doctor and model test

Playgrounds/doctor.py

@@ -0,0 +1,125 @@
+import random
+import torch
+import pandas as pd
+from pathlib import Path
+import Project_Model.Libs.Embedder as Embedder
+import Project_Model.Libs.BPE as BPE
+import Project_Model.Libs.Transformer as Transformer
+import Project_Model.Libs.TorchShims as torch_shims
+from Project_Model.Libs.Training.learning_rade_shedulers import Custom_lr
+from Project_Model.Libs.Training.logistic_collector import LogitsCollector  # import the external collector
+
+# set a fixed seed
+torch.manual_seed(0)
+random.seed(0)
+DEVICE = torch_shims.get_default_device()
+torch.set_default_device(DEVICE)
+
+# BPE Init
+VOCABULARY_PATH = Path("Assets/Model/toy_10/toy_dictionary.json")
+SPECIAL_VOC = BPE.default_special_tokens()
+
+VOCABULARY = BPE.load_nanos_vocabulary(VOCABULARY_PATH)
+TOKENANO = BPE.TokeNanoCore(VOCABULARY, SPECIAL_VOC)
+
+# Constants
+TOKEN_SPACE_SIZE = TOKENANO.vocabulary_size + 1
+EMBEDDED_SIZE = 256
+FEED_FORWARD_MULTIPLIER = 4
+ATTENTION_HEADS = 4
+SENTENCE_LENGTH = 256
+NUMBER_OF_BLOCKS = 2
+MAX_EPOCHS = int(1e3)
+
+PAD_TOKEN = TOKENANO.encode("<PAD>")[0]
+END_TOKEN = TOKENANO.encode("<END>")[0]
+
+# Load CSV
+TOY_DATASET_PATH = Path("Assets/Dataset/1-hop/toy/rdf_text.csv")
+TOY_DATASET = pd.read_csv(TOY_DATASET_PATH)
+
+TOY_BATCH_INPUT_LIST: list[list[int]] = []
+TOY_BATCH_PADDING_LIST: list[list[bool]] = []
+TOY_BATCH_TARGET_LIST: list[list[int]] = []
+TOY_BATCH_DECODER_DEFAULT: list[list[int]] = []
+
+for index, row in TOY_DATASET.iterrows():
+    RDFs: str = row["RDFs"]
+    Abstract: str = row["Abstract"]
+
+    input_tokens = TOKENANO.encode(RDFs)  # encoder input ids
+    output_tokens = TOKENANO.encode(Abstract)[1:]  # decoder target ids (shifted left)
+    decoder_default_tokens = TOKENANO.encode("<SOS>")  # decoder input starts with <SOS>
+
+    input_tokens, padding = Transformer.normalize_sequence(
+        input_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN
+    )  # pad/trim + end token
+    output_tokens, _ = Transformer.normalize_sequence(
+        output_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN
+    )  # pad/trim + end token
+    decoder_default_tokens = Transformer.pad_sequence(
+        decoder_default_tokens, SENTENCE_LENGTH, PAD_TOKEN
+    )  # pad with PAD up to SENTENCE_LENGTH
+
+    TOY_BATCH_INPUT_LIST.append(input_tokens)
+    TOY_BATCH_PADDING_LIST.append(padding)
+    TOY_BATCH_TARGET_LIST.append(output_tokens)
+    TOY_BATCH_DECODER_DEFAULT.append(decoder_default_tokens)
+
+# Training loop
+LOSS_HISTORY = []
+NANOSOCRATES = Transformer.TrainingModel(
+    TOKEN_SPACE_SIZE,
+    EMBEDDED_SIZE,
+    FEED_FORWARD_MULTIPLIER,
+    ATTENTION_HEADS,
+    NUMBER_OF_BLOCKS,
+)
+
+collector = LogitsCollector(PAD_TOKEN, END_TOKEN, TOKENANO)  # collects logits and decodes
+
+NANOSOCRATES.train()
+cross_entropy = torch.nn.CrossEntropyLoss(ignore_index=PAD_TOKEN)
+optimizer = torch.optim.AdamW(NANOSOCRATES.parameters())
+scheduler = Custom_lr(EMBEDDED_SIZE, 4000)  # step each optimizer step
+
+current_epoch = 0
+BATCH_SIZE = min(32, len(TOY_BATCH_INPUT_LIST))  # small batch to stabilize
+
+while current_epoch < MAX_EPOCHS:
+    # simple fixed mini-batch from the top; later you can shuffle/slice
+    enc = torch.tensor(TOY_BATCH_INPUT_LIST[:BATCH_SIZE], dtype=torch.long)        # [B,T] encoder token ids
+    pad = torch.tensor(TOY_BATCH_PADDING_LIST[:BATCH_SIZE], dtype=torch.bool)      # [B,T] True where encoder PAD is present
+    tgt = torch.tensor(TOY_BATCH_TARGET_LIST[:BATCH_SIZE], dtype=torch.long)       # [B,T] decoder targets (ground-truth)
+
+    # decoder prefix buffer: <SOS> at pos 0, PAD elsewhere (no shift here)        # we will fill it step by step
+    dec = torch.tensor(TOY_BATCH_DECODER_DEFAULT[:BATCH_SIZE], dtype=torch.long)   # [B,T]
+
+    total_loss = 0.0
+    collector.reset()  # start fresh for this epoch
+
+    T = tgt.size(1)  # sequence length
+    for t in range(T):
+        optimizer.zero_grad(set_to_none=True)                                       # clear grads for this token step
+
+        prefix = dec[:, : t + 1]                                                    # [B, t+1] current decoder prefix
+        dec_pad_mask = prefix.eq(PAD_TOKEN)                                         # [B, t+1] True where PAD inside prefix
+
+        # one-step logits given prefix (trainer model expects 4 args now)
+        logits_t: torch.Tensor = NANOSOCRATES((enc, pad, prefix, dec_pad_mask))     # [B,V] logits for step t
+        collector.add(logits_t)                                                     # store logits for decoding later
+
+        loss_t = cross_entropy(logits_t, tgt[:, t])                                 # CE expects raw logits; PAD ignored
+        loss_t.backward()                                                           # backprop for this step
+        optimizer.step()                                                            # update params
+        scheduler.step()                                                            # Noam/warmup: step per optimizer step
+
+        total_loss = float(loss_t.detach())                                         # keep last step loss for logging
+
+        # teacher forcing: reveal the correct token for next position
+        if t < T - 1:
+            dec[:, t + 1] = tgt[:, t]                                               # write ground-truth into next slot
+
+    current_epoch += 1
+    print(f"EPOCH {current_epoch}\n\tLoss: {total_loss:.6f}")                        # simple log
+    collector.print_decoded()                                                       # print decoded predictions for the batch

Playgrounds/locistic_test.ipynb

@@ -0,0 +1,221 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c8741a8f",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "EPOCH 1\n",
+      "\tLoss: 7.424792\n",
+      "[0] \n",
+      "[1] \n",
+      "[2] \n",
+      "[3] \n",
+      "[4] \n",
+      "[5] \n",
+      "[6] \n",
+      "[7] \n",
+      "[8] \n",
+      "[9] \n"
+     ]
+    }
+   ],
+   "source": [
+    "import random\n",
+    "import torch\n",
+    "import pandas as pd\n",
+    "from pathlib import Path\n",
+    "import Project_Model.Libs.Embedder as Embedder\n",
+    "import Project_Model.Libs.BPE as BPE\n",
+    "import Project_Model.Libs.Transformer as Transformer\n",
+    "import Project_Model.Libs.TorchShims as torch_shims\n",
+    "from Project_Model.Libs.Training.learning_rade_shedulers import Custom_lr\n",
+    "\n",
+    "import torch\n",
+    "\n",
+    "class LogitsCollector:\n",
+    "    def __init__(self, pad_token: int, end_token: int, tokenizer) -> None:\n",
+    "        self.__pad_token = pad_token                      # used to skip PAD\n",
+    "        self.__end_token = end_token                      # used to stop at END\n",
+    "        self.__tokenizer = tokenizer                      # exposes .decode(list[int]) -> str\n",
+    "        self.__steps: list[torch.Tensor] = []             # list of per-step logits [B,V]\n",
+    "\n",
+    "    def reset(self) -> None:\n",
+    "        self.__steps.clear()                              # clear history\n",
+    "\n",
+    "    def add(self, logits_step: torch.Tensor) -> None:\n",
+    "        if logits_step.dim() == 3:                        # handle [B,1,V]\n",
+    "            logits_step = logits_step[:, -1, :]           # -> [B,V]\n",
+    "        self.__steps.append(logits_step.detach())         # store raw logits (detached)\n",
+    "\n",
+    "    def tokens(self) -> list[list[int]]:\n",
+    "        if not self.__steps:\n",
+    "            return []\n",
+    "        stack = torch.stack(self.__steps, dim=0)          # [T,B,V]\n",
+    "        probs = torch.softmax(stack, dim=-1)              # softmax over vocab -> [T,B,V]\n",
+    "        ids = probs.argmax(dim=-1).transpose(0, 1)        # greedy ids -> [B,T]\n",
+    "        out: list[list[int]] = []\n",
+    "        for row in ids.tolist():\n",
+    "            seq: list[int] = []\n",
+    "            for tok in row:\n",
+    "                if tok == self.__end_token:               # stop on END\n",
+    "                    break\n",
+    "                if tok == self.__pad_token:               # skip PAD\n",
+    "                    continue\n",
+    "                seq.append(tok)\n",
+    "            out.append(seq)\n",
+    "        return out\n",
+    "\n",
+    "    def print_decoded(self) -> None:\n",
+    "        for i, seq in enumerate(self.tokens()):\n",
+    "            try:\n",
+    "                text = self.__tokenizer.decode(seq)       # decode tokens to string\n",
+    "            except Exception:\n",
+    "                text = str(seq)                           # fallback to ids\n",
+    "            print(f\"[{i}] {text}\")                        # simple print\n",
+    "\n",
+    "\n",
+    "# set a fixed seed\n",
+    "torch.manual_seed(0)\n",
+    "random.seed(0)\n",
+    "DEVICE = torch_shims.get_default_device()\n",
+    "torch.set_default_device(DEVICE)\n",
+    "\n",
+    "# BPE Init\n",
+    "VOCABULARY_PATH = Path(\"Assets/Model/toy_10/toy_dictionary.json\")\n",
+    "SPECIAL_VOC = BPE.default_special_tokens()\n",
+    "\n",
+    "VOCABULARY = BPE.load_nanos_vocabulary(VOCABULARY_PATH)\n",
+    "TOKENANO = BPE.TokeNanoCore(VOCABULARY, SPECIAL_VOC)\n",
+    "\n",
+    "# Constants\n",
+    "TOKEN_SPACE_SIZE = TOKENANO.vocabulary_size + 1\n",
+    "EMBEDDED_SIZE = 256\n",
+    "FEED_FORWARD_MULTIPLIER = 4\n",
+    "ATTENTION_HEADS = 4\n",
+    "SENTENCE_LENGTH = 256\n",
+    "NUMBER_OF_BLOCKS = 2\n",
+    "MAX_EPOCHS = int(1e3)\n",
+    "\n",
+    "PAD_TOKEN = TOKENANO.encode(\"<PAD>\")[0]\n",
+    "END_TOKEN = TOKENANO.encode(\"<END>\")[0]\n",
+    "\n",
+    "# Load CSV\n",
+    "TOY_DATASET_PATH = Path(\"Assets/Dataset/1-hop/toy/rdf_text.csv\")\n",
+    "TOY_DATASET = pd.read_csv(TOY_DATASET_PATH)\n",
+    "\n",
+    "TOY_BATCH_INPUT_LIST: list[list[int]] = []\n",
+    "TOY_BATCH_PADDING_LIST: list[list[bool]] = []\n",
+    "TOY_BATCH_TARGET_LIST: list[list[int]] = []\n",
+    "TOY_BATCH_DECODER_DEFAULT: list[list[int]] = []\n",
+    "\n",
+    "for index, row in TOY_DATASET.iterrows():\n",
+    "    RDFs: str = row[\"RDFs\"]\n",
+    "    Abstract: str = row[\"Abstract\"]\n",
+    "\n",
+    "    input_tokens = TOKENANO.encode(RDFs)  # encoder input ids\n",
+    "    output_tokens = TOKENANO.encode(Abstract)[1:]  # decoder target ids (shifted left)\n",
+    "    decoder_default_tokens = TOKENANO.encode(\"<SOS>\")  # decoder input starts with <SOS>\n",
+    "\n",
+    "    input_tokens, padding = Transformer.normalize_sequence(\n",
+    "        input_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN\n",
+    "    )  # pad/trim + end token\n",
+    "    output_tokens, _ = Transformer.normalize_sequence(\n",
+    "        output_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN\n",
+    "    )  # pad/trim + end token\n",
+    "    decoder_default_tokens = Transformer.pad_sequence(\n",
+    "        decoder_default_tokens, SENTENCE_LENGTH, PAD_TOKEN\n",
+    "    )  # pad with PAD up to SENTENCE_LENGTH\n",
+    "\n",
+    "    TOY_BATCH_INPUT_LIST.append(input_tokens)\n",
+    "    TOY_BATCH_PADDING_LIST.append(padding)\n",
+    "    TOY_BATCH_TARGET_LIST.append(output_tokens)\n",
+    "    TOY_BATCH_DECODER_DEFAULT.append(decoder_default_tokens)\n",
+    "\n",
+    "# Training loop\n",
+    "LOSS_HISTORY = []\n",
+    "NANOSOCRATES = Transformer.TrainingModel(\n",
+    "    TOKEN_SPACE_SIZE,\n",
+    "    EMBEDDED_SIZE,\n",
+    "    FEED_FORWARD_MULTIPLIER,\n",
+    "    ATTENTION_HEADS,\n",
+    "    NUMBER_OF_BLOCKS,\n",
+    ")\n",
+    "\n",
+    "collector = LogitsCollector(PAD_TOKEN, END_TOKEN, TOKENANO)  # collects logits and decodes\n",
+    "\n",
+    "NANOSOCRATES.train()\n",
+    "cross_entropy = torch.nn.CrossEntropyLoss(ignore_index=PAD_TOKEN)\n",
+    "optimizer = torch.optim.AdamW(NANOSOCRATES.parameters())\n",
+    "scheduler = Custom_lr(EMBEDDED_SIZE, 4000)  # step each optimizer step\n",
+    "\n",
+    "current_epoch = 0\n",
+    "BATCH_SIZE = min(32, len(TOY_BATCH_INPUT_LIST))  # small batch to stabilize\n",
+    "\n",
+    "while current_epoch < MAX_EPOCHS:\n",
+    "    # simple fixed mini-batch from the top; later you can shuffle/slice\n",
+    "    enc = torch.tensor(TOY_BATCH_INPUT_LIST[:BATCH_SIZE], dtype=torch.long)        # [B,T] encoder token ids\n",
+    "    pad = torch.tensor(TOY_BATCH_PADDING_LIST[:BATCH_SIZE], dtype=torch.bool)      # [B,T] True where encoder PAD is present\n",
+    "    tgt = torch.tensor(TOY_BATCH_TARGET_LIST[:BATCH_SIZE], dtype=torch.long)       # [B,T] decoder targets (ground-truth)\n",
+    "\n",
+    "    # decoder prefix buffer: <SOS> at pos 0, PAD elsewhere (no shift here)        # we will fill it step by step\n",
+    "    dec = torch.tensor(TOY_BATCH_DECODER_DEFAULT[:BATCH_SIZE], dtype=torch.long)   # [B,T]\n",
+    "\n",
+    "    total_loss = 0.0\n",
+    "    collector.reset()  # start fresh for this epoch\n",
+    "\n",
+    "    T = tgt.size(1)  # sequence length\n",
+    "    for t in range(T):\n",
+    "        optimizer.zero_grad(set_to_none=True)                                       # clear grads for this token step\n",
+    "\n",
+    "        prefix = dec[:, : t + 1]                                                    # [B, t+1] current decoder prefix\n",
+    "        dec_pad_mask = prefix.eq(PAD_TOKEN)                                         # [B, t+1] True where PAD inside prefix\n",
+    "\n",
+    "        # one-step logits given prefix (trainer model expects 4 args now)\n",
+    "        logits_t: torch.Tensor = NANOSOCRATES((enc, pad, prefix, dec_pad_mask))     # [B,V] logits for step t\n",
+    "        collector.add(logits_t)                                                     # store logits for decoding later\n",
+    "\n",
+    "        loss_t = cross_entropy(logits_t, tgt[:, t])                                 # CE expects raw logits; PAD ignored\n",
+    "        loss_t.backward()                                                           # backprop for this step\n",
+    "        optimizer.step()                                                            # update params\n",
+    "        scheduler.step()                                                            # Noam/warmup: step per optimizer step\n",
+    "\n",
+    "        total_loss = float(loss_t.detach())                                         # keep last step loss for logging\n",
+    "\n",
+    "        # teacher forcing: reveal the correct token for next position\n",
+    "        if t < T - 1:\n",
+    "            dec[:, t + 1] = tgt[:, t]                                               # write ground-truth into next slot\n",
+    "\n",
+    "    current_epoch += 1\n",
+    "    print(f\"EPOCH {current_epoch}\\n\\tLoss: {total_loss:.6f}\")                        # simple log\n",
+    "    collector.print_decoded()                                                       # print decoded predictions for the batch\n"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "deep_learning",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.13.7"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

Playgrounds/model-teacher-forcing.ipynb

@@ -0,0 +1,205 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0afbf498",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "EPOCH 1\n",
+      "\tLoss: 9.174470901489258\n",
+      "EPOCH 2\n",
+      "\tLoss: 9.20919132232666\n",
+      "EPOCH 3\n",
+      "\tLoss: 9.227106094360352\n",
+      "EPOCH 4\n",
+      "\tLoss: 9.172086715698242\n",
+      "EPOCH 5\n",
+      "\tLoss: 9.180150985717773\n"
+     ]
+    },
+    {
+     "ename": "KeyboardInterrupt",
+     "evalue": "",
+     "output_type": "error",
+     "traceback": [
+      "\u001b[31m---------------------------------------------------------------------------\u001b[39m",
+      "\u001b[31mKeyboardInterrupt\u001b[39m                         Traceback (most recent call last)",
+      "\u001b[36mCell\u001b[39m\u001b[36m \u001b[39m\u001b[32mIn[1]\u001b[39m\u001b[32m, line 116\u001b[39m\n\u001b[32m    113\u001b[39m step_target = target_logits[:, i]             \u001b[38;5;66;03m# [B]\u001b[39;00m\n\u001b[32m    115\u001b[39m loss = cross_entropy(step_logits,step_target) \u001b[38;5;66;03m# now loss is without softmax\u001b[39;00m\n\u001b[32m--> \u001b[39m\u001b[32m116\u001b[39m \u001b[43mloss\u001b[49m\u001b[43m.\u001b[49m\u001b[43mbackward\u001b[49m\u001b[43m(\u001b[49m\u001b[43m)\u001b[49m  \u001b[38;5;66;03m# DAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAMN\u001b[39;00m\n\u001b[32m    117\u001b[39m last_loss = loss\n\u001b[32m    118\u001b[39m optimizer.step()\n",
+      "\u001b[36mFile \u001b[39m\u001b[32m~/miniconda3/envs/deep_learning/lib/python3.13/site-packages/torch/_tensor.py:638\u001b[39m, in \u001b[36mTensor.backward\u001b[39m\u001b[34m(self, gradient, retain_graph, create_graph, inputs)\u001b[39m\n\u001b[32m    595\u001b[39m \u001b[38;5;250m\u001b[39m\u001b[33mr\u001b[39m\u001b[33;03m\"\"\"Computes the gradient of current tensor wrt graph leaves.\u001b[39;00m\n\u001b[32m    596\u001b[39m \n\u001b[32m    597\u001b[39m \u001b[33;03mThe graph is differentiated using the chain rule. If the tensor is\u001b[39;00m\n\u001b[32m   (...)\u001b[39m\u001b[32m    635\u001b[39m \u001b[33;03m        used to compute the :attr:`tensors`. Defaults to ``None``.\u001b[39;00m\n\u001b[32m    636\u001b[39m \u001b[33;03m\"\"\"\u001b[39;00m\n\u001b[32m    637\u001b[39m \u001b[38;5;28;01mif\u001b[39;00m has_torch_function_unary(\u001b[38;5;28mself\u001b[39m):\n\u001b[32m--> \u001b[39m\u001b[32m638\u001b[39m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[43mhandle_torch_function\u001b[49m\u001b[43m(\u001b[49m\n\u001b[32m    639\u001b[39m \u001b[43m        \u001b[49m\u001b[43mTensor\u001b[49m\u001b[43m.\u001b[49m\u001b[43mbackward\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    640\u001b[39m \u001b[43m        \u001b[49m\u001b[43m(\u001b[49m\u001b[38;5;28;43mself\u001b[39;49m\u001b[43m,\u001b[49m\u001b[43m)\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    641\u001b[39m \u001b[43m        \u001b[49m\u001b[38;5;28;43mself\u001b[39;49m\u001b[43m,\u001b[49m\n\u001b[32m    642\u001b[39m \u001b[43m        \u001b[49m\u001b[43mgradient\u001b[49m\u001b[43m=\u001b[49m\u001b[43mgradient\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    643\u001b[39m \u001b[43m        \u001b[49m\u001b[43mretain_graph\u001b[49m\u001b[43m=\u001b[49m\u001b[43mretain_graph\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    644\u001b[39m \u001b[43m        \u001b[49m\u001b[43mcreate_graph\u001b[49m\u001b[43m=\u001b[49m\u001b[43mcreate_graph\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    645\u001b[39m \u001b[43m        \u001b[49m\u001b[43minputs\u001b[49m\u001b[43m=\u001b[49m\u001b[43minputs\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    646\u001b[39m \u001b[43m    \u001b[49m\u001b[43m)\u001b[49m\n\u001b[32m    647\u001b[39m torch.autograd.backward(\n\u001b[32m    648\u001b[39m     \u001b[38;5;28mself\u001b[39m, gradient, retain_graph, create_graph, inputs=inputs\n\u001b[32m    649\u001b[39m )\n",
+      "\u001b[36mFile \u001b[39m\u001b[32m~/miniconda3/envs/deep_learning/lib/python3.13/site-packages/torch/overrides.py:1725\u001b[39m, in \u001b[36mhandle_torch_function\u001b[39m\u001b[34m(public_api, relevant_args, *args, **kwargs)\u001b[39m\n\u001b[32m   1721\u001b[39m \u001b[38;5;28;01mif\u001b[39;00m _is_torch_function_mode_enabled():\n\u001b[32m   1722\u001b[39m     \u001b[38;5;66;03m# if we're here, the mode must be set to a TorchFunctionStackMode\u001b[39;00m\n\u001b[32m   1723\u001b[39m     \u001b[38;5;66;03m# this unsets it and calls directly into TorchFunctionStackMode's torch function\u001b[39;00m\n\u001b[32m   1724\u001b[39m     \u001b[38;5;28;01mwith\u001b[39;00m _pop_mode_temporarily() \u001b[38;5;28;01mas\u001b[39;00m mode:\n\u001b[32m-> \u001b[39m\u001b[32m1725\u001b[39m         result = \u001b[43mmode\u001b[49m\u001b[43m.\u001b[49m\u001b[43m__torch_function__\u001b[49m\u001b[43m(\u001b[49m\u001b[43mpublic_api\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mtypes\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43margs\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mkwargs\u001b[49m\u001b[43m)\u001b[49m\n\u001b[32m   1726\u001b[39m     \u001b[38;5;28;01mif\u001b[39;00m result \u001b[38;5;129;01mis\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;28mNotImplemented\u001b[39m:\n\u001b[32m   1727\u001b[39m         \u001b[38;5;28;01mreturn\u001b[39;00m result\n",
+      "\u001b[36mFile \u001b[39m\u001b[32m~/miniconda3/envs/deep_learning/lib/python3.13/site-packages/torch/utils/_device.py:103\u001b[39m, in \u001b[36mDeviceContext.__torch_function__\u001b[39m\u001b[34m(self, func, types, args, kwargs)\u001b[39m\n\u001b[32m    101\u001b[39m \u001b[38;5;28;01mif\u001b[39;00m func \u001b[38;5;129;01min\u001b[39;00m _device_constructors() \u001b[38;5;129;01mand\u001b[39;00m kwargs.get(\u001b[33m\"\u001b[39m\u001b[33mdevice\u001b[39m\u001b[33m\"\u001b[39m) \u001b[38;5;129;01mis\u001b[39;00m \u001b[38;5;28;01mNone\u001b[39;00m:\n\u001b[32m    102\u001b[39m     kwargs[\u001b[33m\"\u001b[39m\u001b[33mdevice\u001b[39m\u001b[33m\"\u001b[39m] = \u001b[38;5;28mself\u001b[39m.device\n\u001b[32m--> \u001b[39m\u001b[32m103\u001b[39m \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[43mfunc\u001b[49m\u001b[43m(\u001b[49m\u001b[43m*\u001b[49m\u001b[43margs\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43m*\u001b[49m\u001b[43m*\u001b[49m\u001b[43mkwargs\u001b[49m\u001b[43m)\u001b[49m\n",
+      "\u001b[36mFile \u001b[39m\u001b[32m~/miniconda3/envs/deep_learning/lib/python3.13/site-packages/torch/_tensor.py:647\u001b[39m, in \u001b[36mTensor.backward\u001b[39m\u001b[34m(self, gradient, retain_graph, create_graph, inputs)\u001b[39m\n\u001b[32m    637\u001b[39m \u001b[38;5;28;01mif\u001b[39;00m has_torch_function_unary(\u001b[38;5;28mself\u001b[39m):\n\u001b[32m    638\u001b[39m     \u001b[38;5;28;01mreturn\u001b[39;00m handle_torch_function(\n\u001b[32m    639\u001b[39m         Tensor.backward,\n\u001b[32m    640\u001b[39m         (\u001b[38;5;28mself\u001b[39m,),\n\u001b[32m   (...)\u001b[39m\u001b[32m    645\u001b[39m         inputs=inputs,\n\u001b[32m    646\u001b[39m     )\n\u001b[32m--> \u001b[39m\u001b[32m647\u001b[39m \u001b[43mtorch\u001b[49m\u001b[43m.\u001b[49m\u001b[43mautograd\u001b[49m\u001b[43m.\u001b[49m\u001b[43mbackward\u001b[49m\u001b[43m(\u001b[49m\n\u001b[32m    648\u001b[39m \u001b[43m    \u001b[49m\u001b[38;5;28;43mself\u001b[39;49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mgradient\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mretain_graph\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mcreate_graph\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43minputs\u001b[49m\u001b[43m=\u001b[49m\u001b[43minputs\u001b[49m\n\u001b[32m    649\u001b[39m \u001b[43m\u001b[49m\u001b[43m)\u001b[49m\n",
+      "\u001b[36mFile \u001b[39m\u001b[32m~/miniconda3/envs/deep_learning/lib/python3.13/site-packages/torch/autograd/__init__.py:354\u001b[39m, in \u001b[36mbackward\u001b[39m\u001b[34m(tensors, grad_tensors, retain_graph, create_graph, grad_variables, inputs)\u001b[39m\n\u001b[32m    349\u001b[39m     retain_graph = create_graph\n\u001b[32m    351\u001b[39m \u001b[38;5;66;03m# The reason we repeat the same comment below is that\u001b[39;00m\n\u001b[32m    352\u001b[39m \u001b[38;5;66;03m# some Python versions print out the first line of a multi-line function\u001b[39;00m\n\u001b[32m    353\u001b[39m \u001b[38;5;66;03m# calls in the traceback and some print out the last line\u001b[39;00m\n\u001b[32m--> \u001b[39m\u001b[32m354\u001b[39m \u001b[43m_engine_run_backward\u001b[49m\u001b[43m(\u001b[49m\n\u001b[32m    355\u001b[39m \u001b[43m    \u001b[49m\u001b[43mtensors\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    356\u001b[39m \u001b[43m    \u001b[49m\u001b[43mgrad_tensors_\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    357\u001b[39m \u001b[43m    \u001b[49m\u001b[43mretain_graph\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    358\u001b[39m \u001b[43m    \u001b[49m\u001b[43mcreate_graph\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    359\u001b[39m \u001b[43m    \u001b[49m\u001b[43minputs_tuple\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m    360\u001b[39m \u001b[43m    \u001b[49m\u001b[43mallow_unreachable\u001b[49m\u001b[43m=\u001b[49m\u001b[38;5;28;43;01mTrue\u001b[39;49;00m\u001b[43m,\u001b[49m\n\u001b[32m    361\u001b[39m \u001b[43m    \u001b[49m\u001b[43maccumulate_grad\u001b[49m\u001b[43m=\u001b[49m\u001b[38;5;28;43;01mTrue\u001b[39;49;00m\u001b[43m,\u001b[49m\n\u001b[32m    362\u001b[39m \u001b[43m\u001b[49m\u001b[43m)\u001b[49m\n",
+      "\u001b[36mFile \u001b[39m\u001b[32m~/miniconda3/envs/deep_learning/lib/python3.13/site-packages/torch/autograd/graph.py:829\u001b[39m, in \u001b[36m_engine_run_backward\u001b[39m\u001b[34m(t_outputs, *args, **kwargs)\u001b[39m\n\u001b[32m    827\u001b[39m     unregister_hooks = _register_logging_hooks_on_whole_graph(t_outputs)\n\u001b[32m    828\u001b[39m \u001b[38;5;28;01mtry\u001b[39;00m:\n\u001b[32m--> \u001b[39m\u001b[32m829\u001b[39m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[43mVariable\u001b[49m\u001b[43m.\u001b[49m\u001b[43m_execution_engine\u001b[49m\u001b[43m.\u001b[49m\u001b[43mrun_backward\u001b[49m\u001b[43m(\u001b[49m\u001b[43m  \u001b[49m\u001b[38;5;66;43;03m# Calls into the C++ engine to run the backward pass\u001b[39;49;00m\n\u001b[32m    830\u001b[39m \u001b[43m        \u001b[49m\u001b[43mt_outputs\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43m*\u001b[49m\u001b[43margs\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43m*\u001b[49m\u001b[43m*\u001b[49m\u001b[43mkwargs\u001b[49m\n\u001b[32m    831\u001b[39m \u001b[43m    \u001b[49m\u001b[43m)\u001b[49m  \u001b[38;5;66;03m# Calls into the C++ engine to run the backward pass\u001b[39;00m\n\u001b[32m    832\u001b[39m \u001b[38;5;28;01mfinally\u001b[39;00m:\n\u001b[32m    833\u001b[39m     \u001b[38;5;28;01mif\u001b[39;00m attach_logging_hooks:\n",
+      "\u001b[31mKeyboardInterrupt\u001b[39m: "
+     ]
+    }
+   ],
+   "source": [
+    "import random\n",
+    "import torch\n",
+    "import pandas as pd\n",
+    "from pathlib import Path\n",
+    "import Project_Model.Libs.Embedder as Embedder\n",
+    "import Project_Model.Libs.BPE as BPE\n",
+    "import Project_Model.Libs.Transformer as Transformer\n",
+    "import Project_Model.Libs.TorchShims as torch_shims\n",
+    "\n",
+    "# set a fixed seed\n",
+    "torch.manual_seed(0)\n",
+    "random.seed(0)\n",
+    "DEVICE = torch_shims.get_default_device()\n",
+    "torch.set_default_device(DEVICE)\n",
+    "\n",
+    "# set a default device\n",
+    "\n",
+    "# BPE Init\n",
+    "VOCABULARY_PATH = Path(\"Assets/Model/toy_10/toy_dictionary.json\")\n",
+    "SPECIAL_VOC = BPE.default_special_tokens()\n",
+    "\n",
+    "VOCABULARY = BPE.load_nanos_vocabulary(VOCABULARY_PATH)\n",
+    "TOKENANO = BPE.TokeNanoCore(VOCABULARY, SPECIAL_VOC)\n",
+    "\n",
+    "\n",
+    "# Constants\n",
+    "TOKEN_SPACE_SIZE = TOKENANO.vocabulary_size + 1\n",
+    "EMBEDDED_SIZE = 256\n",
+    "FEED_FORWARD_MULTIPLIER =  4\n",
+    "ATTENTION_HEADS = 4\n",
+    "SENTENCE_LENGTH = 256\n",
+    "NUMBER_OF_BLOCKS = 2\n",
+    "MAX_EPOCHS = int(1e3)\n",
+    "\n",
+    "\n",
+    "PAD_TOKEN = TOKENANO.encode(\"<PAD>\")[0]\n",
+    "END_TOKEN = TOKENANO.encode(\"<END>\")[0]\n",
+    "\n",
+    "\n",
+    "# Load CSV\n",
+    "TOY_DATASET_PATH = Path(\"Assets/Dataset/1-hop/toy/rdf_text.csv\")\n",
+    "\n",
+    "TOY_DATASET = pd.read_csv(TOY_DATASET_PATH)\n",
+    "\n",
+    "TOY_BATCH_INPUT_LIST: list[list[int]] = []\n",
+    "TOY_BATCH_PADDING_LIST: list[list[bool]] = []\n",
+    "TOY_BATCH_TARGET_LIST: list[list[int]] = []\n",
+    "TOY_BATCH_DECODER_DEFAULT: list[list[int]]= []\n",
+    "\n",
+    "\n",
+    "for index, row in TOY_DATASET.iterrows():\n",
+    "\n",
+    "    RDFs: str = row[\"RDFs\"]\n",
+    "    Abstract: str = row[\"Abstract\"]\n",
+    "\n",
+    "    input_tokens = TOKENANO.encode(RDFs)\n",
+    "    output_tokens = TOKENANO.encode(Abstract)[1:]\n",
+    "    decoder_default_tokens = TOKENANO.encode(\"<SOS>\")\n",
+    "\n",
+    "    input_tokens, padding = Transformer.normalize_sequence(\n",
+    "        input_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN\n",
+    "    )\n",
+    "    output_tokens, _ = Transformer.normalize_sequence(\n",
+    "        output_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN\n",
+    "    )\n",
+    "    decoder_default_tokens, _ = Transformer.normalize_sequence(\n",
+    "        decoder_default_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN\n",
+    "    )\n",
+    "\n",
+    "    TOY_BATCH_INPUT_LIST.append(input_tokens)\n",
+    "    TOY_BATCH_PADDING_LIST.append(padding)\n",
+    "    TOY_BATCH_TARGET_LIST.append(output_tokens)\n",
+    "    TOY_BATCH_DECODER_DEFAULT.append(decoder_default_tokens)\n",
+    "\n",
+    "# Training loop\n",
+    "LOSS_HISTORY = []\n",
+    "NANOSOCRATES = Transformer.TrainingModel(\n",
+    "    TOKEN_SPACE_SIZE,\n",
+    "    EMBEDDED_SIZE,\n",
+    "    FEED_FORWARD_MULTIPLIER,\n",
+    "    ATTENTION_HEADS,\n",
+    "    NUMBER_OF_BLOCKS\n",
+    ")\n",
+    "\n",
+    "NANOSOCRATES.train() # nothing important, activates dropout etc \n",
+    "cross_entropy = torch.nn.CrossEntropyLoss(ignore_index=PAD_TOKEN)\n",
+    "optimizer = torch.optim.AdamW(NANOSOCRATES.parameters())\n",
+    "scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 4)\n",
+    "\n",
+    "last_loss = 0\n",
+    "\n",
+    "current_epoch = 0\n",
+    "while current_epoch < MAX_EPOCHS:\n",
+    "\n",
+    "    encoder_list = torch.tensor([TOY_BATCH_INPUT_LIST[0]])\n",
+    "    decoder_list = torch.tensor([TOY_BATCH_DECODER_DEFAULT[0]])\n",
+    "    padding_list = torch.tensor([TOY_BATCH_PADDING_LIST[0]], dtype=torch.bool)\n",
+    "    target_logits = torch.tensor([TOY_BATCH_TARGET_LIST[0]]) # Transform target into logits\n",
+    "\n",
+    "    optimizer.zero_grad() # to clear gradient\n",
+    "\n",
+    "    last_loss = 0.0\n",
+    "\n",
+    "    for i in range(0, SENTENCE_LENGTH):\n",
+    "\n",
+    "        # optimizer.zero_grad()\n",
+    "        # forward \n",
+    "        logits: torch.Tensor = NANOSOCRATES((encoder_list, padding_list, decoder_list))\n",
+    "        # probabilities = torch.softmax(logits,2)\n",
+    "        \n",
+    "\n",
+    "        step_logits = logits[:, i, :]                 # [B, V]\n",
+    "        step_target = target_logits[:, i]             # [B]\n",
+    "\n",
+    "        loss = cross_entropy(step_logits,step_target) # now loss is without softmax\n",
+    "        loss.backward()  # DAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAMN\n",
+    "        last_loss = loss\n",
+    "        optimizer.step()\n",
+    "        optimizer.zero_grad()\n",
+    "        scheduler.step()\n",
+    "        \n",
+    "        probabilities = torch.softmax(logits,2)\n",
+    "        most_probable_tokens = torch.argmax(probabilities, 2) \n",
+    "        if i < SENTENCE_LENGTH - 1:\n",
+    "            decoder_list[:,i+1] = most_probable_tokens[:,i]\n",
+    "\n",
+    "\n",
+    "    current_epoch += 1\n",
+    "\n",
+    "    if current_epoch % 1 == 0:\n",
+    "        print(f\"EPOCH {current_epoch}\\n\\tLoss: {last_loss}\")\n",
+    "\n",
+    "\n",
+    "\n",
+    "\n",
+    "\n",
+    "\n"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "deep_learning",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.13.7"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

Playgrounds/prova.py

@@ -1,170 +0,0 @@
-import random
-import torch
-import pandas as pd
-from pathlib import Path
-import Project_Model.Libs.Embedder as Embedder
-import Project_Model.Libs.BPE as BPE
-import Project_Model.Libs.Transformer as Transformer
-import Project_Model.Libs.TorchShims as torch_shims
-
-# set a fixed seed
-torch.manual_seed(0)
-random.seed(0)
-DEVICE = torch_shims.get_default_device()
-torch.set_default_device(DEVICE)
-
-# set a default device
-
-# BPE Init
-VOCABULARY_PATH = Path("Assets/Model/toy_10/toy_dictionary.json")
-SPECIAL_VOC = BPE.default_special_tokens()
-
-VOCABULARY = BPE.load_nanos_vocabulary(VOCABULARY_PATH)
-TOKENANO = BPE.TokeNanoCore(VOCABULARY, SPECIAL_VOC)
-
-
-# Constants
-TOKEN_SPACE_SIZE = TOKENANO.vocabulary_size + 1
-EMBEDDED_SIZE = 256
-FEED_FORWARD_MULTIPLIER =  4
-ATTENTION_HEADS = 8
-SENTENCE_LENGTH = 256
-NUMBER_OF_BLOCKS = 4
-MAX_EPOCHS = int(1e3)
-
-
-PAD_TOKEN = TOKENANO.encode("<PAD>")[0]
-END_TOKEN = TOKENANO.encode("<END>")[0]
-
-
-# Load CSV
-TOY_DATASET_PATH = Path("Assets/Dataset/1-hop/toy/rdf_text.csv")
-
-TOY_DATASET = pd.read_csv(TOY_DATASET_PATH)
-
-TOY_BATCH_INPUT_LIST: list[list[int]] = []
-TOY_BATCH_PADDING_LIST: list[list[bool]] = []
-TOY_BATCH_TARGET_LIST: list[list[int]] = []
-TOY_BATCH_DECODER_DEFAULT: list[list[int]]= []
-
-
-for index, row in TOY_DATASET.iterrows():
-
-    RDFs: str = row["RDFs"]
-    Abstract: str = row["Abstract"]
-
-    input_tokens = TOKENANO.encode(RDFs)
-    output_tokens = TOKENANO.encode(Abstract)[1:]
-    decoder_default_tokens = TOKENANO.encode("<SOS>")
-
-    input_tokens, padding = Transformer.normalize_sequence(
-        input_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN
-    )
-    output_tokens, _ = Transformer.normalize_sequence(
-        output_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN
-    )
-    decoder_default_tokens, _ = Transformer.normalize_sequence(
-        decoder_default_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN, False
-    )
-
-    TOY_BATCH_INPUT_LIST.append(input_tokens)
-    TOY_BATCH_PADDING_LIST.append(padding)
-    TOY_BATCH_TARGET_LIST.append(output_tokens)
-    TOY_BATCH_DECODER_DEFAULT.append(decoder_default_tokens)
-
-    output_tokens = TOKENANO.encode(RDFs)
-    input_tokens = TOKENANO.encode(Abstract)[1:]
-    decoder_default_tokens = TOKENANO.encode("<SOS>")
-
-    input_tokens, padding = Transformer.normalize_sequence(
-        input_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN
-    )
-    output_tokens, _ = Transformer.normalize_sequence(
-        output_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN
-    )
-    decoder_default_tokens, _ = Transformer.normalize_sequence(
-        decoder_default_tokens, SENTENCE_LENGTH, PAD_TOKEN, END_TOKEN, False
-    )
-
-    TOY_BATCH_INPUT_LIST.append(input_tokens)
-    TOY_BATCH_PADDING_LIST.append(padding)
-    TOY_BATCH_TARGET_LIST.append(output_tokens)
-    TOY_BATCH_DECODER_DEFAULT.append(decoder_default_tokens)
-
-# Training loop
-LOSS_HISTORY = []
-NANOSOCRATES = Transformer.TrainingModel(
-    TOKEN_SPACE_SIZE,
-    EMBEDDED_SIZE,
-    FEED_FORWARD_MULTIPLIER,
-    ATTENTION_HEADS,
-    NUMBER_OF_BLOCKS
-)
-cross_entropy = torch.nn.CrossEntropyLoss(ignore_index=PAD_TOKEN)
-optimizer = torch.optim.AdamW(NANOSOCRATES.parameters())
-scheduler = Transformer.WarmupLR(optimizer, 4000, EMBEDDED_SIZE)
-last_loss = 0
-current_epoch = 0
-
-while current_epoch < MAX_EPOCHS:
-
-    optimizer.zero_grad()
-
-    encoder_list = torch.tensor(TOY_BATCH_INPUT_LIST[:])
-    decoder_list = torch.tensor(TOY_BATCH_DECODER_DEFAULT[:])
-    src_padding = torch.tensor(TOY_BATCH_PADDING_LIST[:], dtype=torch.bool)
-
-    # Transform target into logits
-    target_logits = torch.tensor(TOY_BATCH_TARGET_LIST[:])
-
-    last_loss = 0
-    last_prediction: torch.Tensor
-
-    for i in range(0, SENTENCE_LENGTH):
-
-        optimizer.zero_grad()
-        tgt_padding = decoder_list.eq(PAD_TOKEN)
-
-        logits: torch.Tensor = NANOSOCRATES((encoder_list, src_padding, decoder_list, tgt_padding))
-        prob = torch.softmax(logits, 2)
-
-        most_probable_tokens = torch.argmax(prob, 2)
-        last_prediction = most_probable_tokens
-
-        logits = logits[:,:i,:]
-        logits = logits.permute(0, 2, 1)
-
-        loss : torch.Tensor = cross_entropy(logits, target_logits[:, 0:i])
-        # loss : torch.Tensor = cross_entropy(logits, target_logits)
-
-        last_loss = loss
-        loss.backward()
-        optimizer.step()
-        scheduler.step()
-
-        if i < SENTENCE_LENGTH - 1:
-            decoder_list[:,i+1] = target_logits[:,i]
-
-
-
-
-
-
-    current_epoch += 1
-
-    if current_epoch % 1 == 0:
-        print(f"EPOCH {current_epoch}\n\tLoss: {last_loss}")
-
-        for encoded_sentence, expected_sentence in zip(
-            Transformer.tensor2token(last_prediction[:,:], END_TOKEN), # type: ignore
-            Transformer.tensor2token(target_logits[:,:], END_TOKEN)
-        ):
-            decoded_sentence = TOKENANO.decode(encoded_sentence)
-            decoded_target = TOKENANO.decode(expected_sentence)
-            print(f"\tACTUAL:\n\t\t{decoded_sentence}\n\tEXPECTED:\n\t\t{decoded_target}\n")
-
-
-
-
-
-

Project_Model/Libs/BPE/Classes/NanoSocratesBPE.py

@@ -189,7 +189,7 @@ class NanoSocratesBPE(Encoder):
             token_stack.appendleft(right_token)
             token_stack.appendleft(left_token)
 
-        return UTF_8_STRING_ARR.decode("utf-8", errors="ignore")
+        return UTF_8_STRING_ARR.decode("utf-8")
 
     def __token_decode(self, token_id: int) -> tuple[int, int]:
 

Project_Model/Libs/BPE/Classes/TokeNanoCore.py

@@ -31,7 +31,7 @@ class TokeNanoCore:
     def vocabulary_size(self):
         BPE_VOC_SIZE = self.__bpe_encoder.vocabulary_size
         SPECIAL_VOC_SIZE = self.__special_encoder.vocabulary_size
-        return BPE_VOC_SIZE + SPECIAL_VOC_SIZE + 1
+        return BPE_VOC_SIZE + SPECIAL_VOC_SIZE
 
     def encode(self, corpus: str) -> list[int]:
         output: list[int] = []

Project_Model/Libs/Batch/Classes/BatchEmbedder.py

@@ -0,0 +1,11 @@
+from ....Libs.Embedder.Classes.NanoSocratesEmbedder import NanoSocratesEmbedder
+import torch
+
+class BatchEmbedder(torch.nn.Module):
+
+    def __init__(self, vocabulary_size: int, embedding_size: int) -> None:
+        super().__init__()
+        self.__embedder = NanoSocratesEmbedder(vocabulary_size,embedding_size)
+        
+
+    def forward(self, )

Project_Model/Libs/Batch/Classes/Batcher.py

@@ -1,68 +1,49 @@
 import random
-import sys
-from typing import Any, Generator
+from typing import Generator
 import pandas as pd
-from pathlib import Path
-from Project_Model.Libs.Batch.Enums.TaskType import TaskType
+
 import Project_Model.Libs.BPE as BPE
-# from Scripts.Libs.CleaningPipeline.special_token import SpecialToken
-from Project_Model.Libs.Transformer import SpannedMasker, truncate_rdf_list, normalize_sequence
+from Scripts.Libs.CleaningPipeline.special_token import SpecialToken
+from Project_Model.Libs.Transformer.Classes.SpannedMasker import SpannedMasker
 from TokenCompletation import TokenCompletationTransformer
-from Project_Model.Libs.BPE import SpecialToken 
+from Project_Model.Libs.BPE.Enums.SpecialToken import SpecialToken
 
-
-MAX_LENGHT = 128
 class Batcher:
 
-    def __init__(self, dataset_path: Path, tokenizer: BPE.TokeNanoCore, masker: SpannedMasker, seed:int = 0) -> None:
+    def __init__(self, dataset_path: str, batch_size:int, tokenizer: BPE.TokeNanoCore, masker: SpannedMasker) -> None:
         # ABSTRACT, TRIPLE 
         # tasks: 
         #   rdf2text: X: TRIPLE, Y: ABSTRACT 
         #   text2rdf: X: ABSTRACT, X:TRIPLE 
         #   masking ( call masker): X: incomplete_triple Y: complete_triple (as exam) 
         #   completation: X: TRIPLE SUBSET, Y: related TRIPLE SUBSET
-        # it will truncate
-        # it will instantiate spanmaskter and truncator
         self._dataset_path = dataset_path
+        self._batch_size = batch_size
         self._tokenizer = tokenizer
         self._masker = masker
 
-        self._seed = seed
-        # self._token_completation = TokenCompletationTransformer(sotl,eos)
-        self._completation_task_token_truncator = truncate_rdf_list
+        sotl = self._tokenizer.encode(SpecialToken.START_TRIPLE_LIST.value)
+        eos = self._tokenizer.encode(SpecialToken.END_OF_SEQUENCE.value)
+        self._token_completation = TokenCompletationTransformer(sotl,eos)
 
 
-
-
-    def batch(self, batch_size)-> Generator[tuple[list[list[int]], list[list[int]], list[list[int]],list[list[int]], TaskType],Any,Any]:
-        """
-        Yields: X,Y,padding_X
-        """
-        RNG = random.Random(self._seed)
-        self._masker.reseed(self._seed)
-
-        for batch in pd.read_csv(self._dataset_path, chunksize= batch_size):
+    def get_batch(self)-> Generator[pd.DataFrame]:
+        for batch in pd.read_csv(self._dataset_path, chunksize= int(self._batch_size/4)): #now we support 3 task
 
             tokenized_batch = pd.DataFrame()
-            # encode
             tokenized_batch[["Abstract","RDFs"]] = (
                 batch[["Abstract","RDFs"]]
                 .map(lambda t: self._tokenizer.encode(t))
             )
             
-            X,Y, padding_X, padding_Y = self.__rdf2txt_transformation(tokenized_batch)
-            yield X,Y, padding_X, padding_Y, TaskType.RDF2TXT
-            X,Y, padding_X, padding_Y, = self.__txt2rdf_transformation(tokenized_batch)
-            yield X,Y, padding_X, padding_Y, TaskType.TEXT2RDF
-            X,Y, padding_X, padding_Y, = self.__masking_trasformation(tokenized_batch)
-            yield X,Y, padding_X, padding_Y, TaskType.MASKING
-            X,Y, padding_X, padding_Y, = self.__token_completation_task(tokenized_batch, RNG.randint(0,sys.maxsize))
-            yield X,Y, padding_X, padding_Y, TaskType.COMPLETATION
+            rdf2txt_batch = self.__rdf2txt_transformation(tokenized_batch)
+            txt2rdf_batch = self.__txt2rdf_transformation(tokenized_batch)
+            mask_batch = self.__masking_trasformation(tokenized_batch)
+            completation_batch = self.__token_completation_task(tokenized_batch)
 
-            # output = pd.concat([rdf2txt_batch,txt2rdf_batch,completation_batch],ignore_index=True)
-            # output = output.sample(frac=1).reset_index(drop=True)
-            # self.decode_debug(output)
-            # yield output
+            output = pd.concat([rdf2txt_batch,txt2rdf_batch,mask_batch,completation_batch],ignore_index=True)
+            output = output.sample(frac=1).reset_index(drop=True)
+            yield output
             
 
     def __random_subset_rdfs(self, batch: pd.DataFrame, seed = 0):
@@ -76,89 +57,48 @@ class Batcher:
             to_list
         )
 
-    def decode_debug(self, batch: pd.DataFrame):
-        decoded = pd.DataFrame()
-        decoded[["X","Y"]] = (
-                batch[["X","Y"]]
-                .map(lambda t: self._tokenizer.decode(t))
-            )
-        print(decoded)
-
-
-    def __normalization(self, X:list[list[int]], Y: list[list[int]])-> tuple[list[list[int]], list[list[int]], list[list[int]], list[list[int]]]:
-        pad_token = self._tokenizer.encode(SpecialToken.PAD.value)[0]
-        end_token = self._tokenizer.encode(SpecialToken.END_OF_SEQUENCE.value)[0]
-        out_X = []
-        padding_X = []
-        out_Y = []
-        padding_Y = []
-
-        for x in X:
-            out_x, padding_x = normalize_sequence(x,MAX_LENGHT,pad_token,end_token,True)
-            out_X.append(out_x)
-            padding_X.append(padding_x)
-        
-        for y in Y:
-            out_y, padding_y = normalize_sequence(y,MAX_LENGHT,pad_token,end_token,True)
-            out_Y.append(out_y)
-            padding_Y.append(padding_y)
-
-        return out_X,out_Y,padding_X,padding_Y
-
-
     def __rdf2txt_transformation(self, batch: pd.DataFrame):
-        task_token = self._tokenizer.encode(SpecialToken.RDF_TO_TEXT.value)
-        out = batch.rename(columns={"RDFs":"X","Abstract":"Y"})[["X","Y"]]
-        out["X"] = [task_token + x for x in out["X"]]
-        return self.__normalization(out["X"].to_list(),out["Y"].to_list())
+        batch = batch.rename(columns={"RDFs": "X", "Abstract": "Y"})
+        return batch[["X", "Y"]]
 
 
     def __txt2rdf_transformation(self, batch: pd.DataFrame):
-        task_token = self._tokenizer.encode(SpecialToken.TEXT_TO_RDF.value)
-        out = batch.rename(columns={"Abstract":"X","RDFs":"Y"})[["X","Y"]]
-        out["X"] = [task_token + x for x in out["X"]]
-        return self.__normalization(out["X"].to_list(),out["Y"].to_list())
-
+        batch = batch.rename(columns={ "Abstract": "X","RDFs": "Y"})
+        return batch[["X", "Y"]]
     
     def __masking_trasformation(self, batch: pd.DataFrame):
-        X = []
-        Y = []
-        for rdf in batch["RDFs"]:
-            x,y = self._masker.mask_sequence(rdf)
-            X.append(x)
-            Y.append(y)
-        return self.__normalization(X,Y)        
+        # mask_sequence: List[int] -> Tuple[List[int], List[int]]
+        xy_tuples = batch["RDFs"].apply(self._masker.mask_sequence)  # Series of (X, Y)
+
+        output = batch.copy()
+        # Expand into two columns preserving the original index
+        output[["X", "Y"]] = pd.DataFrame(xy_tuples.tolist(), index=batch.index)
+        return output[["X", "Y"]] 
         
     
-    def __token_completation_task(self, batch: pd.DataFrame, minibatch_seed: int):
-        continue_triple_token = self._tokenizer.encode(SpecialToken.CONTINUE_RDF.value)[0]
-        eot = self._tokenizer.encode(SpecialToken.END_TRIPLE.value)[0]
-        X = []
-        Y = []
-        for rdf in batch["RDFs"]:
-            x,y = self._completation_task_token_truncator(rdf, 0.5, continue_triple_token, eot, minibatch_seed)
-            X.append(x)
-            Y.append(y)
-        return self.__normalization(X,Y)
+    def __token_completation_task(self, batch: pd.DataFrame):
+        xy_tuples = batch["RDFs"].apply(self._token_completation.get_completation_tuple)
+        output = batch.copy()
+        output[["X", "Y"]] = pd.DataFrame(xy_tuples.tolist(), index=batch.index)
+        return output[["X", "Y"]] 
 
 
 
+"""
+DATASET_PATH = "Assets/Dataset/Tmp/rdf_text.csv"
+VOCABULARY_path = "Assets/Dataset/Tmp/trimmed.json"
 
-if __name__ == "__main__":
+from pathlib import Path
+VOCABULARY = BPE.load_nanos_vocabulary(Path(VOCABULARY_path))
+SPECIAL_LIST = BPE.default_special_tokens()
+TOKENANO = BPE.TokeNanoCore(VOCABULARY, SPECIAL_LIST)
+SPECIAL_TOKENS: set[int] = set(TOKENANO.encode("".join(SPECIAL_LIST)))
 
-    DATASET_PATH = Path("Assets/Dataset/Tmp/rdf_text.csv")
-    VOCABULARY_path = "Assets/Dataset/Tmp/trimmed.json"
+MASKER = SpannedMasker(TOKENANO.vocabulary_size,SPECIAL_TOKENS)
 
-    from pathlib import Path
-    VOCABULARY = BPE.load_nanos_vocabulary(Path(VOCABULARY_path))
-    SPECIAL_LIST = BPE.default_special_tokens()
-    TOKENANO = BPE.TokeNanoCore(VOCABULARY, SPECIAL_LIST)
-    SPECIAL_TOKENS: set[int] = set(TOKENANO.encode("".join(SPECIAL_LIST)))
-
-    MASKER = SpannedMasker(TOKENANO.vocabulary_size,SPECIAL_TOKENS)
-
-    prova = "<ABS>Cactus Flower is a 1969 American screwball comedy film directed by Gene Saks, and starring Walter Matthau, Ingrid Bergman and Goldie Hawn, who won an Academy Award for her performance.The screenplay was adapted by I. A. L. Diamond from the 1965 Broadway play of the same title written by Abe Burrows, which, in turn, is based on the French play Fleur de cactus by Pierre Barillet and Jean-Pierre Gredy. Cactus Flower was the ninth highest-grossing film of 1969."
-    print(TOKENANO.encode(prova))
-    batcher = Batcher(DATASET_PATH,TOKENANO,MASKER)
-    for batch in batcher.batch(8):
-        print(batch)
+prova = "<ABS>Cactus Flower is a 1969 American screwball comedy film directed by Gene Saks, and starring Walter Matthau, Ingrid Bergman and Goldie Hawn, who won an Academy Award for her performance.The screenplay was adapted by I. A. L. Diamond from the 1965 Broadway play of the same title written by Abe Burrows, which, in turn, is based on the French play Fleur de cactus by Pierre Barillet and Jean-Pierre Gredy. Cactus Flower was the ninth highest-grossing film of 1969."
+print(TOKENANO.encode(prova))
+batcher = Batcher(DATASET_PATH,8,TOKENANO,MASKER)
+for batch in batcher.get_batch():
+    print(batch)
+"""

Project_Model/Libs/Training/learning_rade_shedulers.py

@@ -0,0 +1,41 @@
+import numpy as np
+# custom LR from attention is all you need
+class Custom_lr():
+    def __init__(self, d_model: int, warmup_step:int) -> None:
+
+        self.__d_model = d_model
+        self.__warmup_step = warmup_step
+        self.__epoch = 0
+
+
+    def step(self) -> int:
+        self.__epoch += 1
+        return (self.__d_model ** -0.5) * min(self.__epoch ** -0.5,
+                                   self.__epoch * (self.__warmup_step ** -1.5))
+
+# OTHER LR
+
+# Learning rate schedules (matching visualization parameters)
+def step_lr(epoch, lr):
+    # StepLR: step_size=20, gamma=0.5 (from visualization)
+    return lr * 0.5 if epoch % 20 == 0 and epoch > 0 else lr
+ 
+def exp_lr(epoch, lr):
+    # ExponentialLR: gamma=0.95 (from visualization)
+    return lr * 0.95
+ 
+def cosine_lr(epoch, lr):
+    # CosineAnnealingLR: lr_min=0.001, lr_max=0.1, max_epochs=100 (from visualization)
+    lr_min, lr_max = 0.001, 0.1
+    max_epochs = 100
+    return lr_min + 0.5 * (lr_max - lr_min) * (1 + np.cos(epoch * np.pi / max_epochs))
+ 
+def cyclical_lr(epoch, lr):
+    # CyclicalLR: base_lr=0.001, max_lr=0.1, step_size=20 (from visualization)
+    base_lr = 0.001
+    max_lr = 0.1
+    step_size = 20
+    
+    cycle = np.floor(1 + epoch / (2 * step_size))
+    x = np.abs(epoch / step_size - 2 * cycle + 1)
+    return base_lr + (max_lr - base_lr) * max(0, (1 - x))

Project_Model/Libs/Training/logistic_collector.py

@@ -25,8 +25,8 @@ class LogitsCollector:
         for row in ids.tolist():
             seq: list[int] = []
             for tok in row:
-                # if tok == self.__end_token:               # stop on END
-                #   break
+                if tok == self.__end_token:               # stop on END
+                    break
                 if tok == self.__pad_token:               # skip PAD
                     continue
                 seq.append(tok)
@@ -36,7 +36,6 @@ class LogitsCollector:
     def print_decoded(self) -> None:
         for i, seq in enumerate(self.tokens()):
             try:
-                # text = text + self.__end_token
                 text = self.__tokenizer.decode(seq)       # decode tokens to string
             except Exception:
                 text = str(seq)                           # fallback to ids

Project_Model/Libs/Transformer/Classes/Decoder.py

@@ -46,14 +46,14 @@ class Decoder(nn.Module):
         ]
     ):  # -> list[torch.Tensor]:  # k_x = v_x . While x_q = x
         # WARNING: args is needed to have sequential
-        x, k_x, v_x, src_padding_mask, tgt_padding_mask = args
+        x, k_x, v_x, padding_mask,encoder_padding_mask = args
 
         # build of attention mask
         attention_mask = get_causal_attention_mask(x.size(1))
 
         # 1) Masked Attention
         MASKED_ATTENTION = self.__masked_attention(
-            x, x, x, key_padding_mask=tgt_padding_mask, attention_mask=attention_mask
+            x, x, x, key_padding_mask=padding_mask, attention_mask=attention_mask
         )
 
         # 2) Dropout
@@ -69,7 +69,7 @@ class Decoder(nn.Module):
 
         # 5) Encoder–decoder (cross) attention
         CROSS_ATTENTION = self.__cross_attention(
-            x, k_x, v_x, key_padding_mask=src_padding_mask
+            x, k_x, v_x, key_padding_mask=encoder_padding_mask
         )
 
         # 6) Dropout
@@ -97,7 +97,7 @@ class Decoder(nn.Module):
         # 12) Layer Normalization
         x = self.__layer_norm_3(x)
 
-        return (x, k_x, v_x, src_padding_mask, tgt_padding_mask)
+        return (x, k_x, v_x, padding_mask, encoder_padding_mask)
 
 
 # use eval to disable dropout ecc

Project_Model/Libs/Transformer/Classes/NanoSocrates.py

@@ -0,0 +1,23 @@
+import torch
+from NanoSocratesCore import NanoSocratesCore
+
+class NanoSocrates(torch.nn.Module):
+
+    def __init__(self,
+                embedded_size: int, 
+                feed_forward_dim: int, 
+                encoder_layers: int, 
+                decoder_layers:int,
+                attention_heads: int,
+                vocab_size: int) -> None:
+
+        super().__init__()
+
+        self._model = NanoSocratesCore(            
+                embedded_size, 
+                feed_forward_dim, 
+                encoder_layers, 
+                decoder_layers,
+                attention_heads,
+                vocab_size)
+        

Project_Model/Libs/Transformer/Classes/NanoSocratesCore.py

@@ -16,8 +16,11 @@ class NanoSocratesCore(torch.nn.Module):
         num_encoder_layers: int = 2,
         num_decoder_layers: int = 2,
         num_attention_heads: int = 4,
+        pad_token: int = 0, 
     ) -> None:
 
+        super().__init__() 
+        self.__pad_token = pad_token
         feed_forward_dim = embedding_size * feed_forward_multiplier
 
         self.__sentence_length = sentence_length
@@ -43,69 +46,64 @@ class NanoSocratesCore(torch.nn.Module):
         self.__input_embeder = NanoSocratesEmbedder(vocab_size, embedding_size)
         self.__output_embedder = NanoSocratesEmbedder(vocab_size, embedding_size)
 
+    @torch.no_grad()  # inference only
     def forward(
         self,
         encoder_input: list[list[int]],
-        decoder_input: list[list[int]],
-        encoder_padding_mask: list[list[int]],
+        decoder_input: list[list[int]],           # must start with <SOS> and PAD elsewhere
+        encoder_padding_mask: list[list[bool]],   # True where encoder is PAD
     ):
-
-        if len(encoder_padding_mask) != len(encoder_input):
-            raise Exception("Mismatch in received_dimensions")
-
-        # TODO: check for tensor in input to embedder
         # 1) Embed User-Input for encoders
-        ENCODER_INPUT = self.__input_embeder(encoder_input)
+        ENCODER_INPUT = self.__input_embeder(encoder_input)  # [B,S,E]
 
         # 2) Encode User-Input
-        ENCODER_OUTPUT, _ = self.__encoder_sequence(ENCODER_INPUT, encoder_padding_mask)
+        ENCODER_OUTPUT, encoder_padding_mask = self.__encoder_sequence(
+            (ENCODER_INPUT, encoder_padding_mask) # as tuple
+        )  # [B,S,E], [B,S]
         del ENCODER_INPUT
 
-        exit_loop = False
-        decoder_token_list = decoder_input[:]
+        # 3) Autoregressive Output (greedy)
+        LOGITS_HISTORY: list[torch.Tensor] = []   # keep per-step distributions
+        decoder_token_list = [row[:] for row in decoder_input]  # copy tokens
         decoder_phase = 0
+        exit_loop = False
 
-        LOGITS_HISTORY: list[torch.Tensor] = []
-
-        # 3) Autoregressive Output
         while not exit_loop:
+            decoder_phase += 1  # move to next position
 
-            # 3.0) Increment Counter
-            decoder_phase += 1
+            # 3.1) Build decoder key padding mask from current tokens (True where PAD)
+            DECODER_KEY_PADDING_MASK: list[list[bool]] = [
+                [tok == self.__pad_token for tok in row] for row in decoder_token_list
+            ]  # [B,T]
 
-            # 3.1) Embed Decoder Input
-            decoder_input = self.__output_embedder(decoder_token_list)
+            # 3.2) Embed Decoder Input (full sequence; decoder builds causal mask inside)
+            DECODER_INPUT = self.__output_embedder(decoder_token_list)  # [B,T,E]
 
-            # 3.2) Decode Decoder Input
+            # 3.3) Decode (self-attn uses causal mask internally; we provide PAD masks)
             DECODER_OUTPUT, _, _, _ = self.__decoder_sequence(
-                decoder_input, ENCODER_OUTPUT, ENCODER_OUTPUT
-            )
+                (DECODER_INPUT, ENCODER_OUTPUT, ENCODER_OUTPUT,
+                DECODER_KEY_PADDING_MASK, encoder_padding_mask)
+            )  # [B,T,E]
+            del DECODER_INPUT
 
-            # 3.3) Go back to Token space
-            # TODO: change name
-            LOGITS = self.__linear(DECODER_OUTPUT)
+            # 3.4) Project to token space
+            LOGITS = self.__linear(DECODER_OUTPUT)  # [B,T,V]
             del DECODER_OUTPUT
 
-            # 3.4) Transform in probabilities
-            # TODO: change name
-            TOKEN_PROBABILITIES = torch.softmax(LOGITS, dim=-1)
-            del LOGITS
+            # 3.5) Probabilities and greedy pick at current step
+            TOKEN_PROBABILITIES = torch.softmax(LOGITS, dim=-1)  # [B,T,V]
+            LOGITS_HISTORY.append(TOKEN_PROBABILITIES)  # store for this step
 
-            LOGITS_HISTORY.append(TOKEN_PROBABILITIES)
+            step_idx = decoder_phase - 1  # 0-based
+            TOKEN_IDS = TOKEN_PROBABILITIES[:, step_idx, :].argmax(dim=-1).tolist()  # [B] -> list[int]
 
-            # 3.5) Take most probable tokens
-            TOKEN_IDS = torch.argmax(TOKEN_PROBABILITIES, -1)
+            # 3.6) Write prediction into next slot (the slot is PAD)
+            if step_idx + 1 < self.__sentence_length:
+                for b, tok in enumerate(TOKEN_IDS):
+                    decoder_token_list[b][step_idx + 1] = tok  # feed next position
 
-            # TODO: check for dimensions and for efficiency
-            DECODER_TOKEN_TENSOR = torch.tensor(decoder_token_list)
-            DECODER_TOKEN_TENSOR[:, decoder_phase] = TOKEN_IDS
-            decoder_token_list = DECODER_TOKEN_TENSOR.tolist()
-
-            del TOKEN_IDS
-            del DECODER_TOKEN_TENSOR
-
-            # 3.6) Check if we generated all tokens
+            # 3.7) Stop when we filled the sequence
             if decoder_phase == self.__sentence_length - 1:
                 exit_loop = True
 
-        return LOGITS_HISTORY
+        return LOGITS_HISTORY  # list of [B,T,V] (per step)

Project_Model/Libs/Transformer/Classes/SpannedMasker.py

@@ -25,11 +25,6 @@ class SpannedMasker:
         self.__forbidden_tokens = forbidden_tokens
 
 
-    def reseed(self, seed:int):
-        self.__rng = random.Random(seed)
-
-
-
     def mask_sequence(
         self,
         token_sequence: list[int],

Project_Model/Libs/Transformer/Classes/WarmupLR.py

@@ -1,47 +0,0 @@
-from typing import override
-import torch
-
-
-# custom LR from attention is all you need
-class WarmupLR(torch.optim.lr_scheduler.LRScheduler):
-
-    def __init__(
-        self,
-        optimizer: torch.optim.Optimizer,
-        warmup_steps: int,
-        embedding_size: int,
-        warming_multiplier: float = -1.5,
-        decaying_multiplier: float = -0.5,
-        multiplicative_factor: float = 1.0,
-        last_epoch: int = -1,
-    ) -> None:
-        self.__warmup_steps = warmup_steps
-        self.__embedding_size = embedding_size
-        self.__warming_multiplier = warming_multiplier
-        self.__decaying_multiplier = decaying_multiplier
-        self.__multiplicative_factor = multiplicative_factor
-        super().__init__(optimizer, last_epoch)
-
-    def __scale_at(self, step: int) -> float:
-        step = max(step, 1)
-        return (
-            self.__multiplicative_factor
-            * (self.__embedding_size**self.__decaying_multiplier)
-            * min(
-                step**self.__decaying_multiplier,
-                step * (self.__warmup_steps**self.__warming_multiplier),
-            )
-        )
-
-    @override
-    def get_lr(self) -> list[float]:
-        torch.optim.lr_scheduler._warn_get_lr_called_within_step(self)
-
-        step = max(self.last_epoch, 1)
-        scale = self.__scale_at(step)
-        return [base_lr * scale for base_lr in self.base_lrs]
-
-    def _get_closed_form_lr(self):
-        step = max(self.last_epoch, 1)
-        scale = self.__scale_at(step)
-        return [base_lr * scale for base_lr in self.base_lrs]

Project_Model/Libs/Transformer/Classes/__init__.py

@@ -5,7 +5,6 @@ from .FeedForwardNetwork import FeedForwardNetwork
 from .TorchMultiHeadAttention import TorchMultiHeadAttention
 from .SpannedMasker import SpannedMasker
 from .DeToken import DeToken
-from .WarmupLR import WarmupLR
 
 __all__ = [
     "Decoder",
@@ -13,6 +12,5 @@ __all__ = [
     "FeedForwardNetwork",
     "TorchMultiHeadAttention",
     "SpannedMasker",
-    "DeToken",
-    "WarmupLR"
+    "DeToken"
 ]

Project_Model/Libs/Transformer/Models/TrainingModel.py

@@ -24,32 +24,49 @@ class TrainingModel(torch.nn.Module):
             vocabulary_size, latent_space
         )
 
-        TMP_ENCODERS = [
+        # do NOT share layer weights
+        enc_layers = [
             Encoder(latent_space, feed_forward_latent_space, attention_heads)
-        ] * layer_number
-
-        TMP_DECODERS = [
+            for _ in range(layer_number)
+        ]
+        dec_layers = [
             Decoder(latent_space, feed_forward_latent_space, attention_heads)
-        ] * layer_number
+            for _ in range(layer_number)
+        ]
 
-        self.__encoder = torch.nn.Sequential(*TMP_ENCODERS)
-        self.__decoder = torch.nn.Sequential(*TMP_DECODERS)
+        self.__encoder = torch.nn.Sequential(*enc_layers)
+        self.__decoder = torch.nn.Sequential(*dec_layers)
 
         self.__detokener = DeToken(latent_space, vocabulary_size)
 
-    def forward(self, args: tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]):
+    def forward(
+        self,
+        args: tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]
+    ):
+        # returns logits for the LAST decoder position only -> [B, V]
+        (
+            encoder_embedder_input,      # [B,S] encoder tokens
+            encoder_padding_mask,        # [B,S] True where encoder is PAD
+            decoder_embedder_prefix,     # [B,Tp] decoder prefix (e.g., <SOS> + tokens so far)
+            decoder_padding_mask,        # [B,Tp] True where decoder prefix has PAD
+        ) = args
 
-        encoder_embedder_input, src_padding, decoder_embedder_input, tgt_padding = args
+        # 1) embeddings
+        encoder_tensor = self.__encoder_embedder(encoder_embedder_input)    # [B,S,E]
+        decoder_tensor = self.__decoder_embedder(decoder_embedder_prefix)   # [B,Tp,E]
 
-        encoder_tensor = self.__encoder_embedder(encoder_embedder_input)
-        decoder_tensor = self.__decoder_embedder(decoder_embedder_input)
-
-        encoder_output, _ = self.__encoder((encoder_tensor, src_padding))
+        # 2) encode
+        encoder_output, _ = self.__encoder((encoder_tensor, encoder_padding_mask))  # [B,S,E], [B,S]
 
+        # 3) decode (causal mask is built inside the decoder)
         decoder_output, _, _, _, _ = self.__decoder(
-            (decoder_tensor, encoder_output, encoder_output, src_padding, tgt_padding)
-        )
+            (decoder_tensor, encoder_output, encoder_output,
+             decoder_padding_mask, encoder_padding_mask)
+        )  # [B,Tp,E], ...
 
-        logits: torch.Tensor = self.__detokener(decoder_output)
+        # 4) project only the last time step
+        last_hidden = decoder_output[:, -1:, :]       # [B,1,E]
+        step_logits = self.__detokener(last_hidden)   # [B,1,V]
+        step_logits = step_logits[:, -1, :]           # [B,V]
 
-        return logits
+        return step_logits                             # logits for one token

Project_Model/Libs/Transformer/Utils/__init__.py

@@ -3,7 +3,6 @@ from .task_type import TaskType
 from .post_tokenization import truncate_sequence, pad_sequence, normalize_sequence, create_padding_mask
 from .inference_masking import inference_masking
 from .truncate_rdf_list import truncate_rdf_list
-from .decode_out import tensor2token
 
 __all__ = [
     "TaskType",
@@ -14,6 +13,5 @@ __all__ = [
     "create_padding_mask",
     "normalize_sequence",
     "inference_masking",
-    "truncate_rdf_list",
-    "tensor2token"
+    "truncate_rdf_list"
 ]

Project_Model/Libs/Transformer/Utils/decode_out.py

@@ -1,27 +0,0 @@
-from typing import Generator
-
-import torch
-
-
-def tensor2token(tensor: torch.Tensor, end_token: int) -> Generator[list[int]]:
-
-    if len(tensor.shape) < 1 or len(tensor.shape) > 2:
-        raise ValueError("Shape is not correct")
-
-    if len(tensor.shape) == 1:
-        token_list: list[int] = tensor.tolist()
-        token_list.append(end_token)
-        yield token_list
-        return
-
-    batch_len: int
-    batch_len, _ = tensor.shape
-
-    for i in range(batch_len):
-
-        smaller_tensor = tensor[i, :]
-        token_list: list[int] = smaller_tensor.tolist()
-        token_list.append(end_token)
-        yield token_list
-
-

Project_Model/Libs/Transformer/Utils/post_tokenization.py

@@ -1,20 +1,17 @@
 def truncate_sequence(
-    sequence: list[int], truncate_at: int, end_token: int, add_ending: bool
+    sequence: list[int], truncate_at: int, end_token: int
 ) -> list[int]:
 
     if len(sequence) < truncate_at - 1:
-        if add_ending:
-            sequence.append(end_token)
+        sequence.append(end_token)
         return sequence
 
     if len(sequence) < truncate_at:
-        if add_ending:
-            sequence[-1] = end_token
+        sequence[-1] = end_token
         return sequence
 
     TRUNCATED_SEQUENCE = sequence[:truncate_at]
-    if add_ending:
-        TRUNCATED_SEQUENCE[-1] = end_token
+    TRUNCATED_SEQUENCE[-1] = end_token
 
     return TRUNCATED_SEQUENCE
 
@@ -51,9 +48,8 @@ def normalize_sequence(
     max_length: int,
     pad_token: int,
     end_token: int,
-    add_ending: bool = True
 ) -> tuple[list[int], list[bool]]:
-    new_sequence = truncate_sequence(sequence, max_length, end_token, add_ending)
+    new_sequence = truncate_sequence(sequence, max_length, end_token)
     new_sequence = pad_sequence(new_sequence, max_length, pad_token)
     PADDING_MASK = create_padding_mask(new_sequence, pad_token)
 

requirements.txt

@@ -16,4 +16,3 @@ urllib3==2.5.0
 wheel==0.45.1
 Wikipedia-API==0.8.1
 SQLAlchemy
-torch