From e04489ce4c0fd0095d0c79ef8f504f425e0435a8 Mon Sep 17 00:00:00 2001
From: Shi Xian <40013335+R1ckShi@users.noreply.github.com>
Date: 星期三, 13 三月 2024 16:34:42 +0800
Subject: [PATCH] contextual&seaco ONNX export (#1481)
---
funasr/models/sond/encoder/resnet34_encoder.py | 422 ----------------------------------------------------
1 files changed, 0 insertions(+), 422 deletions(-)
diff --git a/funasr/models/sond/encoder/resnet34_encoder.py b/funasr/models/sond/encoder/resnet34_encoder.py
index 8445feb..8bfe491 100644
--- a/funasr/models/sond/encoder/resnet34_encoder.py
+++ b/funasr/models/sond/encoder/resnet34_encoder.py
@@ -245,147 +245,6 @@
return features, resnet_out_lens
- def gen_tf2torch_map_dict(self):
- tensor_name_prefix_torch = self.tf2torch_tensor_name_prefix_torch
- tensor_name_prefix_tf = self.tf2torch_tensor_name_prefix_tf
- train_steps = self.tf_train_steps
- map_dict_local = {
- # torch: conv1d.weight in "out_channel in_channel kernel_size"
- # tf : conv1d.weight in "kernel_size in_channel out_channel"
- # torch: linear.weight in "out_channel in_channel"
- # tf : dense.weight in "in_channel out_channel"
- "{}.pre_conv.weight".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv/kernel".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": (3, 2, 0, 1),
- },
- "{}.pre_conv_bn.bias".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/beta".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.weight".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/gamma".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.running_mean".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/moving_mean".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.running_var".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/moving_variance".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.num_batches_tracked".format(tensor_name_prefix_torch): train_steps
- }
- for layer_idx in range(3):
- map_dict_local.update({
- "{}.resnet{}_dense.weight".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_dense/kernel".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": (2, 1, 0) if layer_idx == 0 else (1, 0),
- },
- "{}.resnet{}_dense.bias".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_dense/bias".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.weight".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/gamma".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.bias".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/beta".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.running_mean".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/moving_mean".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.running_var".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/moving_variance".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.num_batches_tracked".format(tensor_name_prefix_torch, layer_idx): train_steps
- })
-
- for block_idx in range(len(self.layers_in_block)):
- for layer_idx in range(self.layers_in_block[block_idx]):
- for i in ["1", "2", "_sc"]:
- map_dict_local.update({
- "{}.block_{}.layer_{}.conv{}.weight".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/conv{}/kernel".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": (3, 2, 0, 1),
- },
- "{}.block_{}.layer_{}.bn{}.weight".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/gamma".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.bias".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/beta".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.running_mean".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/moving_mean".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.running_var".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/moving_variance".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.num_batches_tracked".format(tensor_name_prefix_torch, block_idx, layer_idx, i): train_steps,
- })
-
- return map_dict_local
-
- def convert_tf2torch(self,
- var_dict_tf,
- var_dict_torch,
- ):
-
- map_dict = self.gen_tf2torch_map_dict()
-
- var_dict_torch_update = dict()
- for name in sorted(var_dict_torch.keys(), reverse=False):
- if name.startswith(self.tf2torch_tensor_name_prefix_torch):
- if name in map_dict:
- if "num_batches_tracked" not in name:
- name_tf = map_dict[name]["name"]
- data_tf = var_dict_tf[name_tf]
- if map_dict[name]["squeeze"] is not None:
- data_tf = np.squeeze(data_tf, axis=map_dict[name]["squeeze"])
- if map_dict[name]["transpose"] is not None:
- data_tf = np.transpose(data_tf, map_dict[name]["transpose"])
- data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
- assert var_dict_torch[name].size() == data_tf.size(), \
- "{}, {}, {} != {}".format(name, name_tf,
- var_dict_torch[name].size(), data_tf.size())
- var_dict_torch_update[name] = data_tf
- logging.info("torch tensor: {}, {}, loading from tf tensor: {}, {}".format(
- name, data_tf.size(), name_tf, var_dict_tf[name_tf].shape
- ))
- else:
- var_dict_torch_update[name] = torch.Tensor(map_dict[name]).type(torch.int64).to("cpu")
- logging.info("torch tensor: {}, manually assigning to: {}".format(
- name, map_dict[name]
- ))
- else:
- logging.warning("{} is missed from tf checkpoint".format(name))
-
- return var_dict_torch_update
-
class ResNet34Diar(ResNet34):
def __init__(
@@ -477,147 +336,6 @@
endpoints["resnet2_bn"] = features
return endpoints[self.embedding_node], ilens, None
-
- def gen_tf2torch_map_dict(self):
- tensor_name_prefix_torch = self.tf2torch_tensor_name_prefix_torch
- tensor_name_prefix_tf = self.tf2torch_tensor_name_prefix_tf
- train_steps = 300000
- map_dict_local = {
- # torch: conv1d.weight in "out_channel in_channel kernel_size"
- # tf : conv1d.weight in "kernel_size in_channel out_channel"
- # torch: linear.weight in "out_channel in_channel"
- # tf : dense.weight in "in_channel out_channel"
- "{}.pre_conv.weight".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv/kernel".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": (3, 2, 0, 1),
- },
- "{}.pre_conv_bn.bias".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/beta".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.weight".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/gamma".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.running_mean".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/moving_mean".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.running_var".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/moving_variance".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.num_batches_tracked".format(tensor_name_prefix_torch): train_steps
- }
- for layer_idx in range(3):
- map_dict_local.update({
- "{}.resnet{}_dense.weight".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_dense/kernel".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": (3, 2, 0, 1) if layer_idx == 0 else (1, 0),
- },
- "{}.resnet{}_dense.bias".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_dense/bias".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.weight".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/gamma".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.bias".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/beta".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.running_mean".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/moving_mean".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.running_var".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/moving_variance".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.num_batches_tracked".format(tensor_name_prefix_torch, layer_idx): train_steps
- })
-
- for block_idx in range(len(self.layers_in_block)):
- for layer_idx in range(self.layers_in_block[block_idx]):
- for i in ["1", "2", "_sc"]:
- map_dict_local.update({
- "{}.block_{}.layer_{}.conv{}.weight".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/conv{}/kernel".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": (3, 2, 0, 1),
- },
- "{}.block_{}.layer_{}.bn{}.weight".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/gamma".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.bias".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/beta".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.running_mean".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/moving_mean".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.running_var".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/moving_variance".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.num_batches_tracked".format(tensor_name_prefix_torch, block_idx, layer_idx, i): train_steps,
- })
-
- return map_dict_local
-
- def convert_tf2torch(self,
- var_dict_tf,
- var_dict_torch,
- ):
-
- map_dict = self.gen_tf2torch_map_dict()
-
- var_dict_torch_update = dict()
- for name in sorted(var_dict_torch.keys(), reverse=False):
- if name.startswith(self.tf2torch_tensor_name_prefix_torch):
- if name in map_dict:
- if "num_batches_tracked" not in name:
- name_tf = map_dict[name]["name"]
- data_tf = var_dict_tf[name_tf]
- if map_dict[name]["squeeze"] is not None:
- data_tf = np.squeeze(data_tf, axis=map_dict[name]["squeeze"])
- if map_dict[name]["transpose"] is not None:
- data_tf = np.transpose(data_tf, map_dict[name]["transpose"])
- data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
- assert var_dict_torch[name].size() == data_tf.size(), \
- "{}, {}, {} != {}".format(name, name_tf,
- var_dict_torch[name].size(), data_tf.size())
- var_dict_torch_update[name] = data_tf
- logging.info("torch tensor: {}, {}, loading from tf tensor: {}, {}".format(
- name, data_tf.size(), name_tf, var_dict_tf[name_tf].shape
- ))
- else:
- var_dict_torch_update[name] = torch.Tensor(map_dict[name]).type(torch.int64).to("cpu")
- logging.info("torch tensor: {}, manually assigning to: {}".format(
- name, map_dict[name]
- ))
- else:
- logging.warning("{} is missed from tf checkpoint".format(name))
-
- return var_dict_torch_update
class ResNet34SpL2RegDiar(ResNet34_SP_L2Reg):
@@ -711,143 +429,3 @@
return endpoints[self.embedding_node], ilens, None
- def gen_tf2torch_map_dict(self):
- tensor_name_prefix_torch = self.tf2torch_tensor_name_prefix_torch
- tensor_name_prefix_tf = self.tf2torch_tensor_name_prefix_tf
- train_steps = 720000
- map_dict_local = {
- # torch: conv1d.weight in "out_channel in_channel kernel_size"
- # tf : conv1d.weight in "kernel_size in_channel out_channel"
- # torch: linear.weight in "out_channel in_channel"
- # tf : dense.weight in "in_channel out_channel"
- "{}.pre_conv.weight".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv/kernel".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": (3, 2, 0, 1),
- },
- "{}.pre_conv_bn.bias".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/beta".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.weight".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/gamma".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.running_mean".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/moving_mean".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.running_var".format(tensor_name_prefix_torch):
- {"name": "{}/pre_conv_bn/moving_variance".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- "{}.pre_conv_bn.num_batches_tracked".format(tensor_name_prefix_torch): train_steps
- }
- for layer_idx in range(3):
- map_dict_local.update({
- "{}.resnet{}_dense.weight".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_dense/kernel".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": (2, 1, 0) if layer_idx == 0 else (1, 0),
- },
- "{}.resnet{}_dense.bias".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_dense/bias".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.weight".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/gamma".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.bias".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/beta".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.running_mean".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/moving_mean".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.running_var".format(tensor_name_prefix_torch, layer_idx):
- {"name": "{}/resnet{}_bn/moving_variance".format(tensor_name_prefix_tf, layer_idx),
- "squeeze": None,
- "transpose": None,
- },
- "{}.resnet{}_bn.num_batches_tracked".format(tensor_name_prefix_torch, layer_idx): train_steps
- })
-
- for block_idx in range(len(self.layers_in_block)):
- for layer_idx in range(self.layers_in_block[block_idx]):
- for i in ["1", "2", "_sc"]:
- map_dict_local.update({
- "{}.block_{}.layer_{}.conv{}.weight".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/conv{}/kernel".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": (3, 2, 0, 1),
- },
- "{}.block_{}.layer_{}.bn{}.weight".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/gamma".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.bias".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/beta".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.running_mean".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/moving_mean".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.running_var".format(tensor_name_prefix_torch, block_idx, layer_idx, i):
- {"name": "{}/block_{}/layer_{}/bn{}/moving_variance".format(tensor_name_prefix_tf, block_idx, layer_idx, i),
- "squeeze": None,
- "transpose": None,
- },
- "{}.block_{}.layer_{}.bn{}.num_batches_tracked".format(tensor_name_prefix_torch, block_idx, layer_idx, i): train_steps,
- })
-
- return map_dict_local
-
- def convert_tf2torch(self,
- var_dict_tf,
- var_dict_torch,
- ):
-
- map_dict = self.gen_tf2torch_map_dict()
-
- var_dict_torch_update = dict()
- for name in sorted(var_dict_torch.keys(), reverse=False):
- if name.startswith(self.tf2torch_tensor_name_prefix_torch):
- if name in map_dict:
- if "num_batches_tracked" not in name:
- name_tf = map_dict[name]["name"]
- data_tf = var_dict_tf[name_tf]
- if map_dict[name]["squeeze"] is not None:
- data_tf = np.squeeze(data_tf, axis=map_dict[name]["squeeze"])
- if map_dict[name]["transpose"] is not None:
- data_tf = np.transpose(data_tf, map_dict[name]["transpose"])
- data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
- assert var_dict_torch[name].size() == data_tf.size(), \
- "{}, {}, {} != {}".format(name, name_tf,
- var_dict_torch[name].size(), data_tf.size())
- var_dict_torch_update[name] = data_tf
- logging.info("torch tensor: {}, {}, loading from tf tensor: {}, {}".format(
- name, data_tf.size(), name_tf, var_dict_tf[name_tf].shape
- ))
- else:
- var_dict_torch_update[name] = torch.from_numpy(np.array(map_dict[name])).type(torch.int64).to("cpu")
- logging.info("torch tensor: {}, manually assigning to: {}".format(
- name, map_dict[name]
- ))
- else:
- logging.warning("{} is missed from tf checkpoint".format(name))
-
- return var_dict_torch_update
\ No newline at end of file
--
Gitblit v1.9.1