def setUp(self):
self.config()
self.generate_input_data()
paddle.set_default_dtype(self.x_type)
self.__class__.op_type = "fused_attention"
# use autograd to check grad in this unittest.
self.__class__.no_need_check_grad = True
self.q_proj = Linear(self.embed_dim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
self.k_proj = Linear(self.kdim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
self.v_proj = Linear(self.vdim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
self.out_proj = Linear(self.embed_dim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
paddle.set_default_dtype(np.float32)
self.norm1 = LayerNorm(self.embed_dim)
self.norm2 = LayerNorm(self.embed_dim)
paddle.set_default_dtype(self.x_type)
self.dropout = Dropout(self.dropout_prob, mode="upscale_in_train")
def __init__(self, d_model, nhead, dim_feedforward, dropout=0.1,
activation="relu", normalize_before=False,
weight_attr=None, bias_attr=None):
self._config = locals()
self._config.pop("self")
self._config.pop("__class__", None) # py3
super(TransformerDecoderLayer, self).__init__()
self.normalize_before = normalize_before
weight_attrs = _convert_param_attr_to_list(weight_attr, 3)
bias_attrs = _convert_param_attr_to_list(bias_attr, 3)
self.nhead = nhead
self.self_attn = MultiHeadAttention(d_model, nhead, dropout=dropout,
weight_attr=weight_attrs[0], bias_attr=bias_attrs[0])
self.multihead_attn = MultiHeadAttention(d_model, nhead, dropout=dropout,
weight_attr=weight_attrs[1], bias_attr=bias_attrs[1])
self.linear1 = Linear(d_model, dim_feedforward,
weight_attrs[2], bias_attr=bias_attrs[2])
self.dropout = Dropout(dropout, mode="upscale_in_train")
self.linear2 = Linear(dim_feedforward, d_model,
weight_attrs[2], bias_attr=bias_attrs[2])
self.norm1 = LayerNorm(d_model)
self.norm2 = LayerNorm(d_model)
self.norm3 = LayerNorm(d_model)
self.dropout1 = Dropout(dropout, mode="upscale_in_train")
self.dropout2 = Dropout(dropout, mode="upscale_in_train")
self.dropout3 = Dropout(dropout, mode="upscale_in_train")
self.activation = getattr(F, activation)
def setUp(self):
self.config()
self.generate_input_data()
self.rtol = 1e-5
# FIXME(limin29): Because there is a problem with the test precision
# on A100, atol is temporarily set to 1e-2, and it will be
# changed back after the precision problem is solved.
self.atol = 1e-2
# make sure local development precision
if "V100" in paddle.device.cuda.get_device_name():
self.atol = 1e-4
if self.x_type is np.float16:
self.atol = 1e-1
paddle.set_default_dtype(self.x_type)
self.__class__.op_type = "fused_attention"
# use autograd to check grad in this unittest.
self.__class__.no_need_check_grad = True
self.q_proj = Linear(self.embed_dim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
self.k_proj = Linear(self.kdim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
self.v_proj = Linear(self.vdim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
self.out_proj = Linear(self.embed_dim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
paddle.set_default_dtype(np.float32)
self.norm1 = LayerNorm(self.embed_dim)
self.norm2 = LayerNorm(self.embed_dim)
paddle.set_default_dtype(self.x_type)
self.dropout = Dropout(self.dropout_prob, mode="upscale_in_train")
def setUp(self):
paddle.disable_static()
self.__class__.op_type = "fused_feedforward"
#check grad in test_out_and_grad()
self.__class__.no_need_check_grad = True
self.getDtype()
self.getShape()
self.getDiff()
self.getActivation()
self.getNormalizeBefore()
paddle.set_default_dtype(self.dtype)
self.weight_attr = None
self.bias_attr = None
self.weight_attrs = transformer._convert_param_attr_to_list(
self.weight_attr, 2)
self.bias_attrs = transformer._convert_param_attr_to_list(
self.bias_attr, 2)
self.linear1 = Linear(self.d_model,
self.dim_feedforward,
self.weight_attrs[1],
bias_attr=self.bias_attrs[1])
self.linear2 = Linear(self.dim_feedforward,
self.d_model,
self.weight_attrs[1],
bias_attr=self.bias_attrs[1])
paddle.set_default_dtype(self.layer_norm_dtype)
self.norm1 = LayerNorm(self.d_model)
self.norm2 = LayerNorm(self.d_model)
self.dropout = Dropout(0.0, mode="upscale_in_train")
self.dropout1 = Dropout(0.0, mode="upscale_in_train")
self.dropout2 = Dropout(0.0, mode="upscale_in_train")
self.activation = getattr(F, self.act_method)
self.src = np.random.random((self.batch_size, self.query_length,
self.d_model)).astype(self.dtype)
self.dout = np.random.random((self.batch_size, self.query_length,
self.d_model)).astype(self.dtype)
def setUp(self):
self.config()
self.generate_input_data()
self.rtol = 1e-5
# FIXME(wangxi): Because there is a problem with the test precision
# on A100, atol is temporarily set to 1e-2, and it will be
# changed back after the precision problem is solved.
self.atol = 1e-2
# make sure local development precision
if "V100" in paddle.device.cuda.get_device_name():
self.atol = 1e-4
if self.x_type is np.float16:
self.atol = 1e-1
paddle.set_default_dtype(self.x_type)
self.__class__.op_type = "fused_multi_transformer"
# use autograd to check grad in this unittest.
self.__class__.no_need_check_grad = False
bias_attr = paddle.fluid.ParamAttr(
initializer=paddle.fluid.initializer.Constant(value=0.0005))
self.q_proj = Linear(self.embed_dim,
self.embed_dim,
self.weight_attr,
bias_attr=bias_attr)
#bias_attr=self.bias_attr)
self.k_proj = Linear(self.kdim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
self.v_proj = Linear(self.vdim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
self.out_proj = Linear(self.embed_dim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
self.ffn1_proj = Linear(self.embed_dim,
4 * self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
self.ffn2_proj = Linear(4 * self.embed_dim,
self.embed_dim,
self.weight_attr,
bias_attr=self.bias_attr)
paddle.set_default_dtype(np.float32)
self.norm = LayerNorm(self.embed_dim)
self.ffn_norm = LayerNorm(self.embed_dim)
paddle.set_default_dtype(self.x_type)
self.dropout = Dropout(self.dropout_prob, mode="upscale_in_train")
self.activation = getattr(F, self.act_method)