Spaces:
Build error
Build error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,229 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import pickle
|
3 |
+
import cv2
|
4 |
+
import os
|
5 |
+
import numpy as np
|
6 |
+
from PIL import Image
|
7 |
+
from transformers import ViTForImageClassification, AutoImageProcessor, AdamW, ViTImageProcessor, VisionEncoderDecoderModel, AutoTokenizer
|
8 |
+
from torch.utils.data import DataLoader, TensorDataset
|
9 |
+
|
10 |
+
model_path = 'model'
|
11 |
+
train_pickle_path = 'train_data.pickle'
|
12 |
+
valid_pickle_path = 'valid_data.pickle'
|
13 |
+
image_directory = 'images'
|
14 |
+
test_image_path = 'test.jpg'
|
15 |
+
num_epochs = 5 # Fine-tune the model
|
16 |
+
label_list = ["小白", "巧巧", "冏媽", "乖狗", "花捲", "超人", "黑胖", "橘子"]
|
17 |
+
label_dictionary = {"小白": 0, "巧巧": 1, "冏媽": 2, "乖狗": 3, "花捲": 4, "超人": 5, "黑胖": 6, "橘子": 7}
|
18 |
+
num_classes = len(label_dictionary) # Adjust according to your classification task
|
19 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
20 |
+
# device = torch.device("mps")
|
21 |
+
|
22 |
+
def data_generate(dataset):
|
23 |
+
images = []
|
24 |
+
labels = []
|
25 |
+
image_processor = AutoImageProcessor.from_pretrained('google/vit-large-patch16-224-in21k')
|
26 |
+
for folder_name in os.listdir(image_directory):
|
27 |
+
folder_path = os.path.join(image_directory, folder_name)
|
28 |
+
if os.path.isdir(folder_path):
|
29 |
+
for image_file in os.listdir(folder_path):
|
30 |
+
if image_file.startswith(dataset):
|
31 |
+
image_path = os.path.join(folder_path, image_file)
|
32 |
+
# print(image_path)
|
33 |
+
|
34 |
+
img = cv2.imread(image_path)
|
35 |
+
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
36 |
+
|
37 |
+
img = Image.fromarray(img)
|
38 |
+
img = img.resize((224, 224))
|
39 |
+
inputs = image_processor(images=img, return_tensors="pt")
|
40 |
+
images.append(inputs['pixel_values'].squeeze(0).numpy())
|
41 |
+
labels.append(int(folder_name.split('_')[0]))
|
42 |
+
|
43 |
+
images = np.array(images)
|
44 |
+
labels = np.array(labels)
|
45 |
+
|
46 |
+
# Now you can pickle this data
|
47 |
+
train_data = {'img': images, 'label': labels}
|
48 |
+
with open(f'{dataset}_data.pickle', 'wb') as f:
|
49 |
+
pickle.dump(train_data, f)
|
50 |
+
|
51 |
+
def train_model():
|
52 |
+
|
53 |
+
if not os.path.exists(valid_pickle_path):
|
54 |
+
data_generate('valid')
|
55 |
+
if not os.path.exists(train_pickle_path):
|
56 |
+
data_generate('train')
|
57 |
+
|
58 |
+
# Load the train and vaild
|
59 |
+
with open("train_data.pickle", "rb") as f:
|
60 |
+
train_data = pickle.load(f)
|
61 |
+
|
62 |
+
with open("valid_data.pickle", "rb") as f:
|
63 |
+
valid_data = pickle.load(f)
|
64 |
+
|
65 |
+
# Convert the dataset into torch tensors
|
66 |
+
train_inputs = torch.tensor(train_data["img"])
|
67 |
+
train_labels = torch.tensor(train_data["label"])
|
68 |
+
valid_inputs = torch.tensor(valid_data["img"])
|
69 |
+
valid_labels = torch.tensor(valid_data["label"])
|
70 |
+
|
71 |
+
# Create the TensorDataset
|
72 |
+
train_dataset = TensorDataset(train_inputs, train_labels)
|
73 |
+
valid_dataset = TensorDataset(valid_inputs, valid_labels)
|
74 |
+
|
75 |
+
# Create the DataLoader
|
76 |
+
train_loader = DataLoader(train_dataset, batch_size=16, shuffle=True)
|
77 |
+
valid_loader = DataLoader(valid_dataset, batch_size=16, shuffle=True)
|
78 |
+
|
79 |
+
# Define the model and move it to the GPU
|
80 |
+
|
81 |
+
model = ViTForImageClassification.from_pretrained('google/vit-large-patch16-224-in21k', num_labels=num_classes)
|
82 |
+
model.to(device)
|
83 |
+
|
84 |
+
# Define the optimizer
|
85 |
+
optimizer = AdamW(model.parameters(), lr=1e-4)
|
86 |
+
|
87 |
+
for epoch in range(num_epochs):
|
88 |
+
|
89 |
+
model.train()
|
90 |
+
total_loss = 0
|
91 |
+
|
92 |
+
for i, batch in enumerate(train_loader):
|
93 |
+
|
94 |
+
# Move batch to the GPU
|
95 |
+
batch = [r.to(device) for r in batch]
|
96 |
+
|
97 |
+
# Unpack the inputs from our dataloader
|
98 |
+
inputs, labels = batch
|
99 |
+
|
100 |
+
# Clear out the gradients (by default they accumulate)
|
101 |
+
optimizer.zero_grad()
|
102 |
+
|
103 |
+
# Forward pass
|
104 |
+
outputs = model(inputs, labels=labels)
|
105 |
+
|
106 |
+
# Compute loss
|
107 |
+
loss = outputs.loss
|
108 |
+
|
109 |
+
# Backward pass
|
110 |
+
loss.backward()
|
111 |
+
|
112 |
+
|
113 |
+
# Update parameters and take a step using the computed gradient
|
114 |
+
optimizer.step()
|
115 |
+
|
116 |
+
# Update the loss
|
117 |
+
total_loss += loss.item()
|
118 |
+
|
119 |
+
# print(f'{i}/{len(train_loader)} ')
|
120 |
+
|
121 |
+
# Get the average loss for the entire epoch
|
122 |
+
avg_loss = total_loss / len(train_loader)
|
123 |
+
|
124 |
+
# Print the loss
|
125 |
+
print('Epoch:', epoch + 1, 'Training Loss:', avg_loss)
|
126 |
+
|
127 |
+
|
128 |
+
# Evaluate the model on the validation set
|
129 |
+
model.eval()
|
130 |
+
total_correct = 0
|
131 |
+
|
132 |
+
for batch in valid_loader:
|
133 |
+
# Move batch to the GPU
|
134 |
+
batch = [t.to(device) for t in batch]
|
135 |
+
|
136 |
+
# Unpack the inputs from our dataloader
|
137 |
+
inputs, labels = batch
|
138 |
+
|
139 |
+
# Forward pass
|
140 |
+
with torch.no_grad():
|
141 |
+
outputs = model(inputs)
|
142 |
+
|
143 |
+
# Get the predictions
|
144 |
+
predictions = torch.argmax(outputs.logits, dim=1)
|
145 |
+
|
146 |
+
# Update the total correct
|
147 |
+
total_correct += torch.sum(predictions == labels)
|
148 |
+
|
149 |
+
# Calculate the accuracy
|
150 |
+
accuracy = total_correct / len(valid_dataset)
|
151 |
+
print('Validation accuracy:', accuracy.item())
|
152 |
+
|
153 |
+
model.save_pretrained("model")
|
154 |
+
|
155 |
+
def predict():
|
156 |
+
# Load the model
|
157 |
+
model = ViTForImageClassification.from_pretrained(model_path, num_labels=num_classes)
|
158 |
+
|
159 |
+
image_processor = AutoImageProcessor.from_pretrained('google/vit-large-patch16-224-in21k')
|
160 |
+
|
161 |
+
|
162 |
+
# Load the test data
|
163 |
+
# Load the image
|
164 |
+
|
165 |
+
img = cv2.imread(test_image_path)
|
166 |
+
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
167 |
+
|
168 |
+
# Resize the image to 224x224 pixels
|
169 |
+
img = Image.fromarray(img)
|
170 |
+
img = img.resize((224, 224))
|
171 |
+
|
172 |
+
# img to tensor
|
173 |
+
# Preprocess the image and generate features
|
174 |
+
inputs = image_processor(images=img, return_tensors="pt")
|
175 |
+
outputs = model(**inputs)
|
176 |
+
logits = outputs.logits
|
177 |
+
|
178 |
+
probabilities = torch.nn.functional.softmax(logits, dim=-1)
|
179 |
+
predicted_class_idx = logits.argmax(-1).item()
|
180 |
+
|
181 |
+
return label_list[predicted_class_idx] if probabilities.max().item() > 0.90 else '不是校狗'
|
182 |
+
|
183 |
+
def captioning():
|
184 |
+
|
185 |
+
model = VisionEncoderDecoderModel.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
|
186 |
+
feature_extractor = ViTImageProcessor.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
|
187 |
+
tokenizer = AutoTokenizer.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
|
188 |
+
|
189 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
190 |
+
model.to(device)
|
191 |
+
|
192 |
+
max_length = 16
|
193 |
+
num_beams = 4
|
194 |
+
gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
|
195 |
+
|
196 |
+
images = []
|
197 |
+
for image_path in [test_image_path]:
|
198 |
+
i_image = Image.open(image_path)
|
199 |
+
if i_image.mode != "RGB":
|
200 |
+
i_image = i_image.convert(mode="RGB")
|
201 |
+
|
202 |
+
images.append(i_image)
|
203 |
+
|
204 |
+
pixel_values = feature_extractor(images=images, return_tensors="pt").pixel_values
|
205 |
+
pixel_values = pixel_values.to(device)
|
206 |
+
|
207 |
+
output_ids = model.generate(pixel_values, **gen_kwargs)
|
208 |
+
|
209 |
+
preds = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
210 |
+
preds = [pred.strip() for pred in preds]
|
211 |
+
return preds[-1]
|
212 |
+
|
213 |
+
def output(predict_class, caption):
|
214 |
+
conj = ['are', 'is', 'dog']
|
215 |
+
if predict_class == '不是校狗' or caption.find('dog') == -1:
|
216 |
+
print(f'{caption} ({predict_class})')
|
217 |
+
else:
|
218 |
+
for c in conj:
|
219 |
+
if caption.find(c) != -1:
|
220 |
+
print(f'{predict_class} is{caption[caption.find(c) + len(c):]}')
|
221 |
+
return
|
222 |
+
print(f'{caption} ({predict_class})')
|
223 |
+
|
224 |
+
|
225 |
+
if __name__ == '__main__':
|
226 |
+
|
227 |
+
if not os.path.exists(model_path):
|
228 |
+
train_model()
|
229 |
+
output(predict(), captioning())
|