hacky-changes.patch

diff --git c/fb_sol_v1/ocr_lib.py w/fb_sol_v1/ocr_lib.py
index 7f71e92..cd72b0f 100644
--- c/fb_sol_v1/ocr_lib.py
+++ w/fb_sol_v1/ocr_lib.py
@@ -32,14 +32,17 @@ class OCRLib:
         )
 
     def process_rgb_image(self, input_image):
-        boxes, scores = self.detector.detect(input_image)
+        with torch.autograd.profiler.record_function("detect"):
+            boxes, scores = self.detector.detect(input_image)
 
         results = []
-        for box, score in zip(boxes, scores):
-            result = self.recognizer.recognize(input_image, box)
-            result["box"] = box.tolist()
-            result["det_scr"] = score
-            results.append(result)
+        with torch.autograd.profiler.record_function("recognition"):
+            for box, score in zip(boxes, scores):
+                with torch.autograd.profiler.record_function("recognize"):
+                    result = self.recognizer.recognize(input_image, box)
+                result["box"] = box.tolist()
+                result["det_scr"] = score
+                results.append(result)
 
         return results
 
diff --git c/fb_sol_v1/sample_solution.py w/fb_sol_v1/sample_solution.py
index ee48fdc..a83a5ce 100644
--- c/fb_sol_v1/sample_solution.py
+++ w/fb_sol_v1/sample_solution.py
@@ -10,6 +10,7 @@ import sys
 import cv2
 from ocr_lib import OCRLib
 import time
+import torch
 
 def parse_args():
     parser = argparse.ArgumentParser(
@@ -84,8 +85,11 @@ def process_video(input_video, cfg_fn, results_path, sampling_rate):
     success = True
     cnt = 0
     while success:
-        success, img = vidcap.read()
+        with torch.autograd.profiler.record_function("vidcap.read"):
+            success, img = vidcap.read()
         cnt += 1
+        if cnt > 1:
+            break
 
         if not success:
             break
@@ -98,10 +102,12 @@ def process_video(input_video, cfg_fn, results_path, sampling_rate):
         # if os.path.exists(frm_output):
         #    continue
 
-        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
-        # resizing if that could be helpful
-        # img = cv2.resize(img, (640, 480), fx=0, fy=0, interpolation=cv2.INTER_CUBIC)
-        results = reader.process_rgb_image(img)
+        with torch.autograd.profiler.record_function("preprocess"):
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+            # resizing if that could be helpful
+            # img = cv2.resize(img, (640, 480), fx=0, fy=0, interpolation=cv2.INTER_CUBIC)
+        with torch.autograd.profiler.record_function("main-process"):
+            results = reader.process_rgb_image(img)
 
         # Dump bounding Box information
         with open(frm_output, "w") as h_out:
@@ -166,10 +172,13 @@ def main():
     print("OCR Start!")
     start = time.time()
     # query_video(**vars(args))
-    query_video(input_video=sys.argv[1], query_file=sys.argv[2], results_path=results_path)
+    with torch.autograd.profiler.profile() as prof:
+        with torch.autograd.profiler.record_function("whole-program"):
+            query_video(input_video=sys.argv[1], query_file=sys.argv[2], results_path=results_path)
+    prof.export_chrome_trace("/tmp/ocr_prof.json")
     print('ORC Done!')
     end = time.time()
     print("Total time: {} second ".format(end - start))
     
 if __name__ == "__main__":
-    main()
\ No newline at end of file
+    main()
diff --git c/fb_sol_v1/tracing.py w/fb_sol_v1/tracing.py
new file mode 100644
index 0000000..0d60a28
--- /dev/null
+++ w/fb_sol_v1/tracing.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+import os
+import sys
+import json
+import time
+import contextlib
+from queue import Queue
+
+
+@contextlib.contextmanager
+def trace(out_file, clock=None):
+    with contextlib.ExitStack() as handles:
+        if isinstance(out_file, str):
+            out_file = open(out_file, "w")
+            handles.enter_context(out_file)
+
+        tracer = Tracer(out_file, clock or time.monotonic)
+        try:
+            yield tracer
+        finally:
+            tracer.shutdown()
+
+
+class Tracer:
+    def __init__(self, handle, clock):
+        self.handle = handle
+        self.clock = clock
+        self.pid = os.getpid()
+        self.handle.write('[')
+
+    @contextlib.contextmanager
+    def section(self, name):
+        event = dict(
+          name=name,
+          #cat=func_filename,               # Event Category.
+          #tid=self.thread_id,             # Thread ID.
+          ph="B",      # Event Type.
+          pid=self.pid,                   # Process ID.
+          ts=self.clock()*1000000,                   # Timestamp.
+        )
+        self.handle.write(json.dumps(event))
+        self.handle.write(",\n")
+        try:
+            yield
+        finally:
+            event = dict(
+              name=name,
+              #cat=func_filename,               # Event Category.
+              #tid=self.thread_id,             # Thread ID.
+              ph="E",      # Event Type.
+              pid=self.pid,                   # Process ID.
+              ts=self.clock()*1000000,                   # Timestamp.
+            )
+            self.handle.write(json.dumps(event))
+            self.handle.write(",\n")
+
+    def shutdown(self):
+        self.handle.write('{}]')  # empty {} so the final entry doesn't end with a comma