From 2ac38adbe5f4e1374a079e032ed4b504351a207c Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期二, 23 四月 2024 18:08:57 +0800
Subject: [PATCH] Dev gzf exp (#1647)
---
funasr/models/sond/encoder/conv_encoder.py | 102 --------------------------------------------------
1 files changed, 1 insertions(+), 101 deletions(-)
diff --git a/funasr/models/sond/encoder/conv_encoder.py b/funasr/models/sond/encoder/conv_encoder.py
index 4c345cb..2181160 100644
--- a/funasr/models/sond/encoder/conv_encoder.py
+++ b/funasr/models/sond/encoder/conv_encoder.py
@@ -12,7 +12,7 @@
from funasr.models.transformer.layer_norm import LayerNorm
from funasr.models.encoder.abs_encoder import AbsEncoder
import math
-from funasr.models.transformer.repeat import repeat
+from funasr.models.transformer.utils.repeat import repeat
class EncoderLayer(nn.Module):
@@ -172,104 +172,4 @@
outputs = outputs + inputs
return outputs, 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
- 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"
- "{}.cnn_a.0.conv1d.weight".format(tensor_name_prefix_torch):
- {"name": "{}/cnn_a/conv1d/kernel".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": (2, 1, 0),
- },
- "{}.cnn_a.0.conv1d.bias".format(tensor_name_prefix_torch):
- {"name": "{}/cnn_a/conv1d/bias".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
-
- "{}.cnn_a.layeridx.conv1d.weight".format(tensor_name_prefix_torch):
- {"name": "{}/cnn_a/conv1d_layeridx/kernel".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": (2, 1, 0),
- },
- "{}.cnn_a.layeridx.conv1d.bias".format(tensor_name_prefix_torch):
- {"name": "{}/cnn_a/conv1d_layeridx/bias".format(tensor_name_prefix_tf),
- "squeeze": None,
- "transpose": None,
- },
- }
- if self.out_units is not None:
- # add output layer
- map_dict_local.update({
- "{}.conv_out.weight".format(tensor_name_prefix_torch):
- {"name": "{}/cnn_a/conv1d_{}/kernel".format(tensor_name_prefix_tf, self.num_layers),
- "squeeze": None,
- "transpose": (2, 1, 0),
- }, # tf: (1, 256, 256) -> torch: (256, 256, 1)
- "{}.conv_out.bias".format(tensor_name_prefix_torch):
- {"name": "{}/cnn_a/conv1d_{}/bias".format(tensor_name_prefix_tf, self.num_layers),
- "squeeze": None,
- "transpose": None,
- }, # tf: (256,) -> torch: (256,)
- })
-
- 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):
- # process special (first and last) layers
- if name in map_dict:
- 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
- ))
- # process general layers
- else:
- # self.tf2torch_tensor_name_prefix_torch may include ".", solve this case
- names = name.replace(self.tf2torch_tensor_name_prefix_torch, "todo").split('.')
- layeridx = int(names[2])
- name_q = name.replace(".{}.".format(layeridx), ".layeridx.")
- if name_q in map_dict.keys():
- name_v = map_dict[name_q]["name"]
- name_tf = name_v.replace("layeridx", "{}".format(layeridx))
- data_tf = var_dict_tf[name_tf]
- if map_dict[name_q]["squeeze"] is not None:
- data_tf = np.squeeze(data_tf, axis=map_dict[name_q]["squeeze"])
- if map_dict[name_q]["transpose"] is not None:
- data_tf = np.transpose(data_tf, map_dict[name_q]["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:
- logging.warning("{} is missed from tf checkpoint".format(name))
-
- return var_dict_torch_update
--
Gitblit v1.9.1