notebooks/ToL_license_check.py

# ---
# jupyter:
#   jupytext:
#     formats: ipynb,py:percent
#     text_representation:
#       extension: .py
#       format_name: percent
#       format_version: '1.3'
#       jupytext_version: 1.16.0
#   kernelspec:
#     display_name: Python 3 (ipykernel)
#     language: python
#     name: python3
# ---

# %%
import pandas as pd
import seaborn as sns

sns.set_style("whitegrid")
sns.set(rc = {'figure.figsize': (10,10)})

# %% [markdown]
# Load in full images to ease process.

# %%
df = pd.read_csv("../data/predicted-catalog.csv", low_memory = False)

# %%
df.head()

# %%
df.info(show_counts = True)

# %% [markdown]
# The `train_small` is duplicates of `train`, so we will drop those to analyze the full training set plus val.

# %% [markdown]
# `predicted-catalog` doesn't have `train_small`, hence, it's a smaller file.

# %% [markdown]
# Let's add a column indicating the original data source so we can also get some stats by datasource, specifically focusing on EOL since we know licensing for BIOSCAN-1M and iNat21.

# %%
# Add data_source column for easier slicing
df.loc[df['inat21_filename'].notna(), 'data_source'] = 'iNat21'
df.loc[df['bioscan_filename'].notna(), 'data_source'] = 'BIOSCAN'
df.loc[df['eol_content_id'].notna(), 'data_source'] = 'EOL'

# %% [markdown]
# #### Get just EOL CSV for license addition

# %%
eol_df = df.loc[df['data_source'] == 'EOL']

# %%
eol_df.head()

# %% [markdown]
# We don't need the BIOSCAN or iNat21 columns, nor the taxa columns.

# %%
eol_license_cols = eol_df.columns[1:4]
eol_license_cols

# %%
eol_license_df = eol_df[eol_license_cols]
#eol_license_df["license"] = None

# %%
eol_license_df.head()

# %%
#eol_license_df.to_csv("../data/eol_files/eol_licenses.csv", index = False)

# %% [markdown]
# ### Merge with Media Manifest to Check for Licenses
# Previous license files (retained below) are missing files, let's merge with the [media manifest](https://huggingface.co/datasets/imageomics/eol/blob/be7b7e6c372f6547e30030e9576d9cc638320099/data/interim/media_manifest.csv) all these images should have been downloaded from to see if there are any here that don't exist in it. From there we'll check licensing info.

# %%
media = pd.read_csv("../data/media_manifest (july 26).csv", dtype = {"EOL content ID": "int64", "EOL page ID": "int64"}, low_memory = False)
media.info(show_counts = True)

# %%
# Read eol license df back in with type int64 for ID columns
eol_license_df = pd.read_csv("../data/eol_files/eol_licenses.csv", 
                             dtype = {"eol_content_id": "int64", "eol_page_id": "int64"}, 
                             low_memory = False)

# %%
eol_license_df.shape

# %%
eol_df = eol_df.astype({"eol_content_id": "int64", "eol_page_id": "int64"})
eol_df.info()

# %%
eol_license_df = eol_df[eol_license_cols]

# %% [markdown]
# Notice that we have about 300K more entries in the media manifest, which is about expected from the [comparison of predicted-catalog to the original full list](https://huggingface.co/datasets/imageomics/ToL-EDA/blob/main/notebooks/ToL_predicted-catalog_EDA.ipynb).

# %%
media.rename(columns = {"EOL content ID": "eol_content_id"}, inplace = True)

# %%
eol_df_media = pd.merge(eol_license_df, media, how = "left", on = "eol_content_id")

# %%
eol_df_media.info(show_counts = True)

# %% [markdown]
# We have about 620K images missing copyright owner.

# %%
eol_df_media.head()

# %%
eol_df_media.loc[eol_df_media["Copyright Owner"].isna()].nunique()

# %% [markdown]
# The missing info is distributed across 116,609 pages.
#
# There also seems to be a discrepancy in the number of page IDs between these. This lead to duplicated records...definitely something's off.

# %% [markdown]
# Check consistency of merge when matching both `eol_content_id` and `eol_page_id`.

# %%
media.rename(columns = {"EOL page ID": "eol_page_id"}, inplace = True)

# %%
merge_cols = ["eol_content_id", "eol_page_id"]

# %%
eol_df_media_cp = pd.merge(eol_license_df, media, how = "inner", left_on = merge_cols, right_on = merge_cols)
eol_df_media_cp.info(show_counts = True)

# %% [markdown]
# Okay, so we do have a mis-match of about 113K images where the content IDs and page IDs don't both match.

# %%
eol_df_media_cp.to_csv("../data/eol_files/eol_cp_match_media.csv", index = False)

# %%
tol_ids_in_media = list(eol_df_media_cp.treeoflife_id)
tol_ids_in_media[:5]

# %%
eol_license_df.head()

# %% [markdown]
# Let's save a copy of the EOL section with content and page IDs that are mismatched.

# %%
eol_df_missing_media = eol_license_df.loc[~eol_license_df.treeoflife_id.isin(tol_ids_in_media)]
eol_df_missing_media.info(show_counts = True)

# %%
eol_df_missing_media.to_csv("../data/eol_files/eol_cp_not_media.csv", index = False)

# %% [markdown]
# ### Save Record of Missing Content IDs & Compare to Older Media Manifest
# Let's save a record of the missing content IDs, then we'll compare them to the [July 6 media manifest](https://huggingface.co/datasets/imageomics/eol/blob/eaa00a48fa188f12906c5b8074d60aa8e67eb135/data/interim/media_manifest.csv) to see if any are in there. The July 6 media manifest is smaller, but we'll still check.

# %%
eol_missing_content_ids = eol_df_media.loc[eol_df_media["Medium Source URL"].isna()]
eol_missing_content_ids.head()

# %% [markdown]
# The pages exist (`eol.org/pages/<eol_page_id>`), but the content IDs do not (`eol.org/media/<eol_content_id>` produces 404).

# %%
#eol_missing_content_ids.to_csv("../data/eol_files/eol_missing_content_ids.csv", index = False)

# %%
media_old = pd.read_csv("../data/media_manifest.csv", dtype = {"EOL content ID": "int64", "EOL page ID": "int64"}, low_memory = False)
media_old.info(show_counts = True)

# %%
media_old.rename(columns = {"EOL content ID": "eol_content_id"}, inplace = True)

# %%
eol_df_media_old = pd.merge(eol_missing_content_ids[eol_license_cols], media_old, how = "left", on = "eol_content_id")

# %%
eol_df_media_old.info(show_counts = True)

# %% [markdown]
# No, we do not have any of the missing ones in the older media manifest.

# %% [markdown]
# ### Check how this compares to Catalog 
# Let's see if these are all images in TreeOfLife-10M, or a mix between it and Rare Species.

# %%
cat_df = pd.read_csv("../data/catalog.csv", low_memory = False)
# Remove duplicates in train_small
cat_df = cat_df.loc[cat_df.split != 'train_small']

# %%
# Add data_source column for easier slicing
cat_df.loc[cat_df['inat21_filename'].notna(), 'data_source'] = 'iNat21'
cat_df.loc[cat_df['bioscan_filename'].notna(), 'data_source'] = 'BIOSCAN'
cat_df.loc[cat_df['eol_content_id'].notna(), 'data_source'] = 'EOL'

# %%
eol_cat_df = cat_df.loc[cat_df.data_source == "EOL"]

# %%
eol_cat_df_media = pd.merge(eol_cat_df[eol_license_cols], media, how = "left", on = "eol_content_id")
eol_cat_df_media.info(show_counts = True)

# %% [markdown]
# Looks like the problem is distributed across both datasets.

# %%
eol_cat_df_media.loc[eol_cat_df_media["Medium Source URL"].isna()].nunique()

# %% [markdown]
# For `catalog` the missing information is distributed across 9,634 pages, so that's 128 pages (of 400) in the Rare Species dataset that we can't currently match.

# %% [markdown]
# ### What are the taxa of the missing images?
#
# Let's bring back a version with the taxa and see what we're dealing with on that end without needing to open the pages.

# %%
cols_of_interest = ['treeoflife_id', 'eol_content_id', 'eol_page_id', 
                    'kingdom', 'phylum', 'class', 'order', 'family', 
                    'genus', 'species', 'common']

# %%
taxa_cols = ['kingdom', 'phylum', 'class', 'order', 'family', 
                    'genus', 'species', 'common']

# %%
eol_taxa_df_media = pd.merge(eol_df[cols_of_interest], media, how = "left", on = "eol_content_id")

# %%
eol_taxa_df_media.loc[eol_taxa_df_media["Medium Source URL"].isna()].nunique()

# %%
eol_taxa_df_media.loc[eol_taxa_df_media["Medium Source URL"].isna()].info(show_counts = True)

# %%
eol_taxa_df_media.loc[eol_taxa_df_media["Medium Source URL"].isna()].sample(7)

# %% [markdown]
# Save a copy of the missing content IDs with taxa info as well.

# %%
#eol_taxa_df_media.loc[eol_taxa_df_media["Medium Source URL"].isna()].to_csv("../data/eol_files/eol_taxa_missing_content_ids.csv", index = False)

# %% [markdown]
# And in `catalog`...

# %%
eol_cat_df_taxa_media = pd.merge(eol_cat_df[cols_of_interest], media, how = "left", on = "eol_content_id")
eol_cat_df_taxa_media.loc[eol_cat_df_taxa_media["Medium Source URL"].isna()].nunique()

# %% [markdown]
# Alright, so it's 2 orders in Rare species.

# %%
eol_cat_df_taxa_media.loc[eol_cat_df_taxa_media["Medium Source URL"].isna()].info(show_counts = True)

# %%
eol_cat_df_taxa_media.loc[eol_cat_df_taxa_media["Medium Source URL"].isna()].sample(4)

# %% [markdown]
# ## Compare Media Cargo
# Media cargo is all images we downloaded from EOL 29 July 2023, so should match to `predicted-catalog`.

# %%
cargo = pd.read_csv("../data/eol_media_cargo_names.csv", dtype = {"EOL content ID": "int64", "EOL page ID": "int64"})
cargo.info(show_counts = True)

# %%
cargo.nunique()

# %%
cargo.head()

# %%
cargo.rename(columns = {"EOL content ID": "eol_content_id"}, inplace = True)
eol_df_cargo = pd.merge(eol_license_df, cargo, how = "left", on = "eol_content_id")

# %%
eol_df_cargo.info(show_counts = True)

# %% [markdown]
# There seem to be 633 images here that aren't listed in the media cargo.
#
# What about in catalog?

# %%
eol_cat_cargo = pd.merge(eol_cat_df[eol_license_cols], cargo, how = "left", on = "eol_content_id")
eol_cat_cargo.info(show_counts = True)

# %% [markdown]
# Still missing 633 images...so we know it's not part of the Rare Species dataset, but is TreeOfLife-10M...

# %%
media_in_cargo = pd.merge(cargo, media, how = "right", on = "eol_content_id")
media_in_cargo.info(show_counts = True)

# %% [markdown]
# But there are 26,868 images in media manifest that are not in cargo (or at least the content ID's aren't), despite the media cargo having 154K more images listed.

# %% [markdown]
# ## Compare to Newer Media Manifest
#
# We will load in a [new media manifest](https://huggingface.co/datasets/imageomics/eol/blob/3aa274067fc4a18877fb394b1d49a92962c57ed8/data/interim/media_manifest_Dec6.csv) (downloaded Dec. 6) to match up `page_id`s for the missing `content_id`s. This way we can download the images and compare via MD5 checksums to hopefully map the new `content_id`s to the old. (See [discussion #18](https://huggingface.co/datasets/imageomics/eol/discussions/18) in [EOL Repo](https://huggingface.co/datasets/imageomics/eol).)

# %%
media_new = pd.read_csv("../data/media_manifest_Dec6.csv", dtype = {"EOL content ID": "int64", "EOL page ID": "int64"}, low_memory = False)
media_new.info(show_counts = True)

# %%
media_new.head()

# %% [markdown]
# To allow for easier sanity-check on the matches, we'll use the version of missing info list with taxa included.

# %%
eol_taxa_df_missing_media = eol_taxa_df_media.loc[eol_taxa_df_media["Medium Source URL"].isna()]
eol_taxa_df_missing_media.head()

# %% [markdown]
# Rename `EOL content ID` and `EOL page ID` columns to match our `eol_taxa_df_missing_media` for easier merging.

# %%
media_new.rename(columns = {"EOL content ID": "eol_content_id", "EOL page ID": "eol_page_id"}, inplace = True)

# %% [markdown]
# First check for any matching content IDs

# %%
eol_taxa_df_missing_media_new_check = pd.merge(eol_taxa_df_missing_media[cols_of_interest], media_new, how = "left", on = "eol_content_id")
eol_taxa_df_missing_media_new_check.info(show_counts = True)

# %% [markdown]
# Yes, there are no matching content IDs here.
#
# Now, let's get our match on page IDs to check they are all listed still for download.

# %%
pg_ids_missing_content = set(eol_taxa_df_missing_media.eol_page_id)
pg_ids_media_new = set(media_new.eol_page_id)

print(f"There are {len(pg_ids_missing_content)} unique page ids that have missing content ids, and there are {len(pg_ids_media_new)} total page ids in the new media manifest.")
      

# %%
missing_pgs = []
for pg in pg_ids_missing_content:
    if pg not in pg_ids_media_new:
        missing_pgs.append(pg)
print(len(missing_pgs))

# %%
media.rename(columns = {"EOL page ID": "eol_page_id"}, inplace = True)
pg_ids_media = set(media.eol_page_id)

print(f"There are {len(pg_ids_media)} total page ids in the July 26 media manifest.")

missing_pgs_jul26 = []
for pg in pg_ids_missing_content:
    if pg not in pg_ids_media:
        missing_pgs_jul26.append(pg)
print(len(missing_pgs_jul26))

# %% [markdown]
# There seem to be 152 page IDs that don't match to either manifest.

# %%
missing_pgs[:10]

# %%
# Why are these floats...does it matter?
missing_pgs_int = [int(pg) for pg in missing_pgs]

int_missing_pgs = []
for pg in missing_pgs_int:
    if pg not in pg_ids_media_new:
        int_missing_pgs.append(pg)
print(len(int_missing_pgs))
print(int_missing_pgs[:5])

# %% [markdown]
# It does not matter. There seem to be 152 missing pages, let's try making a couple into URLs...
#
# The first has a page (https://eol.org/pages/47186210) without images. Let's compare these 152 page IDs to our `category.csv` page IDs. Maybe these were not added because there were no images (still odd they exist in `predicted-category.csv`, but not in the manifest).

# %%
cat_pgs = set(eol_cat_df.eol_page_id)

print(f"There are {len(cat_pgs)} total page ids in the July 26 media manifest.")

missing_cat_pgs = []
for pg in missing_pgs:
    if pg not in cat_pgs:
        missing_cat_pgs.append(pg)
print(len(missing_cat_pgs))

# %% [markdown]
# Nope, these are all in `category.csv`.
#
# Another no image page (https://eol.org/pages/47186225), [this](https://eol.org/pages/46334362) has more data, but still no images. https://eol.org/pages/47186380 & https://eol.org/pages/47121005 also don't show any images.
#
# Seems the images for all of these were removed or moved to other pages...
#
# Let's make a CSV for the missing pages to check that we do indeed have the images (sanity check), and we can compare the taxa!

# %%
missing_pgs_df = eol_taxa_df_missing_media.loc[eol_taxa_df_missing_media["eol_page_id"].isin(missing_pgs)]
missing_pgs_df = missing_pgs_df[cols_of_interest]
missing_pgs_df.info()

# %%
missing_pgs_df.sample(10)

# %%
#missing_pgs_df.to_csv("../data/eol_files/catalog_missing_media_pages.csv", index = False)

# %% [markdown]
# ### Save File with EOL Page IDs & Number of missing content IDs associated with each

# %%
# Count and record number of content IDs for each page ID
for pg_id in pg_ids_missing_content:
    content_id_list = ['{}'.format(i) for i in eol_taxa_df_missing_media.loc[eol_taxa_df_missing_media['eol_page_id'] == pg_id]['eol_content_id'].unique()]
    eol_taxa_df_missing_media.loc[eol_taxa_df_missing_media['eol_page_id'] == pg_id, "num_content_ids_missing"] = len(content_id_list)

cols_of_interest.append("num_content_ids_missing")
eol_taxa_df_missing_media[cols_of_interest].head()

# %%
#unique page_ids
eol_taxa_df_missing_media['duplicate'] = eol_taxa_df_missing_media.duplicated(subset = "eol_page_id", keep = 'first')
eol_taxa_df_num_missing_pg = eol_taxa_df_missing_media.loc[~eol_taxa_df_missing_media['duplicate']]
eol_taxa_df_num_missing_pg.info()

# %% [markdown]
# This file has the relevant info relating to number of missing content IDs per page id, content ID included is just the first instance of such a content ID.

# %%
#eol_taxa_df_num_missing_pg[cols_of_interest].to_csv("../data/eol_files/eol_taxa_df_num_missing_pg.csv", index = False)

# %%
jul26_page_df = media.loc[media.eol_page_id.isin(pg_ids_missing_content)]
jul26_page_df.info()

# %%
jul26_page_df.nunique()

# %% [markdown]
# Yes, that's the expected number of unique page IDs. Let's save to CSV for download.

# %%
#jul26_page_df.to_csv("../data/eol_files/jul26_pages.csv", index = False)

# %%
dec6_page_df = media_new.loc[media_new.eol_page_id.isin(pg_ids_missing_content)]
dec6_page_df.info()

# %% [markdown]
# Okay, we have 5 more entries here, so let's compare unique counts and consider this one.

# %%
dec6_page_df.nunique()

# %%
#dec6_page_df.to_csv("../data/eol_files/dec6_pages.csv", index = False)

# %% [markdown]
# #### Check Older Media Manifest for Missing Pages
#
# Let's take a look at the July 6th media manifest to see if these pages are there.

# %%
media_old.rename(columns = {"EOL page ID": "eol_page_id"}, inplace = True)
pg_ids_media_old = set(media_old.eol_page_id)

print(f"There are {len(pg_ids_media_old)} total page ids in the July 6 media manifest.")

missing_pgs_jul6 = []
for pg in missing_pgs:
    if pg not in pg_ids_media_old:
        missing_pgs_jul6.append(pg)
print(len(missing_pgs_jul6))

# %% [markdown]
# It seems the missing pages are in the _older_ media manifest!
#
# Let's merge this with the `missing_pgs_df`, so we can get URLs to download from the pages there.

# %%
# Count and record number of content IDs for each page ID
for pg_id in missing_pgs:
    content_id_list_mp = ['{}'.format(i) for i in missing_pgs_df.loc[missing_pgs_df['eol_page_id'] == pg_id]['eol_content_id'].unique()]
    missing_pgs_df.loc[missing_pgs_df['eol_page_id'] == pg_id, "num_content_ids_missing"] = len(content_id_list_mp)

missing_pgs_df.head()

# %%
#unique page_ids
missing_pgs_df['duplicate'] = missing_pgs_df.duplicated(subset = "eol_page_id", keep = 'first')
eol_taxa_num_missing_pgs_df = missing_pgs_df.loc[~missing_pgs_df['duplicate']]
eol_taxa_num_missing_pgs_df.info()

# %%
older_page_df = media_old.loc[media_old.eol_page_id.isin(missing_pgs)]
older_page_df.info()

# %%
older_page_df.loc[older_page_df.eol_page_id.astype(str) == "4446364.0"]

# %% [markdown]
# Looks good, let's save to CSV.

# %%
#older_page_df.to_csv("../data/eol_files/media_old_pages.csv", index = False)

# %% [markdown]
# ## Check EOL License file(s) 
#
# First we'll look at `eol_licenses.csv` from Sam, which only covers `catalog.csv`, so load both these in to make sure we've got full coverage for all included images (Matt's first match attempt from file created above couldn't find ~113K based on `eol_content_id`).

# %%
cat_df = pd.read_csv("../data/catalog.csv", dtype = {"eol_content_id": "int64", "eol_page_id": "int64"}, low_memory = False)
license_df = pd.read_csv("../data/eol_files/eol_licenses.csv", 
                         dtype = {"eol_content_id": "int64", "eol_page_id": "int64"}, 
                         low_memory = False)

# %% [markdown]
# The `train_small` is duplicates of `train`, so we will drop those to analyze the full training set plus val.

# %%
cat_df = cat_df.loc[cat_df.split != 'train_small']

# %%
# Add data_source column for easier slicing
cat_df.loc[cat_df['inat21_filename'].notna(), 'data_source'] = 'iNat21'
cat_df.loc[cat_df['bioscan_filename'].notna(), 'data_source'] = 'BIOSCAN'
cat_df.loc[cat_df['eol_content_id'].notna(), 'data_source'] = 'EOL'

# %%
eol_df = cat_df.loc[cat_df.data_source == "EOL"]

# %%
license_df.head()

# %%
license_df.shape

# %%
eol_df.shape

# %% [markdown]
# Yeah, we're missing about 23K images in the license file.

# %%
license_df.info(show_counts = True)

# %%
license_df.loc[license_df["owner"].isna(), "license"].value_counts()

# %% [markdown]
# CC BY licenses without `owner` indicated is rather problematic.

# %%
license_df.loc[license_df["owner"].isna()].sample(5)

# %% [markdown]
# Tracked down `eol_content_id` [14796160](https://eol.org/media/14796160), original source is [BioImages](https://www.bioimages.org.uk/image.php?id=79950) with copyright Malcolm Storey (like 99% of the images on the site (see their [conditions of use](https://www.bioimages.org.uk/cright.htm))). He is listed as "compiler" on the EOL media page.

# %%
license_df.license.value_counts()

# %%
license_df.loc[license_df["license"] == "No known copyright restrictions"].sample(5)

# %%
#license_df.loc[license_df["owner"].isna()].to_csv("../data/eol_files/eol_licenses_missing_owner.csv", index = False)

# %%