blog - Climate Modeling with SIRENs

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
os.environ["TF_FORCE_GPU_ALLOW_GROWTH"] = "true"

import numpy as np
import xarray as xr
from tqdm.keras import TqdmCallback

import tensorflow as tf
from tensorflow.keras import layers, initializers, activations
from tensorflow.keras.applications.resnet50 import ResNet50
from tensorflow.keras.callbacks import LearningRateScheduler
import tensorflow_addons as tfa

import matplotlib.pyplot as plt

/home/patel_zeel/miniconda3/envs/tensorflow_gpu/lib/python3.10/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
  from .autonotebook import tqdm as notebook_tqdm
2023-07-18 05:13:52.439735: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-07-18 05:13:53.232689: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
/home/patel_zeel/miniconda3/envs/tensorflow_gpu/lib/python3.10/site-packages/tensorflow_addons/utils/tfa_eol_msg.py:23: UserWarning: 

TensorFlow Addons (TFA) has ended development and introduction of new features.
TFA has entered a minimal maintenance and release mode until a planned end of life in May 2024.
Please modify downstream libraries to take dependencies from other repositories in our TensorFlow community (e.g. Keras, Keras-CV, and Keras-NLP). 

For more information see: https://github.com/tensorflow/addons/issues/2807 

  warnings.warn(

def SIREN(input_dim, output_dim, features, activation_scale, dropout):
    first_init = lambda input_dim: initializers.RandomUniform(-1 / input_dim, 1 / input_dim)
    other_init = lambda input_dim: initializers.RandomUniform(-np.sqrt(6 / input_dim) / activation_scale, np.sqrt(6 / input_dim) / activation_scale)
    model = tf.keras.Sequential()
    model.add(layers.Dense(features[0], input_shape=(input_dim,), kernel_initializer=first_init(input_dim), activation=lambda x: tf.sin(activation_scale*x)))
    for i in range(1, len(features)):
        model.add(layers.Dense(features[i], kernel_initializer=other_init(features[i-1]), activation=lambda x: tf.sin(activation_scale*x)))
        model.add(layers.Dropout(dropout))
    model.add(layers.Dense(output_dim, kernel_initializer=other_init(features[-1]), activation='linear'))
    return model

def MLP(input_dim, output_dim, features, dropout):
    model = tf.keras.Sequential()
    model.add(layers.Dense(features[0], input_shape=(input_dim,), activation=activations.relu))
    for i in range(1, len(features)):
        model.add(layers.Dense(features[i], activation=activations.relu))
        model.add(layers.Dropout(dropout))
    model.add(layers.Dense(output_dim, activation='linear'))
    return model
    
def ResNet():
    resnet = ResNet50(include_top=False, weights=None, input_shape=(64, 32, 1), pooling='avg')
    model = tf.keras.Sequential()
    model.add(resnet)
    model.add(layers.Dense(2048, activation='relu'))
    model.add(layers.Dense(32768, activation='linear'))
    return model

data5 = xr.open_dataset("../../super_res/data/era5_low_res/2m_temperature/2m_temperature_2018_5.625deg.nc")
data1 = xr.open_dataset("../../super_res/data/era5_high_res/2m_temperature/2m_temperature_2018_1.40625deg.nc")

data5

time_stamp = slice("2018-01", "2018-03")
train_df = data5.sel(time=time_stamp).to_dataframe().reset_index()
test_df = data1.sel(time=time_stamp).to_dataframe().reset_index()

X = np.stack([train_df.lat.values, train_df.lon.values, train_df.time.astype(np.int64) / 10**9], axis=1)
y = train_df[["t2m"]].values
print(f"{X.shape=}, {y.shape=}")

X_test = np.stack([test_df.lat.values, test_df.lon.values, test_df.time.astype(np.int64) / 10**9], axis=1)
y_test = test_df[["t2m"]].values
print(f"{X_test.shape=}, {y_test.shape=}")

# rff = np.random.normal(size=(2, 16)) * 0.01
# X = np.concatenate([np.sin(X @ rff), np.cos(X @ rff)], axis=1)
# print(f"{sin_cos.shape=}")
# X = X @ sin_cos
# X_test = np.concatenate([np.sin(X_test @ rff), np.cos(X_test @ rff)], axis=1)

print(f"{X.shape=}, {X_test.shape=}")

X.shape=(4423680, 3), y.shape=(4423680, 1)
X_test.shape=(70778880, 3), y_test.shape=(70778880, 1)
X.shape=(4423680, 3), X_test.shape=(70778880, 3)

32*64*24*(31+28+31)

X_max = np.max(X, axis=0, keepdims=True)
X_min = np.min(X, axis=0, keepdims=True)

X_scaled = (X - X_min) / (X_max - X_min)
X_test_scaled = (X_test - X_min) / (X_max - X_min)

# Scaling time
if X.shape[1] == 3:
    X_scaled[:, 2] = X_scaled[:, 2] * 10 - 5
    X_test_scaled[:, 2] = X_test_scaled[:, 2] * 10 - 5

y_min = np.min(y, axis=0, keepdims=True)
y_max = np.max(y, axis=0, keepdims=True)

y_scaled = (y - y_min) / (y_max - y_min)

# y_mean = np.mean(y, axis=0, keepdims=True)
# y_std = np.std(y, axis=0, keepdims=True)

# y_scaled = (y - y_mean) / y_std

model = SIREN(3, 1, [256]*4, 30.0, 0.0)
# model = MLP(3, 1, [256]*4, 0.0)
# model = ResNet()S
# clr = tfa.optimizers.CyclicalLearningRate(initial_learning_rate=1e-3,
#     maximal_learning_rate=1e-2,
#     scale_fn=lambda x: 1/(2.**(x-1)),
#     step_size=2
# )
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=1e-3), loss='mse')

2023-07-18 05:14:34.531498: W tensorflow/core/common_runtime/gpu/gpu_bfc_allocator.cc:47] Overriding orig_value setting because the TF_FORCE_GPU_ALLOW_GROWTH environment variable is set. Original config value was 0.
2023-07-18 05:14:34.531583: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1635] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 78884 MB memory:  -> device: 0, name: NVIDIA A100-SXM4-80GB, pci bus id: 0000:01:00.0, compute capability: 8.0

0.00148

0.00148

callbacks = [TqdmCallback(verbose=1)]
history = model.fit(X_scaled, y_scaled, epochs=1000, batch_size=X_scaled.shape[0], verbose=0, callbacks=callbacks)

  0%|          | 0/1000 [00:00<?, ?epoch/s]2023-07-18 05:14:38.677299: I tensorflow/compiler/xla/stream_executor/cuda/cuda_blas.cc:637] TensorFloat-32 will be used for the matrix multiplication. This will only be logged once.
2023-07-18 05:14:39.357828: I tensorflow/compiler/xla/service/service.cc:169] XLA service 0x7fb81b946130 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices:
2023-07-18 05:14:39.357901: I tensorflow/compiler/xla/service/service.cc:177]   StreamExecutor device (0): NVIDIA A100-SXM4-80GB, Compute Capability 8.0
2023-07-18 05:14:39.363158: I tensorflow/compiler/mlir/tensorflow/utils/dump_mlir_util.cc:269] disabling MLIR crash reproducer, set env var `MLIR_CRASH_REPRODUCER_DIRECTORY` to enable.
2023-07-18 05:14:40.399794: I tensorflow/compiler/xla/stream_executor/cuda/cuda_dnn.cc:424] Loaded cuDNN version 8600
2023-07-18 05:14:40.557542: I ./tensorflow/compiler/jit/device_compiler.h:180] Compiled cluster using XLA!  This line is logged at most once for the lifetime of the process.
100%|██████████| 1000/1000 [04:46<00:00,  3.49epoch/s, loss=0.000295]

plt.plot(history.history['loss'][200:], label='loss');
# plt.plot(history.history['val_loss'][200:], label='val_loss');
plt.legend();

128*256*24

img_index = 0
y_pred = model.predict(X_test_scaled, batch_size=20480) * (y_max - y_min) + y_min
print(y_pred.shape)
plt.imshow(y_pred[img_index*(256*128):(img_index+1)*(256*128)].reshape(256, 128), origin='lower', extent=[-180, 180, -90, 90], cmap='coolwarm', interpolation="none");

3456/3456 [==============================] - 5s 1ms/step
(70778880, 1)

plt.imshow(y.reshape(64, 32), origin='lower', extent=[-180, 180, -90, 90], cmap='coolwarm', interpolation="none");

diff = y_pred.reshape(256, 128) - y_test.reshape(256, 128)
plt.imshow(diff, origin='lower', extent=[-180, 180, -90, 90], cmap='coolwarm', interpolation="none");
plt.colorbar();
plt.title("Diff")

# rmse = np.sqrt(np.mean(np.abs(X_test[:, 0:1])*(y_pred.ravel() - y_test.ravel())**2))/np.mean(y_test.ravel() * np.abs(X_test[:, 0:1]))
def get_lat_weights(lat):
    lat_weights = np.cos(np.deg2rad(lat))
    lat_weights = lat_weights / lat_weights.mean()
    return lat_weights

lat_weights = get_lat_weights(X_test[:, 0])
print(f"{lat_weights.shape=}")

lat_squared_error = lat_weights * (y_pred.ravel() - y_test.ravel())**2
lat_rmse = np.sqrt(lat_squared_error.mean())
print(f"{lat_rmse=}")
# y_pred.shape, lat_weights.shape

lat_weights.shape=(70778880,)
lat_rmse=2.6118446730600438

# lat_rmse=3.4826956884024356

mean_bias = np.mean(y_pred.ravel() - y_test.ravel())
print(f"{mean_bias=}")