(refactor) src

(add) 20 reproducible training noisy random vignere2

evaluation_batchedgpuevaluate.ipynb

@@ -151,9 +151,9 @@
     "# aaaaaa\n",
     "model_ids = {\n",
     "    'caesar': 'slurm_16677',\n",
-    "    'en': 'slurm_17510',\n",
-    "    'de': 'slurm_18065',\n",
-    "    'cs': 'slurm_18066'\n",
+    "    'en_constenigma': 'slurm_17510',\n",
+    "    'de_constenigma': 'slurm_18065',\n",
+    "    'cs_constenigma': 'slurm_18066'\n",
     "}\n",
     "model_id = model_ids[lang]\n",
     "device = torch.device(\"cuda:0\") if torch.cuda.is_available() else torch.device(\"cpu\")\n",

src/ByT5Dataset.py

@@ -48,9 +48,9 @@ def __getitem__(self, idx) -> dict:
             return_tensors="pt",
         )
         return {
-            "input_ids": encoding["input_ids"].squeeze(),
-            "labels": encoding_labels["input_ids"].squeeze(),
-            "attention_mask": encoding["attention_mask"].squeeze(),
+            "input_ids": encoding["input_ids"].squeeze(), # type: ignore
+            "labels": encoding_labels["input_ids"].squeeze(), # type: ignore
+            "attention_mask": encoding["attention_mask"].squeeze(), # type: ignore
             "input_text": input_text,
             "output_text": output_text,
         }
@@ -186,4 +186,22 @@ class ByT5ConstEnigmaDataset(ByT5DatasetOnlyPreprocessCiphertext):
 
     def __init__(self, data, max_length) -> None:
         const_enigma = ciphers.make_const_enigma()
-        super().__init__(const_enigma, data, max_length)
+        super().__init__(const_enigma, data, max_length)
+
+class ByT5NoisyConstEnigmaDataset(ByT5DatasetOnlyPreprocessCiphertext):
+    """
+    Dataset using Enigma cipher with a constant key and noise.
+    """
+
+    def __init__(self, data, max_length, noise_proportion=.1) -> None:
+        noisy_const_enigma = ciphers.make_const_enigma(noise_proportion=noise_proportion)
+        super().__init__(noisy_const_enigma, data, max_length)
+
+
+class ByT5NoisyVignere2Dataset(ByT5DatasetOnlyPreprocessCiphertext):
+    """
+    Dataset using noisy 2-letter Vignere cipher.
+    """
+
+    def __init__(self, data, max_length, noise_proportion=.15) -> None:
+        super().__init__(functools.partial(ciphers.noisy_random_vignere2, noise_proportion=noise_proportion), data, max_length)

src/ciphers.py

@@ -196,7 +196,7 @@ def random_vignere(text: str, key_length: int) -> str:
 
 
 def noisy_random_vignere(
-    text: str, key_length: int, noise_proportion: float = 0.1
+    text: str, key_length: int, noise_proportion: float = 0.15
 ) -> str:
     """Vigenère cipher with a random key of specified length and noise."""
     noisy_text = add_noise_to_text(text, noise_proportion)
@@ -213,11 +213,11 @@ def random_vignere3(text: str) -> str:
     return random_vignere(text, 3)
 
 
-def noisy_random_vignere2(text: str, noise_proportion: float = 0.1) -> str:
+def noisy_random_vignere2(text: str, noise_proportion: float = 0.15) -> str:
     """Vigenère cipher with a random 2-letter key and noise."""
     return noisy_random_vignere(text, 2, noise_proportion)
 
 
-def noisy_random_vignere3(text: str, noise_proportion: float = 0.1) -> str:
+def noisy_random_vignere3(text: str, noise_proportion: float = 0.15) -> str:
     """Vigenère cipher with a random 3-letter key and noise."""
     return noisy_random_vignere(text, 3, noise_proportion)

src/preprocessing.py

@@ -1,6 +1,7 @@
 import random
 import logging
 import re
+from typing import Optional
 from unidecode import unidecode
 
 
@@ -46,7 +47,7 @@ def generate_random_dataset(
 
 
 def load_dataset(
-    rows: int, min_length: int, max_length: int, file_path: str, seed: int = 42, exclude_length: int = None
+    rows: int, min_length: int, max_length: int, file_path: str, seed: int = 42, exclude_length: Optional[int] = None
 ) -> list[str]:
     """Samples a dataset from a file and truncates each rows to a random length generated for it from a range.
 
@@ -165,4 +166,4 @@ def weird(text: str) -> bool:
     if comma_period_count > 15:
         return True
 
-    return False
+    return False

src/utils.py

@@ -16,12 +16,13 @@ def download_newscrawl(year=2012, language="en") -> None:
     newscrawl_url = f"https://data.statmt.org/news-crawl/{language}/"
     filename = f"news.{year}.{language}.shuffled.deduped.gz"
     url = newscrawl_url + filename
-    logging.info(f"Downloading {filename} from {url}")
+    logging.info("Downloading %s from %s", filename, url)
     os.system(f"wget {url}")
     os.system(f"gunzip {filename}")
     logging.info(
-        f"Downloaded and extracted {filename}",
-        f"full path: {os.path.abspath(filename)}",
+        "Downloaded and extracted %s, full path: %s",
+        filename,
+        os.path.abspath(filename),
     )
 
 
@@ -86,7 +87,7 @@ def create_detect_language(lang="en"):
     def detect_language(text):
         # print('\n' in text.strip())
         # return 'aabb'
-        detected = model.predict(text.strip())[0][0][-2:]   # the last two characters are the language in the label
+        detected = model.predict(text.strip())[0][0][-2:]   # type: ignore # the last two characters are the language in the label
         return detected == lang
 
     return detect_language