Spaces:
Build error
Build error
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Import necessary libraries
|
| 2 |
+
import streamlit as st
|
| 3 |
+
from transformers import PretrainedConfig, PreTrainedModel, T5EncoderModel, AutoTokenizer
|
| 4 |
+
import torch
|
| 5 |
+
import torch.nn as nn
|
| 6 |
+
import copy
|
| 7 |
+
|
| 8 |
+
keep_layer_count=6
|
| 9 |
+
|
| 10 |
+
class ByT5ForTextGeotaggingConfig(PretrainedConfig):
|
| 11 |
+
model_type = "byt5_for_text)geotagging"
|
| 12 |
+
|
| 13 |
+
def __init__(self, n_clusters, model_name_or_path, class_to_location=None, **kwargs):
|
| 14 |
+
super(ByT5ForTextGeotaggingConfig, self).__init__(**kwargs)
|
| 15 |
+
self.n_clusters = n_clusters
|
| 16 |
+
self.model_name_or_path = model_name_or_path
|
| 17 |
+
self.class_to_location = class_to_location or {}
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def to_diff_dict(self):
|
| 21 |
+
# Convert the configuration to a dictionary
|
| 22 |
+
config_dict = self.to_dict()
|
| 23 |
+
|
| 24 |
+
# Get the default configuration for comparison
|
| 25 |
+
default_config_dict = PretrainedConfig().to_dict()
|
| 26 |
+
|
| 27 |
+
# Return the differences
|
| 28 |
+
diff_dict = {k: v for k, v in config_dict.items() if k not in default_config_dict or v != default_config_dict[k]}
|
| 29 |
+
|
| 30 |
+
return diff_dict
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def deleteEncodingLayers(model, num_layers_to_keep): # must pass in the full bert model
|
| 34 |
+
oldModuleList = model.encoder.block
|
| 35 |
+
newModuleList = torch.nn.ModuleList()
|
| 36 |
+
|
| 37 |
+
# Now iterate over all layers, only keepign only the relevant layers.
|
| 38 |
+
for i in range(0, num_layers_to_keep):
|
| 39 |
+
newModuleList.append(oldModuleList[i])
|
| 40 |
+
|
| 41 |
+
# create a copy of the model, modify it with the new list, and return
|
| 42 |
+
copyOfModel = copy.deepcopy(model)
|
| 43 |
+
copyOfModel.encoder.block = newModuleList
|
| 44 |
+
|
| 45 |
+
return copyOfModel
|
| 46 |
+
|
| 47 |
+
class ByT5ForTextGeotagging(PreTrainedModel):
|
| 48 |
+
config_class = ByT5ForTextGeotaggingConfig
|
| 49 |
+
|
| 50 |
+
def __init__(self, config):
|
| 51 |
+
super(ByT5ForTextGeotagging, self).__init__(config)
|
| 52 |
+
|
| 53 |
+
self.byt5 = T5EncoderModel.from_pretrained(config.model_name_or_path)
|
| 54 |
+
if keep_layer_count is not None:
|
| 55 |
+
self.byt5 = deleteEncodingLayers(self.byt5, keep_layer_count)
|
| 56 |
+
|
| 57 |
+
hidden_size = self.byt5.config.d_model
|
| 58 |
+
self.fc3 = nn.Linear(hidden_size, config.n_clusters)
|
| 59 |
+
|
| 60 |
+
def forward(self, input, return_coordinates=False):
|
| 61 |
+
input = self.byt5(input[:, 0, :].squeeze(1))['last_hidden_state']
|
| 62 |
+
input = input[:, 0, :].squeeze(1)
|
| 63 |
+
logits = self.fc3(input)
|
| 64 |
+
|
| 65 |
+
if return_coordinates:
|
| 66 |
+
class_idx = torch.argmax(logits, dim=1).item()
|
| 67 |
+
coordinates = self.config.class_to_location.get(str(class_idx))
|
| 68 |
+
return logits, coordinates
|
| 69 |
+
else:
|
| 70 |
+
return logits
|
| 71 |
+
|
| 72 |
+
def geolocate_text_byt5(text):
|
| 73 |
+
input_tensor = byt5_tokenizer(text, return_tensors="pt", truncation=True, max_length=140)['input_ids']
|
| 74 |
+
logits, (lat, lon) = model(input_tensor.unsqueeze(0), return_coordinates=True)
|
| 75 |
+
return lat, lon
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
model = ByT5ForTextGeotagging.from_pretrained("byt5-geotagging-spanish")
|
| 79 |
+
|
| 80 |
+
#text = "¡Barcelona es increíble! #VacacionesEnEspaña"
|
| 81 |
+
|
| 82 |
+
# Streamlit interface
|
| 83 |
+
st.title('GeoTagging using ByT5')
|
| 84 |
+
text_input = st.text_input('Enter your text:')
|
| 85 |
+
if text_input:
|
| 86 |
+
location = geolocate_text_byt5(text_input)
|
| 87 |
+
st.write('Predicted Location: ', location)
|