{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.11.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"tpuV5e8","dataSources":[{"sourceId":106809,"databundleVersionId":13056355,"sourceType":"competition"}],"dockerImageVersionId":31194,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input'):\n    for filename in filenames:\n        print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport h5py\nimport glob\nimport math\nimport random\nimport string\nimport numpy as np\nfrom collections import Counter, defaultdict\nimport torch\nimport torch.nn as nn\nfrom torch.utils.data import Dataset, DataLoader\nimport torch.nn.functional as F\nfrom tqdm import tqdm\nfrom pathlib import Path\nimport csv\n\nDEVICE = \"cuda\" if torch.cuda.is_available() else \"cpu\"\nINPUT_ROOT = \"/kaggle/input/brain-to-text-25\"\nTEST_H5_FILES = sorted(glob.glob(os.path.join(INPUT_ROOT, \"t15_copyTask_neuralData/hdf5_data_final/**/data_test.hdf5\"), recursive=True))\nTRAIN_H5_FILES = sorted(glob.glob(os.path.join(INPUT_ROOT, \"t15_copyTask_neuralData/hdf5_data_final/**/data_train.hdf5\"), recursive=True))\nVAL_H5_FILES = sorted(glob.glob(os.path.join(INPUT_ROOT, \"t15_copyTask_neuralData/hdf5_data_final/**/data_val.hdf5\"), recursive=True))\nOUT_DIR = \"/kaggle/working/brain2text_out\"\nos.makedirs(OUT_DIR, exist_ok=True)\n\ndef load_h5_file_list(file_list, subset=\"train\"):\n    records = []\n    for fp in file_list:\n        with h5py.File(fp, \"r\") as f:\n            keys = list(f.keys())\n            for k in keys:\n                g = f[k]\n                arr = g['input_features'][:]\n                n_time_steps = int(g.attrs.get('n_time_steps', arr.shape[1] if arr.ndim>1 else 0))\n                transcription = None\n                if 'transcription' in g:\n                    t = g['transcription'][:]\n                    try:\n                        transcription = t.astype('U')[0] if isinstance(t, np.ndarray) else t.decode() if isinstance(t, (bytes, bytearray)) else str(t)\n                    except:\n                        transcription = \"\".join([chr(x) for x in t]) if isinstance(t, (list, np.ndarray)) else str(t)\n                sentence_label = g.attrs.get('sentence_label', None)\n                session = g.attrs.get('session', \"\")\n                block_num = int(g.attrs.get('block_num', -1))\n                trial_num = int(g.attrs.get('trial_num', -1))\n                records.append({\n                    \"features\": arr.astype(np.float32),\n                    \"n_time_steps\": n_time_steps,\n                    \"transcription\": transcription if transcription is not None else \"\",\n                    \"sentence_label\": sentence_label if sentence_label is not None else \"\",\n                    \"session\": str(session),\n                    \"block_num\": block_num,\n                    \"trial_num\": trial_num,\n                    \"source_file\": fp\n                })\n    return records\n\ntrain_records = load_h5_file_list(TRAIN_H5_FILES, \"train\")\nval_records = load_h5_file_list(VAL_H5_FILES, \"val\")\ntest_records = load_h5_file_list(TEST_H5_FILES, \"test\")\n\ndef normalize_features(records):\n    all_feats = np.concatenate([r[\"features\"].reshape(-1, r[\"features\"].shape[-1]) for r in records], axis=0)\n    mean = all_feats.mean(axis=0)\n    std = all_feats.std(axis=0) + 1e-6\n    for r in records:\n        r[\"features\"] = (r[\"features\"] - mean) / std\n    return mean, std\n\nnormalize_features(train_records)\n\ndef build_char_vocab(records):\n    cnt = Counter()\n    for r in records:\n        txt = r[\"transcription\"]\n        if txt is None: txt = \"\"\n        txt = txt.lower()\n        for ch in txt:\n            cnt[ch] += 1\n    chars = sorted([c for c in cnt.keys() if ord(c) < 256 and (c.isalpha() or c==\" \" or c==\"'\")])\n    if \" \" not in chars:\n        chars.append(\" \")\n    if \"'\" not in chars:\n        chars.append(\"'\")\n    chars = [\"<blank>\"] + chars\n    idx = {c:i for i,c in enumerate(chars)}\n    return chars, idx\n\nCHARS, CHAR2IDX = build_char_vocab(train_records)\nBLANK_IDX = 0\nNUM_CLASSES = len(CHARS)\n\ndef text_to_indices(text):\n    t = text.lower()\n    res = [CHAR2IDX[c] for c in t if c in CHAR2IDX]\n    return res\n\nclass H5Dataset(Dataset):\n    def __init__(self, records, max_len=None, with_transcripts=True):\n        self.records = records\n        self.with_transcripts = with_transcripts\n        self.max_len = max_len\n    def __len__(self):\n        return len(self.records)\n    def __getitem__(self, idx):\n        r = self.records[idx]\n        x = r[\"features\"].T\n        x = np.nan_to_num(x).astype(np.float32)\n        transcription = r[\"transcription\"] if self.with_transcripts else \"\"\n        y = text_to_indices(transcription) if self.with_transcripts else []\n        return x, np.array(y, dtype=np.int32), r[\"session\"], r[\"block_num\"], r[\"trial_num\"]\n\ndef collate_fn(batch):\n    xs, ys, sessions, blocks, trials = zip(*batch)\n    lengths = [x.shape[0] for x in xs]\n    max_len = max(lengths)\n    feat_size = xs[0].shape[1]\n    xbatch = np.zeros((len(xs), max_len, feat_size), dtype=np.float32)\n    for i,x in enumerate(xs):\n        xbatch[i,:x.shape[0],:] = x\n    y_lengths = [len(y) for y in ys]\n    if sum(y_lengths)==0:\n        ybatch = np.zeros((0,), dtype=np.int32)\n    else:\n        ybatch = np.concatenate(ys) if len(ys)>0 else np.zeros((0,), dtype=np.int32)\n    return torch.from_numpy(xbatch), torch.tensor(lengths, dtype=torch.long), torch.from_numpy(ybatch), torch.tensor(y_lengths, dtype=torch.long), sessions, blocks, trials\n\nclass PositionalEncoding(nn.Module):\n    def __init__(self, d_model, max_len=10000):\n        super().__init__()\n        pe = torch.zeros(max_len, d_model)\n        position = torch.arange(0,max_len,dtype=torch.float).unsqueeze(1)\n        div_term = torch.exp(torch.arange(0,d_model,2).float()*(-math.log(10000.0)/d_model))\n        pe[:,0::2] = torch.sin(position*div_term)\n        pe[:,1::2] = torch.cos(position*div_term)\n        self.pe = pe.unsqueeze(0)\n    def forward(self,x):\n        L = x.size(1)\n        return x + self.pe[:,:L].to(x.device)\n\nclass Encoder(nn.Module):\n    def __init__(self, input_dim, d_model=512, nhead=8, num_layers=6, dim_feedforward=2048, dropout=0.1):\n        super().__init__()\n        self.in_proj = nn.Linear(input_dim, d_model)\n        self.pos = PositionalEncoding(d_model)\n        encoder_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead, dim_feedforward=dim_feedforward, dropout=dropout, activation='relu')\n        self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)\n        self.layer_norm = nn.LayerNorm(d_model)\n    def forward(self,x, lengths):\n        x = self.in_proj(x)\n        x = self.pos(x)\n        x = x.transpose(0,1)\n        mask = torch.zeros(x.size(0), x.size(1), dtype=torch.bool, device=x.device)\n        max_len = x.size(0)\n        for i,l in enumerate(lengths):\n            if l < max_len:\n                mask[l:,i] = True\n        out = self.encoder(x, src_key_padding_mask=mask)\n        out = out.transpose(0,1)\n        out = self.layer_norm(out)\n        return out, mask\n\nclass CTCModel(nn.Module):\n    def __init__(self, input_dim, num_classes):\n        super().__init__()\n        self.enc = Encoder(input_dim, d_model=512, nhead=8, num_layers=6)\n        self.fc = nn.Linear(512, num_classes)\n    def forward(self,x,lengths):\n        enc_out, mask = self.enc(x, lengths)\n        logits = self.fc(enc_out)\n        log_probs = F.log_softmax(logits, dim=-1)\n        return log_probs, mask\n\ndef cer_wer(hyp, ref):\n    h = hyp.strip().split()\n    r = ref.strip().split()\n    if len(r)==0:\n        return len(h), len(r)\n    dp = [[0]*(len(h)+1) for _ in range(len(r)+1)]\n    for i in range(len(r)+1):\n        dp[i][0] = i\n    for j in range(len(h)+1):\n        dp[0][j] = j\n    for i in range(1,len(r)+1):\n        for j in range(1,len(h)+1):\n            if r[i-1]==h[j-1]:\n                dp[i][j]=dp[i-1][j-1]\n            else:\n                dp[i][j]=1+min(dp[i-1][j],dp[i][j-1],dp[i-1][j-1])\n    return dp[len(r)][len(h)], len(r)\n\ndef greedy_decode(log_probs, lengths):\n    preds = []\n    for i, l in enumerate(lengths):\n        lp = log_probs[i,:l,:].cpu().numpy()\n        idxs = lp.argmax(axis=1).tolist()\n        collapsed = []\n        prev = None\n        for t in idxs:\n            if t!=prev and t!=BLANK_IDX:\n                collapsed.append(t)\n            prev = t\n        text = \"\".join([CHARS[t] for t in collapsed])\n        preds.append(text)\n    return preds\n\ntrain_ds = H5Dataset(train_records, with_transcripts=True)\nval_ds = H5Dataset(val_records, with_transcripts=True)\ntest_ds = H5Dataset(test_records, with_transcripts=False)\n\nBATCH = 8\ntrain_loader = DataLoader(train_ds, batch_size=BATCH, shuffle=True, collate_fn=collate_fn, num_workers=2)\nval_loader = DataLoader(val_ds, batch_size=BATCH, shuffle=False, collate_fn=collate_fn, num_workers=2)\ntest_loader = DataLoader(test_ds, batch_size=BATCH, shuffle=False, collate_fn=collate_fn, num_workers=2)\n\nmodel = CTCModel(input_dim=train_records[0][\"features\"].shape[1], num_classes=NUM_CLASSES).to(DEVICE)\noptimizer = torch.optim.AdamW(model.parameters(), lr=3e-4, weight_decay=1e-2)\nscheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=0.5, patience=2, verbose=True)\nctc_loss = nn.CTCLoss(blank=BLANK_IDX, zero_infinity=True)\n\nEPOCHS = 12\n\nbest_wer = 1.0\nfor epoch in range(EPOCHS):\n    model.train()\n    running_loss = 0.0\n    pbar = tqdm(train_loader, desc=f\"train epoch {epoch+1}/{EPOCHS}\")\n    for xb, x_lens, y_concat, y_lens, *_ in pbar:\n        xb = xb.to(DEVICE)\n        x_lens = x_lens.to(DEVICE)\n        optimizer.zero_grad()\n        log_probs, mask = model(xb, x_lens)\n        log_probs_t = log_probs.permute(1,0,2)\n        if y_lens.sum().item() == 0:\n            continue\n        input_lengths = x_lens\n        target_lengths = y_lens\n        loss = ctc_loss(log_probs_t, y_concat.to(DEVICE), input_lengths, target_lengths)\n        loss.backward()\n        torch.nn.utils.clip_grad_norm_(model.parameters(), 5.0)\n        optimizer.step()\n        running_loss += loss.item()\n        pbar.set_postfix(loss=running_loss/ (pbar.n+1))\n    model.eval()\n    total_ed = 0\n    total_words = 0\n    with torch.no_grad():\n        for xb, x_lens, y_concat, y_lens, sessions, blocks, trials in tqdm(val_loader, desc=\"val\"):\n            xb = xb.to(DEVICE)\n            x_lens = x_lens.to(DEVICE)\n            log_probs, mask = model(xb, x_lens)\n            preds = greedy_decode(log_probs, x_lens.cpu().numpy().tolist())\n            idx = 0\n            for i, yl in enumerate(y_lens.tolist()):\n                ref = \"\"\n                if yl>0:\n                    ref = \"\".join([CHARS[int(x)] for x in y_concat[idx:idx+yl].tolist()])\n                idx += yl\n                hyp = preds[i]\n                ed, nwords = cer_wer(hyp, ref)\n                total_ed += ed\n                total_words += nwords\n    val_wer = total_ed / total_words if total_words>0 else 1.0\n    scheduler.step(val_wer)\n    if val_wer < best_wer:\n        best_wer = val_wer\n        torch.save(model.state_dict(), os.path.join(OUT_DIR, \"best_model.pth\"))\n    print(f\"Epoch {epoch+1} val WER: {val_wer:.4f} best WER: {best_wer:.4f}\")\n\nmodel.load_state_dict(torch.load(os.path.join(OUT_DIR, \"best_model.pth\"), map_location=DEVICE))\nmodel.eval()\nall_preds = []\nall_meta = []\nwith torch.no_grad():\n    for xb, x_lens, y_concat, y_lens, sessions, blocks, trials in tqdm(test_loader, desc=\"test infer\"):\n        xb = xb.to(DEVICE)\n        x_lens = x_lens.to(DEVICE)\n        log_probs, mask = model(xb, x_lens)\n        preds = greedy_decode(log_probs, x_lens.cpu().numpy().tolist())\n        for p,s,b,t in zip(preds,sessions,blocks,trials):\n            all_preds.append(p)\n            all_meta.append((s,b,t))\n\ndef make_submission(preds, metas, out_csv=\"/kaggle/working/submission.csv\"):\n    order = list(range(len(preds)))\n    combined = list(zip(preds, metas, order))\n    def meta_key(m):\n        s,b,t = m\n        try:\n            sd = str(s)\n            return (sd, int(b), int(t))\n        except:\n            return (str(s), int(b), int(t))\n    combined_sorted = sorted(combined, key=lambda x: meta_key(x[1]))\n    with open(out_csv, \"w\", newline='') as f:\n        w = csv.writer(f)\n        w.writerow([\"id\",\"text\"])\n        for i,(pred,meta,orig_i) in enumerate(combined_sorted):\n            text = pred.strip()\n            if text==\"\":\n                text = \" \" \n            w.writerow([i, text])\n    return out_csv\n\nsub_path = make_submission(all_preds, all_meta, out_csv=\"/kaggle/working/submission.csv\")\nprint(\"submission saved to\", sub_path)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-12-09T18:54:06.478594Z","iopub.execute_input":"2025-12-09T18:54:06.479381Z","execution_failed":"2025-12-09T18:58:55.412Z"}},"outputs":[],"execution_count":null}]}