Convert float32 -> float64 (default) (#24)

README.md

@@ -5,6 +5,15 @@
 
 [![Build Status](https://github.com/DEEPDIP-project/CoupledNODE.jl/actions/workflows/CI.yml/badge.svg)](https://github.com/DEEPDIP-project/CoupledNODE.jl/actions/workflows/CI.yml)
 
+## Installation
+
+
+```julia
+]
+activate .
+instantiate
+```
+
 ### Context
 
 This repo was created during [this co-working session](https://github.com/DEEPDIP-project/logs/blob/main/meetings/2024-02-20%20Coworking%20session.md).

examples/01.00-Logistic.jl

@@ -57,7 +57,7 @@ f_NODE = create_f_NODE(NN, f_u; is_closed = true);
 θ, st = Lux.setup(rng, f_NODE);
 
 # * We define the NODE
-trange = (0.0f0, 6.0f0)
+trange = (0.0, 6.0)
 u0 = [0.01]
 full_NODE = NeuralODE(f_NODE, trange, Tsit5(), adaptive = false, dt = 0.001, saveat = 0.2);
 
@@ -73,7 +73,7 @@ myloss = create_randloss_MulDtO(u_experiment, nunroll = nunroll, nintervals = ni
 
 # Second, we define this auxiliary NODE that will be used for training
 dt = 0.01 # it has to be as fine as the data
-t_train_range = (0.0f0, dt * (nunroll + 1)) # it has to be as long as unroll
+t_train_range = (0.0, dt * (nunroll + 1)) # it has to be as long as unroll
 training_NODE = NeuralODE(f_NODE,
     t_train_range,
     Tsit5(),

examples/02.00-GrayScott.jl

@@ -65,7 +65,7 @@ f_CNODE = create_f_CNODE(F_u, G_v, grid; is_closed = false);
 length(θ) == 0;
 
 # We now do a short *burnout run* to get rid of the initial artifacts. This allows us to discard the transient dynamics and to have a good initial condition for the data collection run.
-trange_burn = (0.0f0, 10.0f0);
+trange_burn = (0.0, 10.0);
 dt, saveat = (1e-2, 1);
 full_CNODE = NeuralODE(f_CNODE,
     trange_burn,
@@ -79,7 +79,7 @@ burnout_CNODE_solution = Array(full_CNODE(uv0, θ, st)[1]);
 
 # We use the output of the *burnout run* to start a longer simulation
 uv0 = burnout_CNODE_solution[:, :, end];
-trange = (0.0f0, 8000.0f0);
+trange = (0.0, 8000.0);
 # the maximum suggested time step for GS is defined as `1/(4 * Dmax)`
 dt, saveat = (1 / (4 * max(D_u, D_v)), 25);
 full_CNODE = NeuralODE(f_CNODE, trange, Tsit5(), adaptive = false, dt = dt, saveat = saveat);

examples/02.01-GrayScott.jl

@@ -73,7 +73,7 @@ f_CNODE = create_f_CNODE(F_u, G_v, grid; is_closed = false);
 
 # **Burnout run:** to discard the results of the initial conditions.
 # In this case we need 2 burnouts: first one with a relatively large time step and then another one with a smaller time step. This allow us to discard the transient dynamics and to have a good initial condition for the data collection run.
-trange_burn = (0.0f0, 1.0f0)
+trange_burn = (0.0, 1.0)
 dt, saveat = (1e-2, 1)
 burnout_CNODE = NeuralODE(f_CNODE,
     trange_burn,
@@ -83,7 +83,7 @@ burnout_CNODE = NeuralODE(f_CNODE,
     saveat = saveat);
 burnout_CNODE_solution = Array(burnout_CNODE(uv0, θ_0, st_0)[1]);
 # Second burnout with a smaller timestep
-trange_burn = (0.0f0, 500.0f0)
+trange_burn = (0.0, 500.0)
 dt, saveat = (1 / (4 * max(D_u, D_v)), 100)
 burnout_CNODE = NeuralODE(f_CNODE,
     trange_burn,
@@ -95,7 +95,7 @@ burnout_CNODE_solution = Array(burnout_CNODE(burnout_CNODE_solution[:, :, end],
 
 # Data collection run
 uv0 = burnout_CNODE_solution[:, :, end];
-trange = (0.0f0, 2000.0f0);
+trange = (0.0, 2000.0);
 dt, saveat = (1 / (4 * max(D_u, D_v)), 1);
 GS_CNODE = NeuralODE(f_CNODE, trange, Tsit5(), adaptive = false, dt = dt, saveat = saveat);
 GS_sim = Array(GS_CNODE(uv0, θ_0, st_0)[1]);
@@ -243,7 +243,7 @@ display(p)
 # The learned weights look perfect, but let's check what happens if we use them to solve the GS model.
 
 # Let's solve the system, for two different set of parameters, with the trained CNODE and compare with the exact solution
-trange = (0.0f0, 500);
+trange = (0.0, 500);
 dt, saveat = (1, 5);
 
 ## Exact solution

examples/02.02-GrayScott.jl

@@ -68,7 +68,7 @@ f_CNODE = create_f_CNODE(F_u, G_v, grid; is_closed = false);
 θ_0, st_0 = Lux.setup(rng, f_CNODE);
 
 # **Burnout run**
-trange_burn = (0.0f0, 1.0f0)
+trange_burn = (0.0, 1.0)
 dt, saveat = (1e-2, 1)
 burnout_CNODE = NeuralODE(f_CNODE,
     trange_burn,
@@ -78,7 +78,7 @@ burnout_CNODE = NeuralODE(f_CNODE,
     saveat = saveat);
 burnout_CNODE_solution = Array(burnout_CNODE(uv0, θ_0, st_0)[1]);
 # Second burnout with a larger timestep
-trange_burn = (0.0f0, 800.0f0)
+trange_burn = (0.0, 800.0)
 dt, saveat = (1 / (4 * max(D_u, D_v)), 100)
 burnout_CNODE = NeuralODE(f_CNODE,
     trange_burn,
@@ -95,7 +95,7 @@ uv0 = burnout_CNODE_solution[:, :, end];
 # 2. prevent instabilities while training
 # However, this means that the simulation can not be long.
 dt, saveat = (1 / (4 * max(D_u, D_v)), 0.001)
-trange = (0.0f0, 50.0f0)
+trange = (0.0, 50.0)
 GS_CNODE = NeuralODE(f_CNODE, trange, Tsit5(), adaptive = false, dt = dt, saveat = saveat);
 GS_sim = Array(GS_CNODE(uv0, θ_0, st_0)[1])
 
@@ -167,15 +167,15 @@ nsamples = 1;
 # Also, it is important to solve for only the time interval thas is needed at each training step (corresponding to `nunroll` steps)
 dt_train = 0.001;
 saveat_train = saveat
-t_train_range = (0.0f0, saveat_train * nunroll)
+t_train_range = (0.0, saveat_train * nunroll)
 training_CNODE = NeuralODE(f_closed_CNODE,
     t_train_range,
     Tsit5(),
     adaptive = false,
     dt = dt_train,
     saveat = saveat_train);
 # Let's also define a secondary auxiliary CNODE that will be used in case the previous one is unstable
-t_train_range_2 = (0.0f0, saveat_train * 2)
+t_train_range_2 = (0.0, saveat_train * 2)
 t_train_range_2 = t_train_range
 training_CNODE_2 = NeuralODE(f_closed_CNODE,
     t_train_range_2,
@@ -245,7 +245,7 @@ display(p)
 
 # ### Comparison: CNODE vs exact solutions
 
-trange = (0.0f0, 600)
+trange = (0.0, 600)
 dt, saveat = (1, 5)
 
 # Exact solution

examples/02.03-GrayScott.jl

@@ -69,7 +69,7 @@ f_CNODE = create_f_CNODE(F_u, G_v, grid; is_closed = false)
 θ, st = Lux.setup(rng, f_CNODE);
 
 # We now do a short *burnout run* to get rid of the initial artifacts
-trange_burn = (0.0f0, 10.0f0)
+trange_burn = (0.0, 10.0)
 dt_burn, saveat_burn = (1e-2, 1)
 full_CNODE = NeuralODE(f_CNODE,
     trange_burn,
@@ -82,7 +82,7 @@ burnout_CNODE_solution = Array(full_CNODE(uv0, θ, st)[1])
 # **CNODE run** 
 # We use the output of the burnout to start a longer simulations
 uv0 = burnout_CNODE_solution[:, :, end];
-trange = (0.0f0, 7000.0f0)
+trange = (0.0, 7000.0)
 dt, saveat = (0.5, 20)
 full_CNODE = NeuralODE(f_CNODE, trange, Tsit5(), adaptive = false, dt = dt, saveat = saveat)
 untrained_CNODE_solution = Array(full_CNODE(uv0, θ, st)[1])

examples/coupling_functions/functions_example.jl

@@ -1,6 +1,6 @@
 function observation()
     f_ND = create_NODE_obs()
-    trange = (0.0f0, 6.0f0)
+    trange = (0.0, 6.0)
     p0 = [0.01]
     # define the observation from a NeuralODE
     obs_node = NeuralODE(f_ND, trange, Tsit5(), adaptive = false, dt = 0.01, saveat = 0.01)