Spaces:
Sleeping
Sleeping
abhishekrs4
commited on
Commit
•
d0ba6da
1
Parent(s):
39371b7
added flask backend app
Browse files
app.py
ADDED
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import cv2
|
3 |
+
import json
|
4 |
+
import torch
|
5 |
+
import logging
|
6 |
+
|
7 |
+
import numpy as np
|
8 |
+
from flask import Flask, jsonify, request
|
9 |
+
from flask.wrappers import Response
|
10 |
+
|
11 |
+
from iam_line_recognition.model_main import CRNN
|
12 |
+
from iam_line_recognition.utils import ctc_decode
|
13 |
+
from iam_line_recognition.dataset import HWRecogIAMDataset
|
14 |
+
|
15 |
+
app = Flask("IAM_Handwriting_Recognition")
|
16 |
+
logging.basicConfig(level=logging.INFO)
|
17 |
+
|
18 |
+
file_model_local = "artifacts/crnn_H_32_W_768_E_196.pth"
|
19 |
+
file_model_cont = "/data/models/crnn_H_32_W_768_E_196.pth"
|
20 |
+
device = "cpu"
|
21 |
+
num_classes = len(HWRecogIAMDataset.LABEL_2_CHAR) + 1
|
22 |
+
image_height = 32
|
23 |
+
mean_arr = np.array([[0.485, 0.456, 0.406]])
|
24 |
+
std_arr = np.array([[0.229, 0.224, 0.225]])
|
25 |
+
hw_recog_model = CRNN(num_classes, image_height)
|
26 |
+
|
27 |
+
try:
|
28 |
+
logging.info(f"loading model from {file_model_local}")
|
29 |
+
hw_recog_model.load_state_dict(
|
30 |
+
torch.load(file_model_local, map_location=device)
|
31 |
+
)
|
32 |
+
except:
|
33 |
+
logging.info(f"loading model from {file_model_cont}")
|
34 |
+
hw_recog_model.load_state_dict(
|
35 |
+
torch.load(file_model_cont, map_location=device)
|
36 |
+
)
|
37 |
+
hw_recog_model.to(device)
|
38 |
+
hw_recog_model.eval()
|
39 |
+
|
40 |
+
|
41 |
+
def predict_hw(img_test: np.ndarray) -> str:
|
42 |
+
img_test = np.expand_dims(img_test, 0)
|
43 |
+
img_test = img_test.astype(np.float32) / 255.0
|
44 |
+
img_test = (img_test - mean_arr) / std_arr
|
45 |
+
img_test = np.transpose(img_test, axes=[0, 3, 1, 2])
|
46 |
+
img_tensor = torch.tensor(img_test).float()
|
47 |
+
img_tensor = img_tensor.to(device, dtype=torch.float)
|
48 |
+
log_probs = hw_recog_model(img_tensor)
|
49 |
+
pred_labels = ctc_decode(log_probs.detach())
|
50 |
+
str_pred = [HWRecogIAMDataset.LABEL_2_CHAR[i] for i in pred_labels[0]]
|
51 |
+
str_pred = "".join(str_pred)
|
52 |
+
return str_pred
|
53 |
+
|
54 |
+
@app.route("/predict", methods=["POST"])
|
55 |
+
def predict() -> Response:
|
56 |
+
logging.info("IAM Handwriting recognition app")
|
57 |
+
img_file = request.files["image_file"]
|
58 |
+
try:
|
59 |
+
img_arr = np.fromstring(img_file.read(), np.uint8)
|
60 |
+
except:
|
61 |
+
img_arr = np.fromstring(img_file.getvalue(), np.uint8)
|
62 |
+
img_dec = cv2.imdecode(img_arr, cv2.IMREAD_COLOR)
|
63 |
+
img_dec = cv2.cvtColor(img_dec, cv2.COLOR_BGR2RGB)
|
64 |
+
|
65 |
+
img_dec = cv2.resize(img_dec, (768, 32), interpolation = cv2.INTER_LINEAR)
|
66 |
+
|
67 |
+
str_pred = predict_hw(img_dec)
|
68 |
+
|
69 |
+
dict_pred = {
|
70 |
+
"file_name": img_file.filename,
|
71 |
+
"prediction": str_pred,
|
72 |
+
}
|
73 |
+
try:
|
74 |
+
json_pred = jsonify(dict_pred)
|
75 |
+
except TypeError as e:
|
76 |
+
json_pred = jsonify({"error": str(e)})
|
77 |
+
logging.info(json_pred)
|
78 |
+
return json_pred
|
79 |
+
|
80 |
+
if __name__ == "__main__":
|
81 |
+
app.run(host="0.0.0.0", debug=True, port=7860)
|