import matplotlib.pyplot as plt import numpy as np from six import BytesIO from PIL import Image import tensorflow as tf from object_detection.utils import label_map_util from object_detection.utils import visualization_utils as viz_utils from object_detection.utils import ops as utils_op import tarfile import wget import gradio as gr from huggingface_hub import snapshot_download import os import cv2 from tqdm import tqdm PATH_TO_LABELS = 'data/label_map.pbtxt' category_index = label_map_util.create_category_index_from_labelmap(PATH_TO_LABELS, use_display_name=True) def pil_image_as_numpy_array(pilimg): img_array = tf.keras.utils.img_to_array(pilimg) img_array = np.expand_dims(img_array, axis=0) return img_array def load_image_into_numpy_array(path): image = None image_data = tf.io.gfile.GFile(path, 'rb').read() image = Image.open(BytesIO(image_data)) return pil_image_as_numpy_array(image) def load_model(model_repo_id): download_dir = snapshot_download(model_repo_id) saved_model_dir = os.path.join(download_dir, "saved_model") detection_model = tf.saved_model.load(saved_model_dir) return detection_model def predict(pilimg,Threshold): image_np = pil_image_as_numpy_array(pilimg) if Threshold is None or Threshold == 0: Threshold=threshold_d else: Threshold= float(Threshold) return predict2(image_np,Threshold),predict3(image_np,Threshold),Threshold def predict2(image_np,Threshold): results = detection_model(image_np) # if Threshold is None or Threshold == 0: # Threshold=threshold_d # different object detection models have additional results result = {key:value.numpy() for key,value in results.items()} label_id_offset = 0 image_np_with_detections = image_np.copy() viz_utils.visualize_boxes_and_labels_on_image_array( image_np_with_detections[0], result['detection_boxes'][0], (result['detection_classes'][0] + label_id_offset).astype(int), result['detection_scores'][0], category_index, use_normalized_coordinates=True, max_boxes_to_draw=20, min_score_thresh=Threshold,#0.38, agnostic_mode=False, line_thickness=2) result_pil_img2 = tf.keras.utils.array_to_img(image_np_with_detections[0]) return result_pil_img2 def predict3(image_np,Threshold): results = detection_model2(image_np) # different object detection models have additional results result = {key:value.numpy() for key,value in results.items()} label_id_offset = 0 image_np_with_detections = image_np.copy() viz_utils.visualize_boxes_and_labels_on_image_array( image_np_with_detections[0], result['detection_boxes'][0], (result['detection_classes'][0] + label_id_offset).astype(int), result['detection_scores'][0], category_index, use_normalized_coordinates=True, max_boxes_to_draw=20, min_score_thresh=Threshold,#.38, agnostic_mode=False, line_thickness=2) result_pil_img4 = tf.keras.utils.array_to_img(image_np_with_detections[0]) return result_pil_img4 # def detect_video(video): # # Create a video capture object # cap = cv2.VideoCapture(video) # # Process frames in a loop # while cap.isOpened(): # ret, frame = cap.read() # if not ret: # break # # Expand dimensions since model expects images to have shape: [1, None, None, 3] # image_np_expanded = np.expand_dims(frame, axis=0) # # Run inference # output_dict = model(image_np_expanded) # # Extract detections # boxes = output_dict['detection_boxes'][0].numpy() # scores = output_dict['detection_scores'][0].numpy() # classes = output_dict['detection_classes'][0].numpy().astype(np.int64) # # Draw bounding boxes and labels # image_np_with_detections = viz_utils.visualize_boxes_and_labels_on_image_array( # frame, # boxes, # classes, # scores, # category_index, # use_normalized_coordinates=True, # max_boxes_to_draw=20, # min_score_thresh=.5, # agnostic_mode=False) # # Yield the processed frame # yield image_np_with_detections # # Release resources # cap.release() a = os.path.join(os.path.dirname(__file__), "data/c_base_detected.mp4") # Video b = os.path.join(os.path.dirname(__file__), "data/c_tuned_detected.mp4") # Video # def video_demo(video1, video2): # return [video1, video2] label_id_offset = 0 threshold_d= 0.38 REPO_ID = "apailang/mytfodmodel" detection_model = load_model(REPO_ID) REPO_ID2 = "apailang/mytfodmodeltuned" detection_model2 = load_model(REPO_ID2) samples_folder = 'data' # pil_image = Image.open(image_path) # image_arr = pil_image_as_numpy_array(pil_image) # predicted_img = predict(image_arr) # predicted_img.save('predicted.jpg') test1 = os.path.join(os.path.dirname(__file__), "data/test1.jpeg") test2 = os.path.join(os.path.dirname(__file__), "data/test2.jpeg") test3 = os.path.join(os.path.dirname(__file__), "data/test3.jpeg") test4 = os.path.join(os.path.dirname(__file__), "data/test4.jpeg") test5 = os.path.join(os.path.dirname(__file__), "data/test5.jpeg") test6 = os.path.join(os.path.dirname(__file__), "data/test6.jpeg") test7 = os.path.join(os.path.dirname(__file__), "data/test7.jpeg") test8 = os.path.join(os.path.dirname(__file__), "data/test8.jpeg") test9 = os.path.join(os.path.dirname(__file__), "data/test9.jpeg") test10 = os.path.join(os.path.dirname(__file__), "data/test10.jpeg") test11 = os.path.join(os.path.dirname(__file__), "data/test11.jpeg") test12 = os.path.join(os.path.dirname(__file__), "data/test12.jpeg") base_image = gr.Interface( fn=predict, # inputs=[gr.Image(type="pil"),gr.Slider(minimum=0.01, maximum=1, value=0.38 ,label="Threshold",info="[not in used]to set prediction confidence threshold")], inputs=[gr.Image(type="pil"),gr.Slider(minimum=0.05, maximum=1,step=0.05,value=threshold_d ,label="To change default 0.38 prediction confidence Threshold. Range 0.05 to 1",info="Select any image below to start, you may amend threshold after first inference")], outputs=[gr.Image(type="pil",label="Base Model Inference"),gr.Image(type="pil",label="Tuned Model Inference"),gr.Textbox(label="Both images inferenced threshold")], title="Luffy and Chopper Head detection. SSD mobile net V2 320x320 trained with animated characters only", description="Upload a Image for prediction or click on below examples. Prediction confident is defaut to >38%, you may adjust after first inference", examples= [[test1],[test2],[test3],[test4],[test5],[test6],[test7],[test8],[test9],[test10],[test11],[test12],], cache_examples=True,examples_per_page=12 #,label="select image with 0.38 threshold to inference, you may amend threshold after inference" ) # tuned_image = gr.Interface( # fn=predict3, # inputs=gr.Image(type="pil"), # outputs=gr.Image(type="pil"), # title="Luffy and Chopper face detection on images. Result comparison of base vs tuned SSD mobile net V2 320x320", # description="Upload a Image for prediction or click on below examples. Mobile net tuned with data Augmentation. Prediction confident >38%", # examples=[[test1],[test2],[test3],[test4],[test5],[test6],[test7],[test8],[test9],[test10],[test11],[test12],], # cache_examples=True # )#.launch(share=True) # a = os.path.join(os.path.dirname(__file__), "data/a.mp4") # Video # b = os.path.join(os.path.dirname(__file__), "data/b.mp4") # Video # c = os.path.join(os.path.dirname(__file__), "data/c.mp4") # Video # video_out_file = os.path.join(samples_folder,'detected' + '.mp4') # stt_demo = gr.Interface( # fn=display_two_videos, # inputs=gr.Video(), # outputs=gr.Video(type="mp4",label="Detected Video"), # examples=[ # [a], # [b], # [c], # ], # cache_examples=False # ) video = gr.Interface( fn=lambda x,y: [x,y], #video_demo, inputs=[gr.Video(label="Base Model Video",interactive=False),gr.Video(label="Tuned Model Video",interactive=False)], outputs=[gr.Video(label="Base Model Inferenced Video"), gr.Video(label="Tuned Model Inferenced Video")], examples=[ [a, b] ], title="Luffy and Chopper face detection on video Result comparison of base vs tuned SSD mobile net V2 320x320", description="Model has been customed trained to detect Character of Luffy and Chopper with Prediction confident >10%. Videos are pre-inferenced to reduce load time. (Browser zoom out to view right columne - top (base model inference) & bottom(tuned model inference)) " ) demo = gr.TabbedInterface([base_image, video], ["Images", "Video"]) if __name__ == "__main__": demo.launch()