Latest Threat Research:SANDWORM_MODE: Shai-Hulud-Style npm Worm Hijacks CI Workflows and Poisons AI Toolchains.Details →
qmb
Artifacts are distributable versions of a package. It typically includes the source code and metadata
Advanced tools
qmb - npm Package Compare versions
+1
-1
| Metadata-Version: 2.1 | ||
| Name: qmb | ||
| Version: 0.0.7 | ||
| Version: 0.0.8 | ||
| Summary: Quantum Manybody Problem | ||
@@ -5,0 +5,0 @@ Author-email: Hao Zhang <hzhangxyz@outlook.com> |
| Metadata-Version: 2.1 | ||
| Name: qmb | ||
| Version: 0.0.7 | ||
| Version: 0.0.8 | ||
| Summary: Quantum Manybody Problem | ||
@@ -5,0 +5,0 @@ Author-email: Hao Zhang <hzhangxyz@outlook.com> |
+2
-2
@@ -15,3 +15,3 @@ # file generated by setuptools_scm | ||
| __version__ = version = '0.0.7' | ||
| __version_tuple__ = version_tuple = (0, 0, 7) | ||
| __version__ = version = '0.0.8' | ||
| __version_tuple__ = version_tuple = (0, 0, 8) |
+3
-5
@@ -81,7 +81,4 @@ import logging | ||
| optimizer = torch.optim.Adam(network.parameters(), lr=self.learning_rate) | ||
| loss = None | ||
| amplitudes = None | ||
| def closure(): | ||
| nonlocal loss, amplitudes | ||
| optimizer.zero_grad() | ||
@@ -92,2 +89,3 @@ amplitudes = network(configs) | ||
| loss.backward() | ||
| loss.amplitudes = amplitudes | ||
| return loss | ||
@@ -97,3 +95,3 @@ | ||
| for i in range(self.local_step): | ||
| optimizer.step(closure) | ||
| loss = optimizer.step(closure) | ||
| logging.info("local optimizing, step %d, loss %.10f", i, loss.item()) | ||
@@ -110,3 +108,3 @@ if loss < self.local_loss: | ||
| torch.enable_grad(closure)() | ||
| amplitudes = amplitudes.cpu().detach().numpy() | ||
| amplitudes = loss.amplitudes.cpu().detach().numpy() | ||
| final_energy = ((amplitudes.conj() @ (hamiltonian @ amplitudes)) / (amplitudes.conj() @ amplitudes)).real | ||
@@ -113,0 +111,0 @@ logging.info( |
+50
-17
@@ -28,4 +28,4 @@ import logging | ||
| use_lbfgs: typing.Annotated[bool, tyro.conf.arg(aliases=["-2"])] = False | ||
| # Do not calculate deviation or energy when optimizing energy or deviation | ||
| omit_another: typing.Annotated[bool, tyro.conf.arg(aliases=["-i"])] = False | ||
| # Do not calculate deviation when optimizing energy | ||
| omit_deviation: typing.Annotated[bool, tyro.conf.arg(aliases=["-i"])] = False | ||
@@ -40,3 +40,3 @@ def __post_init__(self): | ||
| logging.info( | ||
| "sampling count: %d, learning rate: %f, local step: %d, include outside: %a, use deviation: %a, fix outside: %a, use lbfgs: %a, omit another: %a", | ||
| "sampling count: %d, learning rate: %f, local step: %d, include outside: %a, use deviation: %a, fix outside: %a, use lbfgs: %a, omit deviation: %a", | ||
| self.sampling_count, | ||
@@ -49,3 +49,3 @@ self.learning_rate, | ||
| self.use_lbfgs, | ||
| self.omit_another, | ||
| self.omit_deviation, | ||
| ) | ||
@@ -89,3 +89,10 @@ | ||
| def closure(): | ||
| # Optimizing deviation | ||
| optimizer.zero_grad() | ||
| # Calculate amplitudes i and amplitudes j | ||
| # When including outside, amplitudes j should be calculated individually, otherwise, it equals to amplitudes i | ||
| # It should be notices that sometimes we do not want to optimize small configurations | ||
| # So we calculate amplitudes j in no grad mode | ||
| # but the first several configurations in amplitudes j are duplicated with those in amplitudes i | ||
| # So cat them manually | ||
| amplitudes_i = network(configs_i) | ||
@@ -101,10 +108,19 @@ if self.include_outside: | ||
| amplitudes_j = amplitudes_i | ||
| # <s|H|psi> will be used multiple times, calculate it first | ||
| # as we want to optimize deviation, every value should be calculated in grad mode, so we do not detach anything | ||
| hamiltonian_amplitudes_j = hamiltonian @ amplitudes_j | ||
| deviation = (hamiltonian_amplitudes_j / amplitudes_i).std() | ||
| # energy is just <psi|s> <s|H|psi> / <psi|s> <s|psi> | ||
| energy = (amplitudes_i.conj() @ hamiltonian_amplitudes_j) / (amplitudes_i.conj() @ amplitudes_i) | ||
| # we want to estimate variance of E_s - E with weight <psi|s><s|psi> | ||
| # where E_s = <s|H|psi>/<s|psi> | ||
| # the variance is (E_s - E).conj() @ (E_s - E) * <psi|s> <s|psi> / ... = (E_s <s|psi> - E <s|psi>).conj() @ (E_s <s|psi> - E <s|psi>) / ... | ||
| # so we calculate E_s <s|psi> - E <s|psi> first, which is just <s|H|psi> - <s|psi> E, we name it as `difference' | ||
| difference = hamiltonian_amplitudes_j - amplitudes_i * energy | ||
| # the numerator calculated, the following is the variance | ||
| variance = (difference.conj() @ difference) / (amplitudes_i.conj() @ amplitudes_i) | ||
| # calculate the deviation | ||
| deviation = variance.real.sqrt() | ||
| deviation.backward() | ||
| if self.omit_another: | ||
| deviation.energy = torch.tensor(torch.nan) | ||
| else: | ||
| with torch.no_grad(): | ||
| deviation.energy = ((amplitudes_i.conj() @ hamiltonian_amplitudes_j) / (amplitudes_i.conj() @ amplitudes_i)).real | ||
| # As we have already calculated energy, embed it in deviation for logging | ||
| deviation.energy = energy.real | ||
| return deviation | ||
@@ -119,3 +135,7 @@ | ||
| def closure(): | ||
| # Optimizing energy | ||
| optimizer.zero_grad() | ||
| # Calculate amplitudes i and amplitudes j | ||
| # When including outside, amplitudes j should be calculated individually, otherwise, it equals to amplitudes i | ||
| # Because of gradient formula, we always calculate amplitudes j in no grad mode | ||
| amplitudes_i = network(configs_i) | ||
@@ -126,11 +146,24 @@ if self.include_outside: | ||
| else: | ||
| amplitudes_j = amplitudes_i | ||
| hamiltonian_amplitudes_j = hamiltonian @ amplitudes_j.detach() | ||
| energy = ((amplitudes_i.conj() @ hamiltonian_amplitudes_j) / (amplitudes_i.conj() @ amplitudes_i.detach())).real | ||
| energy.backward() | ||
| if self.omit_another: | ||
| energy.deviation = torch.tensor(torch.nan) | ||
| amplitudes_j = amplitudes_i.detach() | ||
| # <s|H|psi> will be used multiple times, calculate it first | ||
| # it should be notices that this <s|H|psi> is totally detached, since both hamiltonian and amplitudes j is detached | ||
| hamiltonian_amplitudes_j = hamiltonian @ amplitudes_j | ||
| # energy is just <psi|s> <s|H|psi> / <psi|s> <s|psi> | ||
| # we only calculate gradient on <psi|s>, both <s|H|psi> and <s|psi> should be detached | ||
| # since <s|H|psi> has been detached already, we detach <s|psi> here manually | ||
| energy = (amplitudes_i.conj() @ hamiltonian_amplitudes_j) / (amplitudes_i.conj() @ amplitudes_i.detach()) | ||
| # Calculate deviation | ||
| # The variance is (E_s <s|psi> - E <s|psi>).conj() @ (E_s <s|psi> - E <s|psi>) / <psi|s> <s|psi> | ||
| # Calculate E_s <s|psi> - E <s|psi> first and name it as difference | ||
| if self.omit_deviation: | ||
| deviation = torch.tensor(torch.nan) | ||
| else: | ||
| with torch.no_grad(): | ||
| energy.deviation = (hamiltonian_amplitudes_j / amplitudes_i).std() | ||
| difference = hamiltonian_amplitudes_j - amplitudes_i * energy | ||
| variance = (difference.conj() @ difference) / (amplitudes_i.conj() @ amplitudes_i) | ||
| deviation = variance.real.sqrt() | ||
| energy = energy.real | ||
| energy.backward() | ||
| # Embed the deviation which has been calculated in energy for logging | ||
| energy.deviation = deviation | ||
| return energy | ||
@@ -137,0 +170,0 @@ |
Alert delta unavailable
Currently unable to show alert delta for PyPI packages.
Improved metrics
- Total package byte prevSize
140435
2.08%- Lines of code
1419
2.23%