From 74feae9bb9023f50b8557812c234b9cb3eed083d Mon Sep 17 00:00:00 2001 From: Macrobull Date: Mon, 8 Jul 2019 14:48:01 +0800 Subject: [PATCH] force embeddable checking --- README.md | 2 - onnx2fluid/onnx2fluid/symbolic.py | 169 ++++++++++++++++-------------- 2 files changed, 93 insertions(+), 78 deletions(-) diff --git a/README.md b/README.md index 43bc92d..fadfbdd 100644 --- a/README.md +++ b/README.md @@ -8,8 +8,6 @@ X2Paddle支持将Caffe和TensorFlow模型转至PaddlePaddle模型，同时我们 任何使用问题均可通过[ISSUE](https://github.com/PaddlePaddle/X2Paddle/issues)的方式及时反馈，或者也可直接通过pull request的方式一起更新代码和文档。 -> **目前X2Paddle主要支持CV部分模型，对于NLP模型暂未支持。** - ## [caffe2fluid](caffe2fluid) 1. 支持将Caffe模型转至PaddlePaddle fluid可加载预测模型 2. 提供Caffe-PaddlePaddle常用API的对比文档[[doc](caffe2fluid/doc)] diff --git a/onnx2fluid/onnx2fluid/symbolic.py b/onnx2fluid/onnx2fluid/symbolic.py index d5f1167..03a5daf 100644 --- a/onnx2fluid/onnx2fluid/symbolic.py +++ b/onnx2fluid/onnx2fluid/symbolic.py @@ -78,7 +78,6 @@ DEFAULT_OP_MAPPING = { 'Sign': ['sign', ['X'], ['Out']], 'Sin': ['sin', ['X'], ['Out']], 'Squeeze': ['squeeze', ['X'], ['Out']], # attrs bypassed, FIXME: emit squeeze2 - 'Softplus': ['softplus', ['X'], ['Out']], # FIXME: default axis = -1, reshape required before and after 'Softmax': ['softmax', ['X'], ['Out'], dict(axis='')], 'Softplus': ['softplus', ['X'], ['Out']], @@ -305,7 +304,7 @@ def _pad_if_asymmetric(prog, pads, var_input, value_infos): # pads: SSEE if symmetric: return pads[:ndims], var_input - var_padded = var_input + '_padded' # explicit variable + var_padded = var_input + '_pad' # explicit variable prog.Op( '', 'Pad', @@ -317,7 +316,7 @@ def _pad_if_asymmetric(prog, pads, var_input, value_infos): # pads: SSEE 'pads': pads, }, value_infos=value_infos, - name=(var_input + '_pad'), + name=(var_input + '/pad'), ) return [0] * ndims, var_padded @@ -688,13 +687,14 @@ def BatchNormalization(prog, momentum = attrs.get('momentum', .9) # optional epsilon = attrs.get('epsilon', 1e-5) # optional name_attr = ', name={}'.format(repr(name)) if name else '' - if embed_params: - embed_params = _check_embeddable(value_infos, var_scale, var_b, - var_mean, var_var) - if not embed_params and name: - _logger.warning('for op %s(%s -> BatchNormalization -> %s)', name, - inputs, outputs) - _logger.warning('broken Python code will be generated') + embeddable = _check_embeddable(value_infos, var_scale, var_b, var_mean, + var_var) + if not embeddable: + _logger.warning('for op %s(%s -> BatchNormalization -> %s)', name, + inputs, outputs) + _logger.warning('one of the parameters is intermediate value') + _logger.warning('broken Python code will be generated') + embed_params &= embeddable if embed_params: assert name != '' embedded_scale = name + '.w_0' @@ -898,10 +898,10 @@ def ConstantOfShape(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): 'given shape is neither const value nor deductible from output, ' 'this is not supported') attrs = attrs.copy() - attrs.setdefault('value', np.array(0, dtype=np.float32)) + attrs.setdefault('value', _np.array(0, dtype=_np.float32)) attrs.update({'shape': shape}) # pass const - prog.Code('# shape:{}={} # const as literal'.format(var_shape, shape)) + prog.Code('# shape: {} = {} # const as literal'.format(var_shape, shape)) prog.Op( '', 'Constant', @@ -947,13 +947,13 @@ def Conv(prog, pads = attrs.get('pads', [0] * (convnd * 2)) # optional paddings, var_x = _pad_if_asymmetric(prog, pads, var_x, value_infos) name_attr = ', name={}'.format(repr(name)) - if embed_params: - embed_params = _check_embeddable( - value_infos, *([var_w] + ([var_b] if var_b else []))) - if not embed_params: - _logger.warning('for op %s(%s -> Conv -> %s)', name, inputs, - outputs) - _logger.warning('broken Python code will be generated') + embeddable = _check_embeddable(value_infos, + *([var_w] + ([var_b] if var_b else []))) + if not embeddable: + _logger.warning('for op %s(%s -> Conv -> %s)', name, inputs, outputs) + _logger.warning('one of the parameters is intermediate value') + _logger.warning('broken Python code will be generated') + embed_params &= embeddable if embed_params: embedded_w = name + '.w_0' value_infos[var_w]['embedded_as'].append(embedded_w) @@ -1013,7 +1013,7 @@ def Conv(prog, [var_y], {'axis': 1}, value_infos=value_infos, - name=(name + '.bias'), + name=(name + '/bias'), ) else: prog.VarDesc(var_y) @@ -1058,13 +1058,14 @@ def ConvTranspose(prog, pads = attrs.get('pads', [0] * (convnd * 2)) # optional paddings, var_x = _pad_if_asymmetric(prog, pads, var_x, value_infos) name_attr = ', name={}'.format(repr(name)) - if embed_params: - embed_params = _check_embeddable( - value_infos, *([var_w] + ([var_b] if var_b else []))) - if not embed_params: - _logger.warning('for op %s(%s -> ConvTranspose -> %s)', name, - inputs, outputs) - _logger.warning('broken Python code will be generated') + embeddable = _check_embeddable(value_infos, + *([var_w] + ([var_b] if var_b else []))) + if not embeddable: + _logger.warning('for op %s(%s -> ConvTranspose -> %s)', name, inputs, + outputs) + _logger.warning('one of the parameters is intermediate value') + _logger.warning('broken Python code will be generated') + embed_params &= embeddable if embed_params: embedded_w = name + '.w_0' value_infos[var_w]['embedded_as'].append(embedded_w) @@ -1128,7 +1129,7 @@ def ConvTranspose(prog, [var_y], {'axis': 1}, value_infos=value_infos, - name=(name + '.bias'), + name=(name + '/bias'), ) else: prog.VarDesc(var_y) @@ -1148,7 +1149,7 @@ def Gemm(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): trans_a = bool(attrs.get('transA', 0)) # optional trans_b = bool(attrs.get('transB', 0)) # optional - var_mm = var_y if beta == 0 else (name + '_mmed') # explicit variable + var_mm = var_y if beta == 0 else (name + '_mm') # explicit variable prog.Op( '', 'MatMul', @@ -1160,7 +1161,7 @@ def Gemm(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): 'alpha': alpha, }, value_infos=value_infos, - name=(name + '_mm'), + name=(name + '/mm'), ) prog.op_descs[-1].attrs.extend( prog.OpDescAttrs({ @@ -1176,7 +1177,7 @@ def Gemm(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_y], {'axis': 1}, value_infos=value_infos, - name=(name + '_bias'), + name=(name + '/bias'), ) else: var_beta = name + '_beta' # explicit variable @@ -1207,7 +1208,7 @@ def Gemm(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_vm], dict(), value_infos=value_infos, - name=(var_beta + '_scale'), + name=(var_beta + '/scale'), ) prog.Op( '', @@ -1215,7 +1216,7 @@ def Gemm(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_mm, var_vm], [var_y], {'axis': 1}, # - name=(name + '_bias'), + name=(name + '/bias'), ) @@ -1307,6 +1308,9 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): is_reverse = direction == 'reverse' fluid_op = 'dynamic_gru' + _logger.warning('for op (%s -> GRU -> %s)', inputs, outputs) + _logger.warning('one of the parameters is intermediate value') + _logger.warning('broken Python code will be generated') # generation var_x0 = var_x + '_0' # explicit variable @@ -1316,7 +1320,7 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): [var_x], [var_x0], {'axes': [1]}, # index on n - name=(var_x + '_index'), + name=(var_x + '/index'), ) var_w0 = var_w + '_0' # explicit variable prog.Op( @@ -1325,10 +1329,10 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): [var_w], [var_w0], {'axes': [0]}, # index on d - name=(var_w + '_index'), + name=(var_w + '/index'), ) var_fc = var_x0 + '_fc' - var_mm = (var_x0 + '_mmed') if var_b else var_fc + var_mm = (var_x0 + '_mm') if var_b else var_fc prog.Op( '', 'MatMul', @@ -1339,7 +1343,7 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): 'transpose_y': 1, }, value_infos=value_infos, - name=(var_x0 + '_mm'), + name=(var_x0 + '/mm'), ) prog.op_descs[-1].attrs.extend( prog.OpDescAttrs({ @@ -1353,7 +1357,7 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): [var_r], [var_r0], {'axes': [0]}, # index on d - name=(var_r + '_index'), + name=(var_r + '/index'), ) var_r0t = var_r0 + '_t' # explicit variable prog.Op( @@ -1362,7 +1366,7 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): [var_r0], [var_r0t], {'perm': [1, 0]}, # transpose OI->IO - name=(var_r0 + '_transpose'), + name=(var_r0 + '/transpose'), ) if var_b: var_bi = var_b + '_i' # explicit variable @@ -1376,7 +1380,7 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): 'axis': 1, # split on x 'split': [hidden_size * 3, hidden_size * 3], }, - name=(var_b + '_split'), + name=(var_b + '/split'), ) # squeeze bi so Gemm Add can be performed on axis=1 exaclty var_bi0 = var_bi + '_0' # explicit variable @@ -1386,7 +1390,7 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): [var_bi], [var_bi0], {'axes': [0]}, # slice on d - name=(var_bi + '_index'), + name=(var_bi + '/index'), ) prog.Op( '', @@ -1394,7 +1398,7 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): [var_mm, var_bi0], [var_fc], {'axis': 1}, # - name=(var_x0 + '_bias'), + name=(var_x0 + '/bias'), ) if var_xh: var_xh0 = var_xh + '_0' # explicit variable @@ -1404,7 +1408,7 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): [var_xh], [var_xh0], {'axes': [1]}, # index on n - name=(var_xh + '_index'), + name=(var_xh + '/index'), ) var_y00 = var_y + '_00' # explicit variable prog.Code('{} = layers.{}({}, {}, origin_mode=True' @@ -1449,7 +1453,7 @@ def GRU(prog, inputs, outputs, attrs, value_infos, *args, **kwargs): [var_y00], [var_y], {'axes': [1, 1]}, # extrude on dn - name=(var_y + '_reshape'), + name=(var_y + '/reshape'), ) @@ -1511,6 +1515,9 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): fluid_op = 'dynamic_lstm' name_attr = ', name={}'.format(repr(name)) + _logger.warning('for op %s(%s -> LSTM -> %s)', name, inputs, outputs) + _logger.warning('one of the parameters is intermediate value') + _logger.warning('broken Python code will be generated') # generation var_x0 = var_x + '_0' # explicit variable @@ -1520,7 +1527,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_x], [var_x0], {'axes': [1]}, # index on n - name=(var_x + '_index'), + name=(var_x + '/index'), ) var_w0 = var_w + '_0' # explicit variable prog.Op( @@ -1529,10 +1536,10 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_w], [var_w0], {'axes': [0]}, # index on d - name=(var_w + '_index'), + name=(var_w + '/index'), ) var_fc = var_x0 + '_fc' - var_mm = (var_x0 + '_mmed') if var_b else var_fc + var_mm = (var_x0 + '_mm') if var_b else var_fc prog.Op( '', 'MatMul', @@ -1543,7 +1550,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): 'transpose_y': 1, }, value_infos=value_infos, - name=(name + '_mm'), + name=(name + '/mm'), ) prog.op_descs[-1].attrs.extend( prog.OpDescAttrs({ @@ -1557,7 +1564,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_r], [var_r0], {'axes': [0]}, # index on d - name=(var_r + '_index'), + name=(var_r + '/index'), ) var_r0t = var_r0 + '_t' # explicit variable prog.Op( @@ -1566,7 +1573,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_r0], [var_r0t], {'perm': [1, 0]}, # transpose OI->IO - name=(var_r0 + '_transpose'), + name=(var_r0 + '/transpose'), ) if var_b: var_bi = var_b + '_i' # explicit variable @@ -1580,7 +1587,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): 'axis': 1, # split on x 'split': [hidden_size * 4, hidden_size * 4], }, - name=(var_b + '_split'), + name=(var_b + '/split'), ) # squeeze bi so Gemm Add can be performed on axis=1 exaclty var_bi0 = var_bi + '_0' # explicit variable @@ -1590,7 +1597,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_bi], [var_bi0], {'axes': [0]}, # slice on d - name=(var_bi + '_index'), + name=(var_bi + '/index'), ) prog.Op( '', @@ -1598,7 +1605,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_mm, var_bi0], [var_fc], {'axis': 1}, # - name=(name + '_bias'), + name=(name + '/bias'), ) if var_xh: var_xh0 = var_xh + '_0' # explicit variable @@ -1608,7 +1615,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_xh], [var_xh0], {'axes': [1]}, # index on n - name=(var_xh + '_index'), + name=(var_xh + '/index'), ) if var_xc: var_xc0 = var_xc + '_0' # explicit variable @@ -1618,7 +1625,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_xc], [var_xc0], {'axes': [1]}, # index on n - name=(var_xc + '_index'), + name=(var_xc + '/index'), ) var_bhp = var_p if var_b: @@ -1630,7 +1637,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_bh, var_p], [var_bhp], {'axes': [1]}, # cat on x - name=(name + '_concat'), + name=(name + '/concat'), ) else: var_bhp = var_bh @@ -1690,7 +1697,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_yh0], [var_y], # var_yh {'axes': [1, 1]}, # extrude on dn - name=(var_y + '_reshape'), + name=(var_y + '/reshape'), ) if var_yc: prog.Op( @@ -1699,7 +1706,7 @@ def LSTM(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): [var_yc0], [var_yc], {'axes': [1, 1]}, # extrude on dn - name=(var_yc + '_reshape'), + name=(var_yc + '/reshape'), ) @@ -1811,12 +1818,12 @@ def PRelu(prog, mode = 'element' fluid_op = 'prelu' name_attr = ', name={}'.format(repr(name)) if name else '' - if embed_params: - embed_params = _check_embeddable(value_infos, var_slope) - if not embed_params and name: - _logger.warning('for op %s(%s -> PRelu -> %s)', name, inputs, - outputs) - _logger.warning('broken Python code will be generated') + embeddable = _check_embeddable(value_infos, var_slope) + if not embeddable: + _logger.warning('for op %s(%s -> PRelu -> %s)', name, inputs, outputs) + _logger.warning('one of the parameters is intermediate value') + _logger.warning('broken Python code will be generated') + embed_params &= embeddable if embed_params: assert name != '' embedded_slope = name + '.w_0' @@ -1880,12 +1887,20 @@ def Reshape(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): 'input "shape" not inferred, use [1, -1] as dummy value, ' 'the behavior of Paddle fluid maybe undefined', name, inputs, outputs) + shape_dtype = _dtype_or_none(value_infos, var_shape) + if shape_dtype is None: + _logger.warning( + 'in op %s(%s -> Reshape -> %s): ' + 'dtype of input "shape" not inferred, int32 assumed', name, inputs, + outputs) + shape_dtype = _np.dtype('int32') fluid_op = 'reshape' name_attr = ', name={}'.format(repr(name)) # generation - var_shape_int32 = var_shape + '_int32' # explicit variable - prog.Code('# shape:{}={} # const as literal'.format(var_shape, shape)) + var_shape_int32 = var_shape + ('_int32' if shape_dtype != _np.int32 else '' + ) # explicit variable + prog.Code('# shape: {} = {} # const as literal'.format(var_shape, shape)) if is_const_shape: prog.Code('{} = layers.{}({}' ', shape={}' @@ -1898,15 +1913,16 @@ def Reshape(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs): name_attr, )) else: - prog.Op( - '', - 'Cast', - [var_shape], - [var_shape_int32], - {'to': _np.dtype('int32')}, # use np.dtype - value_infos=value_infos, - name=(name + '_cast'), - ) + if shape_dtype != _np.int32: + prog.Op( + '', + 'Cast', + [var_shape], + [var_shape_int32], + {'to': _np.dtype('int32')}, # use np.dtype + value_infos=value_infos, + name=(name + '/cast'), + ) prog.Code('{} = layers.{}({}' ', shape={}' ', actual_shape={}' @@ -2121,7 +2137,8 @@ def Tile(prog, inputs, outputs, attrs, value_infos, name='', *args, **kwargs): name_attr = ', name={}'.format(repr(name)) if name else '' # generation - prog.Code('# repeats:{}={} # const as literal'.format(var_repeats, repeats)) + prog.Code('# repeats: {} = {} # const as literal'.format( + var_repeats, repeats)) prog.Code('{} = layers.{}({}' ', expand_times={}' '{})'.format( -- GitLab