{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.10","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceId":52950,"databundleVersionId":5973250,"sourceType":"competition"},{"sourceId":6182721,"sourceType":"datasetVersion","datasetId":3497052},{"sourceId":6342181,"sourceType":"datasetVersion","datasetId":3651580},{"sourceId":192633,"sourceType":"modelInstanceVersion","isSourceIdPinned":true,"modelInstanceId":164262,"modelId":186601}],"dockerImageVersionId":30512,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import os\nimport gc\nimport json\nimport math\nimport pickle\nimport numpy as np\nimport pandas as pd\nfrom tqdm import tqdm\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nimport tensorflow_addons as tfa\nIS_INTERACTIVE = os.environ['KAGGLE_KERNEL_RUN_TYPE'] == 'Interactive'","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:11.946884Z","iopub.execute_input":"2024-12-08T14:26:11.947233Z","iopub.status.idle":"2024-12-08T14:26:11.952084Z","shell.execute_reply.started":"2024-12-08T14:26:11.947206Z","shell.execute_reply":"2024-12-08T14:26:11.95115Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"with open (\"/kaggle/input/asl-fingerspelling/character_to_prediction_index.json\", \"r\") as f:\n char_to_num = json.load(f)\n\npad_token = '^'\npad_token_idx = 59\n\nchar_to_num[pad_token] = pad_token_idx\n\nnum_to_char = {j:i for i,j in char_to_num.items()}\ndf = pd.read_csv('/kaggle/input/asl-fingerspelling/train.csv')\n\nLIP = [\n 61, 185, 40, 39, 37, 0, 267, 269, 270, 409,\n 291, 146, 91, 181, 84, 17, 314, 405, 321, 375,\n 78, 191, 80, 81, 82, 13, 312, 311, 310, 415,\n 95, 88, 178, 87, 14, 317, 402, 318, 324, 308,\n]\nLPOSE = [13, 15, 17, 19, 21]\nRPOSE = [14, 16, 18, 20, 22]\nPOSE = LPOSE + RPOSE\n\nX = [f'x_right_hand_{i}' for i in range(21)] +[f'x_left_hand_{i}' for i in range(21)] + [f'x_pose_{i}' for i in POSE] + [f'x_face_{i}' for i in LIP] #+ \nY = [f'y_right_hand_{i}' for i in range(21)] +[f'y_left_hand_{i}' for i in range(21)]+ [f'y_pose_{i}' for i in POSE] + [f'y_face_{i}' for i in LIP] #+\nZ = [f'z_right_hand_{i}' for i in range(21)] + [f'z_left_hand_{i}' for i in range(21)]+ [f'z_pose_{i}' for i in POSE] + [f'z_face_{i}' for i in LIP] #+ \n\nSEL_COLS = X + Y + Z\nFRAME_LEN = 128 + 48\nMAX_PHRASE_LENGTH = 64\n\nLIP_IDX_X = [i for i, col in enumerate(SEL_COLS) if \"face\" in col and \"x\" in col]\nRHAND_IDX_X = [i for i, col in enumerate(SEL_COLS) if \"right\" in col and \"x\" in col]\nLHAND_IDX_X = [i for i, col in enumerate(SEL_COLS) if \"left\" in col and \"x\" in col]\nRPOSE_IDX_X = [i for i, col in enumerate(SEL_COLS) if \"pose\" in col and int(col[-2:]) in RPOSE and \"x\" in col]\nLPOSE_IDX_X = [i for i, col in enumerate(SEL_COLS) if \"pose\" in col and int(col[-2:]) in LPOSE and \"x\" in col]\n\nLIP_IDX_Y = [i for i, col in enumerate(SEL_COLS) if \"face\" in col and \"y\" in col]\nRHAND_IDX_Y = [i for i, col in enumerate(SEL_COLS) if \"right\" in col and \"y\" in col]\nLHAND_IDX_Y = [i for i, col in enumerate(SEL_COLS) if \"left\" in col and \"y\" in col]\nRPOSE_IDX_Y = [i for i, col in enumerate(SEL_COLS) if \"pose\" in col and int(col[-2:]) in RPOSE and \"y\" in col]\nLPOSE_IDX_Y = [i for i, col in enumerate(SEL_COLS) if \"pose\" in col and int(col[-2:]) in LPOSE and \"y\" in col]\n\nLIP_IDX_Z = [i for i, col in enumerate(SEL_COLS) if \"face\" in col and \"z\" in col]\nRHAND_IDX_Z = [i for i, col in enumerate(SEL_COLS) if \"right\" in col and \"z\" in col]\nLHAND_IDX_Z = [i for i, col in enumerate(SEL_COLS) if \"left\" in col and \"z\" in col]\nRPOSE_IDX_Z = [i for i, col in enumerate(SEL_COLS) if \"pose\" in col and int(col[-2:]) in RPOSE and \"z\" in col]\nLPOSE_IDX_Z = [i for i, col in enumerate(SEL_COLS) if \"pose\" in col and int(col[-2:]) in LPOSE and \"z\" in col]\n\nRHM = np.load(\"/kaggle/input/aslfr-dataset-tfrecords/mean_std/rh_mean.npy\")\nLHM = np.load(\"/kaggle/input/aslfr-dataset-tfrecords/mean_std/lh_mean.npy\")\nRPM = np.load(\"/kaggle/input/aslfr-dataset-tfrecords/mean_std/rp_mean.npy\")\nLPM = np.load(\"/kaggle/input/aslfr-dataset-tfrecords/mean_std/lp_mean.npy\")\nLIPM = np.load(\"/kaggle/input/aslfr-dataset-tfrecords/mean_std/lip_mean.npy\")\n\nRHS = np.load(\"/kaggle/input/aslfr-dataset-tfrecords/mean_std/rh_std.npy\")\nLHS = np.load(\"/kaggle/input/aslfr-dataset-tfrecords/mean_std/lh_std.npy\")\nRPS = np.load(\"/kaggle/input/aslfr-dataset-tfrecords/mean_std/rp_std.npy\")\nLPS = np.load(\"/kaggle/input/aslfr-dataset-tfrecords/mean_std/lp_std.npy\")\nLIPS = np.load(\"/kaggle/input/aslfr-dataset-tfrecords/mean_std/lip_std.npy\")","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:15.498336Z","iopub.execute_input":"2024-12-08T14:26:15.499154Z","iopub.status.idle":"2024-12-08T14:26:15.726731Z","shell.execute_reply.started":"2024-12-08T14:26:15.499124Z","shell.execute_reply":"2024-12-08T14:26:15.725794Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def load_relevant_data_subset(pq_path):\n return pd.read_parquet(pq_path, columns=SEL_COLS)\n\nfile_id = df.file_id.iloc[0]\ninpdir = \"/kaggle/input/asl-fingerspelling/train_landmarks\"\npqfile = f\"{inpdir}/{file_id}.parquet\"\nseq_refs = df.loc[df.file_id == file_id]\nseqs = load_relevant_data_subset(pqfile)\n\nseq_id = seq_refs.sequence_id.iloc[0]\nframes = seqs.iloc[seqs.index == seq_id]\nphrase = str(df.loc[df.sequence_id == seq_id].phrase.iloc[0])","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:18.909112Z","iopub.execute_input":"2024-12-08T14:26:18.909976Z","iopub.status.idle":"2024-12-08T14:26:21.489079Z","shell.execute_reply.started":"2024-12-08T14:26:18.909947Z","shell.execute_reply":"2024-12-08T14:26:21.488387Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# First convert numpy arrays to tensorflow constants\nRHM = tf.constant(RHM, dtype=tf.float32)\nLHM = tf.constant(LHM, dtype=tf.float32)\nRPM = tf.constant(RPM, dtype=tf.float32)\nLPM = tf.constant(LPM, dtype=tf.float32)\nLIPM = tf.constant(LIPM, dtype=tf.float32)\n\nRHS = tf.constant(RHS, dtype=tf.float32)\nLHS = tf.constant(LHS, dtype=tf.float32)\nRPS = tf.constant(RPS, dtype=tf.float32)\nLPS = tf.constant(LPS, dtype=tf.float32)\nLIPS = tf.constant(LIPS, dtype=tf.float32)\n\n@tf.function()\ndef resize_pad(x):\n if tf.shape(x)[0] < FRAME_LEN:\n padding = [[0, FRAME_LEN-tf.shape(x)[0]], [0, 0], [0, 0]]\n x = tf.pad(x, padding, constant_values=float(\"NaN\"))\n else:\n x = tf.image.resize(x, (FRAME_LEN, tf.shape(x)[1]))\n return x\n\n#@tf.function()\ndef pre_process0(x):\n lip_x = tf.gather(x, LIP_IDX_X, axis=1)\n lip_y = tf.gather(x, LIP_IDX_Y, axis=1)\n lip_z = tf.gather(x, LIP_IDX_Z, axis=1)\n\n rhand_x = tf.gather(x, RHAND_IDX_X, axis=1)\n rhand_y = tf.gather(x, RHAND_IDX_Y, axis=1)\n rhand_z = tf.gather(x, RHAND_IDX_Z, axis=1)\n \n lhand_x = tf.gather(x, LHAND_IDX_X, axis=1)\n lhand_y = tf.gather(x, LHAND_IDX_Y, axis=1)\n lhand_z = tf.gather(x, LHAND_IDX_Z, axis=1)\n\n rpose_x = tf.gather(x, RPOSE_IDX_X, axis=1)\n rpose_y = tf.gather(x, RPOSE_IDX_Y, axis=1)\n rpose_z = tf.gather(x, RPOSE_IDX_Z, axis=1)\n \n lpose_x = tf.gather(x, LPOSE_IDX_X, axis=1)\n lpose_y = tf.gather(x, LPOSE_IDX_Y, axis=1)\n lpose_z = tf.gather(x, LPOSE_IDX_Z, axis=1)\n \n lip = tf.concat([lip_x[..., tf.newaxis], lip_y[..., tf.newaxis], lip_z[..., tf.newaxis]], axis=-1)\n rhand = tf.concat([rhand_x[..., tf.newaxis], rhand_y[..., tf.newaxis], rhand_z[..., tf.newaxis]], axis=-1)\n lhand = tf.concat([lhand_x[..., tf.newaxis], lhand_y[..., tf.newaxis], lhand_z[..., tf.newaxis]], axis=-1)\n rpose = tf.concat([rpose_x[..., tf.newaxis], rpose_y[..., tf.newaxis], rpose_z[..., tf.newaxis]], axis=-1)\n lpose = tf.concat([lpose_x[..., tf.newaxis], lpose_y[..., tf.newaxis], lpose_z[..., tf.newaxis]], axis=-1)\n \n # TIME AUGMENTATION\n if tf.random.uniform(shape=(), minval=0, maxval=1) < 0.2:\n new_width = tf.shape(lip)[1]\n height = tf.cast(tf.shape(lip)[0], tf.float32)\n min_height = tf.cast(height / 2.0, tf.int32)\n max_height = tf.cast(height * 1.5, tf.int32)\n \n new_height = tf.random.uniform(\n shape=(), \n minval=min_height,\n maxval=max_height,\n dtype=tf.int32\n )\n \n resized_lip = tf.image.resize(lip, (new_height, new_width))\n resized_rhand = tf.image.resize(rhand, (new_height, new_width))\n resized_lhand = tf.image.resize(lhand, (new_height, new_width))\n resized_rpose = tf.image.resize(rpose, (new_height, new_width))\n resized_lpose = tf.image.resize(lpose, (new_height, new_width))\n \n return resized_lip, resized_rhand, resized_lhand, resized_rpose, resized_lpose\n \n return lip, rhand, lhand, rpose, lpose\n\n@tf.function(jit_compile=True)\ndef pre_process00(x):\n lip_x = tf.gather(x, LIP_IDX_X, axis=1)\n lip_y = tf.gather(x, LIP_IDX_Y, axis=1)\n lip_z = tf.gather(x, LIP_IDX_Z, axis=1)\n\n rhand_x = tf.gather(x, RHAND_IDX_X, axis=1)\n rhand_y = tf.gather(x, RHAND_IDX_Y, axis=1)\n rhand_z = tf.gather(x, RHAND_IDX_Z, axis=1)\n \n lhand_x = tf.gather(x, LHAND_IDX_X, axis=1)\n lhand_y = tf.gather(x, LHAND_IDX_Y, axis=1)\n lhand_z = tf.gather(x, LHAND_IDX_Z, axis=1)\n\n rpose_x = tf.gather(x, RPOSE_IDX_X, axis=1)\n rpose_y = tf.gather(x, RPOSE_IDX_Y, axis=1)\n rpose_z = tf.gather(x, RPOSE_IDX_Z, axis=1)\n \n lpose_x = tf.gather(x, LPOSE_IDX_X, axis=1)\n lpose_y = tf.gather(x, LPOSE_IDX_Y, axis=1)\n lpose_z = tf.gather(x, LPOSE_IDX_Z, axis=1)\n \n lip = tf.concat([lip_x[..., tf.newaxis], lip_y[..., tf.newaxis], lip_z[..., tf.newaxis]], axis=-1)\n rhand = tf.concat([rhand_x[..., tf.newaxis], rhand_y[..., tf.newaxis], rhand_z[..., tf.newaxis]], axis=-1)\n lhand = tf.concat([lhand_x[..., tf.newaxis], lhand_y[..., tf.newaxis], lhand_z[..., tf.newaxis]], axis=-1)\n rpose = tf.concat([rpose_x[..., tf.newaxis], rpose_y[..., tf.newaxis], rpose_z[..., tf.newaxis]], axis=-1)\n lpose = tf.concat([lpose_x[..., tf.newaxis], lpose_y[..., tf.newaxis], lpose_z[..., tf.newaxis]], axis=-1)\n \n hand = tf.concat([rhand, lhand], axis=1)\n hand = tf.where(tf.math.is_nan(hand), 0.0, hand)\n input_tensor = tf.math.not_equal(tf.reduce_sum(hand, axis=[1, 2]), 0.0)\n alternating_tensor = tf.math.equal( tf.cumsum(tf.ones_like( tf.reduce_sum(hand, axis=[1, 2]) ))%2, 1.0 )\n mask = tf.math.logical_or(input_tensor, alternating_tensor)\n \n lip = lip[mask]\n rhand = rhand[mask]\n lhand = lhand[mask]\n rpose = rpose[mask]\n lpose = lpose[mask]\n\n return lip, rhand, lhand, rpose, lpose\n\n@tf.function()\ndef pre_process1(lip, rhand, lhand, rpose, lpose):\n lip = (resize_pad(lip) - LIPM) / LIPS\n rhand = (resize_pad(rhand) - RHM) / RHS\n lhand = (resize_pad(lhand) - LHM) / LHS\n rpose = (resize_pad(rpose) - RPM) / RPS\n lpose = (resize_pad(lpose) - LPM) / LPS\n\n x = tf.concat([lip, rhand, lhand, rpose, lpose], axis=1)\n shape = tf.shape(x)\n x = tf.reshape(x, (shape[0], shape[1]*shape[2]))\n x = tf.where(tf.math.is_nan(x), tf.zeros_like(x), x)\n return x\n\npre0 = pre_process0(frames)\npre1 = pre_process1(*pre0)\nINPUT_SHAPE = list(pre1.shape)\nprint(INPUT_SHAPE)","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:21.490555Z","iopub.execute_input":"2024-12-08T14:26:21.490848Z","iopub.status.idle":"2024-12-08T14:26:21.909095Z","shell.execute_reply.started":"2024-12-08T14:26:21.490825Z","shell.execute_reply":"2024-12-08T14:26:21.908244Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def decode_fn(record_bytes):\n schema = {\n \"lip\": tf.io.VarLenFeature(tf.float32),\n \"rhand\": tf.io.VarLenFeature(tf.float32),\n \"lhand\": tf.io.VarLenFeature(tf.float32),\n \"rpose\": tf.io.VarLenFeature(tf.float32),\n \"lpose\": tf.io.VarLenFeature(tf.float32),\n \"phrase\": tf.io.VarLenFeature(tf.int64)\n }\n x = tf.io.parse_single_example(record_bytes, schema)\n\n lip = tf.reshape(tf.sparse.to_dense(x[\"lip\"]), (-1, 40, 3))\n rhand = tf.reshape(tf.sparse.to_dense(x[\"rhand\"]), (-1, 21, 3))\n lhand = tf.reshape(tf.sparse.to_dense(x[\"lhand\"]), (-1, 21, 3))\n rpose = tf.reshape(tf.sparse.to_dense(x[\"rpose\"]), (-1, 5, 3))\n lpose = tf.reshape(tf.sparse.to_dense(x[\"lpose\"]), (-1, 5, 3))\n phrase = tf.sparse.to_dense(x[\"phrase\"])\n\n return lip, rhand, lhand, rpose, lpose, phrase #lhand,\n\ndef pre_process_fn(lip, rhand,lhand, rpose, lpose, phrase): #lhand,\n phrase = tf.pad(phrase, [[0, MAX_PHRASE_LENGTH-tf.shape(phrase)[0]]], constant_values=pad_token_idx)\n return pre_process1(lip, rhand, lhand, rpose, lpose), phrase #lhand,\n \ntffiles = [f\"/kaggle/input/chris-tf-v9/{file_id}.tfrecord\" for file_id in df.file_id.unique()]\nval_len = 1\ntrain_batch_size = 64\nval_batch_size = 64\n\ntrain_dataset = tf.data.TFRecordDataset(tffiles[val_len:]).prefetch(tf.data.AUTOTUNE).shuffle(5000)\\\n .map(decode_fn, num_parallel_calls=tf.data.AUTOTUNE)\\\n .map(pre_process_fn, num_parallel_calls=tf.data.AUTOTUNE)\\\n .batch(train_batch_size)\\\n .prefetch(tf.data.AUTOTUNE)\n\nval_dataset = tf.data.TFRecordDataset(tffiles[:val_len]).prefetch(tf.data.AUTOTUNE)\\\n .map(decode_fn, num_parallel_calls=tf.data.AUTOTUNE)\\\n .map(pre_process_fn, num_parallel_calls=tf.data.AUTOTUNE)\\\n .batch(val_batch_size)\\\n .prefetch(tf.data.AUTOTUNE)\n\nbatch = next(iter(val_dataset))\nbatch[0].shape, batch[1].shape","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:25.711435Z","iopub.execute_input":"2024-12-08T14:26:25.712299Z","iopub.status.idle":"2024-12-08T14:26:26.428005Z","shell.execute_reply.started":"2024-12-08T14:26:25.712257Z","shell.execute_reply":"2024-12-08T14:26:26.427086Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class ECA(tf.keras.layers.Layer):\n def __init__(self, kernel_size=5, **kwargs):\n super().__init__(**kwargs)\n self.supports_masking = True\n self.kernel_size = kernel_size\n self.conv = tf.keras.layers.Conv1D(1, kernel_size=kernel_size, strides=1, padding=\"same\", use_bias=False)\n\n def call(self, inputs, mask=None):\n nn = tf.keras.layers.GlobalAveragePooling1D()(inputs, mask=mask)\n nn = tf.expand_dims(nn, -1)\n nn = self.conv(nn)\n nn = tf.squeeze(nn, -1)\n nn = tf.nn.sigmoid(nn)\n nn = nn[:,None,:]\n return inputs * nn\n\nclass CausalDWConv1D(tf.keras.layers.Layer):\n def __init__(self, \n kernel_size=17,\n dilation_rate=1,\n use_bias=False,\n depthwise_initializer='glorot_uniform',\n name='', **kwargs):\n super().__init__(name=name,**kwargs)\n self.causal_pad = tf.keras.layers.ZeroPadding1D((dilation_rate*(kernel_size-1),0),name=name + '_pad')\n self.dw_conv = tf.keras.layers.DepthwiseConv1D(\n kernel_size,\n strides=1,\n dilation_rate=dilation_rate,\n padding='valid',\n use_bias=use_bias,\n depthwise_initializer=depthwise_initializer,\n name=name + '_dwconv')\n self.supports_masking = True\n \n def call(self, inputs):\n x = self.causal_pad(inputs)\n x = self.dw_conv(x)\n return x\n\ndef Conv1DBlock(channel_size,\n kernel_size,\n dilation_rate=1,\n drop_rate=0.0,\n expand_ratio=2,\n se_ratio=0.25,\n activation='swish',\n name=None):\n '''\n efficient conv1d block, @hoyso48\n '''\n if name is None:\n name = str(tf.keras.backend.get_uid(\"mbblock\"))\n # Expansion phase\n def apply(inputs):\n channels_in = tf.keras.backend.int_shape(inputs)[-1]\n channels_expand = channels_in * expand_ratio\n\n skip = inputs\n\n x = tf.keras.layers.Dense(\n channels_expand,\n use_bias=True,\n activation=activation,\n name=name + '_expand_conv')(inputs)\n\n # Depthwise Convolution\n x = CausalDWConv1D(kernel_size,\n dilation_rate=dilation_rate,\n use_bias=False,\n name=name + '_dwconv')(x)\n\n x = tf.keras.layers.BatchNormalization(momentum=0.95, name=name + '_bn')(x)\n\n x = ECA()(x)\n\n x = tf.keras.layers.Dense(\n channel_size,\n use_bias=True,\n name=name + '_project_conv')(x)\n\n if drop_rate > 0:\n x = tf.keras.layers.Dropout(drop_rate, noise_shape=(None,1,1), name=name + '_drop')(x)\n\n if (channels_in == channel_size):\n x = tf.keras.layers.add([x, skip], name=name + '_add')\n return x\n\n return apply\n\nclass MultiHeadSelfAttention(tf.keras.layers.Layer):\n def __init__(self, dim=256, num_heads=4, dropout=0, **kwargs):\n super().__init__(**kwargs)\n self.dim = dim\n self.scale = self.dim ** -0.5\n self.num_heads = num_heads\n self.qkv = tf.keras.layers.Dense(3 * dim, use_bias=False)\n self.drop1 = tf.keras.layers.Dropout(dropout)\n self.proj = tf.keras.layers.Dense(dim, use_bias=False)\n self.supports_masking = True\n\n def call(self, inputs, mask=None):\n qkv = self.qkv(inputs)\n qkv = tf.keras.layers.Permute((2, 1, 3))(tf.keras.layers.Reshape((-1, self.num_heads, self.dim * 3 // self.num_heads))(qkv))\n q, k, v = tf.split(qkv, [self.dim // self.num_heads] * 3, axis=-1)\n\n attn = tf.matmul(q, k, transpose_b=True) * self.scale\n\n if mask is not None:\n mask = mask[:, None, None, :]\n\n attn = tf.keras.layers.Softmax(axis=-1)(attn, mask=mask)\n attn = self.drop1(attn)\n\n x = attn @ v\n x = tf.keras.layers.Reshape((-1, self.dim))(tf.keras.layers.Permute((2, 1, 3))(x))\n x = self.proj(x)\n return x\n\nclass SqueezeExcite(tf.keras.layers.Layer):\n def __init__(self, channels, reduction_ratio=8, **kwargs):\n super().__init__(**kwargs)\n self.supports_masking = True\n self.gap = tf.keras.layers.GlobalAveragePooling1D()\n reduced_channels = max(1, channels // reduction_ratio)\n self.fc1 = tf.keras.layers.Dense(reduced_channels, activation='swish')\n self.fc2 = tf.keras.layers.Dense(channels, activation='sigmoid')\n \n def call(self, inputs, mask=None):\n x = self.gap(inputs, mask=mask)\n x = self.fc1(x)\n x = self.fc2(x)\n return inputs * tf.expand_dims(x, 1)\n\nclass ConvModule(tf.keras.layers.Layer):\n def __init__(self, dim, kernel_size, expansion_factor=2, **kwargs):\n super().__init__(**kwargs)\n self.supports_masking = True\n self.norm = tf.keras.layers.LayerNormalization(epsilon=1e-6)\n self.conv1 = tf.keras.layers.Conv1D(dim * expansion_factor, 1)\n self.conv2 = CausalDWConv1D(kernel_size=kernel_size)\n self.conv3 = tf.keras.layers.Conv1D(dim, 1)\n self.se = SqueezeExcite(dim)\n \n def call(self, inputs, mask=None):\n x = self.norm(inputs)\n x = self.conv1(x)\n x = tf.keras.activations.swish(x)\n x = self.conv2(x)\n x = tf.keras.activations.swish(x)\n x = self.conv3(x)\n x = self.se(x, mask=mask)\n return x + inputs\n\nclass SqueezeformerBlock(tf.keras.layers.Layer):\n def __init__(self, dim, num_heads=8, expansion_factor=4, kernel_size=31, dropout=0.1, **kwargs):\n super().__init__(**kwargs)\n self.supports_masking = True\n \n # Feed Forward Module 1\n self.norm1 = tf.keras.layers.LayerNormalization(epsilon=1e-6)\n self.ffn1 = tf.keras.Sequential([\n tf.keras.layers.Dense(dim * expansion_factor, activation='swish'),\n tf.keras.layers.Dropout(dropout),\n tf.keras.layers.Dense(dim)\n ])\n \n # Multi-head Self Attention\n self.norm2 = tf.keras.layers.LayerNormalization(epsilon=1e-6)\n self.mha = MultiHeadSelfAttention(dim=dim, num_heads=num_heads, dropout=dropout)\n \n # Convolution Module\n self.conv = ConvModule(dim, kernel_size, expansion_factor)\n \n # Feed Forward Module 2\n self.norm3 = tf.keras.layers.LayerNormalization(epsilon=1e-6)\n self.ffn2 = tf.keras.Sequential([\n tf.keras.layers.Dense(dim * expansion_factor, activation='swish'),\n tf.keras.layers.Dropout(dropout),\n tf.keras.layers.Dense(dim)\n ])\n \n self.dropout = tf.keras.layers.Dropout(dropout)\n \n def call(self, inputs, mask=None):\n # First FFN\n residual = inputs\n x = self.norm1(inputs)\n x = self.ffn1(x)\n x = residual + self.dropout(x)\n \n # Self Attention\n residual = x\n x = self.norm2(x)\n x = self.mha(x, mask=mask)\n x = residual + self.dropout(x)\n \n # Convolution\n x = self.conv(x, mask=mask)\n \n # Second FFN\n residual = x\n x = self.norm3(x)\n x = self.ffn2(x)\n x = residual + self.dropout(x)\n \n return x\n \ndef TransformerBlock(dim=256, num_heads=6, expand=4, attn_dropout=0.2, drop_rate=0.2, activation='swish'):\n def apply(inputs):\n x = inputs\n x = tf.keras.layers.LayerNormalization(epsilon=1e-6)(x)\n x = MultiHeadSelfAttention(dim=dim,num_heads=num_heads,dropout=attn_dropout)(x)\n x = tf.keras.layers.Dropout(drop_rate, noise_shape=(None,1,1))(x)\n x = tf.keras.layers.Add()([inputs, x])\n attn_out = x\n\n x = tf.keras.layers.LayerNormalization(epsilon=1e-6)(x)\n x = tf.keras.layers.Dense(dim*expand, use_bias=False, activation=activation)(x)\n x = tf.keras.layers.Dense(dim, use_bias=False)(x)\n x = tf.keras.layers.Dropout(drop_rate, noise_shape=(None,1,1))(x)\n x = tf.keras.layers.Add()([attn_out, x])\n return x\n return apply\n\ndef positional_encoding(maxlen, num_hid):\n depth = num_hid/2\n positions = tf.range(maxlen, dtype = tf.float32)[..., tf.newaxis]\n depths = tf.range(depth, dtype = tf.float32)[np.newaxis, :]/depth\n angle_rates = tf.math.divide(1, tf.math.pow(tf.cast(10000, tf.float32), depths))\n angle_rads = tf.linalg.matmul(positions, angle_rates)\n pos_encoding = tf.concat(\n [tf.math.sin(angle_rads), tf.math.cos(angle_rads)],\n axis=-1)\n return pos_encoding","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:28.87619Z","iopub.execute_input":"2024-12-08T14:26:28.876499Z","iopub.status.idle":"2024-12-08T14:26:28.905117Z","shell.execute_reply.started":"2024-12-08T14:26:28.876474Z","shell.execute_reply":"2024-12-08T14:26:28.904223Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def CTCLoss(labels, logits):\n label_length = tf.reduce_sum(tf.cast(labels != pad_token_idx, tf.int32), axis=-1)\n logit_length = tf.ones(tf.shape(logits)[0], dtype=tf.int32) * tf.shape(logits)[1]\n loss = tf.nn.ctc_loss(\n labels=labels,\n logits=logits,\n label_length=label_length,\n logit_length=logit_length,\n blank_index=pad_token_idx,\n logits_time_major=False\n )\n loss = tf.reduce_mean(loss)\n return loss","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:37.575498Z","iopub.execute_input":"2024-12-08T14:26:37.575804Z","iopub.status.idle":"2024-12-08T14:26:37.58108Z","shell.execute_reply.started":"2024-12-08T14:26:37.575782Z","shell.execute_reply":"2024-12-08T14:26:37.580159Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def get_model(dim=384):\n inp = tf.keras.Input(INPUT_SHAPE)\n x = tf.keras.layers.Masking(mask_value=0.0)(inp)\n \n # Initial processing with larger dimension\n x = tf.keras.layers.Dense(dim, use_bias=False, name='stem_conv')(x)\n pe = tf.cast(positional_encoding(INPUT_SHAPE[0], dim), dtype=x.dtype)\n x = x + pe\n x = tf.keras.layers.BatchNormalization(momentum=0.95, name='stem_bn')(x)\n \n # Increase number of blocks and use residual connections\n num_blocks = 8 \n drop_rate = 0.3 # Adjusted dropout\n \n for i in range(num_blocks):\n res = x # Store residual\n \n x = SqueezeformerBlock(\n dim=dim,\n num_heads=8,\n expansion_factor=4,\n kernel_size=31,\n dropout=drop_rate\n )(x)\n \n # Add residual connection with layer scaling\n scale = tf.Variable(0.1, trainable=True, name=f'layer_scale_{i}')\n x = x * scale + res\n \n # Add layer normalization after each block\n x = tf.keras.layers.LayerNormalization(epsilon=1e-6)(x)\n \n # Enhanced top layers\n x = tf.keras.layers.Dense(dim*2, activation='swish', name='top_conv')(x)\n x = tf.keras.layers.LayerNormalization(epsilon=1e-6)(x)\n x = tf.keras.layers.Dropout(0.4)(x)\n \n # Additional processing before final classification\n x = tf.keras.layers.Dense(dim, activation='swish')(x)\n x = tf.keras.layers.LayerNormalization(epsilon=1e-6)(x)\n x = tf.keras.layers.Dropout(0.3)(x)\n \n x = tf.keras.layers.Dense(len(char_to_num))(x)\n \n model = tf.keras.Model(inp, x)\n \n # Enhanced optimizer configuration\n optimizer = tfa.optimizers.RectifiedAdam(\n learning_rate=4e-3,\n beta_1=0.9,\n beta_2=0.999,\n epsilon=1e-7,\n weight_decay=1e-5,\n sma_threshold=4,\n total_steps=N_EPOCHS * (len(df) // train_batch_size),\n warmup_proportion=0.1,\n min_lr=1e-6,\n )\n optimizer = tfa.optimizers.Lookahead(optimizer, sync_period=5)\n \n # Add gradient clipping\n optimizer = tf.keras.mixed_precision.LossScaleOptimizer(optimizer)\n \n model.compile(\n loss=CTCLoss,\n optimizer=optimizer\n )\n \n return model","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:44.777909Z","iopub.execute_input":"2024-12-08T14:26:44.77824Z","iopub.status.idle":"2024-12-08T14:26:44.788076Z","shell.execute_reply.started":"2024-12-08T14:26:44.778214Z","shell.execute_reply":"2024-12-08T14:26:44.787215Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def num_to_char_fn(y):\n return [num_to_char.get(x, \"\") for x in y]\n\n@tf.function()\ndef decode_phrase(pred):\n x = tf.argmax(pred, axis=1)\n diff = tf.not_equal(x[:-1], x[1:])\n adjacent_indices = tf.where(diff)[:, 0]\n x = tf.gather(x, adjacent_indices)\n mask = x != pad_token_idx\n x = tf.boolean_mask(x, mask, axis=0)\n return x\n\n# A utility function to decode the output of the network\ndef decode_batch_predictions(pred):\n output_text = []\n for result in pred:\n result = \"\".join(num_to_char_fn(decode_phrase(result).numpy()))\n output_text.append(result)\n return output_text","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:51.273137Z","iopub.execute_input":"2024-12-08T14:26:51.273537Z","iopub.status.idle":"2024-12-08T14:26:51.280002Z","shell.execute_reply.started":"2024-12-08T14:26:51.27351Z","shell.execute_reply":"2024-12-08T14:26:51.279215Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"\n# A callback class to output a few transcriptions during training\nclass CallbackEval(tf.keras.callbacks.Callback):\n \"\"\"Displays a batch of outputs after every epoch.\"\"\"\n\n def __init__(self, dataset):\n super().__init__()\n self.dataset = dataset\n\n def on_epoch_end(self, epoch: int, logs=None):\n model.save_weights(\"model.h5\")\n predictions = []\n targets = []\n for batch in self.dataset:\n X, y = batch\n batch_predictions = model(X)\n batch_predictions = decode_batch_predictions(batch_predictions)\n predictions.extend(batch_predictions)\n for label in y:\n label = \"\".join(num_to_char_fn(label.numpy()))\n targets.append(label)\n print(\"-\" * 100)\n # for i in np.random.randint(0, len(predictions), 2):\n for i in range(32):\n print(f\"Target : {targets[i]}\")\n print(f\"Prediction: {predictions[i]}, len: {len(predictions[i])}\")\n print(\"-\" * 100)\n\n# Callback function to check transcription on the val set.\nvalidation_callback = CallbackEval(val_dataset.take(1))","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:53.825088Z","iopub.execute_input":"2024-12-08T14:26:53.825455Z","iopub.status.idle":"2024-12-08T14:26:53.836912Z","shell.execute_reply.started":"2024-12-08T14:26:53.82543Z","shell.execute_reply":"2024-12-08T14:26:53.836288Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"N_EPOCHS = 20\nN_WARMUP_EPOCHS = 5\nLR_MAX = 4e-3\nWD_RATIO = 0.05\nWARMUP_METHOD = \"exp\"\n\nearly_stopping = tf.keras.callbacks.EarlyStopping(\n monitor='val_loss',\n patience=5,\n restore_best_weights=True\n)\n\nreduce_lr = tf.keras.callbacks.ReduceLROnPlateau(\n monitor='val_loss',\n factor=0.2,\n patience=3,\n min_lr=1e-6\n)","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:26:57.673085Z","iopub.execute_input":"2024-12-08T14:26:57.674014Z","iopub.status.idle":"2024-12-08T14:26:57.680835Z","shell.execute_reply.started":"2024-12-08T14:26:57.673971Z","shell.execute_reply":"2024-12-08T14:26:57.679717Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def lrfn(current_step, num_warmup_steps, lr_max, num_cycles=0.50, num_training_steps=N_EPOCHS):\n \n if current_step < num_warmup_steps:\n if WARMUP_METHOD == 'log':\n return lr_max * 0.10 ** (num_warmup_steps - current_step)\n else:\n return lr_max * 2 ** -(num_warmup_steps - current_step)\n else:\n progress = float(current_step - num_warmup_steps) / float(max(1, num_training_steps - num_warmup_steps))\n\n return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress))) * lr_max\n \ndef plot_lr_schedule(lr_schedule, epochs):\n fig = plt.figure(figsize=(20, 10))\n plt.plot([None] + lr_schedule + [None])\n # X Labels\n x = np.arange(1, epochs + 1)\n x_axis_labels = [i if epochs <= 40 or i % 5 == 0 or i == 1 else None for i in range(1, epochs + 1)]\n plt.xlim([1, epochs])\n plt.xticks(x, x_axis_labels) # set tick step to 1 and let x axis start at 1\n \n # Increase y-limit for better readability\n plt.ylim([0, max(lr_schedule) * 1.1])\n \n # Title\n schedule_info = f'start: {lr_schedule[0]:.1E}, max: {max(lr_schedule):.1E}, final: {lr_schedule[-1]:.1E}'\n plt.title(f'Step Learning Rate Schedule, {schedule_info}', size=18, pad=12)\n \n # Plot Learning Rates\n for x, val in enumerate(lr_schedule):\n if epochs <= 40 or x % 5 == 0 or x is epochs - 1:\n if x < len(lr_schedule) - 1:\n if lr_schedule[x - 1] < val:\n ha = 'right'\n else:\n ha = 'left'\n elif x == 0:\n ha = 'right'\n else:\n ha = 'left'\n plt.plot(x + 1, val, 'o', color='black');\n offset_y = (max(lr_schedule) - min(lr_schedule)) * 0.02\n plt.annotate(f'{val:.1E}', xy=(x + 1, val + offset_y), size=12, ha=ha)\n \n plt.xlabel('Epoch', size=16, labelpad=5)\n plt.ylabel('Learning Rate', size=16, labelpad=5)\n plt.grid()\n plt.show()\n\n# Learning rate for encoder\nLR_SCHEDULE = [lrfn(step, num_warmup_steps=N_WARMUP_EPOCHS, lr_max=LR_MAX, num_cycles=0.50) for step in range(N_EPOCHS)]\n# Plot Learning Rate Schedule\nplot_lr_schedule(LR_SCHEDULE, epochs=N_EPOCHS)\n# Learning Rate Callback\nlr_callback = tf.keras.callbacks.LearningRateScheduler(lambda step: LR_SCHEDULE[step], verbose=0)\n\n# Custom callback to update weight decay with learning rate\nclass WeightDecayCallback(tf.keras.callbacks.Callback):\n def __init__(self, wd_ratio=WD_RATIO):\n self.step_counter = 0\n self.wd_ratio = wd_ratio\n \n def on_epoch_begin(self, epoch, logs=None):\n model.optimizer.weight_decay = model.optimizer.learning_rate * self.wd_ratio\n print(f'learning rate: {model.optimizer.learning_rate.numpy():.2e}, weight decay: {model.optimizer.weight_decay.numpy():.2e}')","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:27:01.809585Z","iopub.execute_input":"2024-12-08T14:27:01.810297Z","iopub.status.idle":"2024-12-08T14:27:02.307105Z","shell.execute_reply.started":"2024-12-08T14:27:01.810268Z","shell.execute_reply":"2024-12-08T14:27:02.306317Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Clear any existing models\ntf.keras.backend.clear_session()\n\n# Create the model\nmodel = get_model()\n\n# Train the model\nhistory = model.fit(\n train_dataset,\n validation_data=val_dataset,\n epochs=N_EPOCHS,\n callbacks=[\n validation_callback,\n lr_callback,\n WeightDecayCallback(),\n early_stopping,\n reduce_lr\n ]\n)","metadata":{"execution":{"iopub.status.busy":"2023-07-28T15:16:56.799121Z","iopub.execute_input":"2023-07-28T15:16:56.799851Z","iopub.status.idle":"2023-07-28T15:17:30.607822Z","shell.execute_reply.started":"2023-07-28T15:16:56.799813Z","shell.execute_reply":"2023-07-28T15:17:30.604703Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class TFLiteModel(tf.keras.Model):\n def __init__(self, model):\n super(TFLiteModel, self).__init__()\n self.model = model\n \n def call(self, inputs, training=False):\n # Preprocess Data\n x = tf.cast(inputs, tf.float32)\n x = x[None]\n x = tf.cond(tf.shape(x)[1] == 0, lambda: tf.zeros((1, 1, len(SEL_COLS))), lambda: tf.identity(x))\n x = x[0]\n x = pre_process00(x)\n x = pre_process1(*x)\n x = tf.reshape(x, INPUT_SHAPE)\n x = x[None]\n x = self.model(x, training=False)\n x = x[0]\n x = decode_phrase(x)\n # POST PROCESS. IF PRED LESS THAN 3, USE CONSTANT PREDICTION\n x = tf.cond(tf.shape(x)[0] < 3, lambda: tf.constant(\n [17, 0, 32, 12, 36, 0, 12, 32, 49, 46, 36], tf.int64), lambda: tf.identity(x))\n x = tf.one_hot(x, 59)\n return x\n\n # Add get_config method for serialization\n def get_config(self):\n return {\"model\": self.model}\n \n # Add from_config class method\n @classmethod\n def from_config(cls, config):\n return cls(**config)\n\ntflitemodel_base = TFLiteModel(model)\ntflitemodel_base(frames)[\"outputs\"].shape","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:27:20.21388Z","iopub.execute_input":"2024-12-08T14:27:20.214688Z","iopub.status.idle":"2024-12-08T14:27:24.102822Z","shell.execute_reply.started":"2024-12-08T14:27:20.214652Z","shell.execute_reply":"2024-12-08T14:27:24.101399Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"keras_model_converter = tf.lite.TFLiteConverter.from_keras_model(tflitemodel_base)\nkeras_model_converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS]#, tf.lite.OpsSet.SELECT_TF_OPS]\nkeras_model_converter.optimizations = [tf.lite.Optimize.DEFAULT]\nkeras_model_converter.target_spec.supported_types = [tf.float16]\ntflite_model = keras_model_converter.convert()\nwith open('model.tflite', 'wb') as f:\n f.write(tflite_model)\n \nwith open('inference_args.json', \"w\") as f:\n json.dump({\"selected_columns\" : SEL_COLS}, f)\n \n!zip submission.zip './model.tflite' './inference_args.json'","metadata":{"execution":{"iopub.status.busy":"2024-12-08T14:27:24.103856Z","iopub.status.idle":"2024-12-08T14:27:24.104308Z","shell.execute_reply.started":"2024-12-08T14:27:24.104068Z","shell.execute_reply":"2024-12-08T14:27:24.104088Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"with open (\"inference_args.json\", \"r\") as f:\n SEL_COLS = json.load(f)[\"selected_columns\"]\n \ndef load_relevant_data_subset(pq_path):\n return pd.read_parquet(pq_path, columns=SEL_COLS)\n\ndef create_data_gen(file_ids, y_mul=1):\n def gen():\n for file_id in file_ids:\n pqfile = f\"{inpdir}/{file_id}.parquet\"\n seq_refs = df.loc[df.file_id == file_id]\n seqs = load_relevant_data_subset(pqfile)\n\n for seq_id in seq_refs.sequence_id:\n x = seqs.iloc[seqs.index == seq_id].to_numpy()\n y = str(df.loc[df.sequence_id == seq_id].phrase.iloc[0])\n \n r_nonan = np.sum(np.sum(np.isnan(x[:, RHAND_IDX_X]), axis = 1) == 0)\n l_nonan = np.sum(np.sum(np.isnan(x[:, LHAND_IDX_X]), axis = 1) == 0)\n no_nan = max(r_nonan, l_nonan)\n \n if y_mul*len(y)