Spaces:
Running
Running
Add sample image return functionality to app
Browse filespulls sample image from EOL subset of TreeOfLife-10M
app.py
CHANGED
@@ -6,12 +6,14 @@ import logging
|
|
6 |
|
7 |
import gradio as gr
|
8 |
import numpy as np
|
|
|
9 |
import torch
|
10 |
import torch.nn.functional as F
|
11 |
from open_clip import create_model, get_tokenizer
|
12 |
from torchvision import transforms
|
13 |
|
14 |
from templates import openai_imagenet_template
|
|
|
15 |
|
16 |
log_format = "[%(asctime)s] [%(levelname)s] [%(name)s] %(message)s"
|
17 |
logging.basicConfig(level=logging.INFO, format=log_format)
|
@@ -19,6 +21,12 @@ logger = logging.getLogger()
|
|
19 |
|
20 |
hf_token = os.getenv("HF_TOKEN")
|
21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
model_str = "hf-hub:imageomics/bioclip"
|
23 |
tokenizer_str = "ViT-B-16"
|
24 |
|
@@ -123,12 +131,14 @@ def format_name(taxon, common):
|
|
123 |
|
124 |
|
125 |
@torch.no_grad()
|
126 |
-
def open_domain_classification(img, rank: int
|
127 |
"""
|
128 |
Predicts from the entire tree of life.
|
129 |
If targeting a higher rank than species, then this function predicts among all
|
130 |
species, then sums up species-level probabilities for the given rank.
|
131 |
"""
|
|
|
|
|
132 |
img = preprocess_img(img).to(device)
|
133 |
img_features = model.encode_image(img.unsqueeze(0))
|
134 |
img_features = F.normalize(img_features, dim=-1)
|
@@ -136,21 +146,36 @@ def open_domain_classification(img, rank: int) -> dict[str, float]:
|
|
136 |
logits = (model.logit_scale.exp() * img_features @ txt_emb).squeeze()
|
137 |
probs = F.softmax(logits, dim=0)
|
138 |
|
139 |
-
# If predicting species, no need to sum probabilities.
|
140 |
if rank + 1 == len(ranks):
|
141 |
topk = probs.topk(k)
|
142 |
-
|
143 |
format_name(*txt_names[i]): prob for i, prob in zip(topk.indices, topk.values)
|
144 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
145 |
|
146 |
-
# Sum up by the rank
|
147 |
output = collections.defaultdict(float)
|
148 |
for i in torch.nonzero(probs > min_prob).squeeze():
|
149 |
output[" ".join(txt_names[i][0][: rank + 1])] += probs[i]
|
150 |
|
151 |
topk_names = heapq.nlargest(k, output, key=output.get)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
|
153 |
-
|
|
|
|
|
154 |
|
155 |
|
156 |
def change_output(choice):
|
@@ -179,9 +204,22 @@ if __name__ == "__main__":
|
|
179 |
status_msg = f"{done}/{total} ({done / total * 100:.1f}%) indexed"
|
180 |
|
181 |
with gr.Blocks() as app:
|
182 |
-
|
183 |
-
|
184 |
with gr.Tab("Open-Ended"):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
185 |
with gr.Row():
|
186 |
with gr.Column():
|
187 |
rank_dropdown = gr.Dropdown(
|
@@ -199,32 +237,36 @@ if __name__ == "__main__":
|
|
199 |
show_label=True,
|
200 |
value=None,
|
201 |
)
|
202 |
-
|
203 |
|
204 |
with gr.Row():
|
205 |
gr.Examples(
|
206 |
examples=open_domain_examples,
|
207 |
inputs=[img_input, rank_dropdown],
|
208 |
cache_examples=True,
|
209 |
-
fn=open_domain_classification,
|
210 |
outputs=[open_domain_output],
|
211 |
)
|
212 |
-
|
213 |
-
#
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
|
|
|
227 |
with gr.Tab("Zero-Shot"):
|
|
|
|
|
|
|
228 |
with gr.Row():
|
229 |
with gr.Column():
|
230 |
classes_txt = gr.Textbox(
|
@@ -240,43 +282,44 @@ if __name__ == "__main__":
|
|
240 |
zero_shot_output = gr.Label(
|
241 |
num_top_classes=k, label="Prediction", show_label=True
|
242 |
)
|
243 |
-
|
244 |
|
245 |
with gr.Row():
|
246 |
gr.Examples(
|
247 |
examples=zero_shot_examples,
|
248 |
-
inputs=[
|
249 |
cache_examples=True,
|
250 |
fn=zero_shot_classification,
|
251 |
outputs=[zero_shot_output],
|
252 |
)
|
253 |
-
|
254 |
-
#
|
255 |
-
|
256 |
-
|
257 |
-
|
258 |
-
|
259 |
-
|
260 |
-
|
261 |
-
|
262 |
-
|
263 |
-
|
264 |
-
|
265 |
-
|
266 |
-
|
|
|
267 |
rank_dropdown.change(
|
268 |
fn=change_output, inputs=rank_dropdown, outputs=[open_domain_output]
|
269 |
)
|
270 |
|
271 |
open_domain_btn.click(
|
272 |
-
fn=open_domain_classification,
|
273 |
inputs=[img_input, rank_dropdown],
|
274 |
-
outputs=[open_domain_output],
|
275 |
)
|
276 |
|
277 |
zero_shot_btn.click(
|
278 |
fn=zero_shot_classification,
|
279 |
-
inputs=[
|
280 |
outputs=zero_shot_output,
|
281 |
)
|
282 |
|
@@ -291,4 +334,4 @@ if __name__ == "__main__":
|
|
291 |
)
|
292 |
|
293 |
app.queue(max_size=20)
|
294 |
-
app.launch()
|
|
|
6 |
|
7 |
import gradio as gr
|
8 |
import numpy as np
|
9 |
+
import polars as pl
|
10 |
import torch
|
11 |
import torch.nn.functional as F
|
12 |
from open_clip import create_model, get_tokenizer
|
13 |
from torchvision import transforms
|
14 |
|
15 |
from templates import openai_imagenet_template
|
16 |
+
from components.query import get_sample
|
17 |
|
18 |
log_format = "[%(asctime)s] [%(levelname)s] [%(name)s] %(message)s"
|
19 |
logging.basicConfig(level=logging.INFO, format=log_format)
|
|
|
21 |
|
22 |
hf_token = os.getenv("HF_TOKEN")
|
23 |
|
24 |
+
# For sample images
|
25 |
+
METADATA_PATH = "components/metadata.csv"
|
26 |
+
# Read page ID as int and filter out smaller ablation duplicated training split
|
27 |
+
metadata_df = pl.read_csv(METADATA_PATH, low_memory = False)
|
28 |
+
metadata_df = metadata_df.with_columns(pl.col("eol_page_id").cast(pl.Int64))
|
29 |
+
|
30 |
model_str = "hf-hub:imageomics/bioclip"
|
31 |
tokenizer_str = "ViT-B-16"
|
32 |
|
|
|
131 |
|
132 |
|
133 |
@torch.no_grad()
|
134 |
+
def open_domain_classification(img, rank: int, return_all=False):
|
135 |
"""
|
136 |
Predicts from the entire tree of life.
|
137 |
If targeting a higher rank than species, then this function predicts among all
|
138 |
species, then sums up species-level probabilities for the given rank.
|
139 |
"""
|
140 |
+
|
141 |
+
logger.info(f"Starting open domain classification for rank: {rank}")
|
142 |
img = preprocess_img(img).to(device)
|
143 |
img_features = model.encode_image(img.unsqueeze(0))
|
144 |
img_features = F.normalize(img_features, dim=-1)
|
|
|
146 |
logits = (model.logit_scale.exp() * img_features @ txt_emb).squeeze()
|
147 |
probs = F.softmax(logits, dim=0)
|
148 |
|
|
|
149 |
if rank + 1 == len(ranks):
|
150 |
topk = probs.topk(k)
|
151 |
+
prediction_dict = {
|
152 |
format_name(*txt_names[i]): prob for i, prob in zip(topk.indices, topk.values)
|
153 |
}
|
154 |
+
logger.info(f"Top K predictions: {prediction_dict}")
|
155 |
+
top_prediction_name = format_name(*txt_names[topk.indices[0]]).split("(")[0]
|
156 |
+
logger.info(f"Top prediction name: {top_prediction_name}")
|
157 |
+
sample_img, taxon_url = get_sample(metadata_df, top_prediction_name, rank)
|
158 |
+
if return_all:
|
159 |
+
return prediction_dict, sample_img, taxon_url
|
160 |
+
return prediction_dict
|
161 |
|
|
|
162 |
output = collections.defaultdict(float)
|
163 |
for i in torch.nonzero(probs > min_prob).squeeze():
|
164 |
output[" ".join(txt_names[i][0][: rank + 1])] += probs[i]
|
165 |
|
166 |
topk_names = heapq.nlargest(k, output, key=output.get)
|
167 |
+
prediction_dict = {name: output[name] for name in topk_names}
|
168 |
+
logger.info(f"Top K names for output: {topk_names}")
|
169 |
+
logger.info(f"Prediction dictionary: {prediction_dict}")
|
170 |
+
|
171 |
+
top_prediction_name = topk_names[0]
|
172 |
+
logger.info(f"Top prediction name: {top_prediction_name}")
|
173 |
+
sample_img, taxon_url = get_sample(metadata_df, top_prediction_name, rank)
|
174 |
+
logger.info(f"Sample image and taxon URL: {sample_img}, {taxon_url}")
|
175 |
|
176 |
+
if return_all:
|
177 |
+
return prediction_dict, sample_img, taxon_url
|
178 |
+
return prediction_dict
|
179 |
|
180 |
|
181 |
def change_output(choice):
|
|
|
204 |
status_msg = f"{done}/{total} ({done / total * 100:.1f}%) indexed"
|
205 |
|
206 |
with gr.Blocks() as app:
|
207 |
+
|
|
|
208 |
with gr.Tab("Open-Ended"):
|
209 |
+
with gr.Row(variant = "panel", elem_id = "images_panel"):
|
210 |
+
with gr.Column():
|
211 |
+
img_input = gr.Image(height = 400, sources=["upload"])
|
212 |
+
|
213 |
+
with gr.Column():
|
214 |
+
# display sample image of top predicted taxon
|
215 |
+
sample_img = gr.Image(label = "Sample Image of Predicted Taxon",
|
216 |
+
height = 400,
|
217 |
+
show_download_button = False)
|
218 |
+
|
219 |
+
taxon_url = gr.HTML(label = "More Information",
|
220 |
+
elem_id = "url"
|
221 |
+
)
|
222 |
+
|
223 |
with gr.Row():
|
224 |
with gr.Column():
|
225 |
rank_dropdown = gr.Dropdown(
|
|
|
237 |
show_label=True,
|
238 |
value=None,
|
239 |
)
|
240 |
+
# open_domain_flag_btn = gr.Button("Flag Mistake", variant="primary")
|
241 |
|
242 |
with gr.Row():
|
243 |
gr.Examples(
|
244 |
examples=open_domain_examples,
|
245 |
inputs=[img_input, rank_dropdown],
|
246 |
cache_examples=True,
|
247 |
+
fn=lambda img, rank: open_domain_classification(img, rank, return_all=False),
|
248 |
outputs=[open_domain_output],
|
249 |
)
|
250 |
+
'''
|
251 |
+
# Flagging Code
|
252 |
+
open_domain_callback = gr.HuggingFaceDatasetSaver(
|
253 |
+
hf_token, "bioclip-demo-open-domain-mistakes", private=True
|
254 |
+
)
|
255 |
+
open_domain_callback.setup(
|
256 |
+
[img_input, rank_dropdown, open_domain_output],
|
257 |
+
flagging_dir="bioclip-demo-open-domain-mistakes/logs/flagged",
|
258 |
+
)
|
259 |
+
open_domain_flag_btn.click(
|
260 |
+
lambda *args: open_domain_callback.flag(args),
|
261 |
+
[img_input, rank_dropdown, open_domain_output],
|
262 |
+
None,
|
263 |
+
preprocess=False,
|
264 |
+
)
|
265 |
+
'''
|
266 |
with gr.Tab("Zero-Shot"):
|
267 |
+
with gr.Row():
|
268 |
+
img_input_zs = gr.Image(height = 400, sources=["upload"])
|
269 |
+
|
270 |
with gr.Row():
|
271 |
with gr.Column():
|
272 |
classes_txt = gr.Textbox(
|
|
|
282 |
zero_shot_output = gr.Label(
|
283 |
num_top_classes=k, label="Prediction", show_label=True
|
284 |
)
|
285 |
+
# zero_shot_flag_btn = gr.Button("Flag Mistake", variant="primary")
|
286 |
|
287 |
with gr.Row():
|
288 |
gr.Examples(
|
289 |
examples=zero_shot_examples,
|
290 |
+
inputs=[img_input_zs, classes_txt],
|
291 |
cache_examples=True,
|
292 |
fn=zero_shot_classification,
|
293 |
outputs=[zero_shot_output],
|
294 |
)
|
295 |
+
'''
|
296 |
+
# Flagging Code
|
297 |
+
zero_shot_callback = gr.HuggingFaceDatasetSaver(
|
298 |
+
hf_token, "bioclip-demo-zero-shot-mistakes", private=True
|
299 |
+
)
|
300 |
+
zero_shot_callback.setup(
|
301 |
+
[img_input, zero_shot_output], flagging_dir="bioclip-demo-zero-shot-mistakes/logs/flagged"
|
302 |
+
)
|
303 |
+
zero_shot_flag_btn.click(
|
304 |
+
lambda *args: zero_shot_callback.flag(args),
|
305 |
+
[img_input, zero_shot_output],
|
306 |
+
None,
|
307 |
+
preprocess=False,
|
308 |
+
)
|
309 |
+
'''
|
310 |
rank_dropdown.change(
|
311 |
fn=change_output, inputs=rank_dropdown, outputs=[open_domain_output]
|
312 |
)
|
313 |
|
314 |
open_domain_btn.click(
|
315 |
+
fn=lambda img, rank: open_domain_classification(img, rank, return_all=True),
|
316 |
inputs=[img_input, rank_dropdown],
|
317 |
+
outputs=[open_domain_output, sample_img, taxon_url],
|
318 |
)
|
319 |
|
320 |
zero_shot_btn.click(
|
321 |
fn=zero_shot_classification,
|
322 |
+
inputs=[img_input_zs, classes_txt],
|
323 |
outputs=zero_shot_output,
|
324 |
)
|
325 |
|
|
|
334 |
)
|
335 |
|
336 |
app.queue(max_size=20)
|
337 |
+
app.launch(share=True)
|