Added HistogramByStep panel

Author: dsblank

panels/HistogramByStep/HistogramByStep.py

@@ -0,0 +1,186 @@
+from comet_ml import API, ui
+from comet_ml.data_structure import Histogram
+import random
+import numpy as np
+import plotly.express as px
+
+# Options:
+
+max_xbins = st.sidebar.slider(
+    "Maximum X bins",  # Label for the slider
+    min_value=5,                # Minimum allowed value
+    max_value=20,              # Maximum allowed value
+    value=10,                   # Default starting value
+    step=1                      # Increment step for the slider
+)
+max_ybins = st.sidebar.slider(
+    "Maximum Y bins",  # Label for the slider
+    min_value=5,                # Minimum allowed value
+    max_value=100,              # Maximum allowed value
+    value=50,                   # Default starting value
+    step=1                      # Increment step for the slider
+)
+start = None
+stop = None
+#max_ybins = 50
+#max_xbins = 50
+# Colors are scaled from highest to lowest. You can add
+# additional values between 0 and 1 to add color ranges.
+colorScale = [
+    [0, "white"], # lower values
+    [0.5, "gray"], # middle value
+    [1, "blue"] # higher values
+]
+showScale = False
+layout = {
+    "title": "Histograms by Step",
+    "xaxis": {
+        "ticks": "",
+        "side": "bottom",
+        "title": "Steps"
+    },
+    "yaxis": {
+        "ticks": "",
+        "ticksuffix": " ",
+        "autosize": True,
+        "title": "Weights"
+    }
+}
+
+def collect(histogram, start=None, stop=None, bins=50):
+    """
+    Collect the counts for the given range into bins.
+
+    Args:
+        start: optional, float, start of range to display
+        stop: optional, float, end of range to display
+        bins: optional, int, number of bins
+
+    Returns a list of dicts containing details on each
+    virtual bin.
+    """
+    counts_compressed = histogram.counts_compressed()
+    if start is None:
+        if len(counts_compressed) > 0:
+            start = histogram.values[counts_compressed[0][0]]
+        else:
+            start = -1.0
+    if stop is None:
+        if len(counts_compressed) > 1:
+            stop = histogram.values[counts_compressed[-1][0]]
+        else:
+            stop = 1.0
+
+    step = (stop - start) / bins
+
+    counts = histogram.get_counts(start, stop + step, step)
+    current = start
+    bins = []
+    next_one = current + step
+    i = 0
+    while next_one <= stop + step and i < len(counts):
+        start_bin = histogram.get_bin_index(current)
+        stop_bin = histogram.get_bin_index(next_one)
+        current_bin = {
+            "value_start": current,
+            "value_stop": next_one,
+            "bin_index_start": start_bin,
+            "bin_index_stop": stop_bin,
+            "count": counts[i],
+        }
+        bins.append(current_bin)
+        current = next_one
+        next_one = current + step
+        i += 1
+    return bins
+
+def get_histogram_indices(length, max_xbins):
+    """
+    Get indices from list of histograms, sampling if necessary.
+    """
+    if (length > max_xbins):
+        return (
+            [0] +
+            random.sample(
+                list(range(1, length - 1)),
+                max_xbins - 2) +
+            [length - 1])
+    else:
+        return list(range(length))
+
+def get_histogram_data(
+    experiment,
+    asset,
+):
+    assetJSON = experiment.get_asset(asset["assetId"], return_type="json")
+    histograms = []
+    # {"histograms": [{"step": num, "histogram": {"index_values"}}, ...]
+    selected = get_histogram_indices(len(assetJSON["histograms"]), max_xbins)
+    for index in selected:
+        hist = assetJSON["histograms"][index]
+        histogram = Histogram.from_json(hist["histogram"])
+        histogram.logged_at_step = hist["step"]
+        histograms.append(histogram)
+
+    # First, find the overall min/max of all histograms:
+    min_val, max_val = float("+inf"), float("-inf")
+    for histogram in histograms:
+        # {"value_start", "value_stop", "bin_index_start", "bin_index_stop", "count"}
+        data = collect(histogram, start=None, stop=None, bins=max_ybins)
+        min_val = min(min_val, data[0]["value_start"])
+        max_val = max(max_val, data[-1]["value_stop"])
+        
+    zValues = []
+    yValues = []
+    span = (max_val - min_val)/(max_ybins)
+    xValues = np.arange(min_val, max_val + span, span)
+    for h, histogram in enumerate(histograms):
+        # {"value_start", "value_stop", "bin_index_start", "bin_index_stop", "count"}
+        data = collect(histogram, start=min_val, stop=max_val, bins=max_ybins)
+        zValues.append([bin["count"] for bin in data])
+        yValues.append(histogram.logged_at_step)
+    return [xValues, yValues, np.transpose(zValues), min_val, max_val]
+
+def plot_histogram(experiment, asset):
+    x, y, data, xmin, xmax = get_histogram_data(
+        experiment,
+        asset,
+    )
+    if (len(data[0]) == 0):
+        print("<h1>No histogram data available<h1>")
+        return None
+
+    # Transposed, so x is y, y is x:
+    fig = px.imshow(
+        data,
+        x=y,
+        y=x[:len(data)],
+        aspect="auto",
+        color_continuous_scale=colorScale,
+    )
+    st.plotly_chart(fig)
+
+
+api = API()
+experiments = api.get_panel_experiments()
+if len(experiments) == 1:
+    selected_experiment = experiments[0]
+else:
+    selected_experiment = ui.dropdown("Experiments: ", experiments)
+
+if selected_experiment:
+    assets = sorted(
+        selected_experiment.get_asset_list('histogram_combined_3d'),
+        key=lambda item: item["fileName"]
+    )
+    selected_histogram = ui.dropdown(
+        "Histogram: ",
+        assets,
+        format_func=lambda item: item["fileName"]
+    )
+    if selected_histogram:
+        plot_histogram(selected_experiment, selected_histogram)
+    else:
+        print("No histograms available.")    
+else:
+    print("No experiments available.")