hobs
commited on
Commit
·
cc9cfea
1
Parent(s):
ca70154
load state_dict
Browse files
app.py
CHANGED
@@ -2,9 +2,240 @@
|
|
2 |
|
3 |
import gradio as gr
|
4 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
|
6 |
def greet_nationality(name):
|
7 |
-
nationality =
|
8 |
return f"Hello {name}!!\n Your name seems to be from {nationality}. Am I right?"
|
9 |
|
10 |
|
|
|
2 |
|
3 |
import gradio as gr
|
4 |
|
5 |
+
import os
|
6 |
+
from pathlib import Path
|
7 |
+
# import random
|
8 |
+
# import time
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
|
12 |
+
import pandas as pd
|
13 |
+
from nlpia2.init import SRC_DATA_DIR, maybe_download
|
14 |
+
|
15 |
+
from nlpia2.string_normalizers import Asciifier, ASCII_NAME_CHARS
|
16 |
+
|
17 |
+
name_char_vocab_size = len(ASCII_NAME_CHARS) + 1 # Plus EOS marker
|
18 |
+
|
19 |
+
# Transcode Unicode str ASCII without embelishments, diacritics (https://stackoverflow.com/a/518232/2809427)
|
20 |
+
asciify = Asciifier(include=ASCII_NAME_CHARS)
|
21 |
+
|
22 |
+
|
23 |
+
def find_files(path, pattern):
|
24 |
+
return Path(path).glob(pattern)
|
25 |
+
|
26 |
+
|
27 |
+
# all_letters = ''.join(set(ASCII_NAME_CHARS).union(set(" .,;'")))
|
28 |
+
char2i = {c: i for i, c in enumerate(ASCII_NAME_CHARS)}
|
29 |
+
|
30 |
+
# !curl -O https://download.pytorch.org/tutorial/data.zip; unzip data.zip
|
31 |
+
|
32 |
+
print(f'asciify("O’Néàl") => {asciify("O’Néàl")}')
|
33 |
+
|
34 |
+
# Build the category_lines dictionary, a list of names per language
|
35 |
+
category_lines = {}
|
36 |
+
all_categories = []
|
37 |
+
labeled_lines = []
|
38 |
+
categories = []
|
39 |
+
for filepath in find_files(SRC_DATA_DIR / 'names', '*.txt'):
|
40 |
+
filename = Path(filepath).name
|
41 |
+
filepath = maybe_download(filename=Path('names') / filename)
|
42 |
+
with filepath.open() as fin:
|
43 |
+
lines = [asciify(line.rstrip()) for line in fin]
|
44 |
+
category = Path(filename).with_suffix('')
|
45 |
+
categories.append(category)
|
46 |
+
labeled_lines += list(zip(lines, [category] * len(lines)))
|
47 |
+
|
48 |
+
n_categories = len(categories)
|
49 |
+
|
50 |
+
df = pd.DataFrame(labeled_lines, columns=('name', 'category'))
|
51 |
+
|
52 |
+
|
53 |
+
def readLines(filename):
|
54 |
+
lines = open(filename, encoding='utf-8').read().strip().split('\n')
|
55 |
+
return [asciify(line) for line in lines]
|
56 |
+
|
57 |
+
|
58 |
+
for filename in find_files(path='data/names', pattern='*.txt'):
|
59 |
+
category = os.path.splitext(os.path.basename(filename))[0]
|
60 |
+
all_categories.append(category)
|
61 |
+
lines = readLines(filename)
|
62 |
+
category_lines[category] = lines
|
63 |
+
|
64 |
+
n_categories = len(all_categories)
|
65 |
+
|
66 |
+
|
67 |
+
######################################################################
|
68 |
+
# Now we have ``category_lines``, a dictionary mapping each category
|
69 |
+
# (language) to a list of lines (names). We also kept track of
|
70 |
+
# ``all_categories`` (just a list of languages) and ``n_categories`` for
|
71 |
+
# later reference.
|
72 |
+
#
|
73 |
+
|
74 |
+
print(category_lines['Italian'][:5])
|
75 |
+
|
76 |
+
|
77 |
+
######################################################################
|
78 |
+
# Turning Names into Tensors
|
79 |
+
# --------------------------
|
80 |
+
#
|
81 |
+
# Now that we have all the names organized, we need to turn them into
|
82 |
+
# Tensors to make any use of them.
|
83 |
+
#
|
84 |
+
# To represent a single letter, we use a "one-hot vector" of size
|
85 |
+
# ``<1 x n_letters>``. A one-hot vector is filled with 0s except for a 1
|
86 |
+
# at index of the current letter, e.g. ``"b" = <0 1 0 0 0 ...>``.
|
87 |
+
#
|
88 |
+
# To make a word we join a bunch of those into a 2D matrix
|
89 |
+
# ``<line_length x 1 x n_letters>``.
|
90 |
+
#
|
91 |
+
# That extra 1 dimension is because PyTorch assumes everything is in
|
92 |
+
# batches - we're just using a batch size of 1 here.
|
93 |
+
#
|
94 |
+
|
95 |
+
# Find letter index from all_letters, e.g. "a" = 0
|
96 |
+
|
97 |
+
|
98 |
+
def letterToIndex(c):
|
99 |
+
return char2i[c]
|
100 |
+
|
101 |
+
# Just for demonstration, turn a letter into a <1 x n_letters> Tensor
|
102 |
+
|
103 |
+
|
104 |
+
def encode_one_hot_vec(letter):
|
105 |
+
tensor = torch.zeros(1, len(ASCII_NAME_CHARS))
|
106 |
+
tensor[0][letterToIndex(letter)] = 1
|
107 |
+
return tensor
|
108 |
+
|
109 |
+
# Turn a line into a <line_length x 1 x n_letters>,
|
110 |
+
# or an array of one-hot letter vectors
|
111 |
+
|
112 |
+
|
113 |
+
def encode_one_hot_seq(line):
|
114 |
+
tensor = torch.zeros(len(line), 1, len(ASCII_NAME_CHARS))
|
115 |
+
for li, letter in enumerate(line):
|
116 |
+
tensor[li][0][letterToIndex(letter)] = 1
|
117 |
+
return tensor
|
118 |
+
|
119 |
+
|
120 |
+
print(encode_one_hot_vec('A'))
|
121 |
+
|
122 |
+
print(encode_one_hot_seq('Abe').size())
|
123 |
+
|
124 |
+
|
125 |
+
######################################################################
|
126 |
+
# Creating the Network
|
127 |
+
# ====================
|
128 |
+
#
|
129 |
+
# Before autograd, creating a recurrent neural network in Torch involved
|
130 |
+
# cloning the parameters of a layer over several timesteps. The layers
|
131 |
+
# held hidden state and gradients which are now entirely handled by the
|
132 |
+
# graph itself. This means you can implement a RNN in a very "pure" way,
|
133 |
+
# as regular feed-forward layers.
|
134 |
+
#
|
135 |
+
# This RNN module (mostly copied from `the PyTorch for Torch users
|
136 |
+
# tutorial <https://pytorch.org/tutorials/beginner/former_torchies/
|
137 |
+
# nn_tutorial.html#example-2-recurrent-net>`__)
|
138 |
+
# is just 2 linear layers which operate on an input and hidden state, with
|
139 |
+
# a LogSoftmax layer after the output.
|
140 |
+
#
|
141 |
+
# .. figure:: https://i.imgur.com/Z2xbySO.png
|
142 |
+
# :alt:
|
143 |
+
#
|
144 |
+
#
|
145 |
+
|
146 |
+
|
147 |
+
class RNN(nn.Module):
|
148 |
+
def __init__(self, input_size, hidden_size, output_size):
|
149 |
+
super(RNN, self).__init__()
|
150 |
+
|
151 |
+
self.hidden_size = hidden_size
|
152 |
+
|
153 |
+
self.i2h = nn.Linear(input_size + hidden_size, hidden_size)
|
154 |
+
self.i2o = nn.Linear(input_size + hidden_size, output_size)
|
155 |
+
self.softmax = nn.LogSoftmax(dim=1)
|
156 |
+
|
157 |
+
def forward(self, char_tens, hidden):
|
158 |
+
combined = torch.cat((char_tens, hidden), 1)
|
159 |
+
hidden = self.i2h(combined)
|
160 |
+
output = self.i2o(combined)
|
161 |
+
output = self.softmax(output)
|
162 |
+
return output, hidden
|
163 |
+
|
164 |
+
def initHidden(self):
|
165 |
+
return torch.zeros(1, self.hidden_size)
|
166 |
+
|
167 |
+
|
168 |
+
n_hidden = 128
|
169 |
+
rnn = RNN(len(ASCII_NAME_CHARS), n_hidden, n_categories)
|
170 |
+
|
171 |
+
|
172 |
+
input = encode_one_hot_vec('A')
|
173 |
+
hidden = torch.zeros(1, n_hidden)
|
174 |
+
|
175 |
+
output, next_hidden = rnn(input, hidden)
|
176 |
+
|
177 |
+
|
178 |
+
def categoryFromOutput(output):
|
179 |
+
top_n, top_i = output.topk(1)
|
180 |
+
category_i = top_i[0].item()
|
181 |
+
return all_categories[category_i], category_i
|
182 |
+
|
183 |
+
|
184 |
+
def output_from_str(s):
|
185 |
+
global rnn
|
186 |
+
|
187 |
+
input = encode_one_hot_seq(s)
|
188 |
+
hidden = torch.zeros(1, n_hidden)
|
189 |
+
|
190 |
+
output, next_hidden = rnn(input[0], hidden)
|
191 |
+
print(output)
|
192 |
+
|
193 |
+
return categoryFromOutput(output)
|
194 |
+
|
195 |
+
|
196 |
+
########################################
|
197 |
+
# load/save test for use on the huggingface spaces server
|
198 |
+
# torch.save(rnn.state_dict(), 'rnn_from_scratch_name_nationality.state_dict.pickle')
|
199 |
+
|
200 |
+
state_dict = torch.load('rnn_from_scratch_name_nationality.state_dict.pickle')
|
201 |
+
rnn.load_state_dict(state_dict)
|
202 |
+
|
203 |
+
|
204 |
+
def evaluate(line_tensor):
|
205 |
+
hidden = rnn.initHidden()
|
206 |
+
|
207 |
+
for i in range(line_tensor.size()[0]):
|
208 |
+
output, hidden = rnn(line_tensor[i], hidden)
|
209 |
+
|
210 |
+
return output
|
211 |
+
|
212 |
+
|
213 |
+
def predict(input_line, n_predictions=3):
|
214 |
+
print('\n> %s' % input_line)
|
215 |
+
with torch.no_grad():
|
216 |
+
output = evaluate(encode_one_hot_seq(input_line))
|
217 |
+
|
218 |
+
# Get top N categories
|
219 |
+
topv, topi = output.topk(n_predictions, 1, True)
|
220 |
+
predictions = []
|
221 |
+
|
222 |
+
for i in range(n_predictions):
|
223 |
+
value = topv[0][i].item()
|
224 |
+
category_index = topi[0][i].item()
|
225 |
+
print('(%.2f) %s' % (value, all_categories[category_index]))
|
226 |
+
predictions.append([value, all_categories[category_index]])
|
227 |
+
|
228 |
+
|
229 |
+
predict('Dovesky')
|
230 |
+
predict('Jackson')
|
231 |
+
predict('Satoshi')
|
232 |
+
|
233 |
+
# load/save test for use on the huggingface spaces server
|
234 |
+
########################################
|
235 |
+
|
236 |
|
237 |
def greet_nationality(name):
|
238 |
+
nationality = predict(name)
|
239 |
return f"Hello {name}!!\n Your name seems to be from {nationality}. Am I right?"
|
240 |
|
241 |
|