jit.py

# -*- coding: utf-8 -*-
"""
Using TorchScript to serialize and deploy model
===============================================

Models in TorchANI's model zoo support TorchScript. TorchScript is a way to create
serializable and optimizable models from PyTorch code. It allows users to saved their
models from a Python process and loaded in a process where there is no Python dependency.
"""

###############################################################################
# To begin with, let's first import the modules we will use:
import torch
import torchani
from typing import Tuple, Optional
from torch import Tensor

###############################################################################
# Scripting builtin model directly
# --------------------------------
#
# Let's now load the built-in ANI-1ccx models. The builtin ANI-1ccx contains 8
# models trained with diffrent initialization.
model = torchani.models.ANI1ccx(periodic_table_index=True)

###############################################################################
# It is very easy to compile and save the model using `torch.jit`.
compiled_model = torch.jit.script(model)
torch.jit.save(compiled_model, 'compiled_model.pt')

###############################################################################
# Besides compiling the ensemble, it is also possible to compile a single network
compiled_model0 = torch.jit.script(model[0])
torch.jit.save(compiled_model0, 'compiled_model0.pt')

###############################################################################
# For testing purposes, we will now load the models we just saved and see if they
# produces the same output as the original model:
loaded_compiled_model = torch.jit.load('compiled_model.pt')
loaded_compiled_model0 = torch.jit.load('compiled_model0.pt')


###############################################################################
# We use the molecule below to test:
coordinates = torch.tensor([[[0.03192167, 0.00638559, 0.01301679],
                             [-0.83140486, 0.39370209, -0.26395324],
                             [-0.66518241, -0.84461308, 0.20759389],
                             [0.45554739, 0.54289633, 0.81170881],
                             [0.66091919, -0.16799635, -0.91037834]]])
# In periodic table, C = 6 and H = 1
species = torch.tensor([[6, 1, 1, 1, 1]])

###############################################################################
# And here is the result:
energies_ensemble = model((species, coordinates)).energies
energies_single = model[0]((species, coordinates)).energies
energies_ensemble_jit = loaded_compiled_model((species, coordinates)).energies
energies_single_jit = loaded_compiled_model0((species, coordinates)).energies
print('Ensemble energy, eager mode vs loaded jit:', energies_ensemble.item(), energies_ensemble_jit.item())
print('Single network energy, eager mode vs loaded jit:', energies_single.item(), energies_single_jit.item())


###############################################################################
# Customize the model and script
# ------------------------------
#
# You could also customize the model you want to export. For example, let's do
# the following customization to the model:
#
# - uses double as dtype instead of float
# - don't care about periodic boundary condition
# - in addition to energies, allow returnsing optionally forces, and hessians
# - when indexing atom species, use its index in the periodic table instead of 0, 1, 2, 3, ...
#
# you could do the following:
class CustomModule(torch.nn.Module):

    def __init__(self):
        super().__init__()
        self.model = torchani.models.ANI1x(periodic_table_index=True).double()
        # self.model = torchani.models.ANI1x(periodic_table_index=True)[0].double()
        # self.model = torchani.models.ANI1ccx(periodic_table_index=True).double()

    def forward(self, species: Tensor, coordinates: Tensor, return_forces: bool = False,
                return_hessians: bool = False) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]:
        if return_forces or return_hessians:
            coordinates.requires_grad_(True)

        energies = self.model((species, coordinates)).energies

        forces: Optional[Tensor] = None  # noqa: E701
        hessians: Optional[Tensor] = None
        if return_forces or return_hessians:
            grad = torch.autograd.grad([energies.sum()], [coordinates], create_graph=return_hessians)[0]
            assert grad is not None
            forces = -grad
            if return_hessians:
                hessians = torchani.utils.hessian(coordinates, forces=forces)
        return energies, forces, hessians


custom_model = CustomModule()
compiled_custom_model = torch.jit.script(custom_model)
torch.jit.save(compiled_custom_model, 'compiled_custom_model.pt')
loaded_compiled_custom_model = torch.jit.load('compiled_custom_model.pt')
energies, forces, hessians = custom_model(species, coordinates, True, True)
energies_jit, forces_jit, hessians_jit = loaded_compiled_custom_model(species, coordinates, True, True)

print('Energy, eager mode vs loaded jit:', energies.item(), energies_jit.item())
print()
print('Force, eager mode vs loaded jit:\n', forces.squeeze(0), '\n', forces_jit.squeeze(0))
print()
torch.set_printoptions(sci_mode=False, linewidth=1000)
print('Hessian, eager mode vs loaded jit:\n', hessians.squeeze(0), '\n', hessians_jit.squeeze(0))