg = check_nonlinear(DPUTargetHelper.get_dwconv_engine(target), node) # if not ret: # return ret, msg return True, msg class DPUTargetHelper(object): def get_basic_info(dpu_target): return f"name: {dpu_target.name}\ntype: {dpu_target.type}\nisa_version: {dpu_target.isa_version}" def get_full_info(dpu_target): return dpu_target._legacy_dpu_target_def def parse_range(num_range): new_range = [] token = "," single_pattern = re.compile(r"""\s*(\d+)\s*""") range_pattern = re.compile(r"""\s*(\d+)\s*-\s*(\d+)\s*""") for num_item in num_range.split(token): if "-" in num_item: result = range_pattern.match(num_item) lower = int(result.group(1)) upper = int(result.group(2)) new_range.extend(list(range(lower, upper + 1))) else: result = single_pattern.match(num_item) num = int(result.group(1)) new_range.append(num) return new_range def has_attr(message, member): assert hasattr(message, "ByteSize") return hasattr(message, member) and getattr(message, member).ByteSize() > 0 def get_name(dpu_target): return dpu_target.get_name() def get_type(dpu_target): return dpu_target.get_type() def get_conv_engine(dpu_target): return dpu_target.get_conv_engine() def get_alu_engine(dpu_target): return dpu_target.get_alu_engine() def get_pool_engine(dpu_target): return dpu_target.get_pool_engine() def get_eltwise_engine(dpu_target): return dpu_target.get_eltwise_engine() def has_alu_engine(dpu_target): return hasattr(dpu_target, "alu_engine") and dpu_target.get_alu_engine().ByteSize() > 0 def has_pool_engine(dpu_target): return hasattr(dpu_target, "pool_engine") and dpu_target.get_pool_engine().ByteSize() > 0 def has_dwconv_engine(dpu_target): return hasattr(dpu_target, "dwconv_engine") and dpu_target.get_dwconv_engine().ByteSize() > 0 def has_eltwise_engine(dpu_target): return hasattr(dpu_target, "eltwise_engine") and dpu_target.get_eltwise_engine().ByteSize() > 0 def get_bank_group(dpu_target): return dpu_target.get_bank_group() def get_load_engine(dpu_target): return dpu_target.get_load_engine() def get_dwconv_engine(dpu_target): return dpu_target.get_dwconv_engine() def filter_upsample(node, target): msg = "" if node.node_attr(node.op.AttrName.MODE) == "BILINEAR": prefix_msg = "Try to convert BlinearUpsamle2d to transpose depthwise conv2d failed." ret, check_msg = check_bilinear_upsample_scale(node) if not ret: msg = prefix_msg + check_msg else: transpose_dwconv2d = create_transpose_dwconv2d_from_bilinear_upsample(node) upsample = node.op node.op = transpose_dwconv2d ret, check_msg = filter_transpose_depthwise_conv2d(node, target) node.op = upsample if not ret: msg = prefix_msg + check_msg else: return True, msg align_corners = node.node_attr(node.op.AttrName.ALIGN_CORNERS) if align_corners: msg = "DPU does not support align_corners = True" return False, msg mode = node.node_attr(node.op.AttrName.MODE) if mode == "BILINEAR": msg += f"DPU does not support {mode} mode.(only support NEAREST mode)." return ret, msg load_engine = DPUTargetHelper.get_load_engine(target) channel_parallel = load_engine.channel_parallel ic = node.in_tensors[0].shape[3] ret, msg = check_load_jump_write(ic, channel_parallel) if not ret: return ret, msg return True, msg
null
23,712
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper class DeviceType(AutoName): CPU = auto() DPU = auto() class DPUTargetHelper(object): def get_basic_info(dpu_target): return f"name: {dpu_target.name}\ntype: {dpu_target.type}\nisa_version: {dpu_target.isa_version}" def get_full_info(dpu_target): return dpu_target._legacy_dpu_target_def def parse_range(num_range): new_range = [] token = "," single_pattern = re.compile(r"""\s*(\d+)\s*""") range_pattern = re.compile(r"""\s*(\d+)\s*-\s*(\d+)\s*""") for num_item in num_range.split(token): if "-" in num_item: result = range_pattern.match(num_item) lower = int(result.group(1)) upper = int(result.group(2)) new_range.extend(list(range(lower, upper + 1))) else: result = single_pattern.match(num_item) num = int(result.group(1)) new_range.append(num) return new_range def has_attr(message, member): assert hasattr(message, "ByteSize") return hasattr(message, member) and getattr(message, member).ByteSize() > 0 def get_name(dpu_target): return dpu_target.get_name() def get_type(dpu_target): return dpu_target.get_type() def get_conv_engine(dpu_target): return dpu_target.get_conv_engine() def get_alu_engine(dpu_target): return dpu_target.get_alu_engine() def get_pool_engine(dpu_target): return dpu_target.get_pool_engine() def get_eltwise_engine(dpu_target): return dpu_target.get_eltwise_engine() def has_alu_engine(dpu_target): return hasattr(dpu_target, "alu_engine") and dpu_target.get_alu_engine().ByteSize() > 0 def has_pool_engine(dpu_target): return hasattr(dpu_target, "pool_engine") and dpu_target.get_pool_engine().ByteSize() > 0 def has_dwconv_engine(dpu_target): return hasattr(dpu_target, "dwconv_engine") and dpu_target.get_dwconv_engine().ByteSize() > 0 def has_eltwise_engine(dpu_target): return hasattr(dpu_target, "eltwise_engine") and dpu_target.get_eltwise_engine().ByteSize() > 0 def get_bank_group(dpu_target): return dpu_target.get_bank_group() def get_load_engine(dpu_target): return dpu_target.get_load_engine() def get_dwconv_engine(dpu_target): return dpu_target.get_dwconv_engine() def filter_reshape(node, target): msg = "" if DPUTargetHelper.get_type(target) == "DPUCADF8H": return False, msg input_node = node.owning_graph.parents(node)[0] if not (input_node.target_device and input_node.target_device.get_device_type() == DeviceType.DPU): return False, msg return True, msg
null
23,713
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper def check_load_jump_write(ic, channel_parallel, dilation=None): msg = "" dilation = dilation if dilation is not None else [1, 1] cp_limit = 256 * channel_parallel if ic > cp_limit: msg = f"DPU only supports 'input_channel'({ic}) less than ({cp_limit})" return False, msg return True, msg class DPUTargetHelper(object): def get_basic_info(dpu_target): return f"name: {dpu_target.name}\ntype: {dpu_target.type}\nisa_version: {dpu_target.isa_version}" def get_full_info(dpu_target): return dpu_target._legacy_dpu_target_def def parse_range(num_range): new_range = [] token = "," single_pattern = re.compile(r"""\s*(\d+)\s*""") range_pattern = re.compile(r"""\s*(\d+)\s*-\s*(\d+)\s*""") for num_item in num_range.split(token): if "-" in num_item: result = range_pattern.match(num_item) lower = int(result.group(1)) upper = int(result.group(2)) new_range.extend(list(range(lower, upper + 1))) else: result = single_pattern.match(num_item) num = int(result.group(1)) new_range.append(num) return new_range def has_attr(message, member): assert hasattr(message, "ByteSize") return hasattr(message, member) and getattr(message, member).ByteSize() > 0 def get_name(dpu_target): return dpu_target.get_name() def get_type(dpu_target): return dpu_target.get_type() def get_conv_engine(dpu_target): return dpu_target.get_conv_engine() def get_alu_engine(dpu_target): return dpu_target.get_alu_engine() def get_pool_engine(dpu_target): return dpu_target.get_pool_engine() def get_eltwise_engine(dpu_target): return dpu_target.get_eltwise_engine() def has_alu_engine(dpu_target): return hasattr(dpu_target, "alu_engine") and dpu_target.get_alu_engine().ByteSize() > 0 def has_pool_engine(dpu_target): return hasattr(dpu_target, "pool_engine") and dpu_target.get_pool_engine().ByteSize() > 0 def has_dwconv_engine(dpu_target): return hasattr(dpu_target, "dwconv_engine") and dpu_target.get_dwconv_engine().ByteSize() > 0 def has_eltwise_engine(dpu_target): return hasattr(dpu_target, "eltwise_engine") and dpu_target.get_eltwise_engine().ByteSize() > 0 def get_bank_group(dpu_target): return dpu_target.get_bank_group() def get_load_engine(dpu_target): return dpu_target.get_load_engine() def get_dwconv_engine(dpu_target): return dpu_target.get_dwconv_engine() def filter_pad(node, target): msg = "" mode = node.node_attr(node.op.AttrName.MODE) if mode not in [0, 2]: #DPU only support CONSTANT / SYMMETRIC mode msg = f"DPU only support CONSTANT or SYMMETRIC mode." return False, msg load_engine = DPUTargetHelper.get_load_engine(target) channel_parallel = load_engine.channel_parallel ic = node.in_tensors[0].shape[3] ret, msg = check_load_jump_write(ic, channel_parallel) if not ret: return ret, msg return True, msg
null
23,714
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper class DPUTargetHelper(object): def get_basic_info(dpu_target): return f"name: {dpu_target.name}\ntype: {dpu_target.type}\nisa_version: {dpu_target.isa_version}" def get_full_info(dpu_target): return dpu_target._legacy_dpu_target_def def parse_range(num_range): new_range = [] token = "," single_pattern = re.compile(r"""\s*(\d+)\s*""") range_pattern = re.compile(r"""\s*(\d+)\s*-\s*(\d+)\s*""") for num_item in num_range.split(token): if "-" in num_item: result = range_pattern.match(num_item) lower = int(result.group(1)) upper = int(result.group(2)) new_range.extend(list(range(lower, upper + 1))) else: result = single_pattern.match(num_item) num = int(result.group(1)) new_range.append(num) return new_range def has_attr(message, member): assert hasattr(message, "ByteSize") return hasattr(message, member) and getattr(message, member).ByteSize() > 0 def get_name(dpu_target): return dpu_target.get_name() def get_type(dpu_target): return dpu_target.get_type() def get_conv_engine(dpu_target): return dpu_target.get_conv_engine() def get_alu_engine(dpu_target): return dpu_target.get_alu_engine() def get_pool_engine(dpu_target): return dpu_target.get_pool_engine() def get_eltwise_engine(dpu_target): return dpu_target.get_eltwise_engine() def has_alu_engine(dpu_target): return hasattr(dpu_target, "alu_engine") and dpu_target.get_alu_engine().ByteSize() > 0 def has_pool_engine(dpu_target): return hasattr(dpu_target, "pool_engine") and dpu_target.get_pool_engine().ByteSize() > 0 def has_dwconv_engine(dpu_target): return hasattr(dpu_target, "dwconv_engine") and dpu_target.get_dwconv_engine().ByteSize() > 0 def has_eltwise_engine(dpu_target): return hasattr(dpu_target, "eltwise_engine") and dpu_target.get_eltwise_engine().ByteSize() > 0 def get_bank_group(dpu_target): return dpu_target.get_bank_group() def get_load_engine(dpu_target): return dpu_target.get_load_engine() def get_dwconv_engine(dpu_target): return dpu_target.get_dwconv_engine() def filter_hard_sigmoid(node, target): msg = "" if not DPUTargetHelper.has_alu_engine(target): msg = "This target does not support single hard-sigmoid." return False, msg return True, msg
null
23,715
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper def filter_leaky_relu(node, target): msg = "" alpha = node.node_attr(node.op.AttrName.ALPHA) dpu_alpha = 26.0 / 256 if alpha != dpu_alpha: msg = f"Its alpa is {alpha}, but DPU only support {dpu_alpha}." return False, msg return True, msg
null
23,716
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper class DeviceType(AutoName): CPU = auto() DPU = auto() class DeviceInfo(object): def __init__(self, device_type): assert isinstance(device_type, DeviceType) self._type = device_type self._device_partition_check_msg = None def get_device_type(self): return self._type def set_filter_message(self, msg): self._device_partition_check_msg = msg def get_filter_message(self): return self._device_partition_check_msg def clear_filter_message(self): self._device_partition_check_msg = None def merge_permute_to_matmul(graph, target): def handler(*args, **kwargs): _, node_set = args permute_node = node_set[0] dense_node = node_set[-1] if permute_node.target_device and permute_node.target_device.get_device_type() == DeviceType.DPU: return if permute_node.node_attr(permute_node.op.AttrName.ORDER) == [0, 3, 1, 2] and dense_node.target_device: permute_node.target_device = DeviceInfo(dense_node.target_device.get_device_type()) permute_node.target_device.clear_filter_message() graph_searcher = GraphSearcher(graph) _ = graph_searcher.find_nodes_from_type( [PatternType(pattern=[NNDCT_OP.PERMUTE, NNDCT_OP.FLATTEN, NNDCT_OP.DENSE], action=handler), PatternType(pattern=[NNDCT_OP.PERMUTE, NNDCT_OP.RESHAPE, NNDCT_OP.DENSE], action=handler), ])
null
23,717
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper def filter_concat(node, target): ret = True msg = "" if any([not pn.target_device or pn.target_device.get_device_type() == DeviceType.CPU for pn in node.owning_graph.parents(node)]): msg += "The input of concat is not in DPU subgraph." dimension = node.out_tensors[0].ndim if dimension != 4: msg += "And output dimension is not 4." ret = False else: if node.node_attr(node.op.AttrName.AXIS) != 3: msg += "And it's not a channel-wise concatenation." ret = False if DPUTargetHelper.get_name(target) == "DPUCADF8H": dimension = node.out_tensors[0].ndim if dimension != 4: msg += "Output dimension is not 4." ret = False else: if node.node_attr(node.op.AttrName.AXIS) != 3: msg += "It's not a channel-wise concatenation." ret = False return ret, msg class DeviceType(AutoName): CPU = auto() DPU = auto() class DeviceInfo(object): def __init__(self, device_type): assert isinstance(device_type, DeviceType) self._type = device_type self._device_partition_check_msg = None def get_device_type(self): return self._type def set_filter_message(self, msg): self._device_partition_check_msg = msg def get_filter_message(self): return self._device_partition_check_msg def clear_filter_message(self): self._device_partition_check_msg = None def filter_dpu_interface_concat(graph, target): def handler(*args, **kwargs): _, node_set = args concat = node_set[0] if concat.target_device and concat.target_device.get_device_type() == DeviceType.DPU: return ret, msg = filter_concat(concat, target) if ret: concat.target_device = DeviceInfo(DeviceType.DPU) concat.target_device.clear_filter_message() else: concat.target_device = DeviceInfo(DeviceType.CPU) concat.target_device.set_filter_message(msg) graph_searcher = GraphSearcher(graph) _ = graph_searcher.find_nodes_from_type( [PatternType(pattern=[NNDCT_OP.CONCAT], action=handler)])
null
23,718
import math import numpy as np from nndct_shared.base import NNDCT_OP from nndct_shared.utils import PatternType from nndct_shared.nndct_graph import GraphSearcher, Tensor from pytorch_nndct.parse.torch_op_def import TorchConvTranspose2d, TorchConv2d from .device import DeviceInfo, DeviceType from .target_helper import DPUTargetHelper class DeviceType(AutoName): CPU = auto() DPU = auto() class DeviceInfo(object): def __init__(self, device_type): assert isinstance(device_type, DeviceType) self._type = device_type self._device_partition_check_msg = None def get_device_type(self): return self._type def set_filter_message(self, msg): self._device_partition_check_msg = msg def get_filter_message(self): return self._device_partition_check_msg def clear_filter_message(self): self._device_partition_check_msg = None def filter_dpu_interface_reshape(graph, target): def handler(*args, **kwargs): _, node_set = args reshape = node_set[0] if not (reshape.target_device and reshape.target_device.get_device_type() == DeviceType.DPU): return input_node = reshape.owning_graph.parents(reshape)[0] if input_node.target_device and input_node.target_device.get_device_type() == DeviceType.DPU \ and all([cn.target_device and cn.target_device.get_device_type() == DeviceType.DPU for cn in reshape.owning_graph.children(reshape)]): # an internal reshape, do nothing pass else: if input_node.out_tensors[0].shape[0] != reshape.out_tensors[0].shape[0]: msg = "First dimension is changed." reshape.target_device = DeviceInfo(DeviceType.CPU) reshape.target_device.set_filter_message(msg) graph_searcher = GraphSearcher(graph) _ = graph_searcher.find_nodes_from_type( [PatternType(pattern=[NNDCT_OP.RESHAPE], action=handler), PatternType(pattern=[NNDCT_OP.FLATTEN], action=handler), ])
null
23,719
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape _OP_CONVERTER_DICT = {} def create_op(gm, op, target, args, kwargs): def get_new_target(): if op == "call_module": return gm.get_submodule(target).__class__ elif op == "call_method": return getattr(torch.Tensor, target) else: return target convert_info = _OP_CONVERTER_DICT.get((op, get_new_target()), _OP_CONVERTER_DICT[(op, "")]) normalized_kwargs = kwargs if convert_info.args_kwargs_mapping_fn: if op == "call_module": mod = gm.get_submodule(target) normalized_kwargs = convert_info.args_kwargs_mapping_fn(target, mod, args, kwargs) return convert_info.convert_fn(*args, **normalized_kwargs)
null
23,720
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape _OP_CONVERTER_DICT = {} OpConvertInfo = namedtuple("OpConvertFnInfo", ["convert_fn", "args_kwargs_mapping_fn"]) def register_nndct_op(op, target, args_kwargs_mapping_fn=None): def register(op_convert_func): op_and_target = (op, target) assert op_and_target not in _OP_CONVERTER_DICT.keys() _OP_CONVERTER_DICT[op_and_target] = OpConvertInfo(op_convert_func, args_kwargs_mapping_fn) return op_convert_func return register
null
23,721
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def input_arg(*args, **kwargs): op = TorchBaseOperation(NNDCT_OP.INPUT) return op
null
23,722
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def call_method(*args, **kwargs): op = TorchBaseOperation(NNDCT_OP.CALL_METHOD) return op
null
23,723
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def call_module(*args, **kwargs): op = TorchBaseOperation(NNDCT_OP.CALL_MODULE) return op
null
23,724
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def call_function(*args, **kwargs): op = TorchBaseOperation(NNDCT_OP.CALL_FUNCTION) return op
null
23,725
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def output(*args, **kwargs): op = TorchBaseOperation(NNDCT_OP.RETURN) return op
null
23,726
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def int2tuple(obj, tuple_size): def convert_real_tensor(name, real_tensor): def conv_nd_mapping_fn(mod_qual_name, mod, args, kwargs): weight_tensor = convert_real_tensor(".".join([mod_qual_name, "weight"]), mod.weight) bias_tensor = convert_real_tensor(".".join([mod_qual_name, "bias"]), mod.bias) nd = weight_tensor.ndim - 2 stride = int2tuple(mod.stride, nd) padding = int2tuple(mod.padding, nd) dilation = int2tuple(mod.dilation, nd) groups = mod.groups return {"input": kwargs["input"], "weight": weight_tensor, "bias": bias_tensor, "stride": stride, "padding": padding, "dilation": dilation, "groups": groups }
null
23,727
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def _convolution(input, weight, bias, stride, padding, dilation, transposed, output_padding, groups): weight_size = weight.shape if transposed: weight_size[0], weight_size[1] = weight_size[1], weight_size[0] if weight_size[0] == 1 and groups == weight_size[1]: if weight.ndim == 4: op = TorchConvTranspose2d(NNDCT_OP.DEPTHWISE_CONVTRANSPOSE2D) elif weight.ndim == 5: op = TorchConvTranspose3d(NNDCT_OP.DEPTHWISE_CONVTRANSPOSE3D) elif weight.ndim == 3: raise NotImplementedError("Depthwise_ConvTranpose1D is unsupported") else: if weight.ndim == 4: op = TorchConvTranspose2d(NNDCT_OP.CONVTRANSPOSE2D) elif weight.ndim == 5: op = TorchConvTranspose3d(NNDCT_OP.CONVTRANSPOSE3D) elif weight.ndim == 3: raise NotImplementedError("ConvTranpose1D is unsupported") op.set_config("output_padding", list(output_padding)) op.set_config('in_channels', weight_size[1]) op.set_config('out_channels', weight_size[0] * groups) else: if weight_size[1] == 1 and groups == weight_size[0]: if weight.ndim == 4: op = TorchConv2d(NNDCT_OP.DEPTHWISE_CONV2D) elif weight.ndim == 5: op = TorchConv3d(NNDCT_OP.DEPTHWISE_CONV3D) elif weight.ndim == 3: op = TorchConv1d(NNDCT_OP.DEPTHWISE_CONV1D) else: if weight.ndim == 4: op = TorchConv2d(NNDCT_OP.CONV2D) elif weight.ndim == 5: op = TorchConv3d(NNDCT_OP.CONV3D) elif weight.ndim == 3: op = TorchConv1d(NNDCT_OP.CONV1D) op.set_config('in_channels', weight_size[1] * groups) op.set_config('out_channels', weight_size[0]) # Should add weight first op.set_param(op.ParamName.WEIGHTS, weight) if bias is not None: op.set_config('bias', True) op.set_param(op.ParamName.BIAS, bias) else: op.set_config('bias', False) op.set_config('dilation', list(dilation)) op.set_config('kernel_size', list(weight_size[2:])) op.set_config('stride', list(stride)) op.set_config('groups', groups) op.set_config('padding', list(padding)) return op def conv2d(*, input, weight, bias, stride, padding, dilation, groups): return _convolution(input, weight, bias, stride, padding, dilation, False, None, groups)
null
23,728
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def relu(*, input, inplace): # s op = TorchReLU() op.set_config('inplace', inplace) return op
null
23,729
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def convert_real_tensor(name, real_tensor): if real_tensor is None: return tensor = Tensor( name=name, shape=convert_shape(real_tensor.shape), dtype=convert_dtype(real_tensor.dtype), data = real_tensor.cpu().detach().numpy(), # device=get_tensor_meta_info(input_meta, "device"), requires_grad=real_tensor.requires_grad) return tensor def linear_mapping_fn(mod_qual_name, mod, args, kwargs): weight_tensor = convert_real_tensor(".".join([mod_qual_name, "weight"]), mod.weight) bias_tensor = convert_real_tensor(".".join([mod_qual_name, "bias"]), mod.bias) return {"input": kwargs["input"], "weight": weight_tensor, "bias": bias_tensor }
null
23,730
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def linear(*, input, weight, bias): op = TorchLinear() weight_size = weight.shape op.set_param(op.ParamName.WEIGHTS, weight) if bias is None: op.set_config("bias", False) else: op.set_config("bias", True) op.set_param(op.ParamName.BIAS, bias) op.set_config('out_features', weight_size[0]) op.set_config('in_features', weight_size[1]) return op
null
23,731
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def flatten(*, input, start_dim, end_dim): op = TorchFlatten() op.set_config('input', input) op.set_config("start_dim", start_dim) op.set_config("end_dim", end_dim) return op
null
23,732
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def adaptive_avg_pool2d(*, input, output_size): op = TorchAdaptiveAvgPool() op.set_config("output_size", output_size) return op
null
23,733
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def int2tuple(obj, tuple_size): if isinstance(obj, numbers.Integral): return (obj, ) * tuple_size else: return obj def maxpool2d(*, input, kernel_size, stride=None, padding=0, dilation=1, ceil_mode=False, return_indices=False): op = TorchMaxPool() op.set_config("return_indices", return_indices) op.set_config('kernel_size', list(int2tuple(kernel_size, 2))) if not stride: op.set_config('stride', list(int2tuple(kernel_size, 2))) else: op.set_config('stride', list(int2tuple(stride, 2))) if ceil_mode: op.set_config('ceil_mode', True) else: op.set_config('ceil_mode', False) op.set_config('padding', list(int2tuple(padding, 2))) op.set_config('dilation', list(int2tuple(dilation, 2))) return op
null
23,734
import numbers from collections import namedtuple import torch from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Tensor from pytorch_nndct.parse.torch_op_def import * from pytorch_nndct.fx.translator_utils import convert_dtype, convert_shape def add(*, input, other, alpha): op = TorchAdd() op.set_config('input', input) op.set_config('other', other) op.set_config('alpha', alpha) return op
null
23,735
import torch from collections import OrderedDict from nndct_shared.nndct_graph import Graph, Block, Node, Tensor from pytorch_nndct.fx.convert_op import create_op from pytorch_nndct.fx.translator_utils import convert_dtype, get_meta_info, convert_shape import nndct_shared.utils.tensor_util as tensor_util from nndct_shared.base.key_names import FrameworkType def short_name(gm, node: torch.fx.Node): if node.op == "call_function": return node.target.__name__ elif node.op in ["placeholder", "call_method"]: return node.target elif node.op == "call_module": return gm.get_submodule(node.target).__class__.__name__ elif node.op == "get_attr": return node.target elif node.op == "output": return "output" raise AssertionError(node.op)
null
23,736
import torch from collections import OrderedDict from nndct_shared.nndct_graph import Graph, Block, Node, Tensor from pytorch_nndct.fx.convert_op import create_op from pytorch_nndct.fx.translator_utils import convert_dtype, get_meta_info, convert_shape import nndct_shared.utils.tensor_util as tensor_util from nndct_shared.base.key_names import FrameworkType def get_meta_info(meta, info_type): return getattr(meta, info_type) def convert_dtype(dtype): r"""convert torch dtype to nndct dtype""" return { 'torch.float': 'float32', 'torch.float32': 'float32', 'torch.double': 'float64', 'torch.int': 'int32', 'torch.long': 'int64' }.get(dtype.__str__(), dtype) def convert_shape(shape): return list(shape) def make_tensor(name, meta): tensor = Tensor( name=name, shape=convert_shape(get_meta_info(meta, "shape")), dtype=convert_dtype(get_meta_info(meta, "dtype")), # device=get_tensor_meta_info(input_meta, "device"), requires_grad=get_meta_info(meta, "requires_grad")) return tensor
null
23,737
import torch from collections import OrderedDict from nndct_shared.nndct_graph import Graph, Block, Node, Tensor from pytorch_nndct.fx.convert_op import create_op from pytorch_nndct.fx.translator_utils import convert_dtype, get_meta_info, convert_shape import nndct_shared.utils.tensor_util as tensor_util from nndct_shared.base.key_names import FrameworkType class FrameworkType(object): # Frontend types TORCH = 'torch' CAFFE = 'caffe' TENSORFLOW = 'tensorflow' TF_KERAS = 'tf_keras' # NNDCT as a bridge NNDCT = 'nndct' def convert_op_param(op): for param_name, tensor in op.params.items(): tensor = tensor_util.convert_parameter_tensor_format( tensor, FrameworkType.TORCH, FrameworkType.NNDCT)
null
23,738
import operator import itertools import torch from pytorch_nndct.fx.optimization.utils import replace_node def always_true(*args, **kwargs): return True
null
23,739
import copy import os from typing import Any, Optional, Sequence, Union, List, Dict, Tuple import torch from nndct_shared.utils import NndctScreenLogger, NndctOption, QError, QWarning, QNote from .qproc import TorchQuantProcessor from .qproc import base as qp from .qproc import LSTMTorchQuantProcessor, RNNQuantProcessor from .qproc import vaiq_system_info from .quantization import QatProcessor from .parse.rich_in_out_helper import StandardInputData def dump_xmodel(output_dir="quantize_result", deploy_check=False, app_deploy="CV"): if app_deploy == "CV": lstm_app = False elif app_deploy == "NLP": lstm_app = True NndctScreenLogger().warning(f"The function dump_xmodel() will retire in future version. Use torch_quantizer.export_xmodel() reversely.") qp.dump_xmodel(output_dir, deploy_check, lstm_app)
null
23,740
from nndct_shared.utils import set_option_value, NndctOption from pytorch_nndct.qproc.utils import (get_deploy_graph_list, prepare_quantizable_module, register_output_hook, set_outputs_recorder_status, to_device) def register_output_hook(module: torch.nn.Module, record_once: bool = True) -> NoReturn: def set_outputs_recorder_status(module, turn_on) -> NoReturn: def prepare_quantizable_module( module: torch.nn.Module, input_args, export_folder: str, state_dict_file: Optional[str] = None, quant_mode: int = 1, device: "torch.device" = torch.device("cuda"), connect_qm_with_graph=True) -> Tuple[torch.nn.Module, Graph]: def get_deploy_graph_list(quant_model, nndct_graph, need_partition=True): def prepare_deployable_graph(module, input_args, device, output_dir): module, input_args = to_device(module, input_args, device) quant_module, graph = prepare_quantizable_module( module=module, input_args=input_args, export_folder=output_dir, device=device) if len(graph.all_blocks()) > 1: quant_module.from_script(True) else: quant_module.from_script(False) quant_off_stat = NndctOption.nndct_quant_off.value param_corr_stat = NndctOption.nndct_param_corr.value set_option_value("nndct_quant_off", True) set_option_value("nndct_param_corr", False) register_output_hook(quant_module, record_once=True) set_outputs_recorder_status(quant_module, True) quant_module.eval() if isinstance(input_args, tuple): _ = quant_module.to(device)(*input_args) else: _ = quant_module.to(device)(input_args) deploy_graphs, dev_graph = get_deploy_graph_list(quant_module, graph, need_partition=False) set_option_value("nndct_quant_off", quant_off_stat) set_option_value("nndct_param_corr", param_corr_stat) return dev_graph, deploy_graphs
null
23,741
from nndct_shared.utils import set_option_value, NndctOption from pytorch_nndct.qproc.utils import (get_deploy_graph_list, prepare_quantizable_module, register_output_hook, set_outputs_recorder_status, to_device) def assign_device_to_node(node_device_map): for node, device in node_device_map.items(): node.target_device = device
null
23,742
import importlib.util import os import sys from typing import Any, NoReturn, Optional, Sequence, Tuple, Union, List import itertools import torch import torch.nn as nn import torch.nn.functional as F import nndct_shared.utils as nndct_utils import pytorch_nndct.utils.jit_utils as jit_utils import pytorch_nndct.utils.module_util as module_util from nndct_shared.base import FrameworkType from nndct_shared.compile import DevGraphOptimizer from nndct_shared.nndct_graph import Graph, convert_block_node_to_graph from nndct_shared.optimization import QuantOptimizer from nndct_shared.utils import (GLOBAL_MAP, NNDCT_KEYS, NNDCT_OP, NndctDebugLogger, NndctOption, NndctScreenLogger, permute_data, permute_axes, QError, QWarning, QNote) from nndct_shared.utils.dpu_utils import get_avgpool_dpu_coeff from pytorch_nndct.export import get_script_writer from pytorch_nndct.nn import stacked_lstm from pytorch_nndct.nn.modules import channel_scale, reluk from pytorch_nndct.parse import TorchParser from pytorch_nndct.utils import TorchSymbol, TorchGraphSymbol from pytorch_nndct.utils.module_util import to_device, get_module_name from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version from .ModuleHooker import ModuleHooker class ModuleHooker(object): def name_modules(module): def add_outputs(module, record_once): def add_output_intime(module): def add_input_dump_time(module): def apply_to_children(module, func): def turn_on_record_outputs(module): def turn_off_record_outputs(module): def clear_record_outputs(module): def turn_on_output_intime(module): def turn_off_output_intime(module): def clear_op_called_times(module): def turn_on_input_dump(module): def turn_off_input_dump(module): def clear_input_called_times(module): def register_state_dict_hook(cls, module): def _resume_state_dict_key(op, destination, prefix, local_meta_data): def _state_dict_hooker(op, hooker): def register_output_hook(cls, module, record_once=True): def _record_outputs(op, inputs, outputs): def get_outputs_value(outputs): def _output_hooker(op, record_func): def register_input_dump_hook(cls, module): def dump_input_data(module, inputs): def inputs_flatten(tensors, tensors_flatten): def register_output_intime_hook(cls, module): def _dump_intime(op, inputs, outputs): def get_outputs_value(outputs): def dump_output_data(op, output_data, index=None): def dump_output(op, output_data): def _output_hooker(op, output_func): def detach_node_from_module(cls, module): def _del_node(op): def hook_module_with_node(cls, module, graph): def _add_node_on_module(op): def update_parameters(cls, module, graph, graph2module): def safe_torch_nn_Parameter(torch_tensor, requires_grad): def _graph2module(op): def _module2graph(op): def _get_output_data(outptus, outptus_name): def _get_output_shape(outputs, outputs_name): def update_blobs_once(cls, module, graph=None, time_step=None, update_shape_only=False): def _updata_tensor_data(node, nndct_tensor, torch_tensor): def _updata_tensor_shape(node, nndct_tensor, torch_tensor): def _update_node_outputs(op): def clone_quant_module(cls, quant_module, quant_graph): def hook_module_with_quantizer(cls, module, quantizer): def _add_quantizer_on_module(op): def hook_module_with_input_device_checker(cls, module, module_gen_from_script): def _check_input_args(op, input): def register_tensor_dtype_and_shape_hook(cls, module): def _record_dtype_and_shape(op, inputs, outputs): def _get_output_dtype_and_shape(output): def _hooker(op, record_func): def disconnect_modeule_with_graph(module): ModuleHooker.detach_node_from_module(module)
null