diff --git a/eval/plot.py b/eval/plot.py
index 0b7b66c448d4ad75281c729d18f47bf7840da1c6..f3f82c77ad55872c562a06ffb81b04f4b7269c65 100644
--- a/eval/plot.py
+++ b/eval/plot.py
@@ -3,8 +3,8 @@ import os
import sys
import numpy as np
-from matplotlib import pyplot as plt
import pandas as pd
+from matplotlib import pyplot as plt
def get_stats(l):
@@ -62,20 +62,50 @@ def plot_results(path):
plt.figure(1)
means, stdevs, mins, maxs = get_stats([x["train_loss"] for x in results])
plot(means, stdevs, mins, maxs, "Training Loss", folder, "upper right")
- df = pd.DataFrame({"mean": list(means.values()), "std": list(stdevs.values()), "nr_nodes": [len(results)]*len(means)}, list(means.keys()), columns=["mean", "std", "nr_nodes"])
- df.to_csv(os.path.join(path, "train_loss_" + folder + ".csv"))
+ df = pd.DataFrame(
+ {
+ "mean": list(means.values()),
+ "std": list(stdevs.values()),
+ "nr_nodes": [len(results)] * len(means),
+ },
+ list(means.keys()),
+ columns=["mean", "std", "nr_nodes"],
+ )
+ df.to_csv(
+ os.path.join(path, "train_loss_" + folder + ".csv"), index_label="rounds"
+ )
# Plot Testing loss
plt.figure(2)
means, stdevs, mins, maxs = get_stats([x["test_loss"] for x in results])
plot(means, stdevs, mins, maxs, "Testing Loss", folder, "upper right")
- df = pd.DataFrame({"mean": list(means.values()), "std": list(stdevs.values()), "nr_nodes": [len(results)]*len(means)}, list(means.keys()), columns=["mean", "std", "nr_nodes"])
- df.to_csv(os.path.join(path, "test_loss_" + folder + ".csv"))
+ df = pd.DataFrame(
+ {
+ "mean": list(means.values()),
+ "std": list(stdevs.values()),
+ "nr_nodes": [len(results)] * len(means),
+ },
+ list(means.keys()),
+ columns=["mean", "std", "nr_nodes"],
+ )
+ df.to_csv(
+ os.path.join(path, "test_loss_" + folder + ".csv"), index_label="rounds"
+ )
# Plot Testing Accuracy
plt.figure(3)
means, stdevs, mins, maxs = get_stats([x["test_acc"] for x in results])
plot(means, stdevs, mins, maxs, "Testing Accuracy", folder, "lower right")
- df = pd.DataFrame({"mean": list(means.values()), "std": list(stdevs.values()), "nr_nodes": [len(results)]*len(means)}, list(means.keys()), columns=["mean", "std", "nr_nodes"])
- df.to_csv(os.path.join(path, "test_acc_" + folder + ".csv"))
+ df = pd.DataFrame(
+ {
+ "mean": list(means.values()),
+ "std": list(stdevs.values()),
+ "nr_nodes": [len(results)] * len(means),
+ },
+ list(means.keys()),
+ columns=["mean", "std", "nr_nodes"],
+ )
+ df.to_csv(
+ os.path.join(path, "test_acc_" + folder + ".csv"), index_label="rounds"
+ )
plt.figure(6)
means, stdevs, mins, maxs = get_stats([x["grad_std"] for x in results])
plot(
diff --git a/src/decentralizepy/models/Model.py b/src/decentralizepy/models/Model.py
index 643eec5c17416aff0735307277acef8352df6e3d..196560812b9e5cfa324f360aab3adec8853408b8 100644
--- a/src/decentralizepy/models/Model.py
+++ b/src/decentralizepy/models/Model.py
@@ -56,4 +56,4 @@ class Model(nn.Module):
"""
if self.accumulated_changes is not None:
- self.accumulated_changes[indices] = 0.0
\ No newline at end of file
+ self.accumulated_changes[indices] = 0.0
diff --git a/src/decentralizepy/training/ChangeAccumulator.py b/src/decentralizepy/training/ChangeAccumulator.py
index c3bc81abeedd4f173fec3808276747cc8f910665..5e55621124eb21d7cf8d52ad250d9896cc757730 100644
--- a/src/decentralizepy/training/ChangeAccumulator.py
+++ b/src/decentralizepy/training/ChangeAccumulator.py
@@ -167,7 +167,7 @@ class ChangeAccumulator(Training):
else:
flats = [v.data.flatten() for _, v in self.init_model.items()]
flat = torch.cat(flats)
- self.model.accumulated_changes += (flat - self.prev)
+ self.model.accumulated_changes += flat - self.prev
self.prev = flat
super().train(dataset)
@@ -181,7 +181,7 @@ class ChangeAccumulator(Training):
flat_change = torch.cat(flats_change)
# flatten does not copy data if input is already flattened
# however cat copies
- change = {"flat" : self.model.accumulated_changes + flat_change}
+ change = {"flat": self.model.accumulated_changes + flat_change}
self.model.accumulated_gradients.append(change)
diff --git a/src/decentralizepy/training/FrequencyAccumulator.py b/src/decentralizepy/training/FrequencyAccumulator.py
index 7d7c9ab08bfe6eb7a7df80ec05dc88c2f54fb9b8..91e74b35eff495401725483d006300f4cdf62fab 100644
--- a/src/decentralizepy/training/FrequencyAccumulator.py
+++ b/src/decentralizepy/training/FrequencyAccumulator.py
@@ -88,7 +88,9 @@ class FrequencyAccumulator(Training):
"""
with torch.no_grad():
self.model.accumulated_gradients = []
- tensors_to_cat = [v.data.flatten() for _, v in self.model.state_dict().items()]
+ tensors_to_cat = [
+ v.data.flatten() for _, v in self.model.state_dict().items()
+ ]
concated = torch.cat(tensors_to_cat, dim=0)
self.init_model = fft.rfft(concated)
if self.accumulation:
@@ -96,17 +98,19 @@ class FrequencyAccumulator(Training):
self.model.accumulated_changes = torch.zeros_like(self.init_model)
self.prev = self.init_model
else:
- self.model.accumulated_changes += (self.init_model - self.prev)
+ self.model.accumulated_changes += self.init_model - self.prev
self.prev = self.init_model
super().train(dataset)
with torch.no_grad():
- tensors_to_cat = [v.data.flatten() for _, v in self.model.state_dict().items()]
+ tensors_to_cat = [
+ v.data.flatten() for _, v in self.model.state_dict().items()
+ ]
concated = torch.cat(tensors_to_cat, dim=0)
end_model = fft.rfft(concated)
change = end_model - self.init_model
if self.accumulation:
change += self.model.accumulated_changes
- self.model.accumulated_gradients.append(change)
\ No newline at end of file
+ self.model.accumulated_gradients.append(change)
diff --git a/src/decentralizepy/training/FrequencyWaveletAccumulator.py b/src/decentralizepy/training/FrequencyWaveletAccumulator.py
index ee36894d3d91de9a9cfeca0cc18af56d6c6d0bbb..54238abe2cb902468f230200ad86108d78856c20 100644
--- a/src/decentralizepy/training/FrequencyWaveletAccumulator.py
+++ b/src/decentralizepy/training/FrequencyWaveletAccumulator.py
@@ -93,7 +93,9 @@ class FrequencyWaveletAccumulator(Training):
# this looks at the change from the last round averaging of the frequencies
with torch.no_grad():
self.model.accumulated_gradients = []
- tensors_to_cat = [v.data.flatten() for _, v in self.model.state_dict().items()]
+ tensors_to_cat = [
+ v.data.flatten() for _, v in self.model.state_dict().items()
+ ]
concated = torch.cat(tensors_to_cat, dim=0)
coeff = pywt.wavedec(concated.numpy(), self.wavelet, level=self.level)
data, coeff_slices = pywt.coeffs_to_array(coeff)
@@ -103,13 +105,15 @@ class FrequencyWaveletAccumulator(Training):
self.model.accumulated_changes = torch.zeros_like(self.init_model)
self.prev = self.init_model
else:
- self.model.accumulated_changes += (self.init_model - self.prev)
+ self.model.accumulated_changes += self.init_model - self.prev
self.prev = self.init_model
super().train(dataset)
with torch.no_grad():
- tensors_to_cat = [v.data.flatten() for _, v in self.model.state_dict().items()]
+ tensors_to_cat = [
+ v.data.flatten() for _, v in self.model.state_dict().items()
+ ]
concated = torch.cat(tensors_to_cat, dim=0)
coeff = pywt.wavedec(concated.numpy(), self.wavelet, level=self.level)
data, coeff_slices = pywt.coeffs_to_array(coeff)