Spaces:
Runtime error
Runtime error
File size: 4,961 Bytes
de071e9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 |
import pandas as pd
import matplotlib.pyplot as plt
import datetime
def plot_traces(
results_path,
metric,
plot_path,
model_a_name,
model_b_name,
unpermuted_res=False,
normalize=True,
alpha_step=0.1,
end_points=True,
):
df = pd.read_csv(results_path)
alphas = [round(alpha * alpha_step, 2) for alpha in range(int(1 / alpha_step + 1))]
if end_points is False:
alphas = alphas[1:-1]
if metric == "loss":
plt.figure(figsize=(8, 6))
for index, row in df.iterrows():
row = row[int(len(row) - (len(row) - 2) / 2) :]
if normalize:
row = normalize_trace(row, alpha_step)
plt.plot(alphas, row, "o-")
plt.xlabel("Alpha")
plt.ylabel("Loss")
plt.title(f"{model_a_name} (Left) vs {model_b_name} (Right)")
plot_filename = f"{plot_path}_{datetime.datetime.now().timestamp()}.png"
if metric == "perplexity":
plt.figure(figsize=(8, 6))
for index, row in df.iterrows():
row = row[2 : int(2 + (len(row) - 2) / 2)]
if normalize:
row = normalize_trace(row, alpha_step)
plt.plot(alphas, row, "o-")
plt.xlabel("Alpha")
plt.ylabel("Perplexity")
plt.title(f"{model_a_name} (Left) vs {model_b_name} (Right)")
plot_filename = f"{plot_path}_{datetime.datetime.now().timestamp()}.png"
if unpermuted_res is not False:
plt.plot(alphas, normalize_trace(unpermuted_res, alpha_step))
# Save the plot as a PNG file
plt.savefig(plot_filename, dpi=300, bbox_inches="tight")
plt.close()
def plot_trace(losses, alpha_step, normalize, model_a_name, model_b_name, plot_path):
plt.figure(figsize=(8, 6))
if normalize:
losses = normalize_trace(losses, alpha_step)
alphas = [round(alpha * alpha_step, 2) for alpha in range(int(1 / alpha_step + 1))]
plt.plot(alphas, losses, "o-")
plt.xlabel("Alpha")
plt.ylabel("Loss")
plt.title(f"{model_a_name} (Left) vs {model_b_name} (Right)")
plot_filename = f"{plot_path}_{datetime.datetime.now().timestamp()}.png"
plt.savefig(plot_filename, dpi=300, bbox_inches="tight")
plt.close()
def normalize_trace(trace, alpha_step):
slope = trace[-1] - trace[0]
start = trace[0]
for i in range(len(trace)):
trace[i] -= slope * alpha_step * i
trace[i] -= start
return trace
def normalize_trace_2(trace, alphas):
slope = trace[-1] - trace[0]
start = trace[0]
for i in range(len(trace)):
trace[i] -= slope * alphas[i]
trace[i] -= start
return trace
def max_loss_ahmed(results_path, num_points=5, normalize=True, alphas=[0.0, 0.3, 0.5, 0.7, 1.0]):
df = pd.read_csv(results_path)
max_losses = []
for index, row in df.iterrows():
row = row[-num_points:]
if normalize:
row = normalize_trace_2(row, alphas)
max_losses.append(max(row))
return max_losses
def max_loss_compare(results_path, unpermuted_loss, num_points, normalize=True, alpha_step=0.1):
df = pd.read_csv(results_path)
alphas = [round(alpha * alpha_step, 2) for alpha in range(int(1 / alpha_step + 1))]
permuted_max_losses = []
for index, row in df.iterrows():
row = row[-num_points:]
if normalize:
row = normalize_trace(row, alpha_step)
permuted_max_losses.append(max(row))
if normalize:
unpermuted_loss = normalize_trace(unpermuted_loss, alpha_step)
unpermuted_max_loss = max(unpermuted_loss)
counter = 0
for m in permuted_max_losses:
if m > unpermuted_max_loss:
counter += 1
return counter, len(permuted_max_losses)
def avg_loss_compare(results_path, unpermuted_loss, num_points, normalize=True, alpha_step=0.1):
df = pd.read_csv(results_path)
alphas = [round(alpha * alpha_step, 2) for alpha in range(int(1 / alpha_step + 1))]
permuted_avg_losses = []
for index, row in df.iterrows():
row = row[-num_points:]
if normalize:
row = normalize_trace(row, alpha_step)
permuted_avg_losses.append(sum(row) / len(row))
if normalize:
unpermuted_loss = normalize_trace(unpermuted_loss, alpha_step)
unpermuted_avg_loss = sum(unpermuted_loss) / len(unpermuted_loss)
counter = 0
for m in permuted_avg_losses:
if m > unpermuted_avg_loss:
counter += 1
return counter, len(permuted_avg_losses)
def avg_loss_ahmed(results_path, num_points=5, normalize=True, alphas=[0.0, 0.3, 0.5, 0.7, 1.0]):
df = pd.read_csv(results_path)
avg_losses = []
for index, row in df.iterrows():
row = row[-num_points:]
if normalize:
row = normalize_trace_2(row, alphas)
avg_losses.append(sum(row) / len(row))
return avg_losses
def compute_p_value(counter, total):
return (total - counter - 1) / total
|