Spaces:
Build error
Build error
autonomous019
commited on
Commit
•
79e0b51
1
Parent(s):
2ac8087
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,153 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import ViTConfig, ViTForImageClassification
|
2 |
+
from transformers import ViTFeatureExtractor
|
3 |
+
from PIL import Image
|
4 |
+
import requests
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
import gradio as gr
|
7 |
+
from gradio.mix import Parallel
|
8 |
+
from transformers import ImageClassificationPipeline, PerceiverForImageClassificationConvProcessing, PerceiverFeatureExtractor
|
9 |
+
from transformers import VisionEncoderDecoderModel
|
10 |
+
from transformers import AutoTokenizer
|
11 |
+
import torch
|
12 |
+
from transformers import (
|
13 |
+
AutoModelForCausalLM,
|
14 |
+
LogitsProcessorList,
|
15 |
+
MinLengthLogitsProcessor,
|
16 |
+
StoppingCriteriaList,
|
17 |
+
MaxLengthCriteria,
|
18 |
+
)
|
19 |
+
|
20 |
+
# https://github.com/NielsRogge/Transformers-Tutorials/blob/master/HuggingFace_vision_ecosystem_overview_(June_2022).ipynb
|
21 |
+
# option 1: load with randomly initialized weights (train from scratch)
|
22 |
+
|
23 |
+
#tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-j-6B")
|
24 |
+
#model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-j-6B")
|
25 |
+
|
26 |
+
|
27 |
+
config = ViTConfig(num_hidden_layers=12, hidden_size=768)
|
28 |
+
model = ViTForImageClassification(config)
|
29 |
+
|
30 |
+
#print(config)
|
31 |
+
|
32 |
+
feature_extractor = ViTFeatureExtractor()
|
33 |
+
# or, to load one that corresponds to a checkpoint on the hub:
|
34 |
+
#feature_extractor = ViTFeatureExtractor.from_pretrained("google/vit-base-patch16-224")
|
35 |
+
|
36 |
+
#the following gets called by classify_image()
|
37 |
+
feature_extractor = PerceiverFeatureExtractor.from_pretrained("deepmind/vision-perceiver-conv")
|
38 |
+
model = PerceiverForImageClassificationConvProcessing.from_pretrained("deepmind/vision-perceiver-conv")
|
39 |
+
#google/vit-base-patch16-224, deepmind/vision-perceiver-conv
|
40 |
+
image_pipe = ImageClassificationPipeline(model=model, feature_extractor=feature_extractor)
|
41 |
+
|
42 |
+
def create_story(text_seed):
|
43 |
+
#tokenizer = AutoTokenizer.from_pretrained("gpt2")
|
44 |
+
#model = AutoModelForCausalLM.from_pretrained("gpt2")
|
45 |
+
|
46 |
+
#eleutherAI gpt-3 based
|
47 |
+
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neo-125M")
|
48 |
+
model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-neo-125M")
|
49 |
+
|
50 |
+
# set pad_token_id to eos_token_id because GPT2 does not have a EOS token
|
51 |
+
model.config.pad_token_id = model.config.eos_token_id
|
52 |
+
|
53 |
+
#input_prompt = "It might be possible to"
|
54 |
+
input_prompt = text_seed
|
55 |
+
input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids
|
56 |
+
|
57 |
+
# instantiate logits processors
|
58 |
+
logits_processor = LogitsProcessorList(
|
59 |
+
[
|
60 |
+
MinLengthLogitsProcessor(10, eos_token_id=model.config.eos_token_id),
|
61 |
+
]
|
62 |
+
)
|
63 |
+
stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=100)])
|
64 |
+
|
65 |
+
outputs = model.greedy_search(
|
66 |
+
input_ids, logits_processor=logits_processor, stopping_criteria=stopping_criteria
|
67 |
+
)
|
68 |
+
|
69 |
+
result_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
70 |
+
return result_text
|
71 |
+
|
72 |
+
|
73 |
+
|
74 |
+
|
75 |
+
|
76 |
+
|
77 |
+
def self_caption(image):
|
78 |
+
repo_name = "ydshieh/vit-gpt2-coco-en"
|
79 |
+
#test_image = "cats.jpg"
|
80 |
+
test_image = image
|
81 |
+
#url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
|
82 |
+
#test_image = Image.open(requests.get(url, stream=True).raw)
|
83 |
+
#test_image.save("cats.png")
|
84 |
+
|
85 |
+
feature_extractor2 = ViTFeatureExtractor.from_pretrained(repo_name)
|
86 |
+
tokenizer = AutoTokenizer.from_pretrained(repo_name)
|
87 |
+
model2 = VisionEncoderDecoderModel.from_pretrained(repo_name)
|
88 |
+
pixel_values = feature_extractor2(test_image, return_tensors="pt").pixel_values
|
89 |
+
print("Pixel Values")
|
90 |
+
print(pixel_values)
|
91 |
+
# autoregressively generate text (using beam search or other decoding strategy)
|
92 |
+
generated_ids = model2.generate(pixel_values, max_length=16, num_beams=4, return_dict_in_generate=True)
|
93 |
+
|
94 |
+
# decode into text
|
95 |
+
preds = tokenizer.batch_decode(generated_ids[0], skip_special_tokens=True)
|
96 |
+
preds = [pred.strip() for pred in preds]
|
97 |
+
print("Predictions")
|
98 |
+
print(preds)
|
99 |
+
print("The preds type is : ",type(preds))
|
100 |
+
pred_keys = ["Prediction"]
|
101 |
+
pred_value = preds
|
102 |
+
|
103 |
+
pred_dictionary = dict(zip(pred_keys, pred_value))
|
104 |
+
print("Pred dictionary")
|
105 |
+
print(pred_dictionary)
|
106 |
+
#return(pred_dictionary)
|
107 |
+
preds = ' '.join(preds)
|
108 |
+
story = create_story(preds)
|
109 |
+
story = ' '.join(story)
|
110 |
+
return story
|
111 |
+
|
112 |
+
|
113 |
+
def classify_image(image):
|
114 |
+
results = image_pipe(image)
|
115 |
+
|
116 |
+
print("RESULTS")
|
117 |
+
print(results)
|
118 |
+
# convert to format Gradio expects
|
119 |
+
output = {}
|
120 |
+
for prediction in results:
|
121 |
+
predicted_label = prediction['label']
|
122 |
+
score = prediction['score']
|
123 |
+
output[predicted_label] = score
|
124 |
+
print("OUTPUT")
|
125 |
+
print(output)
|
126 |
+
return output
|
127 |
+
|
128 |
+
|
129 |
+
image = gr.inputs.Image(type="pil")
|
130 |
+
label = gr.outputs.Label(num_top_classes=5)
|
131 |
+
examples = [ ["cats.jpg"], ["batter.jpg"],["drinkers.jpg"] ]
|
132 |
+
title = "Generate a Story from an Image"
|
133 |
+
description = "Demo for classifying images with Perceiver IO. To use it, simply upload an image and click 'submit', a story is autogenerated as well"
|
134 |
+
article = "<p style='text-align: center'></p>"
|
135 |
+
|
136 |
+
img_info1 = gr.Interface(
|
137 |
+
fn=classify_image,
|
138 |
+
inputs=image,
|
139 |
+
outputs=label,
|
140 |
+
)
|
141 |
+
|
142 |
+
img_info2 = gr.Interface(
|
143 |
+
fn=self_caption,
|
144 |
+
inputs=image,
|
145 |
+
#outputs=label,
|
146 |
+
outputs = [
|
147 |
+
gr.outputs.Textbox(label = 'Story')
|
148 |
+
],
|
149 |
+
)
|
150 |
+
|
151 |
+
Parallel(img_info1,img_info2, inputs=image, title=title, description=description, examples=examples, enable_queue=True).launch(debug=True)
|
152 |
+
#Parallel(img_info1,img_info2, inputs=image, outputs=label, title=title, description=description, examples=examples, enable_queue=True).launch(debug=True)
|
153 |
+
|